Feature Selection on Heterogeneous Graph

Heterogeneous graph based information recommendation have been proved useful in recent studies. Given a heterogeneous graph scheme, there are many possible meta paths between the query node and the result node, and each meta path addresses a hypothesis-based ranking function. In prior researches, meta paths are manually selected by domain experts. However, when the graph scheme becomes complex, this method can be inefficient. In this study, we propose feature generation tree, a novel feature selection method for heterogeneous graph mining based recommendation algorithms, which adds graph structure information into the original “feature selection for ranking” algorithm and saves a fair amount of time for feature computation. In our preliminary experiment, the proposed method outperforms the original “feature selection for ranking” algorithm in both efficiency and effectiveness.