Explainable Item Valuation for Recommendation Algorithms: Framework, Acceleration, and Explainability

Description

Personalized recommender systems have gained substantial attention in modern society due to their ability to alleviate information and choice overload, cater to individual user preferences, and enhance the revenue of service and product providers. Despite being helpful and valuable, training an industrial recommender system is still cumbersome, which requires tons of high-quality data and tedious training time.  Although recent studies have managed to identify high-quality and valuable user-item interactions in recommendation models, there are still several challenges: 1) lacking a model-agnostic valuation framework to measure the importance of user-item interactions; 2) lacking effective approaches to accelerate the valuation process; 3) lacking a general scheme to explain the generated data values. To tackle the aforementioned challenges, we will conduct systematic investigations on the item valuation for recommendation algorithms. Generally, our project consists of three major tasks. 1)We plan to design a general item valuation framework for recommendation algorithms. We will build a bilevel optimization framework to assign continuous/discrete values to each item that users have accessed. Due to the complexity of unrolled backpropagation, we will resort to the implicit function theorem to simplify the computation. 2)We plan to accelerate the valuation by advanced techniques. The bilevel optimization typically involves the calculation of a huge Hessian matrix, which is tedious and hardly realistic in the real world. To speed up the process, we will design policy gradient-based methods for the discrete data values and equilibrium propagation for continuous ones. 3) We plan to incorporate the explainability into the generated data values. We will build a model-agnostic scheme with Shapley values to explain the generated data values. Moreover, we will investigate training a surrogate model to facilitate the computation of Shapley values. With the support from industrial collaborators, our outcomes will establish a practical foundation for the analysis of data valuation in recommender systems, provide effective item valuation techniques, enhance the training efficiency of recommender systems, and push forward the scientific frontier of this research area.