This is a corrigendum for the article "Score-based Graph Learning for Urban Flow Prediction" published in ACM Trans. Intell. Syst. Technol. 15, 3, Article 59 (May 2024), 25 pages.

Accurate traffic flow measurement is essential for the development of smart cities, yet the deployment of ubiquitous monitoring sensors using traditional methods is often cost-prohibitive. This paper proposes an innovative entropy-aware masked autoencoder framework, namely TMAE, for low-cost traffic flow inference. TMAE leverages a small number of selectively measured regions with few deployed sensors to infer traffic flow across entire urban areas, incorporating prior knowledge from road distribution maps. Specifically, TMAE employs a shared encoder to process traffic flow context, using self-attention scores to identify the importance of each region and guide a masking policy that retains regions rich in traffic flow information. The road distribution map, reflecting inherent traffic flow patterns, is incorporated as prior knowledge by substituting masked tokens during training. A cross-attention mechanism in the decoder further refines inference, where embeddings from the road distribution map serve as queries, and retained visible patches act as keys and values. Additionally, regional traffic entropy is introduced to quantify the information richness of each region, enabling the selection of minimal measurement regions to optimize inference for other areas. Extensive experiments conducted on datasets from various cities demonstrate the effectiveness and efficiency of TMAE, highlighting its potential as a scalable solution for low-cost traffic flow inference in urban environments. The source code of this work is released at https://github.com/TextGraph/TMAE.

Recent advancements in generative modeling have successfully integrated denoising diffusion probabilistic models (DDPMs) into the domain of speech enhancement (SE). Despite their considerable advantages in generalizability, ensuring semantic consistency of the generated samples with the condition signal remains a formidable challenge. Inspired by techniques addressing posterior collapse in variational autoencoders, we explore skip connections within diffusion-based SE models to improve consistency with condition signals. However, experiments reveal that simply adding skip connections is ineffective and even counterproductive. We argue that the independence between the predictive target and the condition signal causes this failure. To address this, we modify the training objective from predicting random Gaussian noise to predicting clean speech and propose a progressive skip connection strategy to mitigate the decrease in mutual information between the layer's output and the condition signal as network depth increases. Experiments on two standard datasets demonstrate the effectiveness of our approach in both seen and unseen scenarios. The code is publicly available at https://github.com/ICDM-UESTC/SCSE.

Accurate urban flow prediction (UFP) is crucial for a range of smart city applications such as traffic management, urban planning, and risk assessment. To capture the intrinsic characteristics of urban flow, recent efforts have utilized spatial and temporal graph neural networks to deal with the complex dependence between the traffic in adjacent areas. However, existing graph neural network based approaches suffer from several critical drawbacks, including improper graph representation of urban traffic data, lack of semantic correlation modeling among graph nodes, and coarse-grained exploitation of external factors. To address these issues, we propose DiffUFP , a novel probabilistic graph-based framework for UFP. DiffUFP consists of two key designs: (1) a semantic region dynamic extraction method that effectively captures the underlying traffic network topology, and (2) a conditional denoising score-based adjacency matrix generator that takes spatial, temporal, and external factors into account when constructing the adjacency matrix rather than simply concatenation in existing studies. Extensive experiments conducted on real-world datasets demonstrate the superiority of DiffUFP over the state-of-the-art UFP models and the effect of the two specific modules.

Inferring the citywide urban traffic flows is critical for numerous smart city applications such as urban planning, traffic control, and transportation management. Urban traffic flow inference problem aims to generate fine-grained flow maps from the coarse-grained ones. It is still challenging due to the lack of handling uncertainties of flow distributions and complex external factors that affect the inference performance. In this work, we propose a diffusion probabilistic augmentation-based network for considering the uncertainties of urban flows with a relaxed structural constraint and a disentangled scheme for flow map and external factor learning. Experiments are conducted on four large-scale urban flow datasets, and the results show that our method achieves significant performance improvements over strong baselines.

Inferring geographic locations via social posts is essential for many practical location-based applications such as product marketing, point-of-interest recommendation, and infector tracking for COVID-19. Unlike image-based location retrieval or social-post text embedding-based location inference, the combined effect of multi-modal information (i.e., post images, text, and hashtags) for social post positioning receives less attention. In this work, we collect real datasets of social posts with images, texts, and hashtags from Instagram and propose a novel Multi-modal Representation Learning Framework (MRLF) capable of fusing different modalities of social posts for location inference. MRLF integrates a multi-head attention mechanism to enhance location-salient information extraction while significantly improving location inference compared with single domain-based methods. To overcome the noisy user-generated textual content, we introduce a novel attention-based character-aware module that considers the relative dependencies between characters of social post texts and hashtags for flexible multi-model information fusion. The experimental results show that MRLF can make accurate location predictions and open a new door to understanding the multi-modal data of social posts for online inference tasks.

Trajectory-user linking is the basis for many applications such as personalized recommendation and urban planning. Although plenty of efforts have been devoted to this topic, the results achieved are still not good enough. Existing methods mainly employ Recurrent Neural Networks (RNNs) to model trajectories semantically due to the inherent sequential attribute of trajectories. However, these approaches are weak at Point of Interest (POI) representation learning and trajectory feature detection. Thus, the performance of existing solutions is far from the requirements of practical applications. In this paper, we propose a novel Trajectory Convolution-based Trajectory-User Linking (TCTUL) method. Firstly, we connect all POI according to trajectories from all users. The result is a connected graph that can be used to generate more informative POI sequences than other approaches. Secondly, we employ the Node2Vec algorithm to encode each POI into a low-dimensional real value vector. Then, we transform each trajectory into an image-like matrix with fixed dimensions. Finally, a CNN is designed to detect features and predict the user of a given trajectory. The CNN can extract informative features from the matrix representations of trajectories by convolutional operations, Batch normalization, and$K$-max pooling operations. Extensive experiments on real datasets demonstrate that TCTUL substantially outperforms existing solutions in terms of macro-Precision, macro-Recall, macro-F1, and accuracy.

Fine-grained urban flow inference (FUFI) aims at enhancing the resolution of traffic flow, which plays an important role in intelligent traffic management. Existing FUFI methods are mainly based on techniques from image super-resolution (SR) models, which cannot fully capture the influence of external factors and face the ill-posed problem in SR tasks. In this paper, we propose UFI-Flow – Urban Flow Inference via normalizing Flow, a novel model for addressing the FUFI problem in a principled manner by using a single probabilistic loss. UFI-Flow therefore directly accounts for the ill-posed nature of the problem and learns spatial correlations between urban flow maps. In addition, an augmented distribution fusion mechanism is further proposed to reinforce the influence of external factors in the joint distribution inference. We conduct comprehensive experiments on real-world datasets to show the superiority of the proposed model compared to the state-of-the-art baseline approaches.

In this work, we propose a novel session-based recommendation model which can fully leverage the intriguing relationships among items. Firstly, a heterogeneous graph with diverse edges is constructed to capture semantic information among items. Meanwhile, two challenges involved in heterogeneous graph are addressed. One is the noisy or conflict knowledge introduced by meta-path based neighbors, and the other is "disconnected graph" which is incurred by sampling from disparate types of relationships. To alleviate these problems, a global-level contrastive learning model on heterogeneous graph is designed, while we also propose an adaptive subgraph sampling algorithm and a new adaptive edge perturbation policy to cope with the isolated node problem on augmentation. Finally, a local-level fine-tuning model is followed to predict users’ next behavior. Extensive experiments are performed on two real-world datasets demonstrating that the performance of our model is superior to the state-of-the-art methods.

One of the main objectives of information cascade popularity prediction is to forecast the future size of a cascade given the observed propagation information. It is an enabling step for many practical applications (e.g., advertisement, academic writing, etc.). Recent advances in neural networks have spurred a few deep learning-based cascade models, which preserve the structural features of information cascades with node embedding and graph neural networks. However, efforts in cascade graph learning as well as its internal temporal dependency, existing methods mainly focus on node-level similarity learning, ignoring the structural equivalence among different sub-graphs that are more informative for information diffusion prediction. Towards this, we present a kernel-based structural-temporal cascade learning model, called CasKernel, to explicitly estimate and encode the structural similarity of cascades with the graph kernels. Moreover, we employ a non sequential process to address the temporal dependency, which can be used to facilitate information popularity prediction. Experiments conducted on both tweets propagation network and academic citation network demonstrate the effectiveness of our method.

The ubiquity of mobile devices use has generated huge volumes of location-aware contextual data, providing opportunities enriching various location-based social network (LBSN) applications - e.g., trip recommendation, ride-sharing allocation and taxi demand prediction etc. Trajectory-based social circle inference (TSCI), which aims at inferring the social relationships among users based on the human mobility data, has received great attention in recent years due to its importance in many LBSN applications. However, existing solutions suffer from three key challenges, including (1) lack of modeling contextual feature in user check-ins; (2) cannot capture the structural information in user motion patterns; (3) and fail to consider the underlying mobility distribution. In this paper, we propose a novel framework ASCI-CAM (Adversity-based Social Circles Inference via Context-Aware Mobility) to address the above challenges. ASCI-CAM is a graph-based model taking into account the contextual information associated with check-ins which, combined with an attentive auto-encoder, allows for semantic trajectory representation. We regularize the learned trajectory embedding with an adversarial learning procedure, which allows us to better understand the user mobility patterns and personalized trajectory distribution. Our extensive experiments on real-world mobility datasets demonstrate that our model achieves significant improvement over the state-of-the-art baselines.

Cross-site account correlation correlates users who have multiple accounts but the same identity across online social networks (OSNs). Being able to identify cross-site users is important for a variety of applications in social networks, security, and electronic commerce, such as social link prediction and cross-domain recommendation. Because of either heterogeneous characteristics of platforms or some unobserved but intrinsic individual factors, the same individuals are likely to behave differently across OSNs, which accordingly causes many challenges for correlating accounts. Traditionally, account correlation is measured by analyzing user-generated content, such as writing style, rules of naming user accounts, or some existing metadata (e.g., account profile, account historical activities). Accounts can be correlated by de-anonymizing user behaviors, which is sometimes infeasible since such data are not often available. In this work, we propose a method, called ACCount eMbedding (ACCM), to go beyond text data and leverage semantics of network structures, a possibility that has not been well explored so far. ACCM aims to correlate accounts with high accuracy by exploiting the semantic information among accounts through random walks. It models and understands latent representations of accounts using an embedding framework similar to sequences of words in natural language models. It also learns a transformation matrix to project node representations into a common dimensional space for comparison. With evaluations on both real-world and synthetic data sets, we empirically demonstrate that ACCM provides performance improvement compared with several state-of-the-art baselines in correlating user accounts between OSNs.

Understanding human trajectory patterns is an important task in many location based social networks (LBSNs) applications, such as personalized recommendation and preference-based route planning. Most of the existing methods classify a trajectory (or its segments) based on spatio-temporal values and activities, into some predefined categories, e.g., walking or jogging. We tackle a novel trajectory classification problem: we identify and link trajectories to users who generate them in the LBSNs, a problem called Trajectory-User Linking (TUL). Solving the TUL problem is not a trivial task because: (1) the number of the classes (i.e., users) is much larger than the number of motion patterns in the common trajectory classification problems; and (2) the location based trajectory data, especially the check-ins, are often extremely sparse. To address these challenges, a Recurrent Neural Networks (RNN) based semi-supervised learning model, called TULER (TUL via Embedding and RNN) is proposed, which exploits the spatio-temporal data to capture the underlying semantics of user mobility patterns. Experiments conducted on real-world datasets demonstrate that TULER achieves better accuracy than the existing methods.

This work presents a novel writing style-based approach to detect multi-account users on User-Generated Content (UGC) sites. Unlike existing works which emphasize feasibility and privacy leakage, we focus on precise writing style-based multi-account detection. Specifically, we leverage a one-class classification-based approach to detect multi-account behaviors, in which a mutual similarity measurement is defined to increase detection precision. In addition to traditional features used in writing style detection, we also extract bigrams, trigrams, part-of-speech, and grammatical relations. We evaluate our methodology based on datasets crawled from 3 popular OSNs (i.e., Twitter, Facebook, and Google+). Experimental results demonstrate that compared with the most recent achievements, our method not only improves the average detection precision to almost 90%, but also increases both recall and F-measure to 90% and even better.

Recent studies have demonstrated that overlay localization can reduce the inter-ISP traffic efficiently, however, fully localized overlays generally impair the streaming quality. In this paper, we first investigate the effects of overlay localization and then present a novel ISP-friendly hierarchical overlay, termed HOPES, to achieve a favorable tradeoff between the inter-ISP traffic and the streaming quality. In HOPES, there are four components: (1) an algorithm for determining the abstract interconnections of all ISP domains; (2) a super peer selection scheme; (3) an algorithm for constructing the inter-ISP connections between super peers; (4) an algorithm for constructing the intra- ISP connections within each same ISP domain. Simulation results indicate that under various scenarios a significant reduction in the inter-ISP traffic is achievable while the streaming quality is enhanced by shrinking the inter-ISP depth and latency of the delivery path of each chunk.

Efficient and effective full-text retrieval in unstructured peer-to-peer networks remains a challenge in the research community. First, it is difficult, if not impossible, for unstructured P2P systems to effectively locate items with guaranteed recall. Second, existing schemes to improve search success rate often rely on replicating a large number of item replicas across the wide area network, incurring a large amount of communication and storage costs. In this paper, we propose BloomCast, an efficient and effective full-text retrieval scheme, in unstructured P2P networks. By leveraging a hybrid P2P protocol, BloomCast replicates the items uniformly at random across the P2P networks, achieving a guaranteed recall at a communication cost of O(√N), where N is the size of the network. Furthermore, by casting Bloom Filters instead of the raw documents across the network, BloomCast significantly reduces the communication and storage costs for replication. We demonstrate the power of BloomCast design through both mathematical proof and comprehensive simulations based on the query logs from a major commercial search engine and NIST TREC WT10G data collection. Results show that BloomCast achieves an average query recall of 91 percent, which outperforms the existing WP algorithm by 18 percent, while BloomCast greatly reduces the search latency for query processing by 57 percent.

Efficient and effective full-text retrieval in unstructured peer-to-peer networks remains a challenge in the research community. First, it is difficult, if not impossible, for unstructured P2P search protocols to effectively locate items with guaranteed recall rate. Second, existing schemes to improve search successful rate often rely on replicating a large number of item replicas across the wide area network, incurring a large amount of communication and storage cost. In this paper we propose BloomCast, an efficient and effective full-text retrieval scheme, in unstructured P2P networks. BloomCast is effective because it guarantees perfect recall rate with high probability. It is efficient because the overall communication cost of full-text search is reduced below a formal bound. Furthermore, by casting Bloom Filters instead of the raw documents across the network, BloomCast significantly reduces the communication cost and storage cost for replication. We demonstrate the power of BloomCast design through both mathematical proof and comprehensive simulations. Results show that BloomCast outperforms existing schemes in terms of both recall rate and communication cost.

Existing replication strategies in unstructured P2P networks, such as square-root principle based replication, can effectively improve search efficiency. How to get optimal replication strategy, however, is not trivial. In this paper we show, through mathematical proof, that random replication strategy achieves the optimal results. By randomly distributing rather small numbers of item and query replicas in the unstructured P2P network, we can guarantee perfect search success rate comparable to exhaustive search with high probability. Our analysis also shows that the cost for such replication strategy is determined by the network size of a P2P system. We propose a hybrid P2P architecture which combines a lightweight DHT with an unstructured P2P overlay to address the problems of network size estimating and random peer sampling. We conduct comprehensive simulation to evaluate this design. Results show that our scheme achieves perfect search success rate with quite small overhead.