Large language models (LLMs) achieve optimal utility when their responses are grounded in external knowledge sources. However, real-world documents, such as annual reports, scientific papers, and clinical guidelines, frequently combine extensive narrative content with complex, hierarchically structured tables. While existing retrieval-augmented generation (RAG) systems effectively integrate LLMs' generative capabilities with external retrieval-based information, their performance significantly deteriorates especially processing such heterogeneous text-table hierarchies. To address this limitation, we formalize the task of Heterogeneous Document RAG, which requires joint retrieval and reasoning across textual and hierarchical tabular data. We propose MixRAG, a novel three-stage framework: (i) hierarchy row-and-column-level (H-RCL) representation that preserves hierarchical structure and heterogeneous relationship, (ii) an ensemble retriever with LLM-based reranking for evidence alignment, and (iii) multi-step reasoning decomposition via a RECAP prompt strategy. To bridge the gap in available data for this domain, we release the dataset DocRAGLib, a 2k-document corpus paired with automatically aligned text-table summaries and gold document annotations. The comprehensive experiment results demonstrate that MixRAG boosts top-1 retrieval by 46% over strong text-only, table-only, and naive-mixture baselines, establishing new state-of-the-art performance for mixed-modality document grounding.

Generative retrieval seeks to replace traditional search index data structures with a single large-scale neural network, offering the potential for improved efficiency and seamless integration with generative large language models. As an end-to-end paradigm, generative retrieval adopts a learned differentiable search index to conduct retrieval by directly generating document identifiers through corpus-specific constrained decoding. The generalization capabilities of generative retrieval on out-of-distribution corpora have gathered significant attention. Recent advances primarily focus on the problems arising from training strategies, and addressing them through various learning techniques. However, the fundamental challenges of generalization arising from constrained auto-regressive decoding still remain unexplored and systematically understudied. In this paper, we examine the inherent limitations of constrained auto-regressive generation from two essential perspectives: constraints and beam search. We begin with the Bayes-optimal setting where the generative retrieval model exactly captures the underlying relevance distribution of all possible documents. Then we apply the model to specific corpora by simply adding corpus-specific constraints. Our main findings are two-fold: (i) For the effect of constraints, we derive a lower bound of the error, in terms of the KL divergence between the ground-truth and the model-predicted step-wise marginal distributions. This error arises due to the unawareness of future constraints during generation and is shown to depend on the average Simpson diversity index of the relevance distribution. (ii) For the beam search algorithm used during generation, we reveal that the usage of marginal distributions may not be an ideal approach. Specifically, we prove that for sparse relevance distributions, beam search can achieve perfect top-1 precision but suffer from poor top-k recall performance. To support our theoretical findings, we conduct experiments on synthetic and real-world datasets, validating the existence of the error from adding constraints and the recall performance drop due to beam search. This paper aims to improve our theoretical understanding of the generalization capabilities of the auto-regressive decoding retrieval paradigm, laying a foundation for its limitations and inspiring future advancements toward more robust and generalizable generative retrieval.

Computer Systems, Imagery and Media

Recently, Electronic Health Records (EHR) have become valuable for enhancing medical decision making, as well as online disease detection and monitoring. Meanwhile, deep learning-based methods have achieved great success in health risk prediction and diagnosis prediction based on EHR. Nevertheless, deep learning-based models usually require high volumes of data due to the vast amount of parameters. In addition, a considerable proportion of medical codes appear rarely in the EHR data which poses huge difficulties for clinical applications. Hence, some works propose to adopt medical ontologies to enhance the prediction performance and provide interpretable prediction results. However, these medical ontologies are often small-scale and coarse-grained, most of diagnoses and medical concepts are not included, lacking many diagnoses and medical concepts, let alone various relationships between these concepts. To overcome this limitation, we propose to incorporate existing large-scale medical knowledge graphs (KGs) into diagnosis prediction and devise a Stage-awareHierarchicalAttentiveRelational Network, namedHAR. Specifically, for each visit, a personalized sub-KG is extracted from the existing medical KG, on which HAR conducts relation-specific message passing and hierarchical message aggregation to refine representations of nodes that correspond to medical codes in visits. HAR takes the specific stage of a patient's disease progression into consideration, which participates in the computation of relation-level and node-level attention. Extensive experiments on two public datasets demonstrate the effectiveness of HAR in improving both the visit-level precision and code-level accuracy of the diagnosis prediction task.

Temporal Knowledge graph (TKG) reasoning aims to predict missing facts based on historical TKG data. Most of the existing methods are incapable of explicitly modeling the long-term time dependencies from history and neglect the adaptive integration of the long- and short-term information. To tackle these problems, we propose a novel method that utilizes a designed Hierarchical Relational Graph Neural Network to learn the Long- and Short-term representations for TKG reasoning, namely HGLS. Specifically, to explicitly associate entities in different timestamps, we first transform the TKG into a global graph. Based on the built graph, we design a Hierarchical Relational Graph Neural Network that executes in two levels: The sub-graph level is to capture the semantic dependencies within concurrent facts of each KG. And the global-graph level aims to model the temporal dependencies between entities. Furthermore, we design a module to extract the long- and short-term information from the output of these two levels. Finally, the long- and short-term representations are fused into a unified one by Gating Integration for entity prediction. Extensive experiments on four datasets demonstrate the effectiveness of HGLS.

Modeling user preference from his historical sequences is one of the core problems of sequential recommendation. Existing methods in this field are widely distributed from conventional methods to deep learning methods. However, most of them only model users' interests within their own sequences and ignore the dynamic collaborative signals among different user sequences, making it insufficient to explore users' preferences. We take inspiration from dynamic graph neural networks to cope with this challenge, modeling the user sequence and dynamic collaborative signals into one framework. We propose a new method named Dynamic Graph Neural Network for Sequential Recommendation (DGSR), which connects different user sequences through a dynamic graph structure, exploring the interactive behavior of users and items with time and order information. Furthermore, we design a Dynamic Graph Recommendation Network to extract user's preferences from the dynamic graph. Consequently, the next-item prediction task in sequential recommendation is converted into a link prediction between the user node and the item node in a dynamic graph. Extensive experiments on four public benchmarks show that DGSR outperforms several state-of-the-art methods. Further studies demonstrate the rationality and effectiveness of modeling user sequences through a dynamic graph.

Multimedia contents are of predominance in the modern Web era. Recent years have witnessed growing research interests in multimedia recommendation, which aims to predict whether a user will interact with an item with multimodal contents. Most previous studies focus on modeling user-item interactions with multimodal features included as side information. However, this scheme is not well-designed for multimedia recommendation. First, onlycollaborativeitem-item relationships are implicitly modeled through high-order item-user-item co-occurrences. Considering that items are associated with rich contents in multiple modalities, we argue that the latentsemanticitem-item structures underlying these multimodal contents could be beneficial for learning better item representations and assist the recommender models to comprehensively discover candidate items. Second, although previous studies consider multiple modalities, their ways of fusing multiple modalities by linear combination or concatenation is insufficient to fully capture content information of items and item relationships. To address these deficiencies, we propose a latent structureMIning withContRastive mOdality fusion model, which we term MICRO for brevity. To be specific, we devise a novel modality-aware structure learning module, which learns item-item relationships for each modality. Based on the learned modality-aware latent item relationships, we perform graph convolutions to explicitly inject item affinities into modality-aware item representations. Additionally, we design a novel multimodal contrastive framework to facilitate item-level multimodal fusion by mining both modality-shared and modality-specific information. Finally, the item representations are plugged into existing collaborative filtering methods to make accurate recommendation. Extensive experiments on three real-world datasets demonstrate the superiority of our method over state-of-arts and rationalize the design choice of our work.

Multiview network embedding aims at projecting nodes in the network to low-dimensional vectors, while preserving their multiple relations and attribute information. Contrastive learning approaches have shown promising performance in this task. However, they neglect the semantic consistency between fused and view representations and have difficulty in modeling complementary information between different views. To deal with these deficiencies, this work presents a novel Contrastive leaRning framEwork for Multiview network Embedding (CREME). In our work, different views can be obtained based on the various relations among nodes. Then, we generate view embeddings via proper view encoders and utilize an attentive multiview aggregator to fuse these representations. Particularly, we design two collaborative contrastive objectives, view fusion InfoMax and inter-view InfoMin, to train the model in a self-supervised manner. The former objective distills information from embeddings generated from different views, while the latter captures complementary information among views to promote distinctive view embeddings. We also show that the two objectives can be unified into one objective for model training. Extensive experiments on three real-world datasets demonstrate that our proposed CREME is able to consistently outperform state-of-the-art methods.

The problem of session-aware recommendation aims to predict users’ next click based on their current session and historical sessions. Existing session-aware recommendation methods have defects in capturing complex item transition relationships. Other than that, most of them fail to explicitly distinguish the effects of different historical sessions on the current session. To this end, we propose a novel method, named Personalized Graph Neural Networks with Attention Mechanism (A-PGNN) for brevity. A-PGNN mainly consists of two components: one is Personalized Graph Neural Network (PGNN), which is used to extract the personalized structural information in each user behavior graph, compared with the traditional Graph Neural Network (GNN) model, which considers the role of the user when the node embedding is updated. The other is Dot-Product Attention mechanism, which draws on the Transformer net to explicitly model the effect of historical sessions on the current session. Extensive experiments conducted on two real-world data sets show that A-PGNN evidently outperforms the state-of-the-art personalized session-aware recommendation methods.

Dynamic recommendation is essential for modern recommender systems to provide real-time predictions based on sequential data. In real-world scenarios, the popularity of items and interests of users change over time. Based on this assumption, many previous works focus on interaction sequences and learn evolutionary embeddings of users and items. However, we argue that sequence-based models are not able to capture collaborative information among users and items directly. Here we propose Dynamic Graph Collaborative Filtering (DGCF), a novel framework leveraging dynamic graphs to capture collaborative and sequential relations of both items and users at the same time. We propose three update mechanisms: zero-order `inheritance', first-order `propagation', and second-order `aggregation', to represent the impact on a user or item when a new interaction occurs. Based on them, we update related user and item embeddings simultaneously when interactions occur in turn, and then use the latest embeddings to make recommendations. Extensive experiments conducted on three public datasets show that DGCF significantly outperforms the state-of-the-art dynamic recommendation methods up to 30%. Our approach achieves higher performance when the dataset contains less action repetition, indicating the effectiveness of integrating dynamic collaborative information.