Collaborative Contrastive Network for Click-Through Rate Prediction (2411.11508v1)

Abstract: E-commerce platforms provide entrances for customers to enter mini-apps to meet their specific shopping needs. At the entrance of a mini-app, a trigger item recommended based on customers' historical preferences, is displayed to attract customers to enter the mini-app. Existing Click-Through Rate (CTR) prediction approaches have two significant weaknesses: (i) A portion of customer entries is driven by their interest in the mini-app itself rather than the trigger item. In such cases, approaches highly hinging on the trigger item tend to recommend similar items, thus misunderstanding the customers' real intention; (ii) Approaches that consider customers' intention toward mini-apps, require the regular existence of mini-apps for customers to cultivate routine shopping habits, making such approaches less robust for mini-apps that are available for only short periods (1 or 3 days) in Explosive Promotional Scenarios (EPS), such as the Black Friday and China's Double 11 Shopping Carnival. To address the above-mentioned issues, we introduce a more general and robust CTR prediction approach, dubbed Collaborative Contrastive Network (CCN). Given a user, CCN learns to identify two item clusters that can represent the user's interests and disinterests, via leveraging the collaborative relationship of co-click/co-non-click or the non-collaborative relationship of mono-click as the supervision signal for contrastive learning. This paradigm does not need to explicitly estimate user's binary entry intention and avoids amplifying the impact of the trigger item. Online A/B testing on large-scale real-world data demonstrates that CCN sets a new state-of-the-art performance on Taobao, boosting CTR by 12.3% and order volume by 12.7%.

• The paper presents a novel CTR prediction approach that employs dual attraction and repulsion losses to refine user interest modeling.
• It integrates collaborative contrastive learning by leveraging co-click, co-non-click, and mono-click signals for precise interest extraction.
• Experiments on Taobao demonstrate a 12.3% CTR increase and 12.7% order boost, confirming its effectiveness in explosive promotional scenarios.

A Study on Collaborative Contrastive Network for Click-Through Rate Prediction

The paper "Collaborative Contrastive Network for Click-Through Rate Prediction" explores optimizing recommendation systems for e-commerce platforms, specifically within mini-app contexts. The authors acknowledge the limitations of existing CTR prediction approaches, particularly those dependent on trigger items, and propose an innovative solution: the Collaborative Contrastive Network (CCN). This model seeks to provide a more robust and precise understanding of user interests within scenarios where mini-apps are available only for short periods, such as during Explosive Promotional Scenarios (EPS).

Key Insights and Methodology

At the core of the CCN is a novel approach to delineate and leverage user interests and disinterests. Instead of solely relying on a user's interaction with trigger items, CCN captures the collaborative relationship of co-click and co-non-click, as well as non-collaborative signals through mono-click behavior—all of which are employed as signals for contrastive learning. The main idea is to classify items into two clusters, positive and negative, using the user's exposure or click logs to improvise CTR predictions.

The architecture of CCN is twofold:

1. CTR Prediction Module: This module estimates the click-through probability of a target item using embeddings derived from user profiles, target items, trigger items, and user behavior sequences. Through the application of strategies like Multi-head Target Attention (MHTA), it refines interest extraction from user data.
2. Collaborative Module: This module enhances the understanding of user preferences by generating item embeddings that reflect user interest and disinterest. It calculates the collaborative degree and applies this in collaborative contrastive learning—a supervised learning task designed to separate items effectively into interest and disinterest clusters.

A critical component of this method is the dual mechanism of attraction and repulsion losses aimed at refining these embeddings: attraction pulling similar interests closer while repulsion pushes dissimilar interest items apart in the latent space. The efficacy of this approach was validated through extensive experiments on a large industrial dataset from Taobao, showing relative improvements over existing models. The CCN outperformed competitors such as SIM, DIHN, and DIAN, achieving a 12.3% increase in CTR and a 12.7% increase in order volume as demonstrated through online A/B testing.

Results and Implications

The findings of this paper have significant implications for the field of recommender systems in e-commerce platforms. The model's state-of-the-art performance as observed on Taobao's EPS context—particularly during high-traffic promotional events like the Double 11 Shopping Carnival—underscores the role of nuanced user interest modeling beyond superficial signal recognition. Additionally, the research provides a framework adaptable to volatile promotional scenarios, contrasting with traditional systems that depend heavily on static, continually available categories.

From a theoretical standpoint, the integration of contrastive learning techniques into CTR prediction could inspire further research into other user interaction models, potentially enhancing recommendation algorithms across diverse domains. Practically, the implementation of CCN in the dynamic EPS context highlights a scalable solution for industries reliant on recommendation systems, ensuring precise customer engagement, and enhanced transactional outcomes.

Future Directions

Future work could explore the refinement of CCN's underlying components to further boost their adaptability and precision in real-time applications. Consideration of ongoing user behavior shifts using online learning methods might also provide benefit, allowing systems to continuously harmonize with evolving consumer patterns. Additionally, extending this framework to incorporate multi-modal data, such as text and visual information, could provide a more holistic understanding of user interests.

In conclusion, the development of the Collaborative Contrastive Network represents a significant contribution to CTR prediction, with practical implications for improving user experience and commercial success in e-commerce platforms.