Combining feature selectors in a product advertisement classification system

Automated product advertisement classification can be utilized in various applications, such as category-specific search. The advertisements on the web contain several kinds of media sources that can be used for classification and the text information may be the most important source to indicate the semantic meaning of a given product advertisement. The text information of a product advertisement is typically noisy and the noisy words can be removed with some feature selection methods. These methods measure the goodness of a word from different perspectives and we think that the combination of them can improve the classification accuracy. We present a two-step algorithm to combine the feature selectors in this paper. The algorithm first intersects two global feature selection results and then performs a local feature selection. We evaluate it on a product advertisement dataset, which contains 3910 products of 100 categories crawled from the “amazon” website and we extract three kinds of textual information for classification. The experimental results show that our algorithm is superior to the existed combination method with a up to 0.019 Macro-F1 improvement.