GFN4Rec: Generative Flow Networks for Listwise Recommendation

Personalized recommendation systems aim to learn a policy that generates a list of items matching user needs. Compared with pointwise models, listwise methods model inter‑item correlations, improving quality but facing a large combinatorial action space.

Existing generative methods struggle to maintain accuracy while increasing diversity. This work adopts the GFlowNet concept (Yoshua Bengio) to propose GFN4Rec, which models list generation as an autoregressive process in a probability‑flow graph, optimizing the overall list probability to be proportional to its reward.

GFlowNet treats generation as a directed acyclic graph where only leaf nodes receive reward; training matches flow consistency using detailed‑balance or trajectory‑balance equations.

The solution represents list generation as sequential item addition, with a forward probability function determining item selection and a backward function assisting learning. The probability‑flow graph forms a tree, ensuring a unique generation path per list.

GFN4Rec encodes user requests with a transformer, uses masked embeddings for variable list lengths, and feeds the combined representation into both the flow‑evaluation and forward‑probability functions.

Experiments on online simulators and offline datasets show GFN4Rec outperforms collaborative filtering, ListCVAE, and PRM baselines, achieving higher reward, NDCG, and MRR while maintaining diversity.

Ablation studies examine hyper‑parameter effects, balance objectives, and greedy versus exploratory strategies, confirming the model’s robustness.

In summary, GFN4Rec demonstrates that aligning list‑generation probability with reward and leveraging incremental forward modeling effectively balances accuracy and diversity, offering a promising direction for recommender and other generative tasks.