Jinwei Zhang, Xucheng Liang, Yu Zhang, Ruijie Yu, Xiaokang Yang, Yaohui Jin, Yanyan Xu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Yu Zhang, Ruijie Yu, Jidong Tian, Feng Zhu, Jiapeng Liu, Xiaokang Yang, Yaohui Jin, Yanyan Xu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Modeling kinetic equations is essential for understanding the mechanisms of chemical reactions, yet a complex and time-consuming task. Kinetic equation prediction is formulated as a problem of dynamical symbolic regression (DSR) subject to physical chemistry constraints. Deep learning (DL) holds the potential to capture reaction patterns and predict kinetic equations from data of chemical species, effectively avoiding empirical bias and improving efficiency compared with traditional analytical methods. Despite numerous studies focusing on DSR and the introduction of Transformers to predict ordinary differential equations, the corresponding models lack generalization abilities across diverse categories of reactions. In this study, we propose KinFormer, a generalizable kinetic equation prediction model. KinFormer utilizes a conditional Transformer to model DSR under physical constraints and employs Monte Carlo Tree Search to apply the model to new types of reactions. Experimental results on 20 types of organic reactions demonstrate that KinFormer not only outperforms classical baselines, but also exceeds Transformer baselines in out-of-domain evaluations, thereby proving its generalization ability.

Facility location problems on graphs are ubiquitous in the real world and hold significant importance, yet their resolution is often impeded by NP-hardness. MIP solvers can find the optimal solutions but fail to handle large instances, while algorithm efficiency has a higher priority in cases of emergency. Recently, machine learning methods have been proposed to tackle such classical problems with fast inference, but they are limited to the myopic constructive pattern and only consider simple cases in Euclidean space. This paper introduces a unified and generalizable approach to tackle facility location problems on weighted graphs with deep reinforcement learning, demonstrating a keen awareness of complex graph structures. Striking a harmonious balance between solution quality and running time, our method stands out with superior efficiency and steady performance. Our model trained on small graphs is highly scalable and consistently generates high-quality solutions, achieving a speedup of more than 2000 times to Gurobi on instances with 1000 nodes. The experiments on Shanghai road networks further demonstrate its practical value in solving real-world problems. The source codes are available at https://github.com/AryaGuo/PPO-swap.

Human mobility studies how people move to access their needed resources and plays a significant role in urban planning and location-based services. As a paramount task of human mobility modeling, next location prediction is challenging because of the diversity of users’ historical trajectories that gives rise to complex mobility patterns and various contexts. Deep sequential models have been widely used to predict the next location by leveraging the inherent sequentiality of trajectory data. However, they do not fully leverage the relationship between locations and fail to capture users’ multilevel preferences. This work constructs a trajectory graph from users’ historical traces and proposes a Trajectory Graph Enhanced Orientation-based Sequential network (TrajGEOS) for next-location prediction tasks. TrajGEOS introduces hierarchical graph convolution to capture location and user embeddings. Such embeddings consider not only the contextual feature of locations but also the relation between them, and serve as additional features in downstream modules. Specifically, TrajGEOS models short-term preferences from recent trajectory sequences via GRU, mid-term preferences from historical trajectories over weeks through attention mechanisms, and long-term preferences from the global trajectory graph. Extensive experiments on three real-world LBSN datasets corroborate the value of graph and orientation-based modules and demonstrate that TrajGEOS outperforms the state-of-the-art methods on the next location prediction task.

Predicting the next destination is a key in human mobility behavior modeling, which is significant in various fields, such as epidemic control, urban planning, traffic management and recommendation. To achieve this, one typical solution is designing modules based on RNN to capture their preferences to various locations. Although these RNN-based methods can effectively learn individual’s hidden personalized preferences to her visited places, the interactions among users can only be weakly learned through the representations of locations. Targeting this, we propose an end-to-end framework named personalized and group preference guided network (PG2Net), considering the users’ preferences to various places at both individual and collective levels. Specifically, PG2Net concatenates Bi-LSTM and attention mechanism to capture each user’s long-term mobility tendency. To learn population’s group preferences, we utilize spatial and temporal information of the visitations to construct a spatial-temporal dependency module. We adopt a graph embedding method to map users’ trajectory into a hidden space, capturing their sequential relation. In addition, we devise an auxiliary loss to learn the vectorial representation of her next location. Experimental results on two Foursquare check-in datasets and one mobile phone dataset indicate the advantages of our model compared to the state-of-the-art baselines. Source code is available at https://github.com/urbanmobility/PG2Net.

Bus bunching is harmful to the efficiency and stability of bus transit systems, consequently delaying the arrival time of passengers and lowering the public transportation’s adoption rate. Traditional solutions adjust the additional holding time of buses at certain stations to mitigate this phenomenon. These methods sacrifice the system efficiency in exchange for even headway between neighboring buses. Little work focuses on optimizing multiple strategies when a single bus line not only has a bus bay to increase bus dwell time but also owns several dedicated bus lanes to accelerate. In this work, we develop a hierarchical multi-agent reinforcement learning (HMARL) framework to combine these two strategies. Speeding up certain buses via dedicated lanes can counteract the negative influence of additional holding time. Next, to support the two strategies, we devise a two-layer policy scheme, one for high-level policy deciding holding or accelerating and the other for low-level policy determining the specific dwell time or increase of speed. Besides, to handle the issue that the controlling actions of agents are asynchronous and temporally extended, we establish a duration-critic module based on the Recurrent Neural Networks (RNN) mechanism to model other agents’ impact during the period between two consecutive control. We evaluate the proposed framework on a simulated bus line with a quasi-real-world pattern to compare the performance of both traditional headway-based control methods and existing MARL methods. Results show that our method outperforms other baselines, not only stabilizing a strongly unstable bus line but also shortening the traveling times of passengers.

Anomaly detection plays a crucial role in various real-world applications, including healthcare and finance systems. Owing to the limited number of anomaly labels in these complex systems, unsupervised anomaly detection methods have attracted great attention in recent years. Two major challenges faced by the existing unsupervised methods are as follows: 1) distinguishing between normal and abnormal data when they are highly mixed together and 2) defining an effective metric to maximize the gap between normal and abnormal data in a hypothesis space, which is built by a representation learner. To that end, this work proposes a novel scoring network with a score-guided regularization to learn and enlarge the anomaly score disparities between normal and abnormal data, enhancing the capability of anomaly detection. With such score-guided strategy, the representation learner can gradually learn more informative representation during the model training stage, especially for the samples in the transition field. Moreover, the scoring network can be incorporated into most of the deep unsupervised representation learning (URL)-based anomaly detection models and enhances them as a plug-in component. We next integrate the scoring network into an autoencoder (AE) and four state-of-the-art models to demonstrate the effectiveness and transferability of the design. These score-guided models are collectively called SG-Models. Extensive experiments on both synthetic and real-world datasets confirm the state-of-the-art performance of SG-Models.

The 1st Workshop on City Brain Research examines the current challenges and recent breakthroughs related to intelligent urban transportation. The workshop will be organized in a novel form --- offering debates on three main components involved in the transportation policy development cycle: data collection, policy learning, and the effects on human behavior. The organizers intend to invite speakers and attendees from different backgrounds, ranging from computer science, transportation, to urban planning. The final outcomes include live discussions of the three consistent topics, a comprehensive annual report summarizing current practices and future directions, and a detailed tutorial on the workshop day.

Travel time estimation is a crucial task for not only personal travel scheduling but also city planning. Previous methods focus on modeling toward road segments or sub-paths, then summing up for a final prediction, which have been recently replaced by deep neural models with end-to-end training. Usually, these methods are based on explicit feature representations, including spatio-temporal features, traffic states, etc. Here, we argue that the local traffic condition is closely tied up with the land-use and built environment, i.e., metro stations, arterial roads, intersections, commercial area, residential area, and etc, yet the relation is time-varying and too complicated to model explicitly and efficiently. Thus, this paper proposes an end-to-end multi-task deep neural model, named Deep Image to Time (DeepI2T), to learn the travel time mainly from the built environment images, a.k.a. the morphological layout images, and showoff the new state-of-the-art performance on real-world datasets in two cities. Moreover, our model is designed to tackle both path-aware and path-blind scenarios in the testing phase. This work opens up new opportunities of using the publicly available morphological layout images as considerable information in multiple geography-related smart city applications.

Current research on traffic flow prediction mainly concentrates on generating accurate prediction results based on intelligent or combined algorithms but ignores the interpretability of the prediction model. In practice, however, the interpretability of the model is equally important for traffic managers to realize which road segment in the road network will affect the future traffic state of the target segment in a specific time interval and when such an influence is expected to happen. In this paper, an interpretable and adaptable spatiotemporal Bayesian multivariate adaptive-regression splines (ST-BMARS) model is developed to predict short-term freeway traffic flow accurately. The parameters in the model are estimated in the way of Bayesian inference, and the optimal models are obtained using a Markov chain Monte Carlo (MCMC) simulation. In order to investigate the spatial relationship of the freeway traffic flow, all of the road segments on the freeway are taken into account for the traffic prediction of the target road segment. In our experiments, actual traffic data collected from a series of observation stations along freeway Interstate 205 in Portland, OR, USA, are used to evaluate the performance of the model. Experimental results indicate that the proposed interpretable ST-BMARS model is robust and can generate superior prediction accuracy in contrast with the temporal MARS model, the parametric model autoregressive integrated moving averaging (ARIMA), the state-of-the-art seasonal ARIMA model, and the kernel method support vector regression.

Accurate short-term traffic flow prediction plays a fundamental role in intelligent transportation systems (ITS), e.g. advanced traffic management systems (ATMS). To generate accurate short-term traffic volume, nonparametric models have gained credit from quantities of researchers. On the basis of the common thought that future traffic states can be predicted according to the similar states in the historical traffic data, this paper presents a novel nonparametric-model-based method to predict the short-term traffic volume. The applied nonparametric model is the classification and regression trees (CART) model. In the application, the CART model first classifies the historical traffic states into plentiful categories. Afterwards, the linear regression model is built corresponding to each traffic state pattern. Finally, the model predicts the short-term traffic state through clustering the current state vector into the most congenial historical pattern and regression model. In the experiments, the proposed method is tested by using the 15 minutes average traffic volumes on freeways and is compared with the classic nonparametric methods k-nearest neighbors (k-NN) model, and the parametric method Kalman filter model. The results indicate that the CART-based prediction method outperforms the k-NN and Kalman filter methods in both the mean absolute percentage error and the mean absolute scaled error.

Aiming to comply with the requirement of parallel-transportation management systems (PtMS), this paper presents a short-term traffic flow prediction method for signal-controlled urban traffic networks (UTNs) based on the macroscopic UTN model. In contrast with other time-series-based or spatio-temporal correlation methods, the proposed method focuses more on using the substantial mechanism of traffic transmission in road networks and the topology model of the entire UTN. Furthermore, this approach employs a speed-density model based on the fundamental diagram (FD) to obtain more accurate travel times in links. In the comparison experiment, the microscopic traffic simulation software CORSIM is adopted to simulate the real urban traffic. The experiment results fully verify the outstanding performances of the proposed prediction method.