Detect Professional Malicious User With Metric Learning in Recommender Systems

In e-commerce, online retailers are usually suffering from professional malicious users (PMUs), who utilize negative reviews and low ratings to their consumed products on purpose threaten the for illegal profits. PMUs difficult be detected because they masking strategies disguise themselves as normal users. Specifically, there three challenges PMU detection: 1) do not conduct any abnormal or interactions (they never concurrently leave too many at same time), themselves. Therefore, conventional outlier detection methods confused by strategies. 2) model should take both into consideration, which makes a multi-modal problem. 3) no datasets with labels in public, an unsupervised learning To this end, we propose model: MMD, employs Metric Malicious Detection reviews. MMD first utilizes modified RNN project informational review sentiment score, jointly considers Then user profiling (MUP) is proposed catch gap between scores ratings. MUP filters builds candidate set. We apply metric learning-based clustering learn proper matrix detection. Finally, can labeled detect PMUs. attention mechanism improve model’s performance. The extensive experiments four demonstrate that our method solve Moreover, performance of state-of-the-art recommender models enhanced taking preprocessing stage.