| Summary: | 碩士 === 國立雲林科技大學 === 資訊管理系碩士班 === 93 === The visualization-induced SOM (ViSOM) is a non-linear multi-dimensional projection method, extended from self-organizing map (SOM). It overcomes the drawbacks that the structure of the clusters may not be apparent and the nodes often spread around the 2-D map in the SOM. The objective of the ViSOM is to preserve the data structure as well as the topology as faithfully as possible. Even so, it still cannot express reasonably the distance or similarity of categorical data and preserve the structure of categorical data. In this study, the extended ViSOM is proposed to overcome these shortcomings. We utilize the concept hierarchies to define and calculate the distance of categorical values and preserve the structure of mixed data as well as the topology of trained EViSOM map as faithfully as possible. In addition, we perform clustering based on the output map generated by the network and evaluate the clustering result. Experimental results on two synthetic and two real datasets demonstrate that the proposed clustering algorithm is able to cluster mixed data better than the traditional SOM and ViSOM do. In addition, our algorithm better reveals the cluster structure and the clustering quality than traditional approaches, with respect to manual or automatic clustering of the trained map.
|
|---|
