论文检索

Yihang Lu 0001, Yangyang Xu, Qitao Qin, Xianwei Meng 0001

Recent deep learning models for Long-term Time Series Forecasting (LTSF) often emphasize complex, handcrafted designs, while simpler architectures like linear models or MLPs have often outperformed these intricate solutions. In this paper, we revisit and organize the core ideas behind several key techniques, such as redundancy reduction and multi-scale modeling, which are frequently employed in advanced LTSF models. Our goal is to streamline these ideas for more efficient deep learning utilization. To this end, we introduce TimeCapsule, a model built around the principle of high-dimensional information compression that unifies these techniques in a generalized yet simplified framework. Specifically, we model time series as a 3D tensor, incorporating temporal, variate, and level dimensions, and leverage mode production to capture multi-mode dependencies while achieving dimensionality compression. We propose an internal forecast within the compressed representation domain, supported by the Joint-Embedding Predictive Architecture (JEPA), to monitor the learning of predictive representations. Extensive experiments on challenging benchmarks demonstrate the versatility of our method, showing that TimeCapsule can achieve state-of-the-art performance. The code is available at: https://github.com/Luoauoa/TimeCapsule.git.

Donald Loveland, Mingxuan Ju, Tong Zhao 0003, Neil Shah, Danai Koutra

Collaborative filtering (CF) enables large-scale recommendation systems by encoding information from user-item interactions into dense ID-embedding tables. However, as embedding tables grow, closed-form solutions become impractical, necessitating the use of mini-batch gradient descent. Despite extensive work on loss function design, we argue that one core component of these pipelines is heavily overlooked: weight decay. Attaining high-performing models typically requires careful tuning of weight decay, regardless of loss, yet its necessity is not well understood. In this work, we question why weight decay is crucial in CF pipelines and how it impacts training. Through theoretical and empirical analysis, we uncover that weight decay's primary function is to encode popularity information into the magnitudes of the embedding vectors. Moreover, we find that tuning weight decay acts as a coarse, non-linear, knob to influence preference towards popular or unpopular items. Based on these findings, we propose PRISM (Popularity-awaRe Initialization Strategy for embedding Magnitudes), a straightforward yet effective solution to simplify the training of CF models. PRISM pre-encodes the popularity information typically learned through weight decay, eliminating its necessity. Our experiments show that PRISM improves performance by up to 4.77% and reduces training times by 38.48%, compared to state-of-the-art training strategies. Additionally, PRISM offers a cost-effective and meaningful strategy to mitigate popularity bias.

Zirui Liu 0002, Zihan Jiang 0004, An Zhang 0017, Zhouran Shi, Yuxuan Tian, Tong Yang 0003

This paper studies an unexplored attribute in data streams - item freshness. The freshness of an item refers to the time interval between its last arrival and the present moment. The information of item freshness is useful in various scenarios like cache, online advertising, computer network, etc. Currently, there is no algorithm tailored for estimating item freshness. We propose a theoretically guaranteed sketch algorithm called RingSketch, which integrates time-agnostic sketch algorithm with time-aware CLOCK algorithm for real-time freshness measurement. With the key idea of tracing the trajectory of the clock pointer, the estimation process of RingSketch is akin to observing the length of the growth rings in a tree trunk. We theoretically derive the average error of RingSketch and validate it with extensive experiments. The results show that RingSketch simultaneously achieves high accuracy (<10-3 average relative error) and fast update speed (>11.4 M/s), outperforming the baseline solutions by at least 13.3x and 1.5x respectively. All codes are open-sourced at GitHub.

Yuhang Liu 0004, Yingxue Zhang 0002, Xin Zhang 0098, Ling Tian, Yanhua Li, Jun Luo 0007

Understanding and predicting urban dynamics is crucial for managing transportation systems, optimizing urban planning, and enhancing public services. While neural network-based approaches have achieved success, they often rely on task-specific architectures and large volumes of data, limiting their ability to generalize across diverse urban scenarios. Meanwhile, Large Language Models (LLMs) offer strong reasoning and generalization capabilities, yet their application to spatial-temporal urban dynamics remains underexplored. Existing LLM-based methods struggle to effectively integrate multifaceted spatial-temporal data and fail to address distributional shifts between training and testing data, limiting their predictive reliability in real-world applications. To bridge this gap, we propose UrbanMind, a novel spatial-temporal LLM framework for multifaceted urban dynamics prediction that ensures both accurate forecasting and robust generalization. At its core, UrbanMind introduces Muffin-MAE, a multifaceted fusion masked autoencoder with specialized masking strategies that capture intricate spatial-temporal dependencies and intercorrelations among multifaceted urban dynamics. Additionally, we design a semantic-aware prompting and fine-tuning strategy that encodes spatial-temporal contextual details into prompts, enhancing LLMs' ability to reason over spatial-temporal patterns. To further improve generalization, we introduce a test time adaptation mechanism with a test data reconstructor, enabling UrbanMind to dynamically adjust to unseen test data by reconstructing LLM-generated embeddings. Extensive experiments on real-world urban dynamics datasets from multiple cities demonstrate the effectiveness of UrbanMind. The results consistently show that UrbanMind outperforms state-of-the-art baselines, achieving superior accuracy and strong generalization, even in zero-shot scenarios with no prior data.

Yuansan Liu, Sudanthi N. R. Wijewickrema, Dongting Hu, Christofer Bester, Stephen J. O'Leary, James Bailey 0001

Recent successes in diffusion probabilistic models have demonstrated their strength in modeling and generating different types of data, paving the way for their application in generative time series forecasting. However, most existing diffusion based approaches rely on sequential models and unimodal latent variables to capture global dependencies and model entire observable data, resulting in difficulties when it comes to highly stochastic time series data. In this paper, we propose a novel Stochastic Diffusion (StochDiff) model that integrates the diffusion process into time series modeling stage and utilizes the representational power of the stochastic latent spaces to capture the variability of the stochastic time series data. Specifically, the model applies diffusion module at each time step within the sequential framework and learns a step-wise, data-driven prior for generative diffusion process. These features enable the model to effectively capture complex temporal dynamics and the multi-modal nature of the highly stochastic time series data. Through extensive experiments on real-world datasets, we demonstrate the effectiveness of our proposed model for probabilistic time series forecasting, particularly in scenarios with high stochasticity. Additionally, with a real-world surgical use case, we highlight the model's potential in a medical application.

Yuanjun Liu 0001, Guanfeng Liu 0001, Qingzhi Ma, Zhixu Li, Lei Zhao 0001, An Liu 0002

Trajectory similarity computation is a fundamental functionality in trajectory data mining, with wide-ranging applications in location-based services. Position embedding, which transforms GPS points into embedding vectors, plays a critical role in learning-based trajectory similarity models. The quality of these embeddings significantly impacts the performance of the models on downstream tasks. Existing methods fail to satisfy all good properties, i.e., global, continuous, unique, and dynamic, thereby limiting the development of trajectory similarity computation in both local and global scenarios. Inspired by linear counting systems, such as the decimal system, we first propose the łambda-base circular system to embed positions on the circle, then introduce the multi-base global embedding method GPE to encode global positions into vectors. Experiments conducted on five real-world datasets with nine baseline methods demonstrate that the GPE achieves state-of-the-art performance across four key evaluations in downstream tasks.

Yuanhao Liu, Shuo Liu 0017, Yimeng Liu, Chanjin Zheng, Wei Zhang 0056, Hong Qian

Cognitive diagnosis model (CDM) is a fundamental component in intelligent education systems which aims to infer students' mastery levels based on historical response logs. However, existing CDMs usually follow the ID-based embedding paradigm, which could often diminish the effectiveness of CDMs in open student learning environments. This is mainly because most of them cannot directly infer new students' ability or utilize new exercises or knowledge concepts without retraining. Textual semantic information, due to its unified feature space and easy accessibility, can help alleviate this issue. However, directly incorporating textual semantic information may not benefit traditional CDMs due to the following challenges: the diversity and complexity of the original text corpus, lack of response-relevant features, and difficulty in integrating multi-source features. To this end, this paper proposes a Dual-Fusion Cognitive Diagnosis Framework (DFCD) to address the above challenges in open student learning environments. Specifically, to standardize the original text corpus and make it easier for CDMs to capture relevant textual semantic information, this paper first proposes the exercise-refiner and concept-refiner to make the exercises and knowledge concepts more coherent and reasonable in educational scenario via large language models. Then, DFCD encodes the refined features using text embedding models to obtain the textual semantic features. To construct response-relevant features, we propose a unified response-relevant feature construction to fully incorporate the information within the response logs. Finally, DFCD designs a dual-fusion module to merge the features from two sources, namely textual semantic features and response-relevant features. The ultimate representations possess the capability of inference in open student learning environments and can be also plugged in existing CDMs. Extensive experiments across three real-world datasets show that DFCD achieves superior performance and strong adaptability by improving the performance in three different scenarios of open student learning environments around 5% on average.

Yu Liu 0016, Weiyao Tao, Tong Xia, Simon Knight 0005, Tingting Zhu 0001

Survival analysis, which estimates the probability of event occurrence over time from censored data, is fundamental in numerous real-world applications, particularly in high-stakes domains such as healthcare and risk assessment. Despite advances in numerous survival models, quantifying the uncertainty of predictions from these models remains underexplored and challenging. The lack of reliable uncertainty quantification limits the interpretability and trustworthiness of survival models, hindering their adoption in clinical decision-making and other sensitive applications. To bridge this gap, in this work, we introduce SurvUnc, a novel meta-model based framework for post-hoc uncertainty quantification for survival models. SurvUnc introduces an anchor-based learning strategy that integrates concordance knowledge into meta-model optimization, leveraging pairwise ranking performance to estimate uncertainty effectively. Notably, our framework is model-agnostic, ensuring compatibility with any survival model without requiring modifications to its architecture or access to its internal parameters. Especially, we design a comprehensive evaluation pipeline tailored to this critical yet overlooked problem. Through extensive experiments on four publicly available benchmarking datasets and five representative survival models, we demonstrate the superiority of SurvUnc across multiple evaluation scenarios, including selective prediction, misprediction detection, and out-of-domain detection. Our results highlight the effectiveness of SurvUnc in enhancing model interpretability and reliability, paving the way for more trustworthy survival predictions in real-world

Yiling Liu, Wei Shi 0011, Chen Fu, Ziyang Jiang, Zhigang Hua, David E. Carlson

Multi-treatment multi-outcome treatment effect estimation plays a vital role in today's industry-level applications. For example, in social media ads, practitioners simultaneously deploy multiple interventions to users' experience and track multi-faceted metrics (e.g., ad performance, engagement, churn). However, existing methods for estimating treatment effects struggle to simultaneously address the complex interplays and ensure robust counterfactual balancing across treatment-outcome pairs. In our paper, we propose MOTTO (Multi-Outcome, Multi-Treat-ment Transfer Model for Treatment Effect Estimation), a Mixture-of-Experts-based framework designed to jointly handle these multi-faceted challenges. MOTTO explicitly learns the relationships across outcomes and distribution overlaps across treatment groups by isolating shared experts within its partitioned architecture. It then employs selective distribution alignment to these treatment-shared experts, optimizing the balance between factual and counterfactual predictions. Importantly, MOTTO can scale to a large number of treatments and outcomes without significantly increasing the number of parameters, as its shared experts can be primarily used for efficiency with comparable performance. We validate MOTTO on synthetic data, demonstrating MOTTO's strong adaptability to varying levels of outcome correlations and confounding. We further highlight its performance improvements across real-world benchmarks and a large-scale advertising ecosystem on one of the world's largest social media platforms.

Yang Liu 0245, Zinan Zheng, Yu Rong 0001, Deli Zhao, Hong Cheng 0001, Jia Li 0009

Accurate weather forecasting on Subseasonal-to-Seasonal (S2S) timescale is critical to human society such as agriculture planning and extreme weather preparation. Although data-driven models have become alternatives to computationally intensive Numerical Weather Prediction (NWP) systems, existing Transformer-based approaches suffer from biases due to planar projections distorting the spherical geometry and inadequate handling of vector-scalar variable interactions (e.g., wind velocity vs. temperature). To address these limitations, we propose a graph-based Equivariant and Invariant Message Passing (EIMP) framework that directly processes spherical grid data. It maintains SO(3) equivariant embeddings for vector data and SO(3) invariant embeddings for scalar data, which are interacted by a shared invariant message embedding. Guaranteed equivariant and invariant message aggregation functions are proposed to update embeddings under strict symmetry constraints. Extensive experiments on the Earth Reanalysis 5 (ERA5) reanalysis dataset of 41 years demonstrate the proposed model achieves significant improvement over advanced data-driven models and skillful numerical ECMWF systems. Additionally, we empirically show that EIMP demonstrates geometrically superior predictions and conduct ablation studies to validate the efficacy of its design.

Yang Liu 0348, Deyu Bo, Wenxuan Cao, Yuan Fang 0001, Yawen Li 0001, Chuan Shi 0001

Graph self-supervised learning seeks to learn effective graph representations without relying on labeled data. Among various approaches, graph autoencoders (GAEs) have gained significant attention for their efficiency and scalability. Typically, GAEs take incomplete graphs as input and predict missing elements, such as masked node features or edges. Although effective, our experimental investigation reveals that traditional feature or edge masking paradigms primarily capture low-frequency signals in the graph and fail to learn expressive structural information. To address these issues, we propose Graph Positional Autoencoders (GraphPAE), which employ a dual-path architecture to reconstruct both node features and positions. Specifically, the feature path uses positional encoding to enhance the message-passing processing, improving the GAEs' ability to predict the corrupted information. The position path, on the other hand, leverages node representations to refine positions and approximate eigenvectors, thereby enabling the encoder to learn diverse frequency information. We conduct extensive experiments to verify the effectiveness of GraphPAE, including heterophilic node classification, graph property prediction, and transfer learning. The results demonstrate that GraphPAE achieves state-of-the-art performance and consistently outperforms the baselines by a large margin.

Rui Liu, Rui Xie 0002, Zijun Yao 0001, Yanjie Fu, Dongjie Wang

Feature selection removes redundant features to enhance both performance and computational efficiency in downstream tasks. Existing methods often struggle to capture complex feature interactions and adapt to diverse scenarios. Recent advances in this domain have incorporated generative intelligence to address these drawbacks by uncovering intricate relationships between features. However, two key limitations remain: 1) embedding feature subsets in a continuous space is challenging due to permutation sensitivity, as changes in feature order can introduce biases and weaken the embedding learning process; 2) gradient-based search in the embedding space assumes convexity, which is rarely guaranteed, leading to reduced search effectiveness and suboptimal subsets. To address these limitations, we propose a new framework that can: 1) preserve feature subset knowledge in a continuous embedding space while ensuring permutation invariance; 2) effectively explore the embedding space without relying on strong convex assumptions. For the first objective, we develop an encoder-decoder paradigm to preserve feature selection knowledge into a continuous embedding space. This paradigm captures feature interactions through pairwise relationships within the subset, removing the influence of feature order on the embedding. Moreover, an inducing point mechanism is introduced to accelerate pairwise relationship computations. For the second objective, we employ a policy-based reinforcement learning (RL) approach to guide the exploration of the embedding space. The RL agent effectively navigates the space by balancing multiple objectives. By prioritizing high-potential regions adaptively and eliminating the reliance on convexity assumptions, this search strategy effectively reduces the risk of converging to local optima. Finally, we conduct extensive experiments to demonstrate the effectiveness, efficiency, robustness and explicitness of our model. Our code and dataset are publicly accessible on GitHub. https://github.com/RayLiu1103/CAPS.

Qidong Liu 0001, Jiurui Lian, Chaoyue Liu 0009, Zhiguang Cao

The Heterogeneous Capacitated Vehicle Routing Problem (HCVRP) is an NP-hard combinatorial optimization problem. State-of-the-art neural solvers face difficulties in generalizing to large-scale scenarios after training on small-scale instances. Our experiments reveal that performance degradation is primarily due to the low-rank nature of attention matrix in large-scale instances. This results in insufficient distinction among node features, impacting the accuracy of Markov Decision Processes. Additionally, these models utilize self-attention for vehicle information interaction, but overly incorporate features from others, which suppresses individual features and leads to a deviation from the optimal route. To address these challenges, we propose the Rank-Augmented Neural Solver (RANS), which introduces two key innovations: 1) A simple yet effective mechanism to increase and approximate the upper bound of the attention matrix's rank, enabling the generation of more distinctive node features. 2) A Dual Cross-Attention Module within the vehicle encoder that accurately captures each vehicle's optimal routes while maintaining balanced vehicle collaboration. The experimental results show that RANS performs favorably against the baselines. Notably, when applied to instances with up to 10,000 nodes, RANS achieves an inference time that is merely 13.42% of the best baseline among the neural solvers, while simultaneously reducing the min-max travel time by 23.72%.

Maoli Liu, Zhuohua Li 0001, Xiangxiang Dai, John C. S. Lui

Conversational recommender systems proactively query users with relevant ''key terms'' and leverage the feedback to elicit users' preferences for personalized recommendations. Conversational contextual bandits, a prevalent approach in this domain, aim to optimize preference learning by balancing exploitation and exploration. However, several limitations hinder their effectiveness in real-world scenarios. First, existing algorithms employ key term selection strategies with insufficient exploration, often failing to thoroughly probe users' preferences and resulting in suboptimal preference estimation. Second, current algorithms typically rely on deterministic rules to initiate conversations, causing unnecessary interactions when preferences are well-understood and missed opportunities when preferences are uncertain. To address these limitations, we propose three novel algorithms: CLiSK, CLiME, and CLiSK-ME. CLiSK introduces smoothed key term contexts to enhance exploration in preference learning, CLiME adaptively initiates conversations based on preference uncertainty, and CLiSK-ME integrates both techniques. We theoretically prove that all three algorithms achieve a tighter regret upper bound of O (√dTlogT) with respect to the time horizon T, improving upon existing methods. Additionally, we provide a matching lower bound Ω(√dT) for conversational bandits, demonstrating that our algorithms are nearly minimax optimal. Extensive evaluations on both synthetic and real-world datasets show that our approaches achieve at least a 14.6% improvement in cumulative regret.

Junhao Liu 0001, Pengpeng Zhang, Martin Renqiang Min, Jing Zhang 0062

Cell fate decisions are highly coordinated processes governed by complex interactions among numerous regulatory genes, while disruptions in these mechanisms can lead to developmental abnormalities and disease. Traditional methods often fail to capture such combinatorial interactions, limiting their ability to fully model cell fate dynamics. Here, we introduce MetaVelo, a global feature explanation framework for identifying key regulatory gene sets influencing cell fate transitions. MetaVelo models these transitions as a black-box function and employs a differentiable neural ordinary differential equation (ODE) surrogate to enable efficient optimization. By reparameterizing the problem as a controllable data generation process, MetaVelo overcomes the challenges posed by the non-differentiable nature of cell fate dynamics. Benchmarking across diverse stand-alone and longitudinal single-cell RNA-seq datasets and three black-box cell fate models demonstrates its superiority over 12 baseline methods in predicting developmental trajectories and identifying combinatorial regulatory gene sets. MetaVelo further distinguishes independent from synergistic regulatory genes, offering novel insights into the gene interactions governing cell fate. With the growing availability of high-resolution single-cell data, MetaVelo provides a scalable and effective framework for advancing developmental biology and therapeutic applications.

Jian-Dong Liu, Zhi-Hao Tan, Zhi-Hua Zhou

The learnware paradigm proposed by Zhou [34] aims to solve machine learning tasks by leveraging numerous existing high-performing models instead of training from scratch. These models are accommodated in a learnware dock system, where each learnware consists of a model and a specification that characterizes the model's utility, enabling it to be identified for future tasks. A critical challenge in this paradigm remains unresolved: determining what models can be or should be admitted to the system. Without well-established admission criteria, the uncontrolled growth of uploaded models could lead to significant redundancy and inefficiency, resulting in higher storage overhead, increased computational costs, and even potential system failure. To address this gap, this paper presents the first attempt to establish learnware admission criteria and dynamically filter redundant learnwares based on model capability coverage. Specifically, we organize task information from all learnwares into a tree-based structure to assess model capabilities across a continuously expanding task set. Using this structure and model capability representation, we develop an efficient and scalable method for detecting redundant learnwares dynamically without traversing the entire system. Theoretical analysis and extensive experiments involving over ten thousand simulated learnwares validate the efficacy and efficiency of our approach.

Huafeng Liu 0001, Yiran Fu, Jingyue Shi, Liping Jing, Jian Yu 0001

Existing work in physical-informed machine learning (PIML) has shown that data-driven learning of solution operators can provide a fast approximate alternative to classical numerical ordinary/partial differential equations (ODEs/PDEs) solvers. Of these, Neural Operators (NOs) have emerged as particularly promising. However, a key challenge in the field of NOs lies in developing methods that can effectively handle out-of-distribution (OOD) forecasting problems. Such problems involve the ability to adaptively learn from observations of the same dynamical system governed by ODEs/PDEs, where the underlying parameters are unknown and vary across instances. These tasks further require precise predictions even when faced with initial conditions and PDEs/ODEs parameters outside the training distribution. In this study, we consider the problem of training models in a risk-reverse manner. We introduce a risk-aware framework aimed at enhancing the OOD robustness of NOs by stochastically optimizing the conditional value-at-risk (CVAR) of a loss distribution. Through experiments on different distinct OOD tasks, our approach demonstrates a significant performance improvement over existing advanced NOs.

Enze Liu 0005, Bowen Zheng 0005, Wayne Xin Zhao, Ji-Rong Wen

In recent years, substantial research efforts have been devoted to enhancing sequential recommender systems by integrating abundant side information with ID-based collaborative information. This study specifically focuses on leveraging the textual metadata (e.g., titles and brands) associated with items. While existing methods have achieved notable success by combining text and ID representations, they often struggle to strike a balance between textual information embedded in text representations and collaborative information from sequential patterns of user behavior. In light of this, we propose CCFRec, a novel Code-based textual and Collaborative semantic Fusion method for sequential Recommendation. The key idea behind our approach is to bridge the gap between textual and collaborative information using semantic codes. Specifically, we generate fine-grained semantic codes from multi-view text embeddings through vector quantization techniques. Subsequently, we develop a code-guided semantic-fusion module based on the cross-attention mechanism to flexibly extract and integrate relevant information from text representations. In order to further enhance the fusion of textual and collaborative semantics, we introduce an optimization strategy that employs code masking with two specific objectives: masked code modeling and masked sequence alignment. The merit of these objectives lies in leveraging mask prediction tasks and augmented item representations to capture code correlations within individual items and enhance the sequence modeling of the recommendation backbone. Extensive experiments conducted on four public datasets demonstrate the superiority of CCFRec, showing significant improvements over various sequential recommendation models. Our code is available at https://github.com/RUCAIBox/CCFRec.

David Liu 0006, Arjun Seshadri, Tina Eliassi-Rad, Johan Ugander

A wide range of graph embedding objectives decompose into two components: one that enforces similarity, attracting the embeddings of nodes that are perceived as similar, and another that enforces dissimilarity, repelling the embeddings of nodes that are perceived as dissimilar. Without repulsion, the embeddings would collapse into trivial solutions. Skip-Gram Negative Sampling (SGNS) is a popular and efficient repulsion approach that prevents collapse by repelling each node from a sample of dissimilar nodes. In this work, we show that when repulsion is most needed and the embeddings approach collapse, SGNS node-wise repulsion is, in the aggregate, an approximate re-centering of the node embedding dimensions. Such dimension operations are more scalable than node operations and produce a simpler geometric interpretation of the repulsion. Our theoretical result establishes dimension regularization as an effective and more efficient, compared to skip-gram node contrast, approach to enforcing dissimilarity among embeddings of nodes. We use this result to propose a flexible algorithm augmentation framework that improves the scalability of any existing algorithm using SGNS. The framework prioritizes node attraction and replaces SGNS with dimension regularization. We instantiate this generic framework for LINE and node2vec and show that the augmented algorithms preserve downstream link-prediction performance while reducing GPU memory usage by up to 33.3% and training time by 23.4%. Moreover, we show that completely removing repulsion (a special case of our augmentation framework) in LINE reduces training time by 70.9% on average, while increasing link prediction performance, especially for graphs that are globally sparse but locally dense. Global sparsity slows down dimensional collapse, while local density ensures that node attraction brings the nodes near their neighbors. In general, however, repulsion is needed, and dimension regularization provides an efficient alternative to SGNS.

Chengkai Liu, Yangtian Zhang, Jianling Wang, Rex Ying, James Caverlee

Generative models have shown great promise in collaborative filtering by capturing the underlying distribution of user interests and preferences. However, existing approaches struggle with inaccurate posterior approximations and misalignment with the discrete nature of recommendation data, limiting their expressiveness and real-world performance. To address these limitations, we propose FlowCF, a novel flow-based recommendation system leveraging flow matching for collaborative filtering. We tailor flow matching to the unique challenges in recommendation through two key innovations: (1) a behavior-guided prior that aligns with user behavior patterns to handle the sparse and heterogeneous user-item interactions, and (2) a discrete flow framework to preserve the binary nature of implicit feedback while maintaining the benefits of flow matching, such as stable training and efficient inference. Extensive experiments demonstrate that FlowCF achieves state-of-the-art recommendation accuracy across various datasets with the fastest inference speed, making it a compelling approach for real-world recommender systems. The code is available at https://github.com/chengkai-liu/FlowCF.