Hypergraph-based spectral clustering for categorical data

Clustering categorical data has attracted much attention in recent years. In this paper, a hypergraph-based spectral clustering algorithm is proposed for categorical data. Firstly, we convert the categorical data to market basket type data by modeling each instance with categorical attributes as a transaction. By using an itemset counting algorithm, a set of patterns (i.e. frequent itemsets) can be discovered. Then we represent each transaction as a set of these patterns. In the hypergraph model, each transaction is represented as a vertex, and each pattern is regarded as a hyperedge. A hyperedge represents an affinity among subsets of transactions and the weight of the hyperedge reflects the strength of the affinity. At last a hypergraph-based spectral clustering algorithm is used to find the clustering results. Experimental results for selected UCI datasets show the effectiveness of the proposed algorithm.