While the "ranked list" has been the standard for recommendation research for decades, modern user interfaces have shifted toward 2D carousels—rows of items grouped by categories like "Trending Now" or "Action Movies." Despite their ubiquity on platforms like Netflix and Amazon, standard evaluation metrics (e.g., NDCG, MRR) and user behavior models (e.g., the Cascade Model) are designed for 1D lists. This paper introduces the Carousel Click Model (CCM), a generative framework that explains why carousels are more efficient and demonstrates how they boost engagement through structured, topic-based browsing.
Modeling User Behavior in 2D
Traditional models assume a linear scan from top to bottom. However, carousels introduce a two-dimensional interaction. The CCM evolves existing click models by formalizing a two-step process:
- Vertical Scan: The user first evaluates the topic label of a carousel (row).
- Horizontal Scan: If the topic is attractive, the user scans items in that row from left to right. If not, they skip the entire row to examine the next topic.
- Termination: Users may leave the interface at any point if they become frustrated by irrelevant content, modeled via a termination probability.
By allowing users to skip entire rows of irrelevant content, the CCM enables them to examine more items with less fatigue, effectively saving their "attention budget" for relevant categories.
Experimental Validation
The authors validated the CCM using two primary sources: real-world carousel interaction data (776 instances) and large-scale simulations using the MovieLens 100K dataset mapped to 19 genres. The model achieved a better empirical fit to real-world data than the traditional Terminating Cascade Model (TCM), with a lower total variation distance (0.128 vs. 0.133).
Key Results and Insights
The Power of Semantic Labels
The efficiency of a carousel relies heavily on semantic labeling. The authors tested a variant called CCM-NL (No Labels), where users cannot effectively prune the search space. As shown in the results, performance drops significantly without clear labels, as the user can no longer make informed decisions about which rows to skip.
Massive Engagement Boost
In simulations across large catalogs, the CCM attracted significantly more clicks than traditional models. This suggests that the carousel structure is vastly superior for navigating large-scale inventories.
Personalization and Generalization
Integrating the CCM with Matrix Factorization to predict personalized attraction probabilities increased click probability by 25% over popularity-based baselines. Furthermore, the model demonstrated strong generalization, with only a 4.34% gap between training and test performance, indicating it effectively captures underlying behavioral patterns rather than just overfitting to history.
Conclusion
Carousels represent a successful human-AI collaboration: the AI provides structured topics, and the human performs the final filtering based on their immediate intent. The CCM provides a formal mathematical foundation for this interaction, proving that the ability to "skip" is a primary driver of user engagement. Future work may explore more complex 2D eye-tracking patterns and dynamic carousels where user interest shifts during a single session.
Summary updated