Generative retrieval (GR) has revolutionized recommendation systems by integrating the content of the items into semantic identifiers. However, existing frameworks predominantly isolate modalities (e.g., relying solely on text), overlooking the inherently multimodal nature of real-world items. This work addresses the underexplored challenge of Multimodal Generative Retrieval (MGR). Through a systematic analysis of Early and Late Fusion strategies, we reveal that naive integration fails due to two critical limitations: modality sensitivity, where one modality dominates the representation, and modality correspondence, where the model fails to align distinct semantic IDs across modalities. To overcome these challenges, we introduce MGR-LF++, an enhanced late fusion framework. MGR-LF++ employs contrastive modality alignment to enforce cross-modal consistency and incorporates special tokens to preserve autoregressive integrity. Extensive experiments demonstrate that MGR-LF++ achieves performance improvements of over 20% compared to unimodal and naive multimodal alternatives.

Generative recommendation plays a crucial role in personalized systems, predicting users' future interactions from their historical behavior sequences. A critical yet underexplored factor in training these models is data augmentation, the process of constructing training data from user interaction histories. By shaping the training distribution, data augmentation directly and often substantially affects model generalization and performance. Nevertheless, in much of the existing work, this process is simplified, applied inconsistently, or treated as a minor design choice, without a systematic and principled understanding of its effects.

Recently, pretrained large language models (LLMs) have been widely adopted in recommendation systems to leverage their textual understanding and reasoning abilities to model user behaviors and suggest future items. A key challenge in this setting is that items on most platforms are not included in the LLM's training data. Therefore, existing methods often fine-tune LLMs by introducing auxiliary item tokens to capture item semantics. However, in real-world applications such as e-commerce and short video platforms, the item space evolves rapidly, which gives rise to a cold-start setting, where many newly introduced items receive little or even no user engagement. This poses challenges in both learning accurate item token embeddings and generalizing efficiently to accommodate the continual influx of new items. In this work, we propose a novel meta-item token learning strategy to address both these challenges simultaneously. Specifically, we introduce MI4Rec, an LLM-based approach for recommendation that uses just a few learnable meta-item tokens and an LLM encoder to dynamically aggregate meta-items based on item content. We show that this paradigm allows highly efficient and accurate learning in such challenging settings. Extensive experiments on Yelp and Amazon reviews datasets demonstrate the effectiveness of MI4Rec in both warm-start and cold-start recommendations. Notably, MI4Rec achieves an average performance improvement of 20.4% in Recall and NDCG compared to the best-performing baselines. The implementation of MI4Rec is available at https://github.com/zhengzaiyi/MI4Rec

Collaborative filtering (CF) enables large-scale recommendation systems by encoding information from user-item interactions into dense ID-embedding tables. However, as embedding tables grow, closed-form solutions become impractical, necessitating the use of mini-batch gradient descent. Despite extensive work on loss function design, we argue that one core component of these pipelines is heavily overlooked: weight decay. Attaining high-performing models typically requires careful tuning of weight decay, regardless of loss, yet its necessity is not well understood. In this work, we question why weight decay is crucial in CF pipelines and how it impacts training. Through theoretical and empirical analysis, we uncover that weight decay's primary function is to encode popularity information into the magnitudes of the embedding vectors. Moreover, we find that tuning weight decay acts as a coarse, non-linear, knob to influence preference towards popular or unpopular items. Based on these findings, we propose PRISM (Popularity-awaRe Initialization Strategy for embedding Magnitudes), a straightforward yet effective solution to simplify the training of CF models. PRISM pre-encodes the popularity information typically learned through weight decay, eliminating its necessity. Our experiments show that PRISM improves performance by up to 4.77% and reduces training times by 38.48%, compared to state-of-the-art training strategies. Additionally, PRISM offers a cost-effective and meaningful strategy to mitigate popularity bias.

The prevalence of unhealthy eating habits has become a growing concern in the United States. However, popular food recommendation platforms, such as Yelp, tend to prioritize users' dietary preferences over the healthiness of their choices. While some efforts have focused on developing health-aware food recommendation systems, personalization based on specific health conditions remains underexplored. Additionally, the lack of interpretability in these systems prevents users from evaluating the reliability of recommendations, limiting their practical adoption. To address these issues, we introduce two large-scale personalized health-aware food recommendation benchmarks at the first attempt. Building on this, we propose a novel framework called the Multi-Objective Personalized Interpretable Health-aware Food Recommendation System (MOPI-HFRS). This system generates food recommendations by jointly optimizing three objectives: user preference, personalized healthiness, and nutritional diversity. It also incorporates a reasoning module enhanced by large language models (LLMs) to provide interpretable recommendations that promote healthy dietary knowledge. The framework integrates descriptive features and health data using two structure learning and pooling modules within a graph learning framework. Pareto optimization is applied to balance the multi-faceted objectives. To further enhance healthy dietary knowledge, the system leverages LLMs by infusing knowledge from the recommendation model, generating meaningful interpretations for the recommendations. Extensive experiments on the proposed benchmarks demonstrate that MOPI-HFRS outperforms state-of-the-art methods by delivering diverse, healthy food recommendations alongside reliable explanations.

Graph knowledge has been proven effective in enhancing item rankings in recommender systems (RecSys), particularly during the retrieval stage.However, its application in the ranking stage, especially when richer contextual information in user-item interactions is available, remains underexplored.A major challenge lies in the substantial computational cost associated with repeatedly retrieving neighborhood information from billions of items stored in distributed systems.This resource-intensive requirement makes it difficult to scale graph-based methods in practical RecSys.To bridge this gap, we first demonstrate that incorporating graphs in the ranking stage improves ranking qualities.Notably, while the improvement is evident, we show that the substantial computational overheads entailed by graphs are prohibitively expensive for real-world recommendations.In light of this, we propose a non-parametric strategy that utilizes graph convolution for re-ranking only during test time.Our strategy circumvents the notorious computational overheads from graph convolution during training, and utilizes structural knowledge hidden in graphs on-the-fly during testing.It can be used as a plug-and-play module and easily employed to enhance the ranking ability of various ranking layers of a real-world RecSys with significantly reduced computational overhead.Through comprehensive experiments across four benchmark datasets with varying levels of sparsity, we demonstrate that our strategy yields noticeable improvements (i.e., 8.1% on average) during testing time with little to no additional computational overheads (i.e., 0.5% on average).

Collaborative Filtering (CF) methods dominate real-world recommender systems given their ability to learn high-quality, sparse ID-embedding tables that effectively capture user preferences. These tables scale linearly with the number of users and items, and are trained to ensure high similarity between embeddings of interacted user-item pairs, while maintaining low similarity for non-interacted pairs. Despite their high performance, encouraging dispersion for non-interacted pairs necessitates expensive regularization (e.g., negative sampling), hurting runtime and scalability. Existing research tends to address these challenges by simplifying the learning process, either by reducing model complexity or sampling data, trading performance for runtime. In this work, we move beyond model-level modifications and study the properties of the embedding tables under different learning strategies. Through theoretical analysis, we find that the singular values of the embedding tables are intrinsically linked to different CF loss functions. These findings are empirically validated on real-world datasets, demonstrating the practical benefits of higher stable rank -- a continuous version of matrix rank which encodes the distribution of singular values. Based on these insights, we propose an efficient warm-start strategy that regularizes the stable rank of the user and item embeddings. We show that stable rank regularization during early training phases can promote higher-quality embeddings, resulting in training speed improvements of up to 65.9%. Additionally, stable rank regularization can act as a proxy for negative sampling, allowing for performance gains of up to 21.2% over loss functions with small negative sampling ratios. Overall, our analysis unifies current CF methods under a new perspective -- their optimization of stable rank -- motivating a flexible regularization method that is easy to implement, yet effective at enhancing CF systems.

Deep recommender systems rely heavily on large embedding tables to handle high-cardinality categorical features such as user/item identifiers, and face significant memory constraints at scale.To tackle this challenge, hashing techniques are often employed to map multiple entities to the same embedding and thus reduce the size of the embedding tables.Concurrently, graph-based collaborative signals have emerged as powerful tools in recommender systems, yet their potential for optimizing embedding table reduction remains unexplored.This paper introduces GraphHash, the first graph-based approach that leverages modularity-based bipartite graph clustering on user-item interaction graphs to reduce embedding table sizes.We demonstrate that the modularity objective has a theoretical connection to message-passing, which provides a foundation for our method.By employing fast clustering algorithms, GraphHash serves as a computationally efficient proxy for message-passing during preprocessing and a plug-andplay graph-based alternative to traditional ID hashing.Extensive experiments show that GraphHash substantially outperforms diverse hashing baselines on both retrieval and click-through-rate prediction tasks.In particular, GraphHash achieves on average a 101.52% improvement in recall when reducing the embedding table size by more than 75%, highlighting the value of graph-based collaborative information for model reduction.

Graph structure learning (GSL), which aims to learn the adjacency matrix for graph neural networks (GNNs), has shown great potential in boosting the performance of GNNs. Most existing GSL works apply a joint learning framework where the estimated adjacency matrix and GNN parameters are optimized for downstream tasks. However, as GSL is essentially a link prediction task, whose goal may largely differ from the goal of the downstream task. The inconsistency of these two goals limits the GSL methods to learn the potential optimal graph structure. Moreover, the joint learning framework suffers from scalability issues in terms of time and space during the process of estimation and optimization of the adjacency matrix. To mitigate these issues, we propose a graph structure refinement (GSR) framework with a pretrain-finetune pipeline. Specifically, The pre-training phase aims to comprehensively estimate the underlying graph structure by a multi-view contrastive learning framework with both intra- and inter-view link prediction tasks. Then, the graph structure is refined by adding and removing edges according to the edge probabilities estimated by the pre-trained model. Finally, the fine-tuning GNN is initialized by the pre-trained model and optimized toward downstream tasks. With the refined graph structure remaining static in the fine-tuning space, GSR avoids estimating and optimizing graph structure in the fine-tuning phase which enjoys great scalability and efficiency. Moreover, the fine-tuning GNN is boosted by both migrating knowledge and refining graphs. Extensive experiments are conducted to evaluate the effectiveness (best performance on six benchmark datasets), efficiency, and scalability (13.8 times faster using 32.8% GPU memory compared to the best GSL baseline on Cora) of the proposed model.

Graph neural networks (GNNs) have been broadly studied on dynamic graphs for their representation learning, majority of which focus on graphs with homogeneous structures in the spatial domain. However, many real-world graphs - i.e., heterogeneous temporal graphs (HTGs) - evolve dynamically in the context of heterogeneous graph structures. The dynamics associated with heterogeneity have posed new challenges for HTG representation learning. To solve this problem, in this paper, we propose heterogeneous temporal graph neural network (HTGNN) to integrate both spatial and temporal dependencies while preserving the heterogeneity to learn node representations over HTGs. Specifically, in each layer of HTGNN, we propose a hierarchical aggregation mechanism, including intra-relation, inter-relation, and across-time aggregations, to jointly model heterogeneous spatial dependencies and temporal dimensions. To retain the heterogeneity, intra-relation aggregation is first performed over each slice of HTG to attentively aggregate information of neighbors with the same type of relation, and then intra-relation aggregation is exploited to gather information over different types of relations; to handle temporal dependencies, across-time aggregation is conducted to exchange information across different graph slices over the HTG. The proposed HTGNN is a holistic framework tailored heterogeneity with evolution in time and space for HTG representation learning. Extensive experiments are conducted on the HTGs built from different real-world datasets and promising results demonstrate the outstanding performance of HTGNN by comparison with state-of-the-art baselines. Our built HTGs and code have been made publicly accessible at: https://github.com/YesLab-Code/HTGNN.

The explosive growth and increasing sophistication of Android malware call for new defensive techniques to protect mobile users against novel threats. To address this challenge, in this paper, we propose and develop an intelligent system named Dr.Droid to jointly model malware propagation and evolution for their detection at the first attempt. In Dr.Droid, we first exploit higher-level semantic and social relations within the ecosystem (e.g., app-market, app-developer, market-developer relations etc.) to characterize app propagation patterns; and then we present a structured heterogeneous graph to model the complex relations among different types of entities. To capture malware evolution, we further consider the temporal dependence and introduce a heterogeneous temporal graph to jointly model malware propagation and evolution by considering heterogeneous spatial dependencies with temporal dimensions. Afterwards, we propose a novel heterogeneous temporal graph transformer framework (denoted as HTGT) to integrate both spatial and temporal dependencies while preserving the heterogeneity to learn node representations for malware detection. Specifically, in our proposed HTGT, to preserve the heterogeneity, we devise a heterogeneous spatial transformer to derive heterogeneous attentions over each node and edge to learn dedicated representations for different types of entities and relations; to model temporal dependencies, we design a temporal transformer into the HTGT to attentively aggregate its historical sequences of a given node (e.g., app); the two transformers work in an iterative manner for representation learning. Promising experimental results based on the large-scale sample collections from anti-malware industry demonstrate the performance of Dr.Droid, by comparison with state-of-the-art baselines and popular mobile security products.

During the pandemic caused by coronavirus disease (COVID-19), social media has played an important role by enabling people to discuss their experiences and feelings of this global crisis. To help combat the prolonged pandemic that has exposed vulnerabilities impacting community resilience, in this paper, based on our established large-scale COVID-19 related social media data, we propose and develop an integrated framework (named Dr.Emotion) to learn disentangled representations of social media posts (i.e., tweets) for emotion analysis and thus to gain deep insights into public perceptions towards COVID-19. In Dr.Emotion, for given social media posts, we first post-train a transformer-based model to obtain the initial post embeddings. Since users may implicitly express their emotions in social media posts which could be highly entangled with other descriptive information in the post content, to address this challenge for emotion analysis, we propose an adversarial disentangler by integrating emotion-independent (i.e., sentiment-neutral) priors of the posts generated by another post-trained transformer-based model to separate and disentangle the implicitly encoded emotions from the content in latent space for emotion classification at the first attempt. Extensive experimental studies are conducted to fully evaluate Dr.Emotion and promising results demonstrate its performance in emotion analysis by comparison with the state-of-the-art baseline methods. By exploiting our developed Dr.Emotion, we further perform emotion analysis over a large number of social media posts and provide in-depth investigation from both temporal and geographical perspectives, based on which additional work can be conducted to extract and transform the constructive ideas, experiences and support into actionable information to improve community resilience in responses to a variety of crises created by COVID-19 and well beyond.

The fast evolving and deadly outbreak of coronavirus disease (COVID-19) has posed grand challenges to human society. To slow the spread of virus infections and better respond with actionable strategies for community mitigation, leveraging the large-scale and real-time pandemic related data generated from heterogeneous sources (e.g., disease related data, demographic data, mobility data, and social media data), in this work, we propose and develop a data-driven system (named α-satellite), as an initial offering, to provide real-time COVID-19 risk assessment in a hierarchical manner in the United States. More specifically, given a location (either user input or automatic positioning), the system will automatically provide risk indices associated with the specific location, the county that location is in and the state as a whole to enable people to select appropriate actions for protection while minimizing disruptions to daily life to the extent possible. In α-satellite, we first construct an attributed heterogeneous information network (AHIN) to model the collected multi-source data in a comprehensive way; and then we utilize meta-path based schemes to model both vertical and horizontal information associated with a given location (i.e., point of interest, POI); finally we devise a novel heterogeneous graph neural network to aggregate its neighborhood information to estimate the risk of the given POI in a hierarchical manner. To comprehensively evaluate the performance of α-satellite in real-time COVID-19 risk assessment, a set of studies are first performed to validate its utility; based on a real-world dataset consisting of 6,538 annotated POIs, the experimental results show that α-satellite achieves the area of under curve (AUC) of 0.9378, which outperforms the state-of-the-art baselines. After we launched the system for public tests, it had attracted 51,190 users as of May 30. Based on the analysis of its large-scale users, we have a key finding that people from more severe regions (i.e., with larger numbers of COVID-19 cases) have stronger interests using the system for actionable information. Our system and generated benchmark datasets have been made publicly accessible through our website.

The goal of the IEEE 2018 FEMH Voice Data Challenge was to develop an effective algorithmic approach to classifying voice samples as normal or pathological, and further subdivide the pathological samples into three types. We adopted a multi-representation ensemble approach to the task. We designed a pipeline with three classification stages, where each stage used a combination of supervised, semi-supervised and multiple-instance learners. This approach was able to achieve a sensitivity of 89% and specificity of 76% in classifying normal from pathological samples and an unweighted average recall (UAR) of 60.67% in subclassifying pathological samples into three types.