Intelligent Data Analysis: An International Journal

Real-world tasks often involve a continuous flow of new information that affects the learned theory, a situation that classical batch (one-step) learning systems are hardly suitable to handle. On the contrary, incremental (also called "on-line") techniques are able to deal with such a situation by exploiting refinement operators. In many cases deep knowledge about the world is not available: Either incomplete information is available at the time of initial theory generation, or the nature of the concepts evolves dynamically. The latter situation is the most difficult to handle since time evolution needs to be considered. This work presents a new approach to learning in presence of concept drift, and in particular a special version of the incremental system INTHELEX purposely designed to implement such a technique. Its behavior in this context has been checked and analyzed by running it on two different datasets.

A knowledge acquisition method Ripple Down Rules (RDR) can directly acquire and encode knowledge from human experts. It is an incremental acquisition method and each new piece of knowledge is added as an exception to the existing knowledge base. Past researches on RDR method assume that the problem domain is stable. This is not the case in reality, especially when an environment changes. Things change over time. This paper proposes an adaptive Ripple Down Rules method based on the Minimum Description Length Principle aiming at knowledge acquisition in a dynamically changing environment. We consider both the change in class distribution on a domain and the change in knowledge source as typical changes in the environment. When class distribution changes, some pieces of knowledge previously acquired become worthless, and the existence of such knowledge may hinder acquisition of new knowledge. In our approach knowledge deletion is carried out as well as knowledge acquisition so that useless knowledge is properly discarded. To cope with the change in knowledge source, RDR knowledge based systems can be constructed adaptively by acquiring knowledge from both domain experts and data. By incorporating inductive learning methods, knowledge can be acquired (learned) even when only either data or experts are available by switching the knowledge source from domain experts to data and vice versa at any time of knowledge acquisition. Since experts need not be available all the time, it contributes to reducing the cost of personnel expenses. Experiments were conducted by simulating the change in knowledge source and the change in class distribution using the datasets in UCI repository. The results show that it is worth following this path.

In many real-world learning problems the data flows continuously and learning algorithms should be able to respond to this circumstance: the induced concept description should gradually change over time. In this paper, we outline some existing incremental learners based on the theory of Formal Concept Analysis: FCA. Then, we introduce a new learning approach that improves incremental concept formation. This approach has the advantage of handling both the problem of data addition, data deletion, data update, attribute addition and attribute deletion. Finally, we apply the proposed approach to the problem of cancer diagnosis. We measure the effect of incrementality on the quality of the discovered rules using cross-validation.

For many learning tasks where data is collected over an extended period of time, its underlying distribution is likely to change. A typical example is information filtering, i.e. the adaptive classification of documents with respect to a particular user interest. Both the interest of the user and the document content change over time. A filtering system should be able to adapt to such concept changes. This paper proposes several methods to handle such concept drifts with support vector machines. The methods either maintain an adaptive time window on the training data [13], select representative training examples, or weight the training examples [15]. The key idea is to automatically adjust the window size, the example selection, and the example weighting, respectively, so that the estimated generalization error is minimized. The approaches are both theoretically well-founded as well as effective and efficient in practice. Since they do not require complicated parameterization, they are simpler to use and more robust than comparable heuristics.

Experiments with simulated concept drift scenarios based on real-world text data compare the new methods with other window management approaches. We show that they can effectively select an appropriate window size, example selection, and example weighting, respectively, in a robust way. We also explain how the proposed example selection and weighting approaches can be turned into incremental approaches. Since most evaluation methods for machine learning, like e.g. cross-validation, assume that the examples are independent and identically distributed, which is clearly unrealistic in the case of concept drift, alternative evaluation schemes are used to estimate and optimize the performance of each learning step within the concept drift handling frameworks as well as to evaluate and compare the different frameworks.

In this review we explore the topic of sequential learning, where information to be learned and retained arrives in separate episodes over time, in the context of artificial neural networks. Most neural networks handle this kind of task very badly, as new learning completely disrupts information previously learned by the network. This problem, known as "catastrophic forgetting", has received a lot of attention in the literature. We illustrate the catastrophic forgetting effect, and summarise possible solutions. In particular, we review the literature relating to the pseudorehearsal mechanism, which is an effective solution to the catastrophic forgetting problem in back propagation type networks. We then review similar issues of capacity, forgetting, and the use of pseudorehearsal in Hopfield type networks. Finally, we briefly discuss these issues in the context of cognition, and summarise interesting topics for further research.