Abstract
Reranking, as the final stage of recommender systems, demands real-time inference, accuracy, and diversity. Recently, generative reranking has gained attention for its ability to model complex item dependencies. However, most existing methods suffer from the "likelihood trap," where high-likelihood sequences are perceived as low-quality by humans due to list homogeneity and repetition of high-frequency items. We propose Consistent Graph-structured Generative Recommendation (Congrats), a novel framework featuring a graph-structured decoder to expand decoding space and a differentiable cascade system to learn directly from user preferences. Extensive experiments, including online A/B tests on Kuaishou, demonstrate that Congrats significantly improves recommendation quality and diversity.
Introduction
In modern recommender systems, reranking is the critical final stage that determines the exact exposure of items to users. While generative reranking has shown promise by treating lists as sequences, it often falls into the likelihood trap. Models trained via Maximum Likelihood Estimation (MLE) tend to favor "safe," high-frequency "hot" items, leading to repetitive and dull recommendations that ultimately harm user engagement.
As shown in the comparison above, traditional methods often cluster "hot" items, whereas our proposed method, Congrats, enables a more diverse and high-quality selection by breaking away from simple statistical likelihood.
The Congrats Framework
The Congrats (Consistent Graph-structured Generative Recommendation) framework introduces two core innovations to address the limitations of standard generative models.
1. Graph-structured Decoder
Unlike traditional linear chain decoders, Congrats structures hidden states as a Directed Acyclic Graph (DAG). By utilizing learnable transition modules and expanding the number of position vertices, the model can explore multiple paths during sequence generation. This design expands the decoding space and allows the model to capture implicit item dependencies through vertex transitions.
2. Consistent Differentiable Training
To align the model with actual user utility (e.g., clicks, watch time), we integrate an evaluator directly into the training loop. Since item sampling is discrete, we employ the Gumbel-Softmax reparameterization trick to make the process differentiable. This allows gradients to flow from the evaluator back to the generator, ensuring the model optimizes for user satisfaction rather than just sequence probability.
Experimental Evaluation
We evaluated Congrats on two large-scale datasets: Kuaishou (230.5M requests) and Avito (search logs for classified ads).
Offline Results
On the Kuaishou dataset, Congrats achieved a 7% increase in Recall@6 over the previous state-of-the-art method, NAR4Rec. It also demonstrated superior performance in diversity metrics and item coverage while maintaining a lower repetition rate.
Online Impact
In live A/B tests on the Kuaishou app, serving over 300 million daily active users, Congrats delivered significant improvements:
- +0.78% in total Views.
- +2.18% in "Long Views" (high watch duration).
- +3.01% in "Complete Views" (full video completion).
Discussion and Limitations
A key insight from this work is that in recommendation, statistical probability does not always equate to quality. By using a DAG-based decoder, the model can "buffer" conflicting items into different vertices, mitigating erroneous predictions.
However, we observed that increasing the graph size factor (λ) beyond a certain point (e.g., λ > 6) leads to diminishing returns and convergence difficulties. Future work will focus on stabilizing training in these larger search spaces and automating the weighting of multiple optimization objectives.
Summary updated