Abstract:
Deep recommender systems are essential for providing personalized recommendations in various domains, such as e-commerce, social media, and entertainment. In deep recomme...Show MoreMetadata
Abstract:
Deep recommender systems are essential for providing personalized recommendations in various domains, such as e-commerce, social media, and entertainment. In deep recommender systems, feature selection plays a vital role as it identifies the features that are the most informative for predicting user preferences. However, most existing deep recommender systems are designed without a systematic approach to feature selection. They typically feed all available features into their sophisticated neural networks, or experts choose features manually or employ existing feature selection algorithms. These approaches might potentially undermine the accuracy and effectiveness of recommender systems, since they execute feature selection separately from the subsequent model of the recommender system, without considering the model’s prediction behavior. Moreover, existing feature selection methods tend to select a fixed set of features, which is not adaptable to the dynamic and complex environments of practical recommender systems, where the importance of a specific feature can vary across user-item interactions. To address these challenges, we propose a novel adaptive feature selection framework, Probabilistic and Adaptive Feature Selection in Deep Recommendation Systems (ProAdaFS), for deep recommender systems. ProAdaFS leverages the power of two existing adaptive feature selection techniques (AdaFS and AutoField) with significant modifications to enhance feature selection. To identify the most informative features corresponding to a subsequent recommendation task model, we design a network controller that dynamically and adaptively adjusts the probability of selecting a feature field, generates scores and re-evaluates feature fields to identify informative features. Our experiments were conducted on two real-world e-commerce recommender systems datasets. The experimental results demonstrate the effectiveness of ProAdaFS in improving the feature selection process in deep recomme...
Date of Conference: 17-19 January 2024
Date Added to IEEE Xplore: 03 July 2024
ISBN Information:
Print on Demand(PoD) ISSN: 1976-7684