Unsupervised Feature Weighting Based on Local Feature Relatedness

Feature weighting plays an important role in text clustering. Traditional feature weighting is determined by the syntactic relationship between feature and document (e.g. TF-IDF). In this paper, a semantically enriched feature weighting approach is proposed by introducing the semantic relationship between feature and document, which is implemented by taking account of the local feature relatedness — the relatedness between feature and its contextual features within each individual document. Feature relatedness is measured by two methods, document collection-based implicit relatedness measure and Wikipedia link-based explicit relatedness measure. Experimental results on benchmark data sets show that the new feature weighting approach surpasses traditional syntactic feature weighting. Moreover, clustering quality can be further improved by linearly combining the syntactic and semantic factors. The new feature weighting approach is also compared with two existing feature relatedness-based approaches which consider the global feature relatedness (feature relatedness in the entire feature space) and the inter-document feature relatedness (feature relatedness between different documents) respectively. In the experiments, the new feature weighting approach outperforms these two related work in clustering quality and costs much less computational complexity.