Abstract
Slate recommendation generates a list of items as a whole instead of ranking each item individually, so as to better model the intra-list positional biases and item relations. In order to deal with the enormous combinatorial space of slates, recent work considers a generative solution so that a slate distribution can be directly modeled. However, we observe that such approaches -- despite their proved effectiveness in computer vision -- suffer from a trade-off dilemma in recommender systems: when focusing on reconstruction, they easily over-fit the data and hardly generate satisfactory recommendations; on the other hand, when focusing on satisfying the user interests, they get trapped in a few items and fail to cover the item variation in slates. In this paper, we propose to enhance the accuracy-based evaluation with slate variation metrics to estimate the stochastic behavior of generative models. We illustrate that instead of reaching to one of the two undesirable extreme cases in the dilemma, a valid generative solution resides in a narrow "elbow" region in between. And we show that item perturbation can enforce slate variation and mitigate the over-concentration of generated slates, which expand the "elbow" performance to an easy-to-find region. We further propose to separate a pivot selection phase from the generation process so that the model can apply perturbation before generation. Empirical results show that this simple modification can provide even better variance with the same level of accuracy compared to post-generation perturbation methods.
Introduction: The Reconstruction-Concentration Dilemma
In modern recommender systems, items are typically presented in slates (e.g., grids or lists) rather than in isolation. Traditional learning-to-rank (LTR) methods often fail to capture complex intra-list interactions, such as complementarity or competition between items. While generative models like Conditional Variational Auto-Encoders (CVAEs) offer a way to model the entire slate distribution, they face a fundamental Reconstruction-Concentration Dilemma (RCD).
As shown in Figure 4, models often fall into two undesirable extremes. At low $\beta$ values (reconstruction focus), models overfit training data and generalize poorly. Conversely, at high $\beta$ values (interest focus), models suffer from "over-concentration," repeatedly generating a few "safe" prototype slates. The "desired behavior" resides in a very narrow "elbow" region that is difficult to target during training.
Methodology
To address the RCD, this work proposes two primary innovations: a new evaluation framework and the Pivot-CVAE architecture.
Enhanced Evaluation Metrics
Standard accuracy metrics (e.g., Hit Rate) are insufficient for evaluating generative slates. The authors propose decomposing total item variance into slate-mean variance (diversity across different slates) and intra-slate variance (diversity within a single slate). Metrics like Item Coverage and Intra-List Diversity (ILD) are used to track these stochastic behaviors.
Pivot-CVAE Architecture
The proposed Pivot-CVAE decomposes the slate generation process into two distinct phases: pivot selection and slate completion. Unlike standard List-CVAE which generates the entire slate at once, Pivot-CVAE first selects a single "pivot" item and then generates the remaining items conditioned on that pivot and the user's latent state.
By introducing sampling-based perturbation at the pivot selection stage, the model can explore diverse starting points. The subsequent completion phase ensures that the generated slate remains coherent and high-quality, avoiding the accuracy drop typically associated with post-generation shuffling.
Experimental Results
The researchers evaluated Pivot-CVAE on the YOOCHOOSE and MovieLens 100K datasets, as well as simulated environments designed to test multi-item interactions.
Key findings include:
- The "Elbow" is Tiny: In standard CVAEs, the optimal $\beta$ range for balancing accuracy and variety is extremely narrow, often less than 2% of the search space.
- Personalization as a Variance Booster: Conditioning on user IDs helps mitigate prototype collapse by learning user-specific clusters, which naturally increases system-wide variance.
- Superiority of Pre-generation Perturbation: Pivot-CVAE consistently outperformed standard List-CVAE and traditional ranking models. On MovieLens, Pivot-CVAE achieved an Item Coverage of 0.250, a massive improvement over the 0.004 achieved by standard List-CVAE, while maintaining competitive accuracy (ENC).
Conclusion
This paper highlights the critical trade-off between reconstruction and concentration in generative slate recommendation. By introducing Pivot-CVAE and a more nuanced evaluation framework, the authors demonstrate that controlled variation via pivot-based perturbation can effectively expand the "elbow" region of performance. Future work will explore varying pivot positions and extending these principles to other generative architectures like GANs.
Summary updated