Adaptive self-organized maps based on bidirectional approximate reasoning and its applications to information filtering

摘要

Similarity measuring is one substantial part in self-organizing maps (SOM) for its direct influence on the mapping results. The common used similarity measuring method – Euclidean distance cannot always express the exact similarity. In this paper, a novel adaptive self-organized maps based on bidirectional approximate reasoning (ASOMBAR) is proposed to improve the competitive and cooperative process based on the similarity measuring. Unlike the SOM, the proposed ASOMBAR employs the novel fuzzy similarity distance and fuzzy matching criterion to replace the Euclidean distance and original matching criterion, respectively. The fuzzy similarity distance describes the similarity relation more precisely than the Euclidean distance does. The fuzzy matching criterion pays more attention on the large weighted elements and less emphasis on the small weighted elements. Moreover, the new compatible topological neighborhood is also modified basing on the new fuzzy similarity distance and fuzzy matching criterion. Since the ASOMBAR network is self-organizing, the weights of the networks change adaptively according to the input changes. Compared with the well-known growing neural gas (GNG) and SOM, ASOMBAR (when λ ⩾ 0.5) converges quicker to a smaller distortion error. An information filtering example is used to show the effectiveness of ASOMBAR.