One Model to Serve All: Star Topology Adaptive Recommender for Multi-Domain CTR Prediction (2101.11427v5)

Abstract: Traditional industrial recommenders are usually trained on a single business domain and then serve for this domain. However, in large commercial platforms, it is often the case that the recommenders need to make click-through rate (CTR) predictions for multiple business domains. Different domains have overlapping user groups and items. Thus, there exist commonalities. Since the specific user groups have disparity and the user behaviors may change in various business domains, there also have distinctions. The distinctions result in domain-specific data distributions, making it hard for a single shared model to work well on all domains. To learn an effective and efficient CTR model to handle multiple domains simultaneously, we present Star Topology Adaptive Recommender (STAR). Concretely, STAR has the star topology, which consists of the shared centered parameters and domain-specific parameters. The shared parameters are applied to learn commonalities of all domains, and the domain-specific parameters capture domain distinction for more refined prediction. Given requests from different business domains, STAR can adapt its parameters conditioned on the domain characteristics. The experimental result from production data validates the superiority of the proposed STAR model. Since 2020, STAR has been deployed in the display advertising system of Alibaba, obtaining averaging 8.0% improvement on CTR and 6.0% on RPM (Revenue Per Mille).

• The paper presents the STAR model, which addresses multi-domain CTR prediction challenges by combining shared and domain-specific networks in a star topology.
• Experiments on Alibaba data show the STAR model significantly improves CTR and RPM, outperforming baseline models like MMoE and Cross-Stitch networks.
• The STAR model offers a scalable, efficient solution for large platforms, reducing model maintenance while fostering cross-domain learning.

Overview of the STAR Model for Multi-Domain CTR Prediction

The paper "One Model to Serve All: Star Topology Adaptive Recommender for Multi-Domain CTR Prediction" presents an innovative approach aimed at enhancing the effectiveness and efficiency of click-through rate (CTR) prediction models on large-scale commercial platforms. These platforms are characterized by their multi-domain nature, where traditional models—typically trained on single domain data—may fall short of performance goals due to their inability to capture cross-domain interdependencies and dynamics.

Problem Context

Within the context of CTR prediction, the challenge lies in effectively utilizing data from multiple business domains, which may share some user groups and items but also present distinct behavior patterns. The conventional strategy of deploying individual models per domain leads to redundancy and suboptimal learning, while training a single unified model tends to overlook domain-specific characteristics.

Introduction to the STAR Model

The Star Topology Adaptive Recommender (STAR) model is proposed to address these challenges. The STAR model innovatively integrates a shared, domain-agnostic network with domain-specific networks to leverage the strengths of both approaches:

• Shared Network: This component learns general behaviors that are applicable across all domains, acting as a centralized knowledge base.
• Domain-Specific Networks: These networks are tailored to capture unique domain characteristics, allowing for more nuanced CTR predictions.

In a distinctive star topology, the STAR model combines these networks by element-wise multiplying the weights, effectively harnessing both commonalities and distinctions.

Experimentation and Results

In extensive experiments involving Alibaba's production data, which consists of a diverse set of business domains, the STAR model demonstrated a marked improvement over several baseline models. Gains of 8.0% in CTR and 6.0% in Revenue Per Mille (RPM) from the deployment of STAR within Alibaba's advertising systems underscore its practical efficacy.

Key experimental findings include:

• The STAR model recorded significant improvement in AUC scores across all business domains when compared to both single-domain training models and multi-task learning approaches like MMoE and Cross-Stitch networks.
• The deployment strategy showcasing batch normalization tailored to specific domain distributions further bolstered model performance by ensuring the model's adaptability to data distribution variation.

Implications and Future Directions

The STAR model offers a scalable solution for large e-commerce platforms to perform CTR predictions with pronounced efficiency and accuracy. By reducing the need for maintaining multiple models and simplifying the learning pipeline, STAR enhances resource utilization while fostering robust cross-domain learning.

From a theoretical standpoint, the element-wise parameter blending approach invites further exploration into more sophisticated combination techniques, potentially leveraging advanced neural architecture search methods. Future research could also extend the STAR model's applicability to other multi-domain machine learning problems, further cementing its place within the AI model ecosystem. Furthermore, extending STAR to involve real-time dynamic adaptation and continual learning scenarios represents a promising avenue that aligns with evolving user and item dynamics on such large platforms.

By offering an effective and efficient approach to multi-domain CTR prediction, the STAR model represents a significant contribution to the field of recommender systems within commercial platforms.