The large language model (LLM) powered recommendation paradigm has been proposed to address the limitations of traditional recommender systems (RecSys), which often struggle to handle cold-start users or items with new IDs. Despite its effectiveness, this study uncovers that LLM-empowered RecSys are vulnerable to reconstruction attacks that can expose both system and user privacy. To thoroughly examine this threat, we present the first systematic study on inversion attacks targeting LLM-empowered RecSys, wherein adversaries attempt to reconstruct original prompts that contain personal preferences, interaction histories, and demographic attributes by exploiting the output logits of recommendation models. We reproduce the vec2text framework and optimize it using our proposed method - Similarity-Guided Refinement, enabling more accurate reconstruction of textual prompts from model-generated logits. Extensive experiments across two domains (movies and books) and two representative LLM-based recommendation models demonstrate that our method achieves high-fidelity reconstructions. Specifically, we can recover nearly 65% of the user-interacted items and correctly infer age and gender in 87% of the cases. The experiments also reveal that privacy leakage is largely insensitive to the victim model's performance but highly dependent on domain consistency and prompt complexity. These findings expose critical and privacy vulnerabilities in LLM-empowered RecSys. The code for reproduction is provided below: https://github.com/xuemingxxx/Attack_RecSys/

Models based on Transformer variants have consistently demonstrated leading performance in long sequence time series forecasting. However, in some complex application scenarios, Transformers tend to capture low-frequency information in the data while overlooking high-frequency information, which often contains rich non-stationary features. This unbalanced feature extraction approach limits the model's ability to effectively handle real-world time series data. To address this issue, we explicitly represent both low-frequency and high-frequency information and propose a model called STCNet, a data-driven scale-adaptive convolutional network that aims to extract diverse features and patterns from the data by learning features across different frequency bands in a balanced manner. Specifically, we propose an entropy-based adaptive wavelet basis selection algorithm, which can adaptively select appropriate wavelet bases based on the data distribution to achieve effective multi-frequency decomposition of complex time series. In addition, we designed a hierarchical scale-adaptive factor that allows for dynamic adjustment of feature weights according to different time scales through refined layered weight adjustment, significantly enhancing the model's capability in handling non-stationary time series features. To further optimize the output features of the model, we introduce a test-time training mechanism, combined with a fast weight update strategy and a weight-sharing strategy to reduce the number of model parameters, effectively mitigating the risk of overfitting. Experimental results on nine datasets demonstrate that STCNet outperforms the current state-of-the-art models in both effectiveness and efficiency.

Deep models for Multivariate Time Series (MTS) forecasting have recently demonstrated significant success. Channel-dependent models capture complex dependencies that channel-independent models cannot capture. However, the number of channels in real-world applications outpaces the capabilities of existing channel-dependent models, and contrary to common expectations, some models underperform the channel-independent models in handling high-dimensional data, which raises questions about the performance of channel-dependent models. To address this, our study first investigates the reasons behind the suboptimal performance of these channel-dependent models on high-dimensional MTS data. Our analysis reveals that two primary issues lie in the introduced noise from unrelated series that increases the difficulty of capturing the crucial inter-channel dependencies, and challenges in training strategies due to high-dimensional data. To address these issues, we propose STHD, the Scalable Transformer for High-Dimensional Multivariate Time Series Forecasting. STHD has three components: a) Relation Matrix Sparsity that limits the noise introduced and alleviates the memory issue; b) ReIndex applied as a training strategy to enable a more flexible batch size setting and increase the diversity of training data; and c) Transformer that handles 2-D inputs and captures channel dependencies. These components jointly enable STHD to manage the high-dimensional MTS while maintaining computational feasibility. Furthermore, experimental results show STHD's considerable improvement on three high-dimensional datasets: Crime-Chicago, Wiki-People, and Traffic. The source code and dataset are publicly available https://github.com/xinzzzhou/ScalableTransformer4HighDimensionMTSF.git.

Ontology matching (OM) enables semantic interoperability between different ontologies and resolves their conceptual heterogeneity by aligning related entities. OM systems currently have two prevailing design paradigms: conventional knowledge-based expert systems and newer machine learning-based predictive systems. While large language models (LLMs) and LLM agents have revolutionised data engineering and have been applied creatively in many domains, their potential for OM remains underexplored. This study introduces a novel agent-powered LLM-based design paradigm for OM systems. With consideration of several specific challenges in leveraging LLM agents for OM, we propose a generic framework, namely Agent-OM (Agent for Ontology Matching), consisting of two Siamese agents for retrieval and matching, with a set of OM tools. Our framework is implemented in a proof-of-concept system. Evaluations of three Ontology Alignment Evaluation Initiative (OAEI) tracks over state-of-the-art OM systems show that our system can achieve results very close to the long-standing best performance on simple OM tasks and can significantly improve the performance on complex and few-shot OM tasks.

Reinforcement learning (RL)-based multi-hop reasoning has become an interpretable way for knowledge graph reasoning owing to its persuasive explanations for the predicted results, but the reasoning performance of these methods drops significantly over few-shot relations (only contain few triplets). To address this problem, recent studies introduce meta-learning into RL-based reasoning methods. However, the performance of these studies is still limited due to the following points: (1) the overall reasoning accuracy is impaired due to the low reasoning accuracies over some hard relations; (2) the reasoning process becomes laborious and ineffective owing to the existence of noisy data; (3) the generalizability is negatively affected due to the lack of knowledge-sharing. To tackle these challenges, we propose a novel modelHMLSconsisting of two modulesHHML(HierarchicalHardness-awareMeta-reinforcementLearning) andHHS(HierarchicalHardness-awareSampling). Specifically,HHMLcontains the following two components: (1) a hardness-aware RL conducts multi-hop reasoning by training hardness-aware batches and reducing noise; (2) a knowledge-sharing meta-learning adapts to few-shot relations by exploiting common features in the hierarchical relation structure. The other moduleHHSgenerates hardness-aware batches from relation and relation-cluster levels. The experimental results demonstrate that this work notably outperforms the state-of-the-art approaches in few-shot scenarios.

Inductive link prediction (ILP) is to predict links for unseen entities in emerging knowledge graphs (KGs), considering the evolving nature of KGs. A more challenging scenario is that emerging KGs consist of only unseen entities without any edge connected to original KGs, called as disconnected emerging KGs (DEKGs). Existing studies for DEKGs only focus on predicting enclosing links, i.e., predicting links inside the emerging KG. The bridging links, which carry the evolutionary information from the original KG to DEKG, have not been investigated by previous work so far. To fill in the gap, we propose a novel model entitled DEKG-ILP (Disconnected Emerging Knowledge Graph Oriented Inductive Link Prediction) that consists of the following two components. (1) The module CLRM (Contrastive Learning-based Relation-specific Feature Modeling) is developed to extract global relation-based semantic features that are shared between original KGs and DEKGs with a novel sampling strategy. (2) The module GSM (GNN-based Subgraph Modeling) is proposed to extract the local subgraph topological information around each link in KGs. The extensive experiments conducted on several benchmark datasets demonstrate that DEKG-ILP has obvious performance improvements compared with state-of-the-art methods for both enclosing and bridging link prediction.

Multi-modal knowledge graphs (MKGs) include not only the relation triplets, but also related multi-modal auxiliary data (i.e., texts and images), which enhance the diversity of knowledge. However, the natural incompleteness has significantly hindered the applications of MKGs. To tackle the problem, existing studies employ the embedding-based reasoning models to infer the missing knowledge after fusing the multi-modal features. However, the reasoning performance of these methods is limited due to the following problems: (1) ineffective fusion of multi-modal auxiliary features; (2) lack of complex reasoning ability as well as inability to conduct the multi-hop reasoning which is able to infer more missing knowledge. To overcome these problems, we propose a novel model entitled MMKGR (Multi-hop Multi-modal Knowledge Graph Reasoning). Specifically, the model contains the following two components: (1) a unified gate-attention network which is designed to generate effective multi-modal complementary features through sufficient attention interaction and noise reduction; (2) a complementary feature-aware reinforcement learning method which is proposed to predict missing elements by performing the multi-hop reasoning process, based on the features obtained in component (1). The experimental results demonstrate that MMKGR outperforms the state-of-the-art approaches in the MKG reasoning task.

With the prevalence of online social platforms, social recommendation has emerged as a promising direction that leverages the social network among users to enhance recommendation performance. However, the available social relations among users are usually extremely sparse and noisy, which may lead to inferior recommendation performance. To alleviate this problem, this paper novelly exploits the implicit higher-order social influence and dependencies among users to enhance social recommendation. In this paper, we propose a novel embedding method for general bipartite graphs, which defines inter-class message passing between explicit relations and intra-class message passing between implicit higher-order relations via a novel sequential modelling paradigm. Inspired by recent advances in self-attention-based sequential modelling, the proposed model features a self-attentive representation learning mechanism for implicit user-user relations. Moreover, this paper also explores the inductive embedding learning for social recommendation problems to improve the recommendation performance in cold-start settings. The proposed inductive learning paradigm for social recommendation enables embedding inference for those cold-start users and items (unseen during training) as long as they are linked to existing nodes in the original network. Extensive experiments on real-world datasets demonstrate the superiority of our method and suggest that higher-order implicit relationship among users is beneficial to improving social recommendation.

Knowledge graph (KG) embedding aims to encode both entities and relations into a continuous vector space. Most existing methods require that all entities should be observed during training while ignoring the evolving nature of KG. Major recent efforts on this issue embed new entities by aggregating neighborhood information from existing entities and relations with Graph Neural Network (GNN). However, these methods rely on the neighbors seen during training and suffer from the embedding of new entities with insufficient triplets or triplets with the unseen-to-unseen form. To relieve this problem, we propose a two-stage learning model referred as Hyper-Relation Feature Learning Network (HRFN) for effective out-of-knowledge-base embedding. For the first stage, HRFN learns pre-representations for emerging entities using hyper-relation features meta-learned from the training set. A novel feature aggregating network that involves an entity-centered Graph Convolutional Network (GCN) and a relation-centered GCN is proposed to aggregate information from both new entities themselves and their neighbors. For stage two, a transductive learning network is employed to learn finer-grained embeddings based on above-mentioned pre-representations of new entities. Experimental results on the link prediction task demonstrate the superiority of our model. Further analysis is also done to validate the effectiveness and efficiency of pre-representing emerging entities with the hyper-relation feature.

Sequential user behavior modeling plays a crucial role in online user-oriented services, such as product purchasing, news feed consumption, and online advertising. The performance of sequential modeling heavily depends on the scale and quality of historical behaviors. However, the number of user behaviors inherently follows a long-tailed distribution, which has been seldom explored. In this work, we argue that focusing on tail users could bring more benefits and address the long tails issue by learning transferrable parameters from both optimization and feature perspectives. Specifically, we propose a gradient alignment optimizer and adopt an adversarial training scheme to facilitate knowledge transfer from the head to the tail. Such methods can also deal with the cold-start problem of new users. Moreover, it could be directly adaptive to various well-established sequential models. Extensive experiments on four real-world datasets verify the superiority of our framework compared with the state-of-the-art baselines.

On-demand cinemas are a new type of offline entertainment venues which have shown the rapid expansion in the recent years. Recommending movies of interest to the potential audiences in on-demand cinemas is keen but challenging because the recommendation scenario is totally different from all the existing recommendation applications including online video recommendation, offline item recommendation and group recommendation. In this paper, we propose a novel spatio-temporal approach called Pegasus. Because of the specific characteristics of on-demand cinema recommendation, Pegasus exploits the POI (Point of Interest) information around cinemas and the content descriptions of movies, apart from the historical movie consumption records of cinemas. Pegasus explores the temporal dynamics and spatial influences rooted in audience behaviors, and captures the similarities between cinemas, the changes of audience crowds, time-varying features and regional disparities of movie popularity. It offers an effective and explainable way to recommend movies to on-demand cinemas. The corresponding Pegasus system has been deployed in some pilot on-demand cinemas. Based on the real-world data from on-demand cinemas, extensive experiments as well as pilot tests are conducted. Both experimental results and post-deployment feedback show that Pegasus is effective.

Accurate user representation learning has been proven fundamental for many social media applications, including community detection, recommendation, etc. A major challenge lies in that, the available data in a single social network are usually very limited and sparse. In real life, many people are members of several social networks in the same time. Constrained by the features and design of each, any single social platform offers only a partial view of a user from a particular perspective. In this paper, we propose MV-URL, a multi-view user representation learning model to enhance user modeling by integrating the knowledge from various networks. Different from the traditional network embedding frameworks where either the whole framework is single-network based or each network involved is a homogeneous network, we focus on multiple social networks and each network in our task is a heterogeneous network. It's very challenging to effectively fuse knowledge in this setting as the fusion depends upon not only the varying relatedness of information sources, but also the target application tasks. MV-URL focuses on two tasks: user account linkage (i.e., to predict the missing true user account linkage across social media) and user attribute prediction. Extensive evaluations have been conducted on two real-world collections of linked social networks, and the experimental results show the superiority of MV-URL compared with existing state-of-art embedding methods. It can be learned online, and is trivially parallelizable. These qualities make it suitable for real world applications.

Sources of complementary information are connected when we link the user accounts belonging to the same user across different domains or devices. The expanded information promotes the development of a wide range of applications, such as cross-domain prediction, cross-domain recommendation, and advertisement. Due to the great significance of user account linkage, there are increasing research works on this study. With the widespread popularization of GPS-enabled mobile devices, linking user accounts with location data has become an important and promising research topic. Being different from most existing studies in this domain that only focus on the effectiveness, we propose novel approaches to improve both effectiveness and efficiency of user account linkage. In this paper, a kernel density estimation (KDE) based method has been proposed to improve the accuracy by alleviating the data sparsity problem in measuring users' similarities. To improve the efficiency, we develop a grid-based structure to organize location data to prune the search space. The extensive experiments conducted on two real-world datasets demonstrate the superiority of the proposed approach in terms of both effectiveness and efficiency compared with the state-of-art methods.

Heterogenous information network embedding aims to embed heterogenous information networks (HINs) into low dimensional spaces, in which each vertex is represented as a low-dimensional vector, and both global and local network structures in the original space are preserved. However, most of existing heterogenous information network embedding models adopt the dot product to measure the proximity in the low dimensional space, and thus they can only preserve the first-order proximity and are insufficient to capture the global structure. Compared with homogenous information networks, there are multiple types of links (i.e., multiple relations) in HINs, and the link distribution w.r.t relations is highly skewed. To address the above challenging issues, we propose a novel heterogenous information network embedding model PME based on the metric learning to capture both first-order and second-order proximities in a unified way. To alleviate the potential geometrical inflexibility of existing metric learning approaches, we propose to build object and relation embeddings in separate object space and relation spaces rather than in a common space. Afterwards, we learn embeddings by firstly projecting vertices from object space to corresponding relation space and then calculate the proximity between projected vertices. To overcome the heavy skewness of the link distribution w.r.t relations and avoid "over-sampling'' or "under-sampling'' for each relation, we propose a novel loss-aware adaptive sampling approach for the model optimization. Extensive experiments have been conducted on a large-scale HIN dataset, and the experimental results show superiority of our proposed PME model in terms of prediction accuracy and scalability.

Studying recommender systems under streaming scenarios has become increasingly important because real-world applications produce data continuously and rapidly. However, most existing recommender systems today are designed in the context of an offline setting. Compared with the traditional recommender systems, large-volume and high-velocity are posing severe challenges for streaming recommender systems. In this paper, we investigate the problem of streaming recommendations being subject to higher input rates than they can immediately process with their available system resources (i.e., CPU and memory). In particular, we provide a principled framework called as SPMF (Stream-centered Probabilistic Matrix Factorization model), based on BPR (Bayesian Personalized Ranking) optimization framework, for performing efficient ranking based recommendations in stream settings. Experiments on three real-world datasets illustrate the superiority of SPMF in online recommendations.

With the rapid development of location-based social networks (LBSNs), spatial item recommendation has become an important way of helping users discover interesting locations to increase their engagement with location-based services. The availability of spatial, temporal, and social information in LBSNs offers an unprecedented opportunity to enhance the spatial item recommendation. Many previous works studied spatial and social influences on spatial item recommendation in LBSNs. Due to the strong correlations between a user’s check-in time and the corresponding check-in location, which include the sequential influence and temporal cyclic effect, it is essential for spatial item recommender system to exploit the temporal effect to improve the recommendation accuracy. Leveraging temporal information in spatial item recommendation is, however, very challenging, considering (1) when integrating sequential influences, users’ check-in data in LBSNs has a low sampling rate in both space and time, which renders existing location prediction techniques on GPS trajectories ineffective, and the prediction space is extremely large, with millions of distinct locations as the next prediction target, which impedes the application of classical Markov chain models; (2) there are various temporal cyclic patterns (i.e., daily, weekly, and monthly) in LBSNs, but existing work is limited to one specific pattern; and (3) there is no existing framework that unifies users’ personal interests, temporal cyclic patterns, and the sequential influence of recently visited locations in a principled manner. In light of the above challenges, we propose a Temporal Personalized Model ( TPM ), which introduces a novel latent variable topic-region to model and fuse sequential influence, cyclic patterns with personal interests in the latent and exponential space. The advantages of modeling the temporal effect at the topic-region level include a significantly reduced prediction space, an effective alleviation of data sparsity, and a direct expression of the semantic meaning of users’ spatial activities. Moreover, we introduce two methods to model the effect of various cyclic patterns. The first method is a time indexing scheme that encodes the effect of various cyclic patterns into a binary code. However, the indexing scheme faces the data sparsity problem in each time slice. To deal with this data sparsity problem, the second method slices the time according to each cyclic pattern separately and explores these patterns in a joint additive model. Furthermore, we design an asymmetric Locality Sensitive Hashing (ALSH) technique to speed up the online top- k recommendation process by extending the traditional LSH. We evaluate the performance of TPM on two real datasets and one large-scale synthetic dataset. The performance of TPM in recommending cold-start items is also evaluated. The results demonstrate a significant improvement in TPM’s ability to recommend spatial items, in terms of both effectiveness and efficiency, compared with the state-of-the-art methods.

Cross-device and cross-domain user linkage have been attracting a lot of attention recently. An important branch of the study is to achieve user linkage with spatio-temporal data generated by the ubiquitous GPS-enabled devices. The main task in this problem is twofold, i.e., how to extract the representative features of a user; how to measure the similarities between users with the extracted features. To tackle the problem, we propose a novel model STUL (Spatio-Temporal User Linkage) that consists of the following two components. 1) Extract users - spatial features with a density based clustering method, and extract the users - temporal features with the Gaussian Mixture Model. To link user pairs more precisely, we assign different weights to the extracted features, by lightening the common features and highlighting the discriminative features. 2) Propose novel approaches to measure the similarities between users based on the extracted features, and return the pair-wise users with similarity scores higher than a predefined threshold. We have conducted extensive experiments on three real-world datasets, and the results demonstrate the superiority of our proposed STUL over the state-of-the-art methods.

With the rapid prevalence of smart mobile devices and the dramatic proliferation of mobile applications (Apps), App recommendation becomes an emergent task that will benefit different stockholders of mobile App ecosystems. Unlike traditional items, Apps have privileges to access a user's sensitive resources (e.g., contacts, messages and locations) which may lead to security risk or privacy leak. Thus, users' choosing of Apps are influenced by not only their personal interests but also their privacy preferences. Moreover, user privacy preferences vary with App categories. In this paper, we propose a mobile sparse additive generative model (Mobi-SAGE) to recommend Apps by considering both user interests and category-aware user privacy preferences. We collected a real-world dataset from 360 App store - the biggest Android App platform in China, and conduct extensive experiments on it. The experimental results show that our Mobi-SAGE consistently and significantly outperforms the state-of-the-art approaches, which implies the importance of exploiting category-aware user privacy preferences.

With the rapid development of location-based social networks (LBSNs), spatial item recommendation has become an important mobile application, especially when users travel away from home. However, this type of recommendation is very challenging compared to traditional recommender systems. A user may visit only a limited number of spatial items, leading to a very sparse user-item matrix. This matrix becomes even sparser when the user travels to a distant place, as most of the items visited by a user are usually located within a short distance from the user’s home. Moreover, user interests and behavior patterns may vary dramatically across different time and geographical regions. In light of this, we propose ST-SAGE, a spatial-temporal sparse additive generative model for spatial item recommendation in this article. ST-SAGE considers both personal interests of the users and the preferences of the crowd in the target region at the given time by exploiting both the co-occurrence patterns and content of spatial items. To further alleviate the data-sparsity issue, ST-SAGE exploits the geographical correlation by smoothing the crowd’s preferences over a well-designed spatial index structure called the spatial pyramid . To speed up the training process of ST-SAGE, we implement a parallel version of the model inference algorithm on the GraphLab framework. We conduct extensive experiments; the experimental results clearly demonstrate that ST-SAGE outperforms the state-of-the-art recommender systems in terms of recommendation effectiveness, model training efficiency, and online recommendation efficiency.

Point-of-interest (POI) recommendation has become an important way to help people discover attractive and interesting places, especially when they travel out of town. However, the extreme sparsity of user-POI matrix and cold-start issues severely hinder the performance of collaborative filtering-based methods. Moreover, user preferences may vary dramatically with respect to the geographical regions due to different urban compositions and cultures. To address these challenges, we stand on recent advances in deep learning and propose a Spatial-Aware Hierarchical Collaborative Deep Learning model (SH-CDL). The model jointly performs deep representation learning for POIs from heterogeneous features and hierarchically additive representation learning for spatial-aware personal preferences. To combat data sparsity in spatial-aware user preference modeling, both the collective preferences of the public in a given target region and the personal preferences of the user in adjacent regions are exploited in the form of social regularization and spatial smoothing. To deal with the multimodal heterogeneous features of the POIs, we introduce a late feature fusion strategy into our SH-CDL model. The extensive experimental analysis shows that our proposed model outperforms the state-of-the-art recommendation models, especially in out-of-town and cold-start recommendation scenarios.

With the rapid development of location-based social networks (LBSNs), spatial item recommendation has become an important way of helping users discover interesting locations to increase their engagement with location-based services. Although human movement exhibits sequential patterns in LBSNs, most current studies on spatial item recommendations do not consider the sequential influence of locations. Leveraging sequential patterns in spatial item recommendation is, however, very challenging, considering 1) users' check-in data in LBSNs has a low sampling rate in both space and time, which renders existing prediction techniques on GPS trajectories ineffective; 2) the prediction space is extremely large, with millions of distinct locations as the next prediction target, which impedes the application of classical Markov chain models; and 3) there is no existing framework that unifies users' personal interests and the sequential influence in a principled manner. In light of the above challenges, we propose a sequential personalized spatial item recommendation framework (SPORE) which introduces a novel latent variable topic-region to model and fuse sequential influence with personal interests in the latent and exponential space. The advantages of modeling the sequential effect at the topic-region level include a significantly reduced prediction space, an effective alleviation of data sparsity and a direct expression of the semantic meaning of users' spatial activities. Furthermore, we design an asymmetric Locality Sensitive Hashing (ALSH) technique to speed up the online top-k recommendation process by extending the traditional LSH. We evaluate the performance of SPORE on two real datasets and one large-scale synthetic dataset. The results demonstrate a significant improvement in SPORE's ability to recommend spatial items, in terms of both effectiveness and efficiency, compared with the state-of-the-art methods.

Point-of-Interest (POI) recommendation has become an important means to help people discover attractive and interesting places, especially when users travel out of town. However, the extreme sparsity of a user-POI matrix creates a severe challenge. To cope with this challenge, we propose a unified probabilistic generative model, the Topic-Region Model (TRM) , to simultaneously discover the semantic, temporal, and spatial patterns of users’ check-in activities, and to model their joint effect on users’ decision making for selection of POIs to visit. To demonstrate the applicability and flexibility of TRM, we investigate how it supports two recommendation scenarios in a unified way, that is, hometown recommendation and out-of-town recommendation. TRM effectively overcomes data sparsity by the complementarity and mutual enhancement of the diverse information associated with users’ check-in activities (e.g., check-in content, time, and location) in the processes of discovering heterogeneous patterns and producing recommendations. To support real-time POI recommendations, we further extend the TRM model to an online learning model, TRM-Online, to track changing user interests and speed up the model training. In addition, based on the learned model, we propose a clustering-based branch and bound algorithm (CBB) to prune the POI search space and facilitate fast retrieval of the top- k recommendations. We conduct extensive experiments to evaluate the performance of our proposals on two real-world datasets, including recommendation effectiveness, overcoming the cold-start problem, recommendation efficiency, and model-training efficiency. The experimental results demonstrate the superiority of our TRM models, especially TRM-Online, compared with state-of-the-art competitive methods, by making more effective and efficient mobile recommendations. In addition, we study the importance of each type of pattern in the two recommendation scenarios, respectively, and find that exploiting temporal patterns is most important for the hometown recommendation scenario, while the semantic patterns play a dominant role in improving the recommendation effectiveness for out-of-town users.

With the rapid development of location-based social networks (LBSNs), spatial item recommendation has become an important means to help people discover attractive and interesting venues and events, especially when users travel out of town. However, this recommendation is very challenging compared to the traditional recommender systems. A user can visit only a limited number of spatial items, leading to a very sparse user-item matrix. Most of the items visited by a user are located within a short distance from where he/she lives, which makes it hard to recommend items when the user travels to a far away place. Moreover, user interests and behavior patterns may vary dramatically across different geographical regions. In light of this, we propose Geo-SAGE, a geographical sparse additive generative model for spatial item recommendation in this paper. Geo-SAGE considers both user personal interests and the preference of the crowd in the target region, by exploiting both the co-occurrence pattern of spatial items and the content of spatial items. To further alleviate the data sparsity issue, Geo-SAGE exploits the geographical correlation by smoothing the crowd's preferences over a well-designed spatial index structure called spatial pyramid. We conduct extensive experiments and the experimental results clearly demonstrate our Geo-SAGE model outperforms the state-of-the-art.