Deep Target Session Interest Network for Click-Through Rate Prediction
Hongjiang Zhong, Junchao Ma, Xiongbao Duan, Shuting Gu, Junmei Yao · 2024
Click-Through Rate (CTR) prediction is one of the core tasks in the recommendation system (RS). In the feature presentation of CTR prediction model, user behavior sequences contain temporal features and implicit interests, which is used for sequential user behavior modeling. There are many deep CTR models on sequential user behavior modeling, while most of them overlook session-based and target-relevant interests. User behavior sequences are composed of multiple sessions, with highly homogeneous user interests within the same session, while user interests across different sessions are heterogeneous. Motivated by the above discoveries, we propose a novel Deep Target Session Interest Network (DTSIN) in this paper. We divide user behavior sequences into multiple original sessions, and append the target item at the end of each original session to obtain target session. Multi-head self-attention is employed to extract target-relevant interests from target sessions. Then, we apply splitting and mean pooling techniques to generate session interests and target interests. Bi-directional GRU (Bi-GRU) is leveraged to capture sequential relationships between session interests. Subsequently, Target Attention (TA) mechanism is used to reallocate the target-related attention scores for session interests, target interests and hidden states. Finally, different features are concatenated and flattened, which is fed into the Multiple Layer Perceptron (MLP) layer for final CTR prediction. Experimental results on the datasets indicate the performance of DTSIN is superior to other state-of-the-art deep CTR models.