Time Interest Network for Click-Through Rate Prediction
Hui Miao, Qinglei Zhou, Huawei Song, Fangjie Wan · 2022
In the era of the information explosion, it is very important to accurately predict the user's next behavior (such as browsing, collecting and commenting) through the user's characteristics. As a key research issue in the traffic distribution process of the Internet industry, CTR is of great significance to content recommendation and online advertising. However, most existing research ignores the intrinsic structure of user behavior traits: User behavior changes over time, leading to changes in user interest characteristics. Toward that, this paper proposes a new CTR model called click-through rate prediction (TIN) based on the network of temporal interest. The model first maps large-scale sparse input features into low-dimensional embedding vectors, the characteristics of the user's interest in each time window are then extracted by the multi-head self-attention mechanism, Bi-LSTM is then utilized to capture changes of interest between various time periods, and then, it is beneficial for the residual network to achieve gradient multiplication and backpropagation process, effectively avoiding the problem of gradient disappearance, and finally connecting to the multilayer perceptron (MLP) to learn the nonlinear relationship between features. Extensive experiments have conducted on the advertising dataset and the results have demonstrated that the proposed model is competitive with superior CTR estimation ability.