Jian Wang and David Hardtke
In the 24th International World Wide Web Conference (WWW 2015)
Downside management is an important topic in the field of recommender systems. User satisfaction increases when good items are recommended, but satisfaction drops significantly when bad recommendations are pushed to them. For example, a parent would be disappointed if violent movies are recommended to their kids and may stop using the recommendation system entirely. A vegetarian would feel steakhouse recommendations useless. A CEO in a mid-sized company would feel offended by receiving intern-level job recommendations. Under circumstances where there is penalty for
a bad recommendation, a bad recommendation is worse than no recommendation at all. While most existing work focuses on upside management (recommending the best items to users), this paper emphasizes on achieving better downside management (reducing the recommendation of irrelevant or offensive items to users). The approach we propose is general and can be applied to any scenario or domain where downside management is key to the system.
To tackle the problem, we design a user latent preference model to predict the user preference in a specific dimension, say, the dietary restrictions of the user, the acceptable level of adult content in a movie, or the geographical preference of a job seeker. We propose to use multinomial regression as the core model and extend it with a hierarchical Bayesian framework to address the problem of data sparsity. After the user latent preference is predicted, we leverage it to filter out downside items. We validate the soundness of our approach by evaluating it with an anonymous job application dataset on LinkedIn. The effectiveness of the latent preference model was demonstrated in both offline experiments and online A/B testings. The user latent preference model helps to improve the VPI (views per impression) and API (applications per impression) significantly which in turn achieves a higher user satisfaction.