Unsupervised node representation learning aims to obtain meaningful node embeddings without relying on node labels. To achieve this, graph convolution, which aggregates information from neighboring nodes, is commonly employed to encode node features and graph topology. However, excessive reliance on graph convolution can be suboptimal—especially in non-homophilic graphs—since it may yield unduly similar embeddings for nodes that differ in their features or topological properties. As a result, adjusting the degree of graph convolution usage has been actively explored in supervised learning settings, whereas such approaches remain underexplored in unsupervised scenarios. To tackle this, we propose FUEL, which adaptively learns the adequate degree of graph convolution usage by aiming to enhance intra-class similarity and inter-class separability in the embedding space. Since classes are unknown, FUEL leverages node features to identify node clusters and treats these clusters as proxies for classes. Through extensive experiments using 15 baseline methods and 14 benchmark datasets, we demonstrate the effectiveness of FUEL in downstream tasks, achieving state-of-the-art performance across graphs with diverse levels of homophily.

Recently, large language models (LLMs) have been widely used as recommender systems, owing to their reasoning capability and effectiveness in handling cold-start items. A common approach prompts an LLM with a target user's purchase history to recommend items from a candidate set, often enhanced with retrieval-augmented generation (RAG). Most existing RAG approaches retrieve purchase histories of users similar to the target user; however, these histories often contain noisy or weakly relevant information and provide little or no useful information for candidate items. To address these limitations, we propose ItemRAG, a novel RAG approach that shifts focus from coarse user-history retrieval to fine-grained item-level retrieval. ItemRAG augments the description of each item in the target user's history or the candidate set by retrieving items relevant to each. To retrieve items not merely semantically similar but informative for recommendation, ItemRAG leverages co-purchase information alongside semantic information. Especially, through their careful combination, ItemRAG prioritizes more informative retrievals and also benefits cold-start items. Through extensive experiments, we demonstrate that ItemRAG consistently outperforms existing RAG approaches under both standard and cold-start item recommendation settings.

Recipe recommendation has become an essential task in web-based food platforms. A central challenge is effectively leveraging rich multimodal features beyond user–recipe interactions. Our analysis shows that even simple uses of multimodal signals yield competitive performance, suggesting that systematic enhancement of these signals is highly promising. We propose TESMR, a 3-stage framework for recipe recommendation that progressively refines raw multimodal features into effective embeddings through: (1) content-based enhancement using foundation models with multimodal comprehension, (2) relation-based enhancement via message propagation over user–recipe interactions, and (3) learning-based enhancement through contrastive learning with learnable embeddings. Experiments on two real-world datasets show that TESMR outperforms existing methods, achieving 7–15% higher Recall@10.

In e-commerce, where users face a vast array of possible item choices, recommender systems are vital for helping them discover suitable items they might otherwise overlook. While many recommender systems primarily rely on a user's purchase history, recent multi-behavior recommender systems incorporate various auxiliary user behaviors, such as item clicks and cart additions, to enhance recommendations. Despite their overall performance gains, their effectiveness varies considerably between visited items (i.e., those a user has interacted with through auxiliary behaviors) and unvisited items (i.e., those with which the user has had no such interactions). Specifically, our analysis reveals that (1) existing multi-behavior recommender systems exhibit a significant gap in recommendation quality between the two item types (visited and unvisited items) and (2) achieving strong performance on both types with a single model architecture remains hallenging. To tackle these issues, we propose a novel multi-behavior recommender system, MEMBER. It employs a mixture-of-experts framework, with experts designed to recommend the two item types, respectively. Each expert is trained using a self-supervised method specialized for its design goal. In our comprehensive experiments, we show the effectiveness of MEMBER across both item types, achieving up to 65.46% performance gain over the best competitor in terms of Hit Ratio@20.

The evolution of Large Language Models (LLMs) has sparked growing concerns about jailbreak, crafted prompts that bypass safety guardrails and lead to the generation of harmful information. While recent research primarily focuses on identifying harmful content within LLM outputs, limited attention has been paid to detecting jailbreak intent directly from user prompts. This paper proposes a novel framework, called Prompt-Crossing, for analyzing user prompts and identifying those harboring potential jailbreak intent based on the response of LLM. Moreover, while existing methods determine jailbreak intent by looking at whether the response of LLM contains harmful information, our proposed framework can determine whether the user prompt, triggering the response of LLM, has jailbreak intent even if the response does not contain harmful information. In addition, our method performs filtering based on the likelihood calculated in a forward-step manner, deviates from traditional inference-based LLM usage, demonstrating a new application for LLMs. To demonstrate our proposed framework, we achieve state-of-the-art on the AdvBench benchmark using two LLMs, LLAMA and Falcon.

Recent advances have demonstrated the effectiveness of graph-based machine learning on relational databases (RDBs) for predictive tasks. Such approaches require transforming RDBs into graphs, a process we refer to as RDB-to-graph modeling, where rows of tables are represented as nodes and foreign-key relationships as edges.Yet, effective modeling of RDBs into graphs remains challenging.Specifically, there exist numerous ways to model RDBs into graphs, and performance on predictive tasks varies significantly depending on the chosen graph model of RDBs.In our analysis, we find that the best-performing graph model can yield up to a 10\% higher performance compared to the common heuristic rule for graph modeling, which remains non-trivial to identify.To foster research on intelligent RDB-to-graph modeling, we introduce RDB2G-Bench, the first benchmark framework for evaluating such methods.We construct extensive datasets covering 5 real-world RDBs and 12 predictive tasks, resulting in around 50k graph model–performance pairs for efficient and reproducible evaluations.Thanks to our precomputed datasets, we were able to benchmark 10 automatic RDB-to-graph modeling methods on the $12$ tasks about 380$\times$ faster than on-the-fly evaluation, which requires repeated GNN training.Our analysis of the datasets and benchmark results reveals key structural patterns affecting graph model effectiveness, along with practical implications for effective graph modeling.Our datasets and code are available at https://github.com/chlehdwon/RDB2G-Bench.

Bundle recommender systems aim to recommend suitable collections (i.e., bundles) of items to each user, meeting their diverse needs with all-in-one convenience. Typically, they utilize three distinct types of information: user-bundle purchase interactions (U-B view), user-item purchase interactions (U-I view), and bundle-item affiliations (B-I view). Our focus is on better integrating these three perspectives (i.e., views) to deliver more accurate bundle recommendations. Our examination of different role (main or sub-views) combinations of the views reveals two key observations: (1) the best combination varies across target users (i.e., who receive recommendations), and (2) the U-I view is relatively weak as the main role. Driven by these observations, we propose PET, which synergizes the three views through (1) personalized view weighting, (2) U-I view enhancement, and (3) two-pronged contrastive learning. Our extensive experiments demonstrate that PET significantly outperforms existing methods in all popular benchmark datasets. Our code and datasets are available at https://github.com/K-Kyungho/PET.

Item popularity in real-world data follows a long-tail distribution, where a few items attract most of the attention, while the majority receive much less. This disparity results in high-quality embeddings for popular (head) items, but lower-quality embeddings for unpopular (tail) items, leading to less accurate recommendations for the latter. Our observations confirm that embeddings of tail items often exhibit (1) magnitudes (i.e., norms) that are less reflective of actual popularity and (2) directions that are less effective in capturing user preferences, compared to those of head items.

In Spoken Question Answering (SQA), automatic speech recognition (ASR) outputs are often relayed to language models for QA. However, constructing such a cascaded framework with large language models (LLMs) in a real-time SQA setting involves realistic challenges, such as noise in the ASR output, the limited context length of LLMs, and latency in processing large models. This paper proposes a novel model-agnostic framework, RT-VQ2A2, to address these challenges. RT-VQ2A2 consists of three steps: codebook preparation, quantized semantic vector extractor, and dual segment selector. We construct a codebook from clustering, removing outliers on a text corpus derived from ASR to mitigate the influence of ASR error. Extracting quantized semantic vectors through a pre-built codebook shows significant speed and performance improvements in relevant context retrieval. Dual segment selector considers both semantic and lexical aspects to deal with ASR error. The efficacy of RT-VQ2A2 is validated on the widely used Spoken-SQuAD dataset.

Existing multi-hop question generation (QG) methods treat answer-irrelevant documents as non-essential and remove them as impurities. However, this approach can create a training-inference discrepancy when impurities cannot be completely removed, which can lead to a decrease in model performance. To overcome this problem, we propose an auxiliary task, called order-agnostic, which leverages non-essential data in the training phase to create a robust model and extract the consistent embeddings in real-world inference environments. Additionally, we use a single LM to perform both ranker and generator through a prompt-based approach without applying additional external modules. Furthermore, we discover that appropriate utilization of the non-essential components can achieve a significant performance increase. Finally, experiments conducted on HotpotQA dataset achieve state-of-the-art.

Near-data processing (NDP), a solution to reduce data movement overhead between host and memory, should not interfere with host access to ensure system fairness. We propose a cost-effective and energy-efficient LRDIMM-based NDP architecture (HAIL-DIMM) that can seamlessly interleave NDP and host access and is a drop-in replacement for existing main memory modules. The proposed NDP exploits the interleaving capability of the memory controller to interleave NDP and host access naturally. To take advantage of bank interleaving, an atomic operation of the proposed NDP, which consists of data movement and computation, is recognized by the memory controller as a DDR READ/WRITE but by the HAIL-DIMM as NDP based on the request's address. We implement a prototype of the proposed NDP architecture on an FPGA platform as proof of concept. The evaluation results show that the NDP system achieves up to 2.19x speedup in latency and up to 45.4% energy saving for data movement over the baseline system for memory-bound workloads.

We present HyperSum, an extractive summarization framework that captures both the efficiency of traditional lexical summarization and the accuracy of contemporary neural approaches. HyperSum exploits the pseudo-orthogonality that emerges when randomly initializing vectors at extremely high dimensions ("blessing of dimensionality") to construct representative and efficient sentence embeddings. Simply clustering the obtained embeddings and extracting their medoids yields competitive summaries. HyperSum often outperforms state-of-the-art summarizers – in terms of both summary accuracy and faithfulness – while being 10 to 100 times faster. We open-source HyperSum as a strong baseline for unsupervised extractive summarization1.

Graph autoencoders (Graph-AEs) learn representations of given graphs by aiming to accurately reconstruct them. A notable application of Graph-AEs is graph-level anomaly detection (GLAD), whose objective is to identify graphs with anomalous topological structures and/or node features compared to the majority of the graph population. Graph-AEs for GLAD regard a graph with a high mean reconstruction error (i.e. mean of errors from all node pairs and/or nodes) as anomalies. Namely, the methods rest on the assumption that they would better reconstruct graphs with similar characteristics to the majority. We, however, report non-trivial counter-examples, a phenomenon we call reconstruction flip, and highlight the limitations of the existing Graph-AE-based GLAD methods. Specifically, we empirically and theoretically investigate when this assumption holds and when it fails. Through our analyses, we further argue that, while the reconstruction errors for a given graph are effective features for GLAD, leveraging the multifaceted summaries of the reconstruction errors, beyond just mean, can further strengthen the features. Thus, we propose a novel and simple GLAD method, named MUSE. The key innovation of MUSE involves taking multifaceted summaries of reconstruction errors as graph features for GLAD. This surprisingly simple method obtains SOTA performance in GLAD, performing best overall among 14 methods across 10 datasets.

Graph Neural Networks (GNNs) have emerged as effective tools in recommender systems. Among various GNN models, LightGCN is distinguished by its simplicity and outstanding performance. Its efficiency has led to widespread adoption across different domains, including social, bundle, and multimedia recommendations. In this paper, we thoroughly examine the mechanisms of LightGCN, focusing on its strategies for scaling embeddings, aggregating neighbors, and pooling embeddings across layers. Our analysis reveals that, contrary to expectations based on its design, LightGCN suffers from inflexibility and inconsistency when applied to real-world data.

Visual Question Answering (VQA) stands to benefit from the boost of increasingly sophisticated Pretrained Language Model (PLM) and Computer Vision-based models. In particular, many language modality studies have been conducted using image captioning or question generation with the knowledge ground of PLM in terms of data augmentation. However, image generation of VQA has been implemented in a limited way to modify only certain parts of the original image in order to control the quality and uncertainty. In this paper, to address this gap, we propose a method that utilizes the diffusion model, pre-trained with various tasks and images, to inject the prior knowledge base into generated images and secure diversity without losing generality about the answer. In addition, we design an effective training strategy by considering the difficulty of questions to address the multiple images per QA pair and to compensate for the weakness of the diffusion model. VQA model trained on our strategy improves significant performance on the dataset that requires factual knowledge without any knowledge information in language modality.

Kyungho Kim, Kyungjae Lee, Seung-won Hwang, Young-In Song, Seungwook Lee. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. 2021.

In this letter, we present an implementation of deblurring hardware to mitigate blur incurred by optical aberrations in a real-time manner to increase resolution for mobile camera modules. As optical aberrations tend to be variant according to spatial location, algorithm should support spatially variant deblurring. Also, a deblurring problem is known to be rank-deficient, thus requires a regularization and a sparse norm such as Lpnorm where 0 ≤ p ≤ 1 has achieved satisfactory solutions. We focus on elaborating algorithm details and a corresponding architecture to efficiently implement with those two requirements. Also, we present a simulation result with the real sensor and this result shows that image resolution can be increased in terms of Modulated Transfer Function (MTF).

The signal regeneration method (SRE) and associated adaptive algorithms are proposed for fast fading channels. The adaptive algorithms such as the leaky least mean square (LLMS) algorithm assisted by the SRE method provide the signal-to-interference-noise-ratio (SINR) performance comparable to the linear minimum mean square error (LMMSE) equalizer, with low numerical complexity and low memory requirements. The proposed SRE-LLMS algorithm is analyzed based on the independence assumption, and shown to provide low excessive mean square error (EMSE) compared to the LMS algorithm without incurring a bias in its estimate. The simulation results for the wireless channel models such as VA30 and VA120 are presented to validate the analysis.

Silicon validation becomes difficult because of rapidly increasing complexity and operation speed of integrated circuits. When an error is found after a chip is fabricated, post-silicon repair is necessary. Full mask revision may significantly increase the cost and time-to-market. In this paper, we describe partial metal revision techniques in which only top-level metal layers are revised to fix ¿small¿ errors with minimal increase of the cost. When an error cannot be fixed by partial metal layer revision, full metal revision or full mask revision is necessary. However, frequently errors are small enough to be fixed by partial metal layer revision. Effective partitioning and pin-extension to top-level metal layers can significantly improve the repairability by using top-level metal revision.

In this paper, we report DRM receiver synchronization methods at the tracking stage. The DRM synchronization tracking processes mainly include symbol timing offset estimation, fine fractional carrier frequency offset estimation, and sampling clock frequency offset estimation. All these estimation methods exploit the characteristics of DRM signal structure and gain reference cells inserted into the signal. Their performance is assessed by simulation with DRM full link-level simulator over different shortwave channel models and verified with DRM hardware platform based on FPGA. Both results demonstrate the robustness of the proposed tracking methods.