Next Point-of-Interest (POI) recommendation task focuses on mining user behavioral preference patterns from historical check-ins to provide personalized suggestions for the next destination. Existing methods primarily rely on shallow contextual information and handcrafted feature interactions to predict the next POI. However, the inherent sparsity and complexity of user mobility patterns limit the computational capacity of non-reasoning models to capture deep intent, while large language models (LLMs) perform suboptimally because they lack a deep understanding of semantic IDs (SIDs) when SIDs are trained separately. To address these limitations, we propose Think2Go, a novel generative next POI recommendation framework, which enhances the model's comprehension of SID representations and explores diverse spatial-temporal patterns via test-time computational scaling. We unify supervised fine-tuning (SFT) and reinforcement learning (RL)-based reasoning within a single architecture, enabling joint optimization of memorization and adaptive reasoning to better retain user behavior patterns while exploring diverse user preferences. To further calibrate policy optimization in adaptive reasoning, we propose two advantage weighting mechanisms that integrate (1) prompt epistemic uncertainty, estimated via kernel density methods to assess the spatial-temporal periodic pattern alignment between queries and user history, promoting increased exploration under high epistemic uncertainty; and (2) reward-informed advantage scaling, captured by normalizing rewards against their maxima to adapt update magnitudes, thereby improving training stability and mitigating overfitting to noisy signals. This joint calibration forms an implicit curriculum learning strategy, delivering fine-grained, instance-aware policy updates that prevent entropy collapse and support robust exploration. Extensive experiments conducted on three real-world datasets demonstrate that Think2Go exhibits strong generalization capabilities and enhances the LLM's understanding of SIDs.

Graphs are prevalent in numerous real-world applications. Previous methods directly model graph structures and achieve significant success. However, these methods encounter bottlenecks due to the inherent irregularity of graphs. An innovative solution is converting graphs into textual representations, thereby harnessing the powerful capabilities of Large Language Models (LLMs) to process and comprehend graphs. In this article, we present a comprehensive review of methodologies for applying LLMs to graphs, termed LLM4graph. The core of LLM4graph lies in transforming graphs into texts for LLMs to understand and analyze. Thus, we propose a novel taxonomy of LLM4graph methods from the view of the transformation. Specifically, existing methods can be divided into two paradigms: Graph2text and Graph2token, which transform graphs into texts or tokens as the input of LLMs, respectively. We point out four challenges during the transformation to systematically present existing methods from a problem-oriented perspective. For practical concerns, we provide a guideline for researchers on selecting appropriate models and LLMs for different graphs and hardware constraints. To empirically evaluate our taxonomy and different technical choices, we conduct experiments with representative methods in Graph2text and Graph2token. We also identify five future research directions for LLM4graph.

Next Point-of-Interest (POI) prediction is important to various human mobility applications, such as route planning and location-based advertising. To address the spatial-temporal sparsity issues arising from users' irregular and inconsistent visit times to different POIs, multi-granular structures can be incorporated to enhance feature representation through hierarchical relationships. However, existing methods often fall short in capturing the comprehensive multi-granularity spatial-temporal correlations due to three primary limitations: (1) users' complex mobility patterns entangled in single trajectory data, (2) limited mobility patterns details due to independent modeling at each granularity, and (3) low inference efficiency in cascaded multi-granularity predictions. To tackle these challenges, we propose a novel approach that models transformations across different granularities in both spatial regions and temporal periods as a diffusion process, leading to the development of the Multi-Granularity Spatial-Temporal Diffusion Network (MGSTDN). In particular, this model adopts a multi-task architecture, where predictions at varying spatial-temporal granularities (i.e., different diffusion steps) are treated as distinct tasks. By employing a multi-granularity diffusion mechanism in both spatial and temporal dimensions, it captures more nuanced spatial-temporal correlations, enhancing the physical constraints and behavioral pattern dependencies across granularities. During the diffusion process's forward stage, coarser-grained regions and periods are derived based on fine-grained features. In the reverse stage, finer-grained regions and periods are recovered from coarse-grained features, guided by encoded historical trajectory information, until the next POI is determined. To improve computational efficiency, we introduce a multi-granularity mapping propagation matrix, enabling parallel computation and accelerating the prediction process across different granularities. We evaluated the effectiveness of MGSTDN through extensive experiments on three datasets, demonstrating significant improvements over existing methods.

Generative self-supervised learning has recently become popular in molecular modeling because it can improve accuracy and generalization. However, existing generative self-supervised tasks often have simplified designs that do not effectively use substructure information. Substructure information is important for molecules because it can provide local semantics and capture analogous semantic information on a graph-level scale. For example, -OH, as one of the substructures, is typically associated with hydrophilicity. To address this limitation, we propose a novel pretraining task that incorporates substructure information into generative self-supervised tasks. This integration involves creating a substructure-based vocabulary and fusing structural insights into the representation learning process. We evaluate our approach on 10 publicly available datasets, covering diverse molecular property prediction tasks. Our results consistently show the effectiveness of incorporating substructure information compared with both contrastive and generative self-supervised pretraining methodologies.

Recently, the Weisfeiler-Lehman (WL) graph isomorphism test was used to measure the expressiveness of graph neural networks (GNNs), showing that the neighborhood aggregation GNNs were at most as powerful as 1-WL test in distinguishing graph structures. There were also improvements proposed in analogy to k-WL test ($k>1$). However, the aggregations in these GNNs are far from injective as required by the WL test, and suffer from weak distinguishing strength, making it become the expression bottleneck. In this paper, we improve the expressiveness by exploring powerful aggregations. We reformulate an aggregation with the corresponding aggregation coefficient matrix, and then systematically analyze the requirements on this matrix for building more powerful and even injective aggregations. We also show the necessity of applying nonlinear units ahead of aggregations, which is different from most existing GNNs. Based on our theoretical analysis, we develop ExpandingConv. Experimental results show that our model significantly boosts performance, especially for large and densely connected graphs.

For learning graph representations, not all detailed structures within a graph are relevant to the given graph tasks. Task-relevant structures can be localized or sparse which are only involved in subgraphs or characterized by the interactions of subgraphs (a hierarchical perspective). A graph neural network should be able to efficiently extract task-relevant structures and be invariant to irrelevant parts, which is challenging for general message passing GNNs. In this work, we propose to learn graph representations from a sequence of subgraphs of the original graph to better capture task-relevant substructures or hierarchical structures and skip noisy parts. To this end, we design soft-mask GNN layer to extract desired subgraphs through the mask mechanism. The soft-mask is defined in a continuous space to maintain the differentiability and characterize the weights of different parts. Compared with existing subgraph or hierarchical representation learning methods and graph pooling operations, the soft-mask GNN layer is not limited by the fixed sample or drop ratio, and therefore is more flexible to extract subgraphs with arbitrary sizes. Extensive experiments on public graph benchmarks show that soft-mask mechanism brings performance improvements. And it also provides interpretability where visualizing the values of masks in each layer allows us to have an insight into the structures learned by the model.

Users’ preferences, and consequently their ratings and reviews to items, change over time. Likewise, characteristics of items are also time-varying. By dividing data into time periods, temporal Recommender Systems (RSs) improve recommendation accuracy by exploring the temporal dynamics in user rating data. However, temporal RSs have to cope with rating sparsity in each time period. Meanwhile, reviews generated by users contain rich information about their preferences, which can be exploited to address rating sparsity and further improve the performance of temporal RSs. In this article, we develop a temporal rating model with topics that jointly mines the temporal dynamics of both user-item ratings and reviews. Studying temporal drifts in reviews helps us understand item rating evolutions and user interest changes over time. Our model also automatically splits the review text in each time period into interim words and intrinsic words. By linking interim words and intrinsic words to short-term and long-term item features, respectively, we jointly mine the temporal changes in user and item latent features together with the associated review text in a single learning stage. Through experiments on 28 real-world datasets collected from Amazon , we show that the rating prediction accuracy of our model significantly outperforms the existing state-of-art RS models. And our model can automatically identify representative interim words in each time period as well as intrinsic words across all time periods. This can be very useful in understanding the time evolution of users’ preferences and items’ characteristics.