PROJECT TITLE :

Meta-Wrapper Differentiable Wrapping Operator for User Interest Selection in CTR Prediction

ABSTRACT:

The click-through rate (CTR) prediction, also known as CTR prediction, is becoming an increasingly important component of recommender systems. CTR prediction attempts to determine the likelihood that a user will click on a particular item. In recent times, a number of deep learning models that are equipped with the capacity to automatically extract the user's interests based on his or her behaviors have achieved a great deal of success. In these works, the attention mechanism is utilized to select the user-interested behaviors from the user's historical behavior data, with the end goal of improving the CTR predictor's overall performance. In a normal situation, these attentive modules can normally be jointly trained with the base predictor through the utilization of gradient descents. In this paper, we take the approach of viewing user interest modeling as a problem of feature selection, which we refer to as user interest selection. Within the context of the wrapper method, we suggest a novel approach that we refer to as the Meta-Wrapper for dealing with this kind of issue. To be more specific, we employ a differentiable module as our wrapping operator and then recast its learning problem as a continuous bilevel optimization. This allows us to achieve our desired results. In addition, we solve the optimization problem by employing a meta-learning algorithm, and we theoretically demonstrate that it converges. In the meantime, we also provide theoretical analysis to show that our proposed method 1) achieves better resistance to overfitting and 2) improves the efficiency of the wrapper-based feature selection. In conclusion, extensive tests conducted on three different publicly available datasets demonstrate that our method is superior in terms of improving the performance of CTR prediction.