Stream-based explainable recommendations via blockchain profiling

Abstract

Explainable recommendations enable users to understand why certain items are suggested and, ultimately, nurture system transparency, trustworthiness, and confidence. Large crowdsourcing recommendation systems ought to crucially promote authenticity and transparency of recommendations. To address such challenge, this paper proposes the use of stream-based explainable recommendations via blockchain profiling. Our contribution relies on chained historical data to improve the quality and transparency of online collaborative recommendation filters – Memory-based and Model-based – using, as use cases, data streamed from two large tourism crowdsourcing platforms, namely Expedia and TripAdvisor. Building historical trust-based models of raters, our method is implemented as an external module and integrated with the collaborative filter through a post-recommendation component. The inter-user trust profiling history, traceability and authenticity are ensured by blockchain, since these profiles are stored as a smart contract in a private Ethereum network. Our empirical evaluation with HotelExpedia and Tripadvisor has consistently shown the positive impact of blockchain-based profiling on the quality (measured as recall) and transparency (determined via explanations) of recommendations.

Keywords

Recommendation systems explainability blockchain data streams historical profiling crowdsourcing intelligent information systems

1. Introduction

Crowdsourcing platforms are open, distributed collaborative systems that enable the general public “users” to carry demanding jobs which otherwise would be carried out by paid employees or contractors. Recommendation engines in large e-commerce crowdsourcing platforms entice users by suggesting products based on past customer feedback, current user preferences and interactions, and inter-user similarities. However, the majority of these recommendations are left unexplained, leaving users clueless about their rationale and could eventually lead to distrust of the platform. Therefore, there is a growing need to provide explainable recommendations in order to increase user confidence, but the formulation of explanations is not a trivial task. For example, Besnard et al. [1] employ an ontology-based approach to represent causal statements and then apply several patterns of inference on the ontological information to generate explanations.

According to Tintarev et al. [2], an explanation is any content added to the recommendation that aims to increase transparency, trust, and decision-making effectiveness and generate high quality recommendations [3]. The explanation of recommendations should arguably be sine qua non for crowdsourcing platforms, as they are largely based on past customer feedback.

Traditional recommendation systems have been benchmarked against single user perceptions only using metrics such as precision, recall, or novelty [4] and applying either collaborative or content-based filtering to generate recommendations. However, it has become self-evident that such evaluations ought to consider a more comprehensive search space and, ideally, take into account additional system parameters and group aspects to cross-validate and eventually explain recommendations [5].

Consequently, modern recommendation filters rely on inter-user and inter-item similarities – also known as user-based or item-based filters [6, 7] – to generate a list of items ordered descendingly according to the predicted ratings, the so-called recommendations. To focus on the perspective of a given active user, the recommendation pipeline may include pre- and post-recommendation filters in order to further tailor recommendations to the user’s profile. Specifically, pre-filters use the current user context to reduce the initial search space, whereas post-filters refine the ordering of the list of items from the perspective of the user. In the latter case, the final items order considers not only the value of the predicted ratings, but also other relevant user profile dimensions such as Trust and Reputation (T&R) [8].

In this context, this paper explores the generation of explanations for different types of recommendation systems, using, as case studies, two data sets from the tourism domain namely HotelExpedia and TripAdvisor, via blockchain profiling. Crowdsourced from the Expedia and TripAdvisor tourism sites, these two data sets have been processed incrementally as data streams.

Blockchain (see Section 2.1) has been considered a promising technology with intrinsic properties such as security, authenticity, traceability, and transparency. Especially suitable for environments that require temporal order storage, blockchain ensures data remain authentic and certified by distributed entities rather than a single central authority [9]. This work relies on blockchain to create authentic, traceable, historical profiles used to explain opaque recommendations by means of a private blockchain distributed network (i.e., including several nodes). This design choice grants: (i) a temporal database where past transactions remain recorded in the chain of blocks; (ii) an immutable database where, once a transaction is written in a block, it remains stored in a Merkle tree connected to previous blocks through interdependent hashes; and (iii) a distributed, replicated database, which is harder to infiltrate thanks to the blockchain consensus.

In order to address the explainability and transparency challenges, we focus on two research questions:

RQ1: RQ1:
Can we improve the quality and transparency of existing online collaborative filters by means of an external a posteriori-filter?
RQ2:
Can we use the evolving incremental trust and reputation (T&R) profiles of raters for this purpose?

Transparency is particularly relevant for existing crowdsourcing platforms which incorporate opaque recommendation engines such as Expedia or TripAdvisor. In addition, the evolving T&R profiles can be stored in a blockchain to provide authenticity, traceability and historical logging. These profiles can then be used to reorder the output of opaque collaborative filters and to explain the top recommendations presented.

This article significantly expands and refines our previous work reported in Leal et al. [10] by furnishing a novel method to marshal and enable trust and reputation on incremental profiles using blockchain smart contracts. Such chained historical data are then used to improve the quality and transparency of recommendations for different recommendation filters, namely, memory-based [8] and model-based [11].

In a similar fashion to Jiang-Adeli’s work [12], our system has adopted an engineering approach to forecast the behaviour of incremental profiles under specific operating conditions by using memory- and model-based filters. Given their tamper-proof nature, smart contracts aim to ensure complete data authenticity, i.e. avoid the manipulation from unethical stakeholders as well as user and data provenance.

The main contribution of this work is to rely on engineering-inspired historical stream-based profiling with the help of blockchain to provide explanations and higher quality recommendations. We have performed a series of experiments on representative data sets from two major tourism crowdsourcing platforms (HotelExpedia and Tripadvisor), used as case studies. Our results show noteworthy advantages when using Recall@10 and Target Recall@10 (TRecall) as evaluation metrics. Moreover, from a systems point of view, we have evaluated the blockchain performance in terms of latency and throughput, measured as transactions per second (t/s). The results show not only a positive impact of historical profiling in recommendation, but that trust chained profiles underpin the creation of authentic, traceable and explainable recommendations.

The rest of this paper is organised as follows. Section 2 reviews the background on explainable recommendations, blockchain, traceability, smart contracts, and the current contribution. Section 3 presents a motivational example concerning explainable recommendations in crowdsourcing platforms. Section 4 describes the proposed method for the creation of explainable recommendations. Section 5 presents the experiments performed and the results of the empirical evaluation. Finally, Section 6 summarises and discusses the outcomes of this work.
2. Literature survey

Explainable recommendations have been discussed in the literature for distinct domains, as part of the ongoing endeavour for Explainable Artificial Intelligence [13, 14] which aims to both improve prediction accuracy and allow humans to understand and have confidence in artificial intelligence systems. Whilst enhancing recommendations with explanations can significantly expand the overall search space by introducing additional user actions, the overall user engagement should arguably be improved ultimately benefiting the given recommendations and user confidence.

Recommendation explainability in fashion has proved to be particularly popular due to the subjective nature of the domain. E-commerce platforms have looked at regions within a picture to explain a given recommendation, e.g., using a multimodal attention convolutional neural network to extract specific features within a fashion shot [15, 16]. Other researchers have employed a semantic extraction network, where two tasks are divided in distinct semantic attributes (items), and then a convolutional neural network classifier is used to represent and classify the items [17]; quantified the influence of different visual regions with a combination of human-interpretable features [18]; used fuzzy logic to build adaptive personalised profiles [19]; or employed proximity alignment of co-purchasing networks to provide online recommendations [20].

Within the food industry, researchers have adopted graphs to describe the hierarchy of preferences of a user [21]. To enhance predictions they used an attention layer where profiles are updated using a hierarchical propagation mechanism and the explanations are generated using a dynamic programming method to determine the usefulness of each feature. Other approaches use deterministic Markov decision process combined with knowledge graph to produce explainable recommendations [22]. Via a reinforcement learning mechanism, their proposed system learns to predict potential items of user interest, such that history can serve as an explanation for the prediction. Other works propose a knowledge-aware path recurrent network model to exploit knowledge graph for recommendation [23]. This graph model can represent semantics and entity relations between users and songs. The sequential dependencies within a path, allows to infer the rationale of a user-item interaction and prediction.

A social network approach has proposed an explanation mechanism for recommendation systems that uses the concept of viewpoints [24]. A viewpoint is the combination of a concept, topic, and sentiment of a sentence, since Natural Language Processing (NLP) techniques can provide useful insights on reviews and explain a recommendation. Similarly, other research groups combine different user sentence vectors into a unified embedding, which are used in the final rating. The attention weights encode the relevance of different review sentences, which can provide rating prediction explainability [25].

In songs and multimedia, recommendations have long employed different techniques including tag-based methods coupled with digital signal processing [26]. To enable explainability, a matrix factorisation combined with a deep learning model have been used to elucidate the recommendations of songs [27]. To produce the explanations the authors used a forward propagation of segments to explain the prediction. Additionally, the explainability of a recommendation can be achieved using rule based learning model from knowledge graph [28]; or using multi-armed bandits to produce recommendations and a reward function that predicts the user engagement of an item with an explanation on a given context [29]. The latter has been developed by the Spotify research team who discusses how their bandits system learns and predicts satisfaction (e.g., click-through rate, consumption probability) for any combination of item, explanation, and context. They also put particular emphasis on the ability to enable logging and contextual retraining in their implementation as a way to “learn from mistakes”.

Specifically, this paper focuses on explaining online recommendations from crowdsourcing platforms. Crowdsourcing platforms not only advertise offers, but also promote the voluntary feedback sharing, which influences the behaviour of the other customers [30]. However, this crowdsourced information, which is voluntarily and freely shared, raises questions about its reliability. Therefore, it is relevant to assess the reliability of crowdsourced information, namely, by using trust and reputation models. Service reputation can be inferred from the reputation of the service contributors, which, in turn, can be based on the analysis of the individual contributions, e.g., reviews or ratings. Therefore, higher reputation indicates higher service quality, allowing, for example, the provider to increase the price [31]. Hence, rater trustworthiness can be derived from the crowdsourced rating stream. This can be achieved together with collaborative recommendation filters, decision trees or neural networks classifiers. Regardless of the technique, the designed system must ensure data authenticity, traceability, and transparency as well as work online and scale well. To maintain authenticity and traceability, the pairwise trust between raters or the stream of ratings can be stored in the blockchain. Considering transparency, the chained pairwise trust and decision trees have a clear advantage over black box neural network models since the explanations are immediately available in the blockchain [10] or are by default embedded in the path from the tree root to the selected leaf [32, 33]. Nevertheless, there are post hoc complex explanation techniques that can be used together with deep neural networks [34]. When it comes to stream processing, while there are online neural network algorithms [35, 36, 37], decision trees [38] and recommendation filters [39] are typically faster. Ultimately, offline neural networks algorithms, such as [40, 41, 42], could be adapted for stream processing with considerable effort. Considering scalability – an essential characteristic for big data processing – there are distributed implementations of both recommendation filters [43], decision trees [44] and deep neural networks [37, 45, 46, 47].

2.1 Blockchain

A blockchain, a data structure built as a persistent linked list of records, groups records into blocks which are concatenated via metadata with blocks of the previous chain into a continuous timeline. Each block contains a timestamp with a cryptographic hash of the previous block, and the transaction data (generally represented as a Merkle tree). Due to the use of cryptographic techniques, data contained in a block can only be altered by modifying all subsequent ones. Such property enables its application in distributed environments, where the blockchain data structure acts as a non-relational public database containing irrefutable historical information.

By using an appropriate consensus protocol, it is possible to keep the integrity of the data across all the network nodes without the need of a central trusted entity [48]. As a result, blockchain technology maintains a reliable system state, achieved and strengthened by the member nodes themselves, even in environments where a minority of nodes pursue some malicious behaviour.

There are several types of protocols to reach consensus in a blockchain network such as Proof of Work, Proof of Stake, and Proof of Authority. In particular, this paper explores Proof of Authority (PoA). In PoA-based networks, transactions and blocks are validated by approved accounts, also known as validators, in an automated process that only requires authority nodes to remain secure. Presently, PoA is considered more robust than Proof of Stake and much more efficient than Proof of Work [49]. Besides, this protocol has been implemented recently in the Ethereum network [50]. This implementation is highly scalable for private networks with diverse requirements in terms of quality of service and service level agreements [51, 52].

2.2 History, traceability and transparency

The development of Artificial Intelligence based systems should ideally be guided by Accountability, Responsibility and Transparency (ART) design principles [53], i.e., systems should explain and justify their decisions (accountability), incorporate human values into technical requirements (responsibility) and describe the decision-making process and how data is used, collected and governed (transparency). With its transparency, traceability, trust, immutability, desintermediation, and security characteristics [54], blockchain is arguably one of the most promising techniques to achieve ART compliant data processing as data history stays automatically imprinted in the chain of blocks. Having been tested in different applications and industries, blockchain technologies have proved to intrinsically grant (i) decentralised immutable and traceable reputation; (ii) unique users, i.e., only registered users can contribute with ratings or reviews; and (iii) portable and transversal reputation [55].

2.3 Smart contracts

Szabo [56] proposed the term smart contract to refer to a computer program which uses algorithmic protocols to ensure the execution of the terms of a contract for securing relationships on public networks. Blockchain technologies enable the implementation of smart contracts in distributed environments without the need of a trusted central authority [57]. A smart contract is composed of dedicated data structures and methods, whose execution are stored as immutable transactions in the blockchain [58].

Due to the intrinsic properties of blockchain technologies such as history, traceability, transparency and security support [59], there is an increased interest in the use of blockchain and smart contracts to implement distributed trust and reputation management systems with the goal to improve the trustworthiness of the information, privacy and security of such systems [60].

Blockchain smart contracts have been employed by trust based systems and/or reputation based systems in a variety of areas such as autonomous systems [61], multi-agent systems [62], fair payments with reputation for cyber physical systems [63], crowdsourcing [64], recommender systems [65] and crowdsourcing-based recommendation systems [10]. Smart contracts have been used to calculate and manage the trust [66] and reputation scores [62, 64], and also to store the trust and reputation values in the blockchain via the contract’s variables [64, 10].

In contrast to the previous works that use smart contracts, our approach employs smart contracts to store historical, traceable inter-user trust profiles with the aim to enable explainable and higher quality recommendations.

2.4 Contribution

This work proposes a novel method to explain and refine stream-based collaborative recommendations, supported by blockchain technology. The research addresses the problem of making collaborative recommendation engines transparent, a key feature for any state-of-the-art recommendation/search engine. The proposed method stores profiles in the blockchain not only to keep the history, immutability and traceabilility of the contents, but also to generate trustworthy explanations and improve the quality of recommendations.

Table 1
Top 3 explainable recommendations for an active user $a$

1.	Hotel Porto: 4 stars \| 4.8 rating (Why?)
Why?	Porto is our top recommendation for you: it was rated 4.8 by a group of seven like-minded users (Who?) with a joint reputation $\mu_{R}=$ 13% and in whom you trust $\mu_{T}=$ 13%
Who?	User 11 ( $u_{11}$ ) rated Porto 4.8 and has a system-wide reputation of $R_{11}=$ 19%; Your trust in $u_{11}$ is $T_{u,11}=$ 20%; In the past you chose 12 items (Which?) based on $u_{11}$ .
	User 82 ( $u_{82}$ ) gave Porto a 5.0 and has a reputation of $R_{82}=$ 16%; Your trust in $u_{82}$ is $T_{u,82}=$ 18%; In the past you have chosen 12 items (Which?) based on $u_{82}$ .
	Etc.
Which?	You chose in the past Hotel Vigo based on $u_{11}$ (4.5 rating); You chose in the past Hotel Dublin based on a 4.6 rating of $u_{11}$ ; etc.
	You chose in the past Hotel Vigo based on $u_{82}$ (4.7 rating); You chose in the past Hotel Dublin based on a 4.8 rating of $u_{82}$ ; etc.
	Etc.
2.	Hotel Vigo: 4 stars \| 4.8 rating (Why?)
3.	Hotel Dublin: 4 stars \| 4.7 rating (Why?)

The application of blockchain together with trust and reputation to crowdsourcing platforms is relatively new [67]. The works found in the literature include trust [68, 69] and reputation [70, 64]. Regarding the blockchain, a few use the Ethereum open source framework [70, 69, 64] and one relies on a proprietary solution [68]. This review shows that most crowdsourcing platforms disregard processing transparency. In particular, in the tourism domain where crowdsourcing is extremely popular, no such mechanisms are found in the prevailing platforms.

The proposed method builds incremental trust and reputation profiles (supported by smart contracts) and explores historical data to achieve both decision-making transparency and increased recommendation quality. Moreover, it was designed to be implemented as a plugable module, to be composed by several distributed nodes, and integrated with existing opaque collaborative filters via a post-recommendation filter.

The main contribution focuses on incremental profiling with look-back refinement to provide authentic and trustworthy explanations and higher quality recommendations, improving the user experience.

Figure 1.

Recommendation and explanation engine. Module A represents the standard recommendation engine. Module B portrays the novel explanation module (Blockchain, T&R profiling with look-back refinement, and post-filtering).

3. Motivational example

Algorithmic transparency is not a mere ethical concern since it contributes to improving the user experience and confidence on AI-based systems. The proposed method relies on T&R incremental profiling and on the transaction history kept by blockchain to improve the quality of recommendations and to explain the top recommendations. Table 1 presents an example concerning explanations based on Incremental Trust & Reputation Profiling. The active user can click on the questions Why?, Who?, and Which? and the system is able to present the corresponding explanations. These explanations present the user with the reasoning collaborative filters apply to generate personalised recommendations, which is largely based on the set of pertinent item co-raters. Consequently, our method uses the chained inter-user trust and the derived system-wide reputation of the pertinent item co-raters to provide the active user with the reasons behind any item recommendation.

The post-filter reorders and provides explanations for the top $N$ recommendations based on the chained inter-user trust profiles. Specifically, the chaining of the incremental trust grants by default: (i) authenticity – the stored information remains unchanged (but does not guarantee its original correctness); (ii) historical record-keeping – the evolution of the trust incremental profiles through time is maintained in the blocks; and (iii) traceability – contributions are traced back to contributors. The user-item trust together with the system-wide user and item reputations are derived from the chained inter-user trust.

4. Proposed method

The proposed method explores blockchain technologies to maintain the authenticity, history and traceability of the trust and reputation models.

In this context, we developed a new plugable module – module B – incorporating blockchain-based profiling and trust and reputation post-filtering. As illustrated in Fig. 1, the proposed module can be deployed together with opaque stream-based collaborative filters – module A – to explain recommendations. The blockchain-based profiling integrates a private Ethereum network adapted from Leal et al. [10].

Module B includes an Incremental T&R profile updater, the smart contract holder (private Ethereum network) and the post-filter, whereas module A contains the collaborative recommendation filter, and the evaluator (Root Mean Square Error and/or Recall@N metrics). Our method applies incremental updating (data streaming), i.e., the profiles and models are updated every time a new rating event occurs. In the case of memory-based filters, module B works independently from module A, whereas, in the case of memory-based filters, every time a new event is processed by the module A, the filter shares with module B the set of pertinent neighbours.

4.1 Profiling

Profiling is a core component of a recommendation system as it has a direct impact on the quality of recommendations. The type and complexity of profiles depend on the available information sources. This work relies on multi-criteria rating profiles to predict personalised ratings (collaborative filters) and on blockchain inter-user trust profiles to reorder and explain predictions. Specifically, in stream-based recommendation, these profiles are continuously updated and refined.

4.1.1 Rating-based

Rating-based crowdsourcing platforms often allow the classification of services according to different criteria, e.g., overall, cleanliness or staff ratings. In this work, the collaborative filters adopt the Personalised Weighted Rating Average (PWRA). Proposed by Leal el al. (2017) [71], this rating-based profiling combines existing multiple criteria ratings into a single rating. Equation (1) displays the PWRA $r_{u,i}$ where $r^{c}_{u,i}$ is the non-null rating given by user $u$ to item $i$ regarding criterion $c$ and $n_{u,c}$ represents the total number of non-null criterion $c$ ratings given by user $u$ .

$\displaystyle r_{u,i}={\displaystyle\frac{\sum_{c=1}^{C}n_{u,c}r^{c}_{u,i}}{% \sum_{c=1}^{C}n_{u,c}}}$ (1)

4.1.2 Trust and reputation

Trust is a one-to-one relationship based on the outcomes of direct interactions, whereas Reputation is a many-to-one relationship built upon indirect interactions such as hearsay. Trust and Reputation profiles can be built and employed with both memory and model-based collaborative filters by means of an a posteriori recommendation filter. Therefore, we apply a cascade post-filter to first sort the predictions by rating, then by trust and, finally, by reputation. The aim is to simultaneously improve the quality of recommendations and support the generation of explainable collaborative recommendations.

The trust network corresponds to a multigraph, i.e., a directed graph with up to two edges between the same pair of nodes, representing a trustor and a trustee. These entities can be two users or one user and one item. A directed edge indicates the one-way trust the source node (trustor) has in the target node (trustee). In this context, three situations may occur between the pair of nodes: (i) there is no edge between the nodes; (ii) there is only one directed edge; or (iii) there are two directed edges, each with its own value.

Memory-based filters

Leal et al. (2018) [8] proposed a trust and reputation model for memory-based filters which quantifies the trustworthiness between the active user and its neighbour users by analysing their mutual influence. It takes into account the number of times the active user selects, from the top ten recommendations, recommendations based on a given neighbour. The social reputation of an user is derived from the trust attributed to the user by the set of pertinent neighbours.

Trustworthiness ( $T_{a,k}$ ) of $k$ from the perspective of active user $a$ is computed through Eq. (2). It increases when $a$ selects recommendations based on neighbour $k$ and decreases when $a$ ignores recommendations based on $k$ . In this context, $a.n_{k}$ represents the number of items actually selected by $a$ due to $k$ and $a.N_{k}$ is the number of times $k$ was a neighbour of $a$ .

$\displaystyle T_{a,k}=\frac{a.n_{k}}{a.N_{k}}$ (2)

Neighbour-based reputation ( $R_{a,k}$ ) quantifies the reputation of neighbour $k$ from the perspective of the active user $a$ . This reputation is based on the trust the set of pertinent neighbours $a.P_{k}$ has on $k$ , where $a.P_{k}$ is the set of neighbours shared by $a$ and $k$ . In Eq. (3), $p\in a.P_{k}$ and represents a pertinent neighbour; $T_{a,p}$ is the trust $a$ has in $p$ ; and $N_{P}$ is the size of $a.P_{k}$ .

$\displaystyle R_{a,k}=\frac{\sum_{s=1}^{N_{P}}T_{a,p}}{N_{P}}$ (3)

System-wide reputation ( $R_{k}$ ) represents the global reputation of neighbour $k$ , considering the subset of pertinent users $k . P$ , i.e. those who share at least one neighbour with $k$ . Equation (4) presents $R_{k}$ , the system-wide reputation of $k$ , where $R_{p,k}$ is given by Eq. (3), and $N_{P}$ is the size of $k . P$ .

$\displaystyle R_{k}=\frac{\sum_{p=1}^{N_{P}}R_{p,k}}{N_{P}}$ (4)

Algorithm Appendix, presented in Appendix Appendix A. Algorithms, describes the Memory-based Incremental Trust and Reputation profiling. The algorithm is invoked whenever a new rating event occurs.

Model-based filters

Leal et al. (2019) [11] designed a trust-based profiling approach for model-based filters that quantifies the relatedness between the active user $a$ and its co-raters. The trustworthiness of $k$ according to $a$ is based on the set of pertinent co-rated items, where a pertinent co-rated item is defined as being similarly classified ( $\pm$ 10%) by both users. The set of pertinent co-raters of a given user includes all those who have co-rated at least one pertinent item.

Trustworthiness ( $T_{a,k}$ ) is calculated through Eq. (5) where $a.n_{k}$ is the number of pertinent co-rated items by both $a$ and $k$ , and $a . N$ is the total number of items rated so far by the active user $a$ .

$\displaystyle T_{a,k}=\frac{a.n_{k}}{a.N}$ (5)

System-wide user reputation ( $R_{k}$ ) corresponds to the average trust given to user $k$ by his co-raters, considering a trustworthiness threshold. In Eq. (6), $T_{c,k}$ represents the trust co-rater $c$ has in $k$ ; $k . C$ is the set of co-raters of $k$ , $T_{p,k}\geqslant\mu_{R}$ is the applied threshold, and $k.N_{R}$ is the number of co-raters with trustworthiness equal or greater than the threshold.

$\displaystyle R_{k}=\frac{\sum_{p=1}^{k.N_{R}}T_{c,k}}{k.N_{R}},T_{c,k}% \geqslant\mu_{R}$ (6)

Algorithm Appendix, included in Appendix Appendix A. Algorithms, describes the Model-based Incremental T&R profiling. Its execution is triggered by the arrival of a new rating event.

Cascade post-filter

sorts sequentially by rating, trust and, finally, reputation, the predictions generated after each event arrival. To do so, it derives the trust and the reputation that the active user has in the predicted item rating from the trust between the active user and the pertinent item raters. The pertinent users from the perspective of the active user are, in the case of memory-based filters, the subset of the neighbours of the active user who co-rated the item, whereas in the case of model-based filters, they are the subset of relevant item co-raters in whom the active user trusts, at least, as much as the average system-wide reputation.

Item trust ( $T_{a,i}$ ) corresponds to the average trustworthiness of the pertinent item co-raters from the perspective of the active user. $T_{a,i}$ is given by Eq. (7) where $a$ is the active user, $i$ is the item, $p$ represents a pertinent co-rater, $T_{a,p}$ is the trustworthiness $a$ has in $p$ , $N_{P}$ represents the number of pertinent co-raters and $\mu_{R}$ is the average system-wide reputation. Equation (7.1) is applied with memory-based filters and Eq. (7.2) with model-based filters.

$\displaystyle T_{a,i}=\begin{cases}\displaystyle\frac{\sum_{p=1}^{N_{P}}T_{a,p% }}{N_{P}}&\text{\@setsize{\small}{11pt}{\ixpt}{\@ixpt}{(1)}}\\ \displaystyle\frac{\sum_{p=1}^{N_{P}}T_{a,p}}{N_{P}},\quad T_{a,p}\geqslant\mu% _{R}&\text{\@setsize{\small}{11pt}{\ixpt}{\@ixpt}{(2)}}\end{cases}$ (7)

Item reputation ( $R_{a,i}$ ) is the average reputation of the pertinent item co-raters from the perspective of the active user $a$ . In Eq. (8), $i$ represents the item, $p\in P_{a}$ , the set of pertinent co-raters of $a$ , and $N_{P}$ the cardinality of $P_{a}$ . Equation (8.1) displays the memory-based item reputation, where $R_{a,p}$ is the neighbour-based reputation of $a$ , and Eq. (8.2) presents the model-based item reputation, where $T_{a,p}$ is the trust $a$ has in $p$ and $\mu_{R}$ is the average system-wide reputation.

$\displaystyle R_{a,i}=\begin{cases}\displaystyle\frac{\sum_{p=1}^{N_{P}}R_{a,p% }}{N_{P}}&\text{\@setsize{\small}{11pt}{\ixpt}{\@ixpt}{(1)}}\\ \displaystyle\frac{\sum_{p=1}^{N_{P}}T_{a,p}}{N_{P}},\quad T_{a,p}\geqslant\mu% _{R}&\text{\@setsize{\small}{11pt}{\ixpt}{\@ixpt}{(2)}}\end{cases}$ (8)

4.2 Smart contracts

Trust and reputation models aim to help the user navigate through a large search space by suggesting relevant items and, thus, improving the user experience. However, when the models are centrally stored, they can be easily manipulated, e.g., to meet hidden interests. Therefore, this paper explores a blockchain-based solution to ensure the authenticity of the profiles supporting explanations. The proposed solution represents T&R models as smart contracts in the blockchain. The smart contract holds a hash table containing the corresponding profiles and supports a collection of transactions to alter and access the stored model as illustrated in Fig. 2.

Figure 2.

Smart contract structure.

Specifically, the trust smart contract stores the trust between neighbours or co-raters obtained with Eq. (2) or Eq. (5), depending on the collaborative predictor employed. These values are updated with each incoming event, therefore allowing tracing over time. Algorithm Appendix A. Algorithms, included in Appendix Appendix A. Algorithms, shows the contents of the Trust Smart Contract.

4.3 Look-back profile refinement

Every time some trustor chooses an item due to a trustee, the incremental trust of the trustor in the trustee is updated and stored in the blockchain. This event-driven storage of the incremental trust in the blockchain constitutes yet another source of relevant information – the trust time series. This historical trust data can be used to refine further individual profile, resulting in the look-back refinement.

Figure 3 plots the incremental trust and reputation with and without look-back refinement for two HotelExpedia users, where user 121 is the trustor and user 104 is the trustee. Figure 3a shows the evolution of the incremental trust user 121 has in his neighbour 104 ( $T_{121{,}104}$ ) together with system-wide incremental reputation of user 104 ( $R_{104}$ ). Figure 3b displays the same information with look-back refinement.

Figure 3.

Incremental T&R profiling with and without look-back refinement. (a) $T_{121{,}104}$ and $R_{104}$ without look-back refinement; (b) $T_{121{,}104}$ and $R_{104}$ with look-back refinement.

The incremental and historical data can be combined to refine the inter-user trust and, after, derive user-item trust and system-wide user and item reputation. This work considers statistical linearisation (Eq. (9)), recursive statistical linearisation (Eq. (10)) and slope sign changes reinforcement (Eq. (11)). The resulting profile is called the incremental profile with look-back refinement or simply refined profile.

Equation (9) provides the linearised statistical trust and reputation profiles ( $S$ ). The refined ${T_{S}}_{a,x}$ is the statistical look-back trust the active user $a$ has in user or item $x$ based on current and past incremental profile values, where $\alpha$ is a linearisation parameter and $T_{a,x}$ , $\mu_{T_{a,x}}$ and $\sigma_{T_{a,x}}$ represent the current, average and standard deviation of the incremental profile value, respectively. The refined ${R_{S}}_{x}$ is the statistical system-wide reputation of user or item $x$ based on current and past incremental profile values, where $\alpha$ is a linearisation parameter and $R_{x}$ , $\mu_{R_{x}}$ and $\sigma_{R_{x}}$ represent the current, average and standard deviation of the incremental profile value, respectively.

$\displaystyle\begin{aligned} \displaystyle{T_{S}}_{a,x}&\displaystyle=\alpha T% _{a,x}+(1-\alpha)|\mu_{T_{a,x}}-\sigma_{T_{a,x}}|\\ \displaystyle{R_{S}}_{x}&\displaystyle=\alpha R_{x}+(1-\alpha)|\mu_{R_{x}}-% \sigma_{R_{x}}|\end{aligned}$ (9)

Equation (10) displays the recursive linearised statistical trust and reputation profiles ( $R$ ). The only difference as compared to Eq. (9) is that this refined profile is recursive, using the incremental and the prior refined profile values.

$\displaystyle\begin{aligned} \displaystyle{T_{R}}_{a,x}&\displaystyle=\alpha T% _{a,x}+(1-\alpha)|\mu_{T_{{R}_{a,x}}}-\sigma_{T_{{R}_{a,x}}}|\\ \displaystyle{R_{R}}_{x}&\displaystyle=\alpha R_{x}+(1-\alpha)|\mu_{R_{{R}_{x}% }}-\sigma_{R_{{R}_{x}}}|\end{aligned}$ (10)

Equation (11) presents the sign changes refined profiles (C) based on the slope sign changes of the incremental profile. In this case, the current incremental profile value is reinforced by the ratio of value increases to value decreases plus equal values, where $i$ is the number of profile value increases, $e+d$ is the number of equal and decreased values events. ${T_{C}}_{a,x}$ represents the refined trust the active user $a$ has on user or item $x$ based on the slope sign changes of the incremental profile, and $T_{a,x}$ the incremental trust value. ${R_{C}}_{x}$ represents the refined system-wide reputation of user or item $x$ based on the slope sign changes of the incremental profile, and $R_{x}$ the incremental system-wide reputation value.

$\displaystyle\begin{aligned} \displaystyle{T_{C}}_{a,x}&\displaystyle=T_{a,x}% \left(1+\frac{i}{e+d}\right)\\ \displaystyle{R_{C}}_{x}&\displaystyle=R_{x}\left(1+\frac{i}{e+d}\right)\end{aligned}$ (11)

4.4 Explainable recommendations

Explainable recommendations address the question of ‘Why has this item been recommended?” by sharing with the user the reasons for the recommendation. In a transparent system, it is important to explain the recommendation process, clarifying if it took into account the user’s preferences or rather hidden interests. In this context, the proposed method relies on the chained models to reorder and explain recommendations following the approach proposed by Leal et al. [10].

The proposed method associates and stores in a smart contract the one-way trust each user (trustor) maintains in each of his influential users (trustees). This set of users corresponds, in the case of memory-based filters, to the top nearest neighbours and, in the case of model-based filters, to those users with a larger number of relevant co-rated items. Once the collaborative filter generates predictions for the active user, the post-filter sorts the list and presents the top $N$ recommendations together with links to explanations. The trust smart contract constitutes a trust-based, historical, traceable, and immutable source of information for explainable recommendations.

The idea of using trust models as part of a broader recommendation explanation was introduced by O’Donovan and Smyth (2005) [72]. Table 2 exemplifies the type of support information available to generate explanations for the active user $a$ , namely the resource info (name and identifier) and its average rating ( $\mu_{r}$ ), and the number of users (#), average trustworthiness ( $\mu_{T}$ ) and system-wide reputation ( $\mu_{R}$ ) of the users behind the resource recommendation. $\mu_{T}$ corresponds to the average look-back trust of $a$ in $k$ ( $T_{b_{a,k}}$ ). Based on these data, the active user obtains, for example, the following explanation for the top recommendation: “Hotel Porto was recommended to you because it was rated with an average of 4.8 by a group of seven co-raters in whom you trust 13% and who have an average system-wide reputation of 13%”.

Table 2
Aggregated explanation data

Hotel			Users
Name	ID	$\mu_{r}$	#	$\mu_{T}$ (%)	$\mu_{R}$ (%)
Porto	$h_{a}$	4.8	7	13	13
Vigo	$h_{b}$	4.8	7	12	13
Dublin	$h_{c}$	4.7	10	10	12

Table 3

Detailed explanation data

Hotel Porto Co-raters
ID	$r_{k,h_{a}}$	$T_{b_{u,k}}$ (%)	$R_{b_{k}}$ (%)	#
11	4.8	20	19	12
82	5.0	16	18	12
34	4.7	14	12	14
14	4.6	12	13	12
15	4.9	10	11	8
67	4.7	10	11	8
20	4.9	10	10	6

These generic explanations can be further detailed by specifying the individual contributions of each co-rater $k$ . For example, Table 3 displays the detailed information supporting the recommendation of hotel Porto ( $h_{a}$ ) to the active user $a$ , namely the identification of co-rater $k$ , the rating ( $r_{k,h_{a}}$ ) given by $k$ to $h_{a}$ , the refined trust ( ${T_{b}}_{a,k}$ ) the active user $a$ has in co-rater $k$ , the system-wide refined reputation ( $R_{b_{k}}$ ) of the co-rater $k$ , and the number of times $a$ chose a recommendation supported by $k$ (#). Based on these data, the active user obtains the following complementary explanation: “Specifically, user 11, with a reputation of 19% and whose tastes you trusted 12 out of 60 times ( $T_{b_{a,k}}=$ 20%), rated this hotel with a 4.8; user 82, with a reputation of 18% and whose recommendations you trust 16%, gave it a 5.0; etc.”. Table 1 provides a more detailed example of the generated explanations.

4.5 Evaluation

The proposed method is empirically evaluated by calculating incrementally for each incoming event the following metrics:

Incremental root mean square (RMSE), defined by Takács et al. (2009) [73].

Incremental Recall@N, proposed by Cremonesi et al. (2010) [74], computes the classification accuracy based on a sample of 1000 randomly selected items never rated by the active user plus the newly rated item by the active user. The sample is then sorted by descending prediction value and the top $N$ predictions are recommended to the active user. If the newly rated item belongs to this final list, it counts as a hit.

Incremental TRecall@N, proposed by Veloso et al. (2017) [75] computes the classification accuracy similarly with the approach of the incremental Recall@N. However, it counts a hit only if the prediction is close to the target rating, i.e., within a radius of $\frac{N}{2}$ of the actual rating of the active user.

The performance of the blockchain is assessed using the following metrics [76]:

Throughput to determine the number of successful blockchain transactions per time unit.

Latency to establish the time delay between the submission and completion of a blockchain transaction.

5. Experiments and results

The empirical evaluation focused on the assessment of stream-based explainable recommendations with smart contract profiling. The implemented stream-based recommendation engines include a collaborative memory-based filter [8] and a collaborative model-based filter [11]. The accuracy and classification accuracy were evaluated using RMSE, Recall@10 and TRecall@10. The aim was to analyse the performance of: (i) PoA consensus algorithm; (ii) rating-based (PWRA); and (iii) post-filtering with trust-based (T&R) profiling. Specifically, the profiling experiments contemplated incremental profiling with and without look-back refinement (with the three look-back refinement versions).

The experiments were conducted on an OpenStack cloud instance with 16 GiB RAM, 8 CPU and 160 GiB of hard-disk space. The blockchain technologies employed were: (i) Go-ethereum1 as Ethereum client; (ii) Solidity2 as smart contracts language; and (iii) Web3J3 as Java Application Programming Interface. The configuration of the implemented private Ethereum network is composed of two miner nodes. In the case of PoA, the private Ethereum network was configured with a block period of 2 s and a block gas limit of 0 $\times$ 8000000. The network provides cost-free processing, i.e., the transactions are submitted using 0 as gas price.

The models are incrementally updated with each incoming stream event. The default trustworthiness is initialised at 0%. The $N$ used for the incremental metrics was set to 10. The predictions were rounded to one decimal place accuracy before being submitted to the cascade post-filter where they are sequentially sorted in descending order of prediction value, item trust (Eq. (7)) and item reputation (Eq. (8)). The refined profiles are built using incremental trust values with a $z\text{-score}=\pm 0.7$ to eliminate the system cold start. The refined profile experiments based on Eq. (9) were performed with $\alpha=$ 0.5.

The implemented application programming interface (API) with Ethereum offers three transactions. The uniform resource identifier (URI) of each transaction, including the input parameters, and responses are summarised in Table 4. Parameter $v$ represents whether the last rating event of $a$ has led to the decrease or increase of trust in $k$ .

Table 4
Blockchain API

URI	Response
bc/sc/updateTrust/{ $a$ }/{ $k$ }/{ $v$ }	–
bc/sc/getTrust/{ $a$ }/{ $k$ }	${T_{a,k}}$
bc/getPastTrust/{ $a$ }/{ $k$ }/{ $n$ }	$[{T_{a,k}}_{i-n},\ldots,{T_{a,k}}_{i}]$

$a$ represents the trustor; $k$ represents the trustee; $v\in\{0,1\}$ ; $T_{a,k}$ represents the current trust $a$ has in $k$ ; $[{T_{a,k}}_{i-n},\ldots,{T_{a,k}}_{i}]$ is the list of the last $n$ values of $T_{a,k}$ .

5.1 Data sets

The proposed method was evaluated with the incremental protocol and different data sets, profiling approaches and recommendation filters. Specifically, we used the HotelExpedia and TripAdvisor data sets to perform the experiments. The data sets were ordered temporally and the events were processed sequentially without initial models.

HotelExpedia was gathered by Leal et al. (2017) [77]. After discarding the anonymous users, the resulting data set contains 50603 hotels, 1090 users and 214342 reviews from 10 different locations. Each user classified at least 10 hotels and each hotel contains at least 10 reviews. The average and standard deviation of the number of hotels rated per user is 197 $\pm$ 370 and reviews per hotel is 38 $\pm$ 97. This work uses the user and hotel identification and, as multi-criteria ratings, the overall, cleanliness, hotel condition, service, and room comfort.

TripAdvisor was collected by Wang et al. (2010) [78]. It includes 9114 hotels, 7453 users and 127517 hotel reviews. Each user classified at least 10 hotels and each hotel contains at least 10 reviews. The average and standard deviation of the number of hotels rated per user is 17 $\pm$ 209 and reviews per hotel is 14 $\pm$ 27. This work uses the user and hotel identification and, as multi-criteria ratings, the overall, value, rooms, location, cleanliness, service, and sleep quality.

5.2 Results

The results of the experiments are depicted in Table 5. They compare the average incremental accuracy, classification accuracy and processing time of the different approaches.

Table 5
Comparison of Memory- and Model-based approaches. The table contains the results of: (i) baseline method (PWRA); (ii) baseline method $+$ T&R post-filtering $+$ blockchain using PoA; and (iii) baseline method $+$ incremental T&R with look-back refinement $+$ blockchain using PoA. The results show a positive impact of incremental profiles in terms of Recall@10 and TRecall@10 in both recommendation models. Since our solution acts a posteriori, RMSE remains unchanged

Memory-based recommendation with HotelExpedia ( $k$ -NN)
Experiment	Consensus	CF profiling	PF	RMSE	Recall@10	TRecall@10	$\overline{\Delta t}$ (s)
1	–	PWRA	–	0.222	0.229	0.184	0.077
2	PoA	PWRA	$T\&R$	0.222	0.632	0.435	0.245
3	PoA	PWRA	$\bm{T_{S}\&R_{S}}$	0.222	0.660	0.450	0.284
		PWRA	$T_{R}\&R_{R}$	0.222	0.651	0.451	0.285
		PWRA	$T_{C}\&R_{C}$	0.222	0.660	0.450	0.289
Model-based recommendation with TripAdvisor (SVD-SDG)
Experiment	Consensus	CF profiling	PF	RMSE	Recall@10	TRecall@10	$\overline{\Delta t}$ (s)
1	–	PWRA	–	0.165	0.334	0.328	0.296
2	PoA	PWRA	$T\&R$	0.165	0.456	0.451	0.329
3	PoA	PWRA	$\bm{T_{S}\&R_{S}}$	0.165	0.655	0.649	0.386
		PWRA	${T_{R}\&R_{R}}$	0.165	0.654	0.649	0.388
		PWRA	$T_{C}\&R_{C}$	0.165	0.654	0.649	0.390

Collaborative filter (CF); Post-filter (PF); Look-back refinement: Statistical (S); Recursive (R); Sign changes (C).

The first set of experiments was performed with the HotelExpedia data set. The memory-based collaborative filter employs $k$ -Nearest Neighbours ( $k$ -NN). It adopts in all experiments a $k$ -NN filter (20 neighbours) with incremental updating and rating-based (PWRA) profiling. This filter corresponds also to the baseline method. The second experiment employs a post-filter supported by incremental trust and reputation (T&R) blockchain profiling with PoA consensus. The third experiment displays the post-filter results with the three refined incremental T&R blockchain profiling approaches.

The second set of experiments was performed with the TripAdvisor data set. The model-based filter employs Singular Value Decomposition with Stochastic Gradient Descent (SVD-SGD). Similarly, the first experiment corresponds to the baseline method, namely SVD-SGD model-based collaborative filter with incremental updating and rating-based (PWRA) profiling. Experiments two and three use this baseline method to generate predictions and apply four post-filter variants, corresponding to the incremental T&R and three refined incremental T&R blockchain profiles.

Table 6

T&R incremental profiling: Improvement with look-back refinement

Data set	Collaborative filter	R@10	TR@10	Latency
			(%)	(%)
Expedia	$k$ -NN	$+$ 4.43	$+$ 3.44	$+$ 15.92
TripAdvisor	SVD SGD	$+$ 43.64	$+$ 43.90	$+$ 17.33

As expected, the prediction accuracy remains unchanged since the collaborative filter is the same. Blockchain profiling slows the execution as it requires time to mine the blocks. In our Ethereum network with PoA the average mining time is 2 s per block.

Overall, the results of coupling a T&R post-filter to collaborative filters show that:

Incremental T&R reordering improves the quality of the default recommendations. Recall@10 increases 176% for HotelExpedia and 37% for TripAdvisor and the TRecall@10 136% for HotelExpedia and 38% for TripAdvisor.

Incremental T&R with Look-back refinement reordering improves further the quality of recommendations. Table 6 displays the improvement obtained with Eq. (9). HotelExpedia shows an approximate increase in Recall@10 and TRecall@10 of 4%, and TripAdvisor an increase in Recall@10 and in TRecall@10 of 44%.

Profiles, when stored as smart contracts, maintain their history immutable and traceable at the cost of time, and can be used to provide trustworthy explanations. When compared with the baseline method, the average increase in processing time per event was: (i) 2.2 and 2.7 times for the incremental and refined versions with memory-based filtering; and (ii) 0.11 and 0.30 times for the incremental and refined versions with model-based filtering.

Table 7

Chi-square test results

Data set	Collaborative filter	Significance
HotelExpedia	$k$ -NN	0.99
TripAdvisor	SVD-SGD	0.99

To statistically analyse the memory- and model-based results with and without look-back profiling, we performed a Chi-Square Test of independence. This test considered, as rows, the five profile categories, and as columns, the R@10, TR@10 and latency metric categories presented in Table 5. The five profiling approaches are the baseline, without post-filtering, and the four PoA blockchain profiling post-filtering variants, resulting in 8 degrees of freedom. These variants include, first, the incremental T&R profiling and, then, the statistical, recursive and sign changes look-back refinements. The Chi-Square Test examines, then, for each data set and collaborative filter, the relationship between the different types of profiles and the three evaluation metrics. Table 7 presents the Chi-Square Test results of the memory and model experiments, which show that, in both cases, the values of the metrics depend of the profiling variant.

5.3 Blockchain performance

Finally, we determined the latency and throughput of the implemented private Ethereum network using PoA. These experiments were repeated four times.

Table 8 displays the results in terms of average latency and throughput as well as the number of transactions on the private Ethereum network with both data sets.

Table 8
Ethereum performance results

Data set	Latency (s)	Throughput (t/s)	Transactions (#)
HotelExpedia	0.0045	222	3420000
TripAdvisor	0.0047	213	3281950

6. Conclusions

Crowdsourcing platforms rely on voluntary contributions, such as ratings, reviews or views, to generate recommendations. While research has shown that richer profiles improve the accuracy of the recommendations, these recommendations are mostly opaque to the end-user. To address this problem, this work explores the generation of trustworthy explanations supported by blockchain technology. Our proposal explores: (i) incremental trust and reputation profiling with look-back refinement; (ii) the storage of inter-user trust in a blockchain smart contract; and (iii) post-filtering as a means to improve and explain collaborative recommendations.

The proposed incremental updating method can be applied to any crowdsourced platform that provides recommendations. In order to evaluate our method, we have used HotelExpedia and TripAdvisor data sets to test and evaluate it using incremental RMSE, Recall@10 and TRecall@10 as evaluation metrics.

The post-filter reorders the recommendations, in the memory-based case, based on the trust between the active user and his neighbours as well as the system-wide reputation of his neighbours; and, in the model-based case, on the trust between the active user and his co-raters together with their derived system-wide reputation.

When we compare the incremental T&R profiling with and without look-back refinement, we can see a clear improvement of the recall-based metrics. This improvement was achieved exclusively by using the new look-back method to reorder the list of recommendations issued by the collaborative filter. In the case of the model-based filter, the increase was 40% higher than with memory-based. This can be explained by the fact that $k$ -NN already generates recommendations based on user similarities (the nearest neighbours), whereas SVD-SGD, being a matrix factorisation technique, does not maintain such a direct relationship between the active user and his co-raters. The RMSE does not improve because the collaborative filter remains unchanged as this method is intended to work with legacy opaque recommendation engines.

Concerning blockchain technology, and as a proof of concept, we have analysed the average latency and throughput per transaction and the average execution time per event. The blockchain latency was less than 5 ms/t with a throughput of at least 210 t/s. These values can be used as an indication of the expected blockchain performance when both the experiments and the blockchain are deployed within the same network. Nevertheless, the obtained latency was a good result for the Ethereum blockchain network. The average execution time per event is considerably higher with memory-based filtering and HotelExpedia (2.2 and 2.7 times higher for the incremental and refined versions) and higher with model based filtering and TripAdvisor (0.38 and 0.44 times higher for the incremental and refined versions). This indicates that the price to pay for a distributed, temporal, traceable and immutable database is latency.

To sum up, this paper presents a novel explanations module for crowdsourcing collaborative filters. This plugable module, supported by incremental trust and reputation profiling with look-back refinement, improves the quality of recommendations and explains the recommendations to the end-user.

As future work, we intend to identify and minimise the impact of malicious users (human or bots) in collaborative recommendation by exploring blockchain as the system-wide data gatekeeper. In this new role, it will hold individual profile smart contracts between each user and the platform, comprising all profile features, instead of just T&R. Finally, decision trees or neural networks classifiers, using single or ensemble models, can be used to predict and classify rater trustworthiness directly from the crowdsourced rating stream. However, to maintain data authenticity and traceability, those classifiers should be supported by blockchain solutions; this is another research direction.

Footnotes

geth.ethereum.org/.

solidity.readthedocs.io/.

docs.web3j.io/.

Acknowledgments

This work was partially financed by: (i) the ERDF – European Regional Development Fund through the Operational Programme for Competitiveness and Internationalisation – COMPETE 2020 Programme within project POCI-01-0145-FEDER-006961, and by National Funds through the Portuguese funding agency, FCT – Fundação para a Ciência e a Tecnologia, within project UIDB/50014/2020; (ii) the Xunta de Galicia (Centro singular de investigación de Galicia accreditation 2019–2022, also financed from ERDF); and (iii) the Irish Research Council within the framework of the EU ERA-NET CHIST-ERA project SPuMoNI: Smart Pharmaceutical Manufacturing www.spumoni.eu.

Appendix

Memory-based incremental trust and reputation

User generated event $r_{a,i}$

$I$ , $U$ , $s c$

Updates $T_{a,k}$ in the blockchain for all $k\in a.K$ ; Derives $R_{a,k}$ and $R_{k}$ for all $k\in a.K$

*[h]Update the trustworthiness of $k\in a.K$ , according to $a$ : $T_{a,k}$

() $k\in a.K$ (*[h]Item $i$ was recommended to $a$ due to $k$ ) $\textit{rater}(k,i)$ $\textit{sc.updateTrust}(a,k,1)$ $\textit{sc.updateTrust}(a,k,0)$

*[h]Update neighbour-based reputation of $k\in a.K$ , according to $a$ : $R_{a,k}$

() $k\in a.K$ $R_{a,k}\leftarrow 0$

$p\in a.P_{k}$ $R_{a,k}\leftarrow R_{a,k}+\textit{sc.getTrust}(a,p)$ $R_{a,k}\leftarrow{\displaystyle\frac{R_{a,k}}{\textit{size}(a.P_{k})}}$ *[h]Update system-wide reputation of $k\in a.K$ : $R_{k}$

() $k\in a.K$ $R_{k}\leftarrow 0$

$p\in k.P$ $R_{k}\leftarrow R_{k}+R_{p,k}$ $R_{k}\leftarrow{\displaystyle\frac{R_{k}}{\textit{size}(k.P)}}$

[h!t] Model-based: Incremental trust and reputationUser generated event $r_{a,i}$

$I$ , $U$ , $s c$ , $\mu_{R}$

Updates $T_{a,k}$ and $R_{k}$ for all $k\in a.C$ ; Recalculates $\mu_{R}$

$a.C\leftarrow a.C\cup i.R$ ; *[h]Update the set of co-raters of $a$ : $a . C$

$a.P\leftarrow a.P\cup getPertinents(a,i)$ *[h]Update the set of pertinent co-raters of $a$ : $a . P$ *[h]Update trustworthiness of $k\in a.C$ , according to $a$ : $T_{a,k}$

() $k\in a.C$ $k\in a.P$ $\textit{sc.updateTrust}(a,k,1)$ $\textit{sc.updateTrust}(a,k,0)$

*[h]Update system-wide reputation of $k\in a.C$ : $R_{k}$

() $k\in a.C$ $k.C\in i.R$ $k.C\leftarrow k.C\cup\{a\}$

$\delta r_{a,k}\leftarrow r_{a,i}-r_{k,i}$

$-10\%\leqslant\delta r_{a,k}\leqslant 10\%$ $k.P\leftarrow k.P\cup\{a\}$

() $c\in k.C$ $R_{k}\leftarrow 0$

$N_{R}\leftarrow 0$

() $T_{c,k}>\mu_{R}$

$R_{k}\leftarrow R_{k}+\textit{sc.getTrust}(c,k)$

$N_{R}\leftarrow N_{R}+1$

$R_{k}\leftarrow{\displaystyle\frac{R_{k}}{N_{R}}}$ *[h]Update global average reputation: $\mu_{R}$

$\mu_{R}\leftarrow 0$

() $u\in U$ $\mu_{R}\leftarrow\mu_{R}+R_{u}$

$\mu_{R}\leftarrow{\displaystyle\frac{\mu_{R}}{\textit{size}(U)}}$

*[h]Update the raters of $i$ : $i . R$

$i.R\leftarrow i.R\cup\{a\}$

*[h]Returns the subset of pertinent co-raters of $a$ due to $i$

FMainupdatePertinents FnFunction:

$a$ , $i$ $P_{a}\leftarrow\varnothing$

() $k\in i.R$ $\delta r_{a,k}\leftarrow r_{a,i}-r_{k,i}$

$-10\%\leqslant\delta r_{a,k}\leqslant 10\%$ $P_{a}\leftarrow P_{a}\cup\{k\}$

$P_{a}$

Appendix A. Algorithms

Algorithm Appendix describes the Memory-based Incremental Trust and Reputation profiling. The algorithm is invoked whenever a new rating event occurs and needs access to the current set of items $I$ , the current set of users $U$ and the trust smart contract $s c$ . In addition, $a . K$ represents the set of neighbours behind the last top $N$ items recommended to $a$ , $a.P_{k}$ is the set of pertinent neighbours between $a$ and $k$ , i.e., neighbours shared by $a$ and $k$ , and $k . P$ is the set of pertinent users of $k$ , those who share at least one neighbour with $k$ . Algorithm Appendix updates $T_{a,k}$ in the blockchain for all $k\in a.K$ and, then, derives $R_{a,k}$ and $R_{k}$ for all $k\in a.K$ .

Algorithm Appendix describes the Model-based Incremental Trust and Reputation profiling, which is triggered by the arrival of a rating event $r_{a,i}$ , where $i\in I$ is the rated item and $a\in U$ is the active user. It requires access to the current set of items $I$ , the current set of users $U$ , the trust smart contract $s c$ , and the average system-wide reputation $\mu_{R}$ . In this case, $i . R$ represents the set of raters of $i$ , $u . C$ is the current set of co-raters of $u$ , and $u . P$ is the current set of pertinent co-raters of $u$ . Algorithm Appendix incrementally updates $T_{a,k}$ in the blockchain for all $k\in a.C$ , derives $R_{k}$ for all $k\in a.C$ and recalculates $\mu_{R}$ .

Algorithm Appendix A. Algorithms describes the Trust Smart Contract. The smart contract maintains a hashtable with as many entries as users. In turn, each user entry keeps an individual hashtable holding trustee data. The smart contract implements two transactions: (i) updates the trust the trustor has in the trustee; and (ii) retrieves the current trust the trustor has in the trustee. In this incremental setup, the trustor corresponds to the active user $a$ and the trustees are, in the case of memory-based filters, neighbours and, in the case of model-based, co-raters.

Structstruct {}

[!t] Trust Smart Contract Store trust transaction: $a$ , $t$ , $T_{a,t}$ Retrieve trust transaction: $a$ , $t$

Store/retrieve $T_{a,t}$ in/from the smart contract

*[h]Trustee data structure

TrusteeDS $t$ *[h]trustee user

$n p r$ *[h]number of pertinent ratings

$n r$ *[h]number of ratings

$T_{a,t}$ *[h]Trust trustor $a$ has in trustee $t$

*[h]Trustor hashtable: holds the trustee data of all trustees of $a$

TrustorHT $a$ *[h]active trustor user mapping(uint => TrusteeDS) trustee;

*[h]User hashtable: aggregates the trustor hashtables of all users mapping(uint => TrustorHT) user;

*[h]Creates or updates $T_{a,t}$ in the user hashtable, where $v\in\{0,1\}$

FMainupdateTrust FnFunction:

$a$ , $t$ , $v$ $\textit{user}[a].a$ = NULL $\textit{user}[a].a\leftarrow a$

$\textit{user}[a].\textit{trustee}[t].t\leftarrow t$

$\textit{user}[a].\textit{trustee}[t].npr\leftarrow\textit{user}[a].\textit{% trustee}[t].npr+v$

$\textit{user}[a].\textit{trustee}[t].nr\leftarrow\textit{user}[a].\textit{% trustee}[t].nr+1$

$\textit{user}[a].\textit{trustee}[t].T_{a,t}\leftarrow{\displaystyle\frac{% \textit{user}[a].\textit{trustee}[t].npr}{\textit{user}[a].\textit{trustee}[t]% .nr}}$

*[h]Retrieves $T_{a,p}$ from the user hashtable

FMaingetTrust FnFunction:

$a$ , $t$ $T_{a,t}\leftarrow\textit{user}[a].\textit{trustee}[t].T_{a,t}$ $T_{a,t}$

References

Besnard

Cordier

Moinard

. Ontology-based inference for causal explanation. Integrated Computer-Aided Engineering. 2008; 15(4): 351-367.

Tintarev

Masthoff

. Explaining recommendations: Design and evaluation. in: Recommender systems handbook. 2nd ed. Boston: Springer US; Ricci

Rokach

Shapira

, eds. 2015. p. 479-510.

Zhang

Chen

. Explainable recommendation: A survey and new perspectives. Foundations and Trends© in Information Retrieval. 2020; 14(1): 1-101.

Hopfgartner

Kille

Heintz

Turrin

. Real-time recommendation of streamed data. in: RecSys’15. Vienna: ACM; 2015. p. 361-362.

Abdollahpouri

Adomavicius

Burke

Guy

Jannach

Kamishima

, et al. Multistakeholder recommendation: Survey and research directions. User Modeling and User-Adapted Interaction. 2020; 30: 127-158.

Bobadilla

Ortega

Hernando

Gutiérrez

. Recommender systems survey. Knowledge-Based Systems. 2013; 46: 109-132.

Sarwar

Karypis

Konstan

Riedl

. Item-based collaborative filtering recommendation algorithms. in: WWW’01. Hong Kong: ACM; 2001. p. 285-295.

Leal

Malheiro

Burguillo

. Trust and reputation modelling for tourism recommendations supported by crowdsourcing. in: WorldCIST’18. vol. 745 of Advances in Intelligent Systems and Computing. Naples: Springer; 2018. p. 829-838.

Di Francesco Maesa

Mori

. Blockchain 3.0 applications survey. Journal of Parallel and Distributed Computing. 2020; 138: 99-114.

10.

Leal

Veloso

Malheiro

González-Vélez

. Trust and reputation smart contracts for explainable recommendations. in: WorldCIST 2020. vol. 1159 of Advances in Intelligent Systems and Computing. Budva: Springer. 2020; p. 124-133.

11.

Leal

Malheiro

Burguillo

. Incremental hotel recommendation with inter-guest trust and similarity post-filtering. in: WorldCIST’19. vol. 930 of Advances in Intelligent Systems and Computing. Galicia: Springer; 2019. p. 262-272.

12.

Jiang

Adeli

. Fuzzy clustering approach for accurate embedding dimension identification in chaotic time series. Integrated Computer-Aided Engineering. 2003; 10(3): 287-302.

13.

Barredo Arrieta

Díaz-Rodríguez

Del Ser

Bennetot

Tabik

Barbado

, et al. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Information Fusion. 2020; 58: 82-115.

14.

Chen

Wang

Xie

Parsana

Soni

, et al. Towards explainable conversational recommendation. in: International Joint Conferences on Artificial Intelligence (IJCAI); 2020..

15.

Chen

Zhang

Cao

Qin

, et al. Personalized fashion recommendation with visual explanations based on multimodal attention network: Towards visually explainable recommendation. in: SIGIR’19. Paris: ACM; 2019. p. 765-774.

16.

Chen

Zhang

Cao

Qin

Zha

. Visually explainable recommendation. arXiv preprint arXiv: 180110288. 2018.

17.

Hou

Chen

Zheng

Liu

. Explainable fashion recommendation: A semantic attribute region guided approach. in: IJCAI 2019. Macao: Ijcai.org; 2019. p. 4681-4688.

18.

Tangseng

Okatani

. Toward explainable fashion recommendation. in: WACV’20. Colorado: IEEE; 2020. p. 2142-2151.

19.

Besbes

Baazaoui-Zghal

. Personalized and context-aware retrieval based on fuzzy ontology profiling. Integrated Computer-Aided Engineering. 2017; 24(1): 87-103.

20.

Yang

Zhao

Huang

Liu

. Community detection for proximity alignment. Integrated Computer-Aided Engineering. 2014; 21(1): 59-76.

21.

Gao

Wang

Xie

. Explainable recommendation through attentive multi-view learning. in: Proceedings of the AAAI Conference on Artificial Intelligence. 2019; 33: p. 3622-3629.

22.

Xian

Muthukrishnan

de Melo

Zhang

. Reinforcement knowledge graph reasoning for explainable recommendation. in: SIGIR’19. Paris: ACM; 2019. p. 285-294.

23.

Wang

Cao

Chua

. Explainable reasoning over knowledge graphs for recommendation. in: Proceedings of the AAAI Conference on Artificial Intelligence. 2019; 33: p. 5329-5336.

24.

Ren

Liang

Wang

de Rijke

. Social collaborative viewpoint regression with explainable recommendations. in: WSDM’17. Cambridge: ACM; 2017. p. 485-494.

25.

Chen

Zhang

Qin

. Dynamic explainable recommendation based on neural attentive models. in: AAAI 2019. Honolulu: AAAI Press; 2019. p. 53-60.

26.

Horsburgh

Craw

Massie

. Learning pseudo-tags to augment sparse tagging in hybrid music recommender systems. Artificial Intelligence. 2015; 219: 25-39.

27.

Damak

Nasraoui

. SeER: An explainable deep learning MIDI-based hybrid song recommender system. arXiv preprint arXiv: 190701640. 2019.

28.

Zhang

Cao

Jin

Wang

Liu

, et al. Jointly learning explainable rules for recommendation with knowledge graph. in: WWW’19. San Francisco: ACM; 2019. p. 1210-1221.

29.

McInerney

, et al. Explore, exploit, and explain: Personalizing explainable recommendations with Bandits. in: RecSys’18. Vancouver: ACM; 2018. p. 31-39.

30.

Leal

Malheiro

González-Vélez

Burguillo

. Trust-based modelling of multi-criteria crowdsourced data. Data Science and Engineering. 2017; 2(3): 199-209.

31.

Önder

Treiblmaier

, et al. Blockchain and tourism: Three research propositions. Annals of Tourism Research. 2018; 72(C): 180-182.

32.

Gunning

Aha

. DARPA’s explainable artificial intelligence (XAI) program. AI Magazine. 2019; 40(2): 44-58.

33.

Samek

. Learning with explainable trees. Nature Machine Intelligence. 2020; 2: 16-17.

34.

Samek

Montavon

Lapuschkin

Anders

Müller

. Explaining deep neural networks and beyond: A review of methods and applications. Proceedings of the IEEE. 2021; 109(3): 247-278.

35.

Ducange

Marcelloni

Pecori

. Fuzzy hoeffding decision tree for data stream classification. International Journal of Computational Intelligence Systems. 2021; 14: 946-964.

36.

Chen

Liu

Tang

. Multi-objective evolutionary triclustering with constraints of time-series gene expression data. Integrated Computer-Aided Engineering. 2019; 26(4): 399-410.

37.

Lara-Benítez

Carranza-García

García-Gutiérrez

Riquelme

. Asynchronous dual-pipeline deep learning framework for online data stream classification. Integrated Computer-Aided Engineering. 2020; 27(2): 101-119.

38.

da Costa

VGT

de Leon Ferreira de Carvalho

ACP

Barbon Junior

. Strict very fast decision tree: A memory conservative algorithm for data stream mining. Pattern Recognition Letters. 2018; 116: 22-28.

39.

Ferrari Dacrema

Parroni

Cremonesi

Jannach

. In: Critically examining the claimed value of convolutions over user-item embedding maps for recommender systems. New York, NY, USA: Association for Computing Machinery; 2020. p. 355-363.

40.

Cañete-Sifuentes

Monroy

Medina-Pérez

. A review and experimental comparison of multivariate decision trees. IEEE Access. 2021.

41.

Pereira

Piteri

Souza

Papa

Adeli

. FEMa: A finite element machine for fast learning. Neural Computing and Applications. 2020; 32(10): 6393-404.

42.

Alam

KMR

Siddique

Adeli

. A dynamic ensemble learning algorithm for neural networks. Neural Computing and Applications. 2020; 32(12): 8675-8690.

43.

Leal

Veloso

Malheiro

González-Vélez

Burguillo

. Scalable modelling and recommendation using wiki-based crowdsourced repositories. Electronic Commerce Research and Applications. 2019; 33: 100817.

44.

Ducange

Fazzolari

Marcelloni

. An overview of recent distributed algorithms for learning fuzzy models in Big Data classification. Journal of Big Data. 2020; 7.

45.

Torres

Galicia de Castro

Troncoso

Martínez-Álvarez

. A scalable approach based on deep learning for big data time series forecasting. Integrated Computer-Aided Engineering. 2018; 25: 1-14.

46.

Zhang

Rong

Paul

Neri

Pérez-Jiménez

. A complete arithmetic calculator constructed from spiking neural P systems and its application to information fusion. International Journal of Neural Systems. 2021; 31(1): 2050055.

47.

Rashvand

Ahmadzadeh

Shayegh

. Design and implementation of a spiking neural network with integrate-and-fire neuron model for pattern recognition. International Journal of Neural Systems. 2021; 31(3): 2050073.

48.

Cao

Zhang

Liu

Zhang

Neri

. Cloud-assisted secure eHealth systems for tamper-proofing EHR via blockchain. Information Sciences. 2019; 485: 427-440.

49.

Angelis

Aniello

Baldoni

Lombardi

Margheri

Sassone

. PBFT vs Proof-of-Authority: Applying the CAP theorem to permissioned blockchain. in: Proceedings of the Second Italian Conference on Cyber Security. vol. 2058 of CEUR Workshop Proceedings. CEUR-WS.org; Ferrari

Baldi

Baldoni

, eds. 2018.

50.

Wood

. Ethereum: A secure decentralised generalised transaction ledger. Parity.io; 2014. Petersburg version 41c1837: 2021-02-14. Online at: https://ethereum.github.io/yellowpaper/paper.pdf (Last Accessed: 5/Mar/21).

51.

Leal

Chis

González-Vélez

. Performance evaluation of private ethereum networks. SN Computer Science. 2020; 1(5): 285: 1-17.

52.

Leal

Chis

González-Vélez

. Multi-service model for blockchain networks. Information Processing & Management. 2021; 58(3): 102525.

53.

Dignum

. Responsible artificial intelligence: Designing AI for human values. ITU Journal: ICT Discoveries. 2017; 1(1): 1-8.

54.

Nam

Dutt

Chathoth

Khan

. Blockchain technology for smart city and smart tourism: Latest trends and challenges. Asia Pacific Journal of Tourism Research. 2019; 1-15.

55.

Casino

Dasaklis

Patsakis

. A systematic literature review of blockchain-based applications: Current status, classification and open issues. Telematics and Informatics. 2019; 36: 55-81.

56.

Szabo

. Formalizing and securing relationships on public networks. First Monday. 1997; 2(9).

57.

Singh

Parizi

Zhang

Choo

KKR

Dehghantanha

. Blockchain smart contracts formalization: Approaches and challenges to address vulnerabilities. Computers & Security. 2020; 88: 101654.

58.

Zheng

Xie

Dai

Chen

Weng

, et al. An overview on smart contracts: Challenges, advances and platforms. Future Generation Computer Systems. 2020; 105: 475-491.

59.

Liu

Han

. Blockchain technology and smart contract for civil structural health monitoring system. Computer-Aided Civil and Infrastructure Engineering. 2021; 36.

60.

Bellini

Iraqi

Damiani

. Blockchain-based distributed trust and reputation management systems: A survey. IEEE Access. 2020; 8: 21127-21151.

61.

Alowayed

Canini

Marcos

Chiesa

Barcellos

. Picking a partner: A fair blockchain based scoring protocol for autonomous systems. in: ANRW’18. Montreal: ACM; 2018. p. 33-39.

62.

Calvaresi

Mattioli

Dubovitskaya

Dragoni

Schumacher

. Reputation management in multi-agent systems using permissioned blockchain technology. in: WI’18. Santiago: ACM; 2018. p. 719-725.

63.

Zhao

Yang

. Secure Pub-Sub: Blockchain-based fair payment with reputation for reliable cyber physical systems. IEEE Access. 2018; 6: 12295-12303.

64.

Weng

Yang

Zhang

Hou

, et al. CrowdBC: A blockchain-based decentralized framework for crowdsourcing. IEEE Transactions on Parallel and Distributed Systems. 2019; 30(6): 1251-1266.

65.

Lisi

De Salve

Mori

Ricci

. A smart contract based recommender system. in: Economics of Grids, Clouds, Systems, and Services. Cham: Springer International Publishing; Djemame

Altmann

Bañares

JÁ

Agmon Ben-Yehuda

Naldi

, eds. 2019. p. 29-42.

66.

Cinque

Esposito

Russo

. Trust management in fog/ edge computing by means of blockchain technologies. in: IThings’18. Halifax: IEEE; 2018. p. 1433-1439.

67.

Sun

Lei

Qin

. A survey of blockchain technology on security, privacy, and trust in crowdsourcing services. World Wide Web. 2020; 23: 393-419.

68.

Buccafurri

Lax

Nicolazzo

Nocera

. Tweetchain: An alternative to blockchain for crowd-based applications. in: ICWE 2017. vol. 10360 of Lecture Notes in Computer Science. Rome: Springer; 2017. p. 386-393.

69.

Tang

Wang

. ZebraLancer: Private and anonymous crowdsourcing system atop open blockchain. in: ICDCS 2018. Vienna: IEEE; 2018. p. 853-865.

70.

Bhatia

Kumaraguru

Dubey

Buduru

Kaulgud

. WorkerRep: Building trust on crowdsourcing platform using blockchain. IIIT-Delhi; 2018.

71.

Leal

González-Vélez

Malheiro

Burguillo

. Profiling and rating prediction from multi-criteria crowd-sourced hotel ratings. in: ECMS’17. Budapest: ECMS; 2017. p. 576-582.

72.

O’Donovan

Smyth

. Trust in recommender systems. in: IUI’05. San Diego: ACM; 2005. p. 167-174.

73.

Takács

Pilászy

Németh

Tikk

. Scalable collaborative filtering approaches for large recommender systems. The Journal of Machine Learning Research. 2009; 10: 623-656.

74.

Cremonesi

Koren

Turrin

. Performance of Recommender Algorithms on Top-n Recommendation Tasks. in: RecSys’10. Barcelona: ACM; 2010. p. 39-46.

75.

Veloso

Malheiro

Burguillo

Foss

. Personalised fading for stream data. in: SAC’17. Marrakech: ACM; 2017. p. 870-872.

76.

Schäffer

, et al. Performance and scalability of private ethereum blockchains. in: BPM 2019. vol. 361 of Lecture Notes in Business Information Processing. Vienna: Springer; 2019. p. 103-118.

77.

Leal

Malheiro

Burguillo

. Prediction and analysis of hotel ratings from crowd-sourced data. in: WorldCIST 2017. vol. 570 of Advances in Intelligent Systems and Computing. Madeira: Springer; 2017. p. 493-502.

78.

Wang

Zhai

. Latent aspect rating analysis on review text data: A rating regression approach. in: KDD ’10. New York, NY, USA: ACM; 2010. p. 783-792.

Stream-based explainable recommendations via blockchain profiling

Abstract

Keywords

1. Introduction

2.1 Blockchain

2.2 History, traceability and transparency

2.3 Smart contracts

2.4 Contribution

Table 1 Top 3 explainable recommendations for an active user a

4. Proposed method

4.1 Profiling

4.1.1 Rating-based

Table 2 Aggregated explanation data

5. Experiments and results

Table 4 Blockchain API

5.2 Results

Table 8 Ethereum performance results

Footnotes

Acknowledgments

Appendix

Appendix A. Algorithms

References

Table 1
Top 3 explainable recommendations for an active user $a$

Table 2
Aggregated explanation data

Table 4
Blockchain API

Table 8
Ethereum performance results