Abstract
A recommender system learns to predict the user-specific preference or intention over many items simultaneously for all users, making personalized recommendations based on a relatively small number of observations. One central issue is how to leverage three-way interactions, referred to as user-item-stage dependencies on a monotonic chain of events, to enhance the prediction accuracy. A monotonic chain of events occurs, for instance, in an article sharing dataset, where a "follow" action implies a "like" action, which in turn implies a "view" action. In this article, we develop a multistage recommender system utilizing a two-level monotonic property characterizing a monotonic chain of events for personalized prediction. Particularly, we derive a large-margin classifier based on a nonnegative additive latent factor model in the presence of a high percentage of missing observations, particularly between stages, reducing the number of model parameters for personalized prediction while guaranteeing prediction consistency. On this ground, we derive a regularized cost function to learn user-specific behaviors at different stages, linking decision functions to numerical and categorical covariates to model user-item-stage interactions. Computationally, we derive an algorithm based on blockwise coordinate descent. Theoretically, we show that the two-level monotonic property enhances the accuracy of learning as compared to a standard method treating each stage individually and an ordinal method utilizing only one-level monotonicity. Finally, the proposed method compares favorably with existing methods in simulations and an article sharing dataset.
Introduction
In modern recommender systems, user behavior is rarely a single isolated event. Instead, it often follows a logical, sequential progression known as a monotonic behavior chain. For example, on a content platform, a user must "view" an article before they can "like" it, and they are unlikely to "follow" the author unless they have liked the content.
Standard recommender systems often treat these stages independently, which leads to two primary issues: logical inconsistency (predicting a "follow" without a "like") and data sparsity (later stages have far fewer positive observations). The framework proposed by Dai, Shen, and Pan addresses these by enforcing "two-level monotonicity" to ensure logical integrity and improve predictive power by leveraging dependencies across the entire chain.
Methodology
The core of the proposed system is the Two-level Monotonic Property, which encompasses both forward and backward logical constraints:
- Forward Monotonicity: If a user does not perform an action at a current stage, they cannot perform it at any subsequent stage.
- Backward Monotonicity: If a user performs an action at a future stage, they must have performed all preceding actions in the chain.
To mathematically guarantee these properties, the authors introduce a nonnegative additive latent factor model. By representing the decision function as a base preference minus a series of nonnegative "costs" for reaching subsequent stages, the model is structurally incapable of producing inconsistent predictions.
As shown in the architecture above, the model integrates continuous user/item features ($u_i, v_j$) and categorical features ($s_i, o_j$) into latent factors. These factors are then combined to model the user-item-stage interaction. The system utilizes a large-margin classifier (adapting hinge loss for multistage pairs) and a blockwise coordinate descent algorithm for efficient optimization.
Experimental Results
The model was evaluated using the Deskdrop dataset, a real-world article-sharing benchmark containing ~73,000 users and ~3,000 articles. The data follows a clear View $\rightarrow$ Like $\rightarrow$ Follow chain and is extremely sparse (less than 0.2% observed interactions).
Key findings include:
- Superior Accuracy: The method outperformed SVD++, Gradient Boosting, and Deep Neural Networks, reducing error rates by 8.2% to 75.3%.
- Guaranteed Consistency: While baseline models produced logically inconsistent predictions in up to 13.7% of cases, the proposed model maintained 0% inconsistency.
- Resilience to Sparsity: By "borrowing strength" from denser early stages (like views), the model significantly improved performance in the data-scarce "Follow" stage.
Conclusion
By treating multistage interactions as a unified, monotonic process rather than independent events, this framework provides a robust solution for complex recommendation environments. While currently optimized for linear chains, the use of consistency as a regularizer offers a powerful template for future work in more complex, branching behavior trees and large-scale industrial applications.
Summary updated