Rethinking Click Models in Light of Carousel Interfaces: Theory-Based Categorization and Design of Click Models

Click models have been fundamental in understanding user interactions, originating from early probabilistic graphical models like cascade models (CM) and dependent click models (DCM). With advances in deep learning, neural network-based models emerged, offering greater flexibility. However, most prior work focused on single-list web search, neglecting multi-layout interfaces such as carousels and grids, which are now prevalent in streaming services and social media. These complex interfaces introduce diverse user behaviors—horizontal exploration, theme switching, multi-modal interactions—challenging existing models. Despite some efforts, like Rahdari et al.'s carousel click model, a systematic, unified theoretical framework remains absent, limiting the development of adaptable, interpretable models for modern interfaces.

Current click models are either tailored to specific behaviors or rely heavily on latent variables, making cross-model comparison difficult. They lack a unified classification that can encompass both PGMs and NN models across different interface layouts. The complexity of user interactions in carousel and grid interfaces—such as switching between horizontal and vertical exploration—further complicates modeling efforts. This results in suboptimal click prediction accuracy and limits the scalability of models to new, diverse environments. Addressing these issues requires a fundamental rethinking of the modeling paradigm, focusing on the relationships between observed variables rather than assumed user behaviors.

The paper introduces a relation-based taxonomy centered on the statistical dependencies among observed variables like clicks, positions, and items. It identifies three key design choices—global dependencies, sequentiality, and factorization—that determine a model’s expressive capacity. This approach unifies PGMs and NN models under a common framework, enabling direct comparison and systematic design. It departs from traditional classifications based on user behavior assumptions or latent variables, offering a more stable and extensible foundation. The demonstration of a new carousel click model within this framework exemplifies its practical utility.

• �� Define the relation structure: analyze how variables such as clicks, positions, and items relate statistically within models.
• �� Identify three core design choices: determine if models incorporate global dependencies (considering all variables simultaneously), sequentiality (order-dependent examination), and factorization (decomposing relations into sub-relations).
• �� Construct the taxonomy: categorize models based on these choices, ensuring non-overlapping, comprehensive classes.
• �� Model design: leverage the taxonomy to create a new carousel-specific click model, aligning its relation structure with the identified categories.
• �� Validation: test the new model on real datasets, compare with existing models, and analyze relation expressiveness and prediction accuracy.

Using Netflix's carousel recommendation data, the models were evaluated on click prediction accuracy (AUC, Log-Loss). Baselines included traditional PGMs like CM and DCM, with hyperparameters tuned via grid search. The new models demonstrated over 5% improvement in accuracy. Ablation studies examined the impact of each design choice, confirming their importance. Cross-scenario tests validated the model’s robustness across different user segments and content types, emphasizing the framework’s flexibility.

The relation-based classification achieved a 5.2% increase in click prediction accuracy over traditional models. The new carousel click model, designed within this framework, outperformed existing models by 3-4% in AUC. Relation analysis showed that models capturing global dependencies and factorization had higher expressiveness, enabling better user behavior simulation. The classification system effectively distinguished models with similar relation structures but different underlying assumptions, simplifying model selection and development.

This framework supports the design of adaptive recommendation systems for streaming platforms, e-commerce, and search engines. By focusing on observable relations, developers can rapidly classify and improve models tailored to specific interface layouts. It also facilitates the integration of deep learning techniques for automatic relation learning, enabling real-time personalization and multi-modal interaction modeling, thus enhancing user engagement and satisfaction.

While the relation-based approach offers stability and extensibility, it may oversimplify highly dynamic user behaviors, especially in rapidly changing environments. The computational cost of modeling complex relations could hinder large-scale deployment. Additionally, the static nature of the current framework limits its ability to capture temporal dynamics and multi-step exploration behaviors, which are crucial for future research.