FreshGraph: A Spam-Aware Recommender System for Cold Start Problem

摘要

Recommender systems provide personalized recommendation to help users levitating from information overload. Collaborative filtering based recommendation methods are playing a dominant role in the industry because of its versatility and simplicity. However, its performance suffers from sparse data, and being less effective in cold-start problem settings. In real world scenario, when users are recommended with items, it is very easy to overwhelm the target users with impersonalized information, which drives away valuable audience. In this paper, we propose a two-steps spam aware recommendation framework to effectively recommend new items to target users. By utilizing heterogeneous information graph structure, we first use item-user Meta-Path similarity measure for user candidate selection. Then we use entropy encoding measurement to identify false positive from candidate list to prevent possible spam from happening. The proposed method leverages the semantic information that persists inside the graph structure, which not only considers item content features, but also take user activeness into account for more effective audience targeting. The proposed method produces an explainable top-K user list for the new item, while K is a trailed number to each given item individually. Meanwhile, the proposed method is also adaptive to data change overtime, while capable of processing requests in a real-time fashion.