TL;DR
BrowseFlow is a novel generative user simulator that tackles the "needle-in-a-haystack" problem of sparse clicks and the inherent randomness of user behavior. By decoupling browsing depth from key events and using conditional discrete flow matching, it generates realistic, diverse interaction trajectories that significantly outperform traditional stepwise simulators in capturing the true distribution of user actions.
What Problem Does This Paper Address?
Simulating how users interact with recommendation feeds is crucial for testing new algorithms without risking real-world revenue. However, current simulators struggle with two major realities of human behavior:
- Sparse Key Events: In a typical session, a user might scroll past dozens of items but only click or "like" one or two. Traditional models try to predict every single action in a dense sequence (e.g., scroll, scroll, click, scroll, exit). Because scrolling is so much more common, the model often fails to learn the nuances of the rare, high-value "key events" like clicks.
- Stochastic Responses: User behavior is inherently random. If you show the same user the same feed twice, they might click different things or stop at different points. Most existing models are evaluated by how well they "match" a single logged trajectory, which ignores the fact that many other outcomes were equally probable.
Methods
The authors propose BrowseFlow, which moves away from dense, step-by-step prediction. Instead, it represents a user's response as a pair: (Browsing Extent, Key Event Sequence).
- Browsing Extent Prediction: First, the model predicts the "extent"—how far down the list the user will actually view. This is modeled as a categorical distribution over all possible depths in the recommendation session.
- Conditional Discrete Flow Matching: Once the extent is fixed, BrowseFlow generates the clicks. Rather than predicting them one by one from top to bottom, it uses a Continuous-Time Markov Chain (CTMC). It starts with an empty sequence and learns a "flow" that stochastically inserts events (clicks) into the sequence.
- Stochastic Insertion: The model learns transition rates to insert events at specific depths and on specific items. This allows it to naturally generate a variable number of events without needing a separate "count" predictor.
Datasets
The researchers evaluated BrowseFlow on three datasets:
- TaobaoBrowse (New): A large-scale benchmark introduced in this paper, featuring real recommendation contexts paired with fine-grained browsing trajectories (scrolling, clicking, and exiting).
- KuaiRand-Like: A public dataset from a short-video platform, adapted to evaluate "like" events over fixed exposure sequences.
- AdSERP: A dataset of search engine interactions, used to test click generation in a search context.
Key Insights and Ideas
- Separation of Concerns: By separating "how far a user scrolls" from "what they click," the model can focus its capacity on the difficult task of event placement without being overwhelmed by the trivial task of predicting repeated scrolls.
- Flow Matching for Sequences: While flow matching is popular in image generation, applying it to discrete, variable-length sequences in user simulation is a clever shift. It allows the model to learn the global "dynamics" of a session rather than just the next-step probability.
- Distributional Evaluation: The authors argue that we shouldn't just measure if a simulator reproduces a specific log. Instead, they use the Energy Score and Wasserstein distance (Count W1) to see if the distribution of generated sessions matches the distribution of real sessions. This accounts for both fidelity (is it realistic?) and variability (is it diverse?).
Key Results
BrowseFlow consistently achieved the best performance across all three datasets:
- Distributional Accuracy: On the TaobaoBrowse dataset, BrowseFlow reduced the Energy Score (ES) by roughly 33% and the Count W1 (event count alignment) by over 70% compared to the best baselines like UserSim and CaseRec.
- Better Extent Recovery: As shown in the behavioral analysis, BrowseFlow is much more accurate at predicting the browsing depth of long sessions, whereas traditional models tend to prematurely "exit" the user.
- Realistic Click Placement: BrowseFlow's "first click" placement closely matches real-world data, while baselines often place clicks too early in the sequence.
Limitations and Future Work
- Computational Complexity: While the paper notes that BrowseFlow generates responses in a single pass (per extent/flow sample), the iterative nature of flow matching inference (discretizing flow time into 32 intervals) may be more computationally expensive than simple one-shot classifiers.
- Fixed Feed Assumption: The current model assumes a fixed recommendation feed is known upfront. Future work could integrate this with "interactive" recommendation where the feed changes based on real-time simulated actions.
- Beyond Clicks: While the model handles clicks and likes, extending it to more complex multi-type events (e.g., add-to-cart, share, long-view) remains an open area for exploration.
Summary updated