An Unbiased User Model for Interest Diffusion in the Heterogeneous Network Recommendation

Interest diffusion in a bipartite network has been verified effective for personality recommendation. User nodes and item nodes compose a heterogeneous bipartite network. Treating user attitude toward items as interest resource and allocating it in the heterogeneous network along the linkage make it possible to recommend the most likely interested items to target users. However, the way it models user interest and quantizes interest lead some bias into the algorithm. Such as, it adopts the positive attitude into the interest diffusion process but without seeing that negative attitude also reflect user flavor and should be introduced into the computation. To overcome the drawbacks, this paper proposed an unbiased user interest model. Firstly, the unbiased model considers both positive and negative attitudes into user interest to eliminate the negative feedback discrimination. Secondly, it adopts a new criterion to divide interest from a perspective of statistic optimization. Last, it treats positive interest and negative interest equally to lower the rating bias by a symmetric formula when quantizing and weighting interest. Extensive experiments conducted on real data validate the advantage of the proposed model even with the same following diffusion process as the baseline algorithm.