1$^{st}$ Place Solution of WWW 2025 EReL@MIR Workshop Multimodal CTR Prediction Challenge

TL;DR

The paper tackles multimodal CTR prediction by leveraging frozen multimodal embeddings within a Transformer-based sequential model enhanced by a DCNv2 cross-feature interaction module. The proposed architecture combines embedding concatenation, short-term user interests via a Transformer, and high-order feature interactions to predict click probability with a sigmoid-activated MLP, achieving an on-leaderboard $AUC$ of $0.9839$. Key contributions include a clear problem formulation, a practical four-component network, and extensive ablations demonstrating the importance of both sequential modeling and interaction learning for multimodal inputs. The approach emphasizes efficiency and deployability, using frozen multimodal embeddings and straightforward integration, with future directions aimed at aligning multimodal embeddings more tightly with CTR tasks and exploring learnable adaptations.

Abstract

The WWW 2025 EReL@MIR Workshop Multimodal CTR Prediction Challenge focuses on effectively applying multimodal embedding features to improve click-through rate (CTR) prediction in recommender systems. This technical report presents our 1$^{st}$ place winning solution for Task 2, combining sequential modeling and feature interaction learning to effectively capture user-item interactions. For multimodal information integration, we simply append the frozen multimodal embeddings to each item embedding. Experiments on the challenge dataset demonstrate the effectiveness of our method, achieving superior performance with a 0.9839 AUC on the leaderboard, much higher than the baseline model. Code and configuration are available in our GitHub repository and the checkpoint of our model can be found in HuggingFace.

Figures (2)

• Figure 1: The overall architecture of our model.
• Figure 2: Parameter sensitivity analysis on hyperparameters.