*= Equal Contributors
Advice techniques in large-scale on-line marketplaces are important to aiding customers in discovering new content material. Nonetheless, state-of-the-art techniques for item-to-item suggestion duties are sometimes based mostly on a shallow degree of contextual relevance, which might make the system inadequate for duties the place merchandise relationships are extra nuanced. Contextually related merchandise pairs can typically have problematic relationships which are complicated and even controversial to finish customers, and so they might degrade person experiences and model notion when beneficial to customers. For instance, the advice of a guide about one sports activities group to somebody studying a guide about that group’s greatest rival may very well be a nasty expertise, regardless of the presumed similarities of the books. On this paper, we suggest a classifier to determine and forestall such problematic item-to-item suggestions and to boost general person experiences. The proposed strategy makes use of lively studying to pattern onerous examples successfully throughout delicate merchandise classes and employs human raters for information labeling. We additionally carry out offline experiments to exhibit the efficacy of this method for figuring out and filtering problematic suggestions whereas sustaining suggestion high quality.