Abstract Session-based recommendation (SBR) predicts potential items of interest by analyzing user behavior within sessions. In this work, we explore the continual learning for SBR task, a challenging and practical task closely aligned with real-world online recommendation system due to (1) periodic updates of the model may trigger catastrophic forgetting, and (2) the continuous emergence of new interactions reflects rapidly changing user interests. Although recent studies have mitigated catastrophic forgetting by replaying a small subset of historical data into the model, these samples fail to represent the distribution of the entire dataset. Moreover, the research on SBR when examining changes in user interests is confined to offline settings and does not adequately consider multiple time-correlated user interests. This limitation makes it challenging to finely model the rapid changes in user interests in continual learning. To overcome the limitations of traditional data replay, we propose a data distillation framework for SBR, which synthesizes information-rich samples for replay from the entire dataset instead of relying on simple sampling. Furthermore, to address the changing user interests in continual learning scenarios, we developed an Interest Enhanced Sub-graph Neural Network (IES-GNN), which is capable of efficiently extracting and dynamically modeling the evolution of user interests within sessions. Testing on three real-world datasets demonstrates that our approach outperforms several advanced methods in continual learning for SBR task.

Session-based recommendation (SBR) systems have gained significant attention in recent years due to their ability to model user–item interaction sequences. However, traditional models struggle to adapt to dynamic environments, particularly when capturing real-time changes in user preferences. To address this, streaming SBR (SSBR) systems have been proposed, which train long-term user preference models offline and update short-term preferences dynamically in a streaming setting. Despite this advancement, existing SSBR methods face challenges in modeling complex inter-session dependencies and capturing the evolution of user interests. To overcome these limitations, we introduce the hierarchical information fusion–graph neural network (HIF-GNN) framework. This framework leverages item-level and session-level graph convolutional networks to capture users’ multilevel behavioral patterns. We also propose a hierarchical attention fusion module, which dynamically adjusts the importance of different information levels, and a diffusion sampling module that utilizes entropy-based sampling to adaptively capture users’ evolving interests and session behaviors. Extensive experiments on three real-world datasets show that HIF-GNN outperforms state-of-the-art methods, demonstrating its superior performance and suitability for dynamic streaming recommendation tasks.

As the integration of artificial intelligence technology with the medical field deepens, medication recommendation, as an important subfield, demonstrates immense potential value. Medication recommendation combines patient medical history with biomedical knowledge to assist doctors in determining medication combinations more accurately and safely. Existing works based on molecular knowledge neglect the 3D geometric structure of molecules and fail to learn the high-dimensional information of medications, leading to structural confusion. Additionally, it does not extract key substructures from a single patient visit, resulting in the failure to identify medication molecules suitable for the current patient visit. To address the above limitations, we propose a bimodal molecular recommendation framework named BiMoRec, which introduces 3D molecular structures to obtain atomic 3D coordinates and edge indices, overcoming the inherent lack of high-dimensional molecular information in 2D molecular structures. To retain the fast training and prediction efficiency of the recommendation system, we use bimodal graph contrastive pretraining to maximize the mutual information between the two molecular modalities, achieving the fusion of 2D and 3D molecular graphs. In addition, we propose a multi-step molecular enhancement mechanism to re-weight the molecules and optimize the utilization of the refined molecular embedding space. Our implementation on the MIMIC-III and MIMIC-IV datasets demonstrates that our method achieves state-of-the-art performance. Compared to the second-best baseline, our model improves accuracy by 0.61%, while maintaining the same level of DDI and model efficiency. Our source code is publicly available at: https://github.com/guangyunms/BiMoRec .

Medication recommendation systems are developed to recommend suitable medications tailored to specific patient. Previous researches primarily focus on learning medication representations, which have yielded notable advances. However, these methods are limited to capturing personalized patient representations due to the following primary limitations: (i) unable to capture the differences in the impact of diseases/procedures on patients across various patient health states; (ii) fail to model the direct causal relationships between medications and specific health state of patients, resulting in an inability to determine which specific disease each medication is treating. To address these limitations, we propose CausalMed, a patient health state-centric model capable of enhancing the personalization of patient representations. Specifically, CausalMed first captures the causal relationship between diseases/procedures and medications through causal discovery and evaluates their causal effects. Building upon this, CausalMed focuses on analyzing the health state of patients, capturing the dynamic differences of diseases/procedures in different health states of patients, and transforming diseases/procedures into medications on direct causal relationships. Ultimately, CausalMed integrates information from longitudinal visits to recommend medication combinations. Extensive experiments on real-world datasets show that our method learns more personalized patient representation and outperforms state-of-the-art models in accuracy and safety.

Recently, causal feature selection (CFS) has attracted considerable attention due to its outstanding interpretability and predictability performance. Such a method primarily includes the Markov blanket (MB) discovery and feature selection based on Granger causality. Representatively, the max-min MB (MMMB) can mine an optimal feature subset, i.e., MB; however, it is unsuitable for streaming features. Online streaming feature selection (OSFS) via online process streaming features can determine parents and children (PC), a subset of MB; however, it cannot mine the MB of the target attribute ( T ), i.e., a given feature, thus resulting in insufficient prediction accuracy. The Granger selection method (GSM) establishes a causal matrix of all features by performing excessively time; however, it cannot achieve a high prediction accuracy and only forecasts fixed multivariate time series data. To address these issues, we proposed an online CFS for streaming features (OCFSSFs) that mine MB containing PC and spouse and adopt the interleaving PC and spouse learning method. Furthermore, it distinguishes between PC and spouse in real time and can identify children with parents online when identifying spouses. We experimentally evaluated the proposed algorithm on synthetic datasets using precision, recall, and distance. In addition, the algorithm was tested on real-world and time series datasets using classification precision, the number of selected features, and running time. The results validated the effectiveness of the proposed algorithm.

Most existing studies on next point-of-interest (POI) recommendation assume that users deliver certain check-ins over individual POIs. In reality, we typically obtain uncertain check-ins due to the presence of collective POIs, which are gathering places of multiple individual POIs (e.g., shopping malls). On one hand, such uncertain check-ins over collective POIs hinder more accurate next POI recommendation for users due to the transition vanishing issue; on the other hand, the presence of collective POIs poses the challenge for businesses to select which collective POIs to locate in due to complicated competition and cooperation relations between businesses. As such, these collective POIs bring an unprecedented opportunity and necessity on recommendation for both users and businesses. Therefore, we propose novel solutions of location service beneficial for users-businesses. For users, we propose the STSP equipped with category- and location-aware encoders, to deliver more accurate next POI prediction by fusing rich context features. Regarding businesses, we explore their competition and cooperation relations from check-in records, based on which we derive theliving environment(LE) of a business. Insight on site selection for businesses is provided by exploiting the LE, aiming to bring in more profits. Extensive empirical studies demonstrate the efficiency of our solutions.

Skeleton-based action recognition has achieved promising performance recently, but there are still many challenges, e.g., the structural relation between joints and the different attention of frames, due to the complex spatial-temporal evolution of skeletal joints. In this paper, we propose a spatial reasoning and context-aware attention network for skeleton-based action recognition, which consists of a spatial reasoning module and a context-aware attention module. The spatial reasoning module can exploit the structural relation between joints to obtain the spatial features within each skeleton frame, followed by the context-aware attention module learning the different attention of frames. We perform experiments on two datasets and verify the effectiveness of each module in the proposed network. The comparison results demonstrate that our method achieves state-of-the-art performance.