Managing Strategies for a Self-Organising Map based Filtering Agent Community
In a time of continuously growing Internet with more and more useless information, an efficient ability of information filtering is desperately needed. Self Organizing Maps provide the opportunity to support this information retrieval. These Self Organizing Maps are a type of unsupervised learning. SOMs align themselves to represent given documents in their lattices of neurons, which are called map units. Each map unit represents a collection of similar documents. More related documents can be found in the neighborhood of such a map unit. The disadvantage of being too slow can be countervailed with combining several improvements. In this thesis these improvements are extracted from existing SOMs and combined in one implementation. These improvements speed up the learning process of this neural net in a significant way. An agent-based filtering community was set up where a manager agent administrates several SOM based filtering agents from the viewpoint of resources saving. The ability of data mining can be sustained while new information is learned by a SOM based filtering agent. Further it is shown how to find the best matching documents for filter requests and how to eliminate duplicates without much additional cost.