Data Interaction Based Click‑Through Rate Model (RIM): Review, Architecture, and Experimental Insights

The presentation begins with an overview of CTR prediction in search, advertising, and recommendation scenarios, describing how the field has progressed from manual feature engineering and logistic regression to factorization machines, deep neural networks, and recent attention to data interaction.

It then reviews deep CTR models, covering feature interaction techniques such as factorization machines (FM), factorization‑machine supported neural networks (FNN), multiplication‑based interactions (inner/outer product, PNN, DeepFM, PIN, DCN), convolution‑based interactions, and attention‑based interactions, highlighting their strengths and limitations.

Next, the talk discusses user‑behavior sequence modeling, including models like DIN, DIEN, hierarchical recurrent networks, and the User Behavior Retrieval (UBR) approach that searches across the entire dataset for relevant historical behaviors.

The core of the article introduces the Retrieval & Interaction Machine (RIM) model, which treats each sample as a query, retrieves a set of related samples from the whole dataset using inverted indexes and BM25 ranking, aggregates them via attention (including label information), and combines the aggregated vector with the target sample for multi‑domain feature interaction and downstream prediction.

Experimental results on multiple public datasets and tasks (CTR, ranking, regression) show that RIM consistently achieves statistically significant AUC improvements over baseline models, and analyses of search strategies, sample quantity, and interaction functions further validate its effectiveness.

Finally, the authors outline future work such as optimizing search strategies, improving online search efficiency, exploring more sophisticated similarity measures, and applying graph neural networks for richer data aggregation.