On incorporating side information into linear recommender systems using adversarial training
Authors: Dora Jambor and Putra Manggala
Presented as a poster at WiML'17 and SOCML'17.
Abstract from the paper:The abundance of side information associated with items in ecommerce applications has provided numerous new possibilities for enriching recommender systems with rich sources of information. A variety of methods have shown positive effects of using side information on the recommendations’ quality, often optimized with accuracy in mind.
In real-world applications, it can be desirable to incorporate different aspects of quality into recommendations such as domain knowledge, diversity, and fairness. A common challenge when optimizing for such qualities is to find an optimal trade-off behavior between accuracy and other metrics.
In this paper, we propose a new framework, AdRec, for incorporating side information into linear recommender systems via adversarial training. The adversarial training along with some proposed tunable hyperparameters allows practitioners to smoothly control the level of side information influence, and thereby achieve a desired trade-off profile. We evaluate AdRec on two datasets with manually generated side information to illustrate this method. We show how AdRec is generally more performant compared to popular baseline in terms of producing an optimal trade-off behavior. Lastly, we discuss some practical considerations for AdRec training.