The Item Category-Oriented User Behavior Path Analysis in CTR Prediction

In E-commerce, Click-Through Rate (CTR) prediction models are often intended for analyzing user behavior patterns, identifying potential user interests, and therefore estimating purchasing intent. Existing models often incorporate various types of interactions between users and items/products (referred to as multi-type user behavior) to address the issue of data sparsity caused by relying on a single type of behavior. While this method introduces more user behavior features, it also brings some irrelevant noise into the prediction model—information not directly related to the predicted item. To address this, we introduce the concept of the Item Category-oriented User behavior Path, or ICUP, which consist of behaviors such as clicks, add-to-cart actions, and favorites, all focused on items within the same category. These behavior paths not only reflect the user’s potential interest in the target category of items but also implicitly indicate the user’s thought processes and decision-making throughout the purchasing process. We employ an adaptive mechanism for depicting the unique characteristics of user behavior preferences. Finally, a contrastive learning mechanism is integrated into the feature representation of behavior paths, enhancing the accuracy of the feature representation. We conducted a series of experiments on the publicly available Taobao dataset, and the results verify that our CTR prediction method yields desirable improvements in AUC and accuracy compared to baseline models.