Style Controlling in Recommendation
摘要
Practical recommender systems often possess different styles, which reflect the general rules and unique values that the system designers attempt to highlight. Conventional systems mostly rely on hard rules to constrain or rerank the results for accomplishing expected styles. However, these empirical rules may be complicated, unstructured, having conflicts and entanglements that even harm the accuracy. In this work, we propose a simple Style controller (StyCon) paradigm to jointly encode four typical styling rules in training via uniformly formatted pair-wise losses, verifying the effectiveness of StyCon in both offline and online.