Deep Session Interest Network for Click-Through Rate Prediction (1905.06482v1)

Abstract: Click-Through Rate (CTR) prediction plays an important role in many industrial applications, such as online advertising and recommender systems. How to capture users' dynamic and evolving interests from their behavior sequences remains a continuous research topic in the CTR prediction. However, most existing studies overlook the intrinsic structure of the sequences: the sequences are composed of sessions, where sessions are user behaviors separated by their occurring time. We observe that user behaviors are highly homogeneous in each session, and heterogeneous cross sessions. Based on this observation, we propose a novel CTR model named Deep Session Interest Network (DSIN) that leverages users' multiple historical sessions in their behavior sequences. We first use self-attention mechanism with bias encoding to extract users' interests in each session. Then we apply Bi-LSTM to model how users' interests evolve and interact among sessions. Finally, we employ the local activation unit to adaptively learn the influences of various session interests on the target item. Experiments are conducted on both advertising and production recommender datasets and DSIN outperforms other state-of-the-art models on both datasets.

• The paper introduces DSIN, which segments user behavior into sessions to improve CTR prediction accuracy.
• It employs a self-attention mechanism with bias encoding and Bi-LSTM to extract and model sequential user interests.
• Empirical evaluations demonstrate DSIN's superiority over models like DIN and DIEN on large-scale datasets.

Deep Session Interest Network for Click-Through Rate Prediction

The paper, "Deep Session Interest Network for Click-Through Rate Prediction," presents a sophisticated approach to enhancing Click-Through Rate (CTR) prediction by integrating the inherent session-based structure of user behavior sequences in recommender systems. The focus on accurately modeling user interests derived from sequential behaviors marks an innovative deviation from traditional CTR prediction models.

Core Contributions

The authors introduce the Deep Session Interest Network (DSIN), a novel model aimed at addressing the limitations of existing CTR models by explicitly considering the session-based organization of user behaviors. The work highlights key innovations:

1. Session Segmentation: DSIN efficiently divides user behavior sequences into sessions based on temporal proximity, capturing homogeneous behavioral patterns within sessions and recognizing the variance across different sessions.
2. Self-Attention Mechanism with Bias Encoding: This component extracts user interests within each session by employing a multi-head self-attention mechanism enhanced with bias encoding. The bias adds a dimension to capture the sequential order and allows for differentiation between representation subspaces.
3. Bi-LSTM for Session Interaction: To handle the sequential dependency between session interests, DSIN utilizes Bi-directional LSTM (Bi-LSTM) which encapsulates the interaction and evolution of session interests over time, thus enhancing the expressive power of interest dynamics.
4. Local Activation Unit: This unit adaptively weights session interests concerning the target item, ensuring relevant interests are prioritized in the final user interest representation.

Strong Numerical Results

The DSIN model demonstrates superior performance in comparison to state-of-the-art CTR models, such as YoutubeNet, DIN, and DIEN. Empirical evaluations conducted on advertising and large-scale recommender datasets indicate DSIN's superiority, with noted improvement in AUC metrics—affirming its efficacy in industrial scenarios, notably in platforms like Alibaba.

Implications and Future Directions

The introduction of DSIN and its architectural innovations significantly impact both practical implementations and theoretical understanding of CTR predictions. Practically, DSIN's adeptness at capturing session-based user interests can directly translate into increased accuracy of recommendation systems in online advertising and retail applications.

From a theoretical perspective, DSIN offers insights into user interest modeling, prompting future exploration into context-aware and knowledge-graph-enhanced CTR models. Future developments could look into incorporating richer contextual data or knowledge graphs to refine the representation of user interests, moving towards more explainable AI models in recommender systems.

Overall, this paper adds substantial value to the literature on CTR prediction by presenting a structured mechanism to model user interests through a session-based approach. The integration of advanced neural architectures and attention mechanisms sets a promising benchmark for future research in this domain.