Long-term conversational agents rely on personal memory to maintain coherence and personalization, yet practical systems must operate under context budgets and cope with evolving or contradictory user information. We frame persona memory as a retrieval problem over a growing memory store, and propose REMAP, a reflection-guided memory editing approach for online alignment of persona facts that selectively writes and revises memory entries based on the current dialogue evidence and retrieved related items. The method aims to preserve salient facts while reducing redundancy and resolving apparent conflicts, enabling more efficient context utilization over extended interaction horizons. Experiments on multi-session dialogue datasets show consistent gains in persona-consistent retrieval and response continuity over commonly used memory strategies, while achieving more selective memory updates under comparable operational overhead.

This paper explores advancements in automated Question-Answer (QA) extraction using large language models (LLMs), addressing challenges in transforming unstructured text into high-quality, retrievable QA pairs. Traditional approaches, whether through segmented question and answer generation or end-to-end extraction, often struggle with efficiency, dataset limitations, and performance consistency. Leveraging recent progress in LLMs, we constructed a large-scale Chinese QA extraction dataset with 143,846 documents and evaluated multiple fine-tuned models on public and private datasets. Surprisingly, code-based English LLMs outperformed Chinese-specialized models on Chinese text with a lower hallucination rate. Building upon this finding, we enhanced the best-performing code-based model with an expanded Chinese vocabulary, creating Code Llama-M, which achieved better results. Integrating Code Llama-M into our internal assistant, Luo Ying, demonstrated notable user satisfaction gains, affirming its practical impact. Key contributions include: (i) creation of a robust Chinese QA extraction instruction dataset; (ii) evidence of cross-lingual efficacy of code-based LLMs for Chinese QA tasks, further enhanced through Code Llama-M's expanded Chinese vocabulary; and (iii) successful application of the fine-tuned LLM in a live assistant system, enhancing user experience.

Generative latent diffusion models (LDMs) have been extensively applied in various fields yet underperform in time-series prediction. Therefore, We propose the Re-Diffusion model, a latent diffusion approach that generates backbone residuals specifically tailored for time-series forecasting. The model comprises a variational autoencoder that compresses the residuals between the actual future values and the predictions from the backbone into latent space. It also includes a conditional diffusion generator to forecast the potential distribution of these residuals. Our findings reveal that this latent-space methodology particularly enhances existing backbone predictors, by effectively reducing prediction bias through an advanced estimation of complex error distributions. While previous diffusion-based models tend to struggle with long-term forecasting, Re-Diffusion integrates the strengths of diffusion methods, leading to improvements in long-term predictions. Our experimental results indicate that the Re-Diffusion model achieves a 10% promotion over state-of-art predictors, marking a significant advancement in the field of time-series forecasting.

Graph Neural Networks (GNNs) have shown remarkable performance in modeling graph data. However, Labeling graph data typically relies on unreliable information, leading to noisy node labels. Existing approaches for GNNs under Label Noise (GLN) employ supervision signals beyond noisy labels for robust learning. While empirically effective, they tend to over-reliance on supervision signals built upon external assumptions, leading to restricted applicability. In this work, we shift the focus to exploring how to extract useful information and learn from the graph itself, thus achieving robust graph learning. From an information theory perspective, we theoretically and empirically demonstrate that the graph itself contains reliable information for graph learning under label noise. Based on these insights, we propose the Topological Feature Reconstruction (TFR) method. Specifically, TFR leverages the fact that the pattern of clean labels can more accurately reconstruct graph features through topology, while noisy labels cannot. TFR is a simple and theoretically guaranteed model for robust graph learning under label noise. We conduct extensive experiments across datasets with varying properties. The results demonstrate the robustness and broad applicability of our proposed TFR compared to state-of-the-art baselines. Codes are available at https://github.com/eaglelab-zju/TFR.

Detecting out-of-distribution (OOD) samples poses a significant safety challenge when deploying models in open-world scenarios. Advanced works assume that OOD and in-distributional (ID) samples exhibit a distribution discrepancy, showing an encouraging direction in estimating the uncertainty with embedding features or predicting outputs. In this work, we propose a data structure-aware approach to mitigate the sensitivity of distances to the “curse of dimensionality”, where high-dimensional features are mapped to the manifold of ID samples, leveraging the well-known manifold assumption. Specifically, we present a novel distance termed as tangent distance, which tackles the issue of generalizing the meaningfulness of distances on testing samples to detect OOD inputs. Extensive experiments show that the tangent distance performs competitively with other post hoc OOD detection baselines on common and large-scale benchmarks.

Recent advancements in test-time adaptation (TTA) offer promising solutions for mitigating performance degradation caused by distribution shifts. However, they may fall short in practical deployment of GNNs during test time, due to the significant reliance on impractical retraining or modifications to online GNN models. To address such challenges, in this work, we propose a novel method, i.e., Test-Time Graph REBirth, dubbed TT-GREB, to effectively generalize the well-trained GNN models to the test-time graphs under distribution shifts by directly manipulating the test graph data. Concretely, we develop an overall framework designed by two principles for obtaining newly reborn test graphs, corresponding to two sub-modules: (1) prototype extractor for re-extracting the environment-invariant features of the test-time graph; and (2) environment refiner for re-fining the environment-varying features to explore the potential shifts. Furthermore, we propose a dual test-time graph contrastive learning objective with an effective iterative optimization strategy to obtain optimal prototype components and environmental components of the test graph. Extensive experiments on real-world graphs under diverse test-time distribution shifts verify the effectiveness of our proposed method, showcasing its superior ability to manipulate test-time graphs for better GNN generalization ability.

Time series anomaly detection aims to identify unusual patterns in data or deviations from systems' expected behavior. The reconstruction-based methods are the mainstream in this task, which learn point-wise representation via unsupervised learning. However, the unlabeled anomaly points in training data may cause these reconstruction-based methods to learn and reconstruct anomalous data, resulting in the challenge of capturing normal patterns. In this paper, we propose a time series anomaly detection method based on implicit neural representation (INR) reconstruction, named TSINR, to address this challenge. Due to the property of spectral bias, TSINR enables prioritizing low-frequency signals and exhibiting poorer performance on high-frequency abnormal data. Specifically, we adopt INR to parameterize time series data as a continuous function and employ a transformer-based architecture to predict the INR of given data. As a result, the proposed TSINR method achieves the advantage of capturing the temporal continuity and thus is more sensitive to discontinuous anomaly data. In addition, we further design a novel form of INR continuous function to learn inter- and intra-channel information, and leverage a pre-trained large language model to amplify the intense fluctuations in anomalies. Extensive experiments demonstrate that TSINR achieves superior overall performance on both univariate and multivariate time series anomaly detection benchmarks compared to other state-of-the-art reconstruction-based methods. Our codes are available here.

Dual-energy X-ray absorptiometry (DXA) enables accurate bone mineral density but requires specialized equipment and protocols. X-ray-based BMD screening offers opportunistic early detection, though prior methods struggle with X-ray intensity variations and demand large datasets. We introduce a cross-modal knowledge distillation BMD prediction framework (CMKD-BMD) fusing X-ray/CT data to enhance X-ray-only BMD prediction. Each single-modal student network employs a multi-scale visual extractor for hierarchical features, an unsupervised graph-based structural learner for anatomical relationships, and an adaptive fusion module to generate a unified representation. The teacher network integrates single-modal representations from students and transfers multimodal knowledge from the teacher to students. Our model outperforms existing methods on both the collected dataset (comprising 1620 X-rays and 280 CT cases) and the publicly available VerSe2019 dataset, demonstrating superior BMD estimation performance. Furthermore, we developed OrthoSim, an orthopedic surgical simulation platform with CMKD-BMD, which has shown promising clinical effectiveness in trial evaluations. Our code is available at https://github.com/KeyueShi/CMKD-BMD.

Detecting out-of-distribution (OOD) samples poses a significant safety challenge when deploying models in open-world scenarios. Advanced works assume that OOD and in-distributional (ID) samples exhibit a distribution discrepancy, showing an encouraging direction in estimating the uncertainty with embedding features or predicting outputs. Besides incorporating auxiliary outlier as decision boundary, quantifying a “meaningful distance” in embedding space as uncertainty measurement is a promising strategy. However, these distances-based approaches overlook the data structure and heavily rely on the high-dimension features learned by deep neural networks, causing unreliable distances due to the “curse of dimensionality”. In this work, we propose a data structure-aware approach to mitigate the sensitivity of distances to the “curse of dimensionality”, where high-dimensional features are mapped to the manifold of ID samples, leveraging the well-known manifold assumption. Specifically, we present a novel distance termed astangent distance, which tackles the issue of generalizing the meaningfulness of distances on testing samples to detect OOD inputs. Inspired by manifold learning for adversarial examples, where adversarial region probability density is close to the orthogonal direction of the manifold, and both OOD and adversarial samples have common characteristic$-$imperceptible perturbations with shift distribution, we propose that OOD samples are relatively far away from the ID manifold, wheretangent distancedirectly computes the Euclidean distance between samples and the nearest submanifold space$-$instantiated as the linear approximation of local region on the manifold. We provide empirical and theoretical insights to demonstrate the effectiveness of OOD uncertainty measurements on the low-dimensional subspace. Extensive experiments show that thetangent distanceperforms competitively with other post hoc OOD detection baselines on common and large-scale benchmarks, and the theoretical analysis supports our claim that ID samples are likely to reside in high-density regions, explaining the effectiveness of internal connections among ID data.

Dynamic graph data mining has gained popularity in recent years due to the rich information contained in dynamic graphs and their widespread use in the real world. Despite the advances in dynamic graph neural networks (DGNNs), the rich information and diverse downstream tasks have posed significant difficulties for the practical application of DGNNs in industrial scenarios. To this end, in this paper, we propose to address them by pre-training and present the Contrastive Pre-Training Method for Dynamic Graph Neural Networks (CPDG). CPDG tackles the challenges of pre-training for DGNNs, including generalization capability and long-short term modeling capability, through a flexible structural-temporal subgraph sampler along with structural-temporal contrastive pre-training schemes. Extensive experiments conducted on both large-scale research and industrial dynamic graph datasets show that CPDG outperforms existing methods in dynamic graph pre-training for various downstream tasks under three transfer settings.

Recent neural information retrieval models using dense text representations generated by pre-trained models commonly face two issues. First, a pre-trained model (e.g., BERT) usually truncates a long document before giving its representation, which may cause the loss of some important semantic information. Second, although pre-training models like BERT have been widely used in generating sentence embeddings, a substantial body of literature has shown that the pre-training models often represent sentence embeddings in a homogeneous and narrow space, known as the problem of representation anisotropy, which hurts the quality of dense vector retrieval. In this paper, we split the query and the document in information retrieval into two sets of natural sentences and generate their sentence embeddings with BERT, the most popular pre-trained model. Before aggregating the sentence embeddings to get the entire embedding representations of the input query and document, to alleviate the usual representation degeneration problem of sentence embeddings from BERT, we sample the variational auto-encoder's latent space distribution to obtain isotropic sentence embeddings and utilize supervised contrastive learning to uniform the distribution of these sentence embeddings in the representation space. Our proposed model undergoes training optimization for both the query and the document in the abovementioned aspects. Our model performs well in evaluating three extensively researched neural information retrieval datasets.

It is significant to accurately predict the epidemic trend of COVID-19 due to its detrimental impact on the global health and economy. Although machine learning based approaches have been applied to predict epidemic trend, standard models have shown low accuracy for long-term prediction due to a high level of uncertainty and lack of essential training data. This paper proposes an improved machine learning framework employing Generative Adversarial Network (GAN) and Long Short-Term Memory (LSTM) for adversarial training to forecast the potential threat of COVID-19 in countries where COVID-19 is rapidly spreading. It also investigates the most updated COVID-19 epidemiological data before October 18, 2020 and model the epidemic trend as time series that can be fed into the proposed model for data augmentation and trend prediction of the epidemic. The proposed model is trained to predict daily numbers of cumulative confirmed cases of COVID-19 in Italy, USA, China, Germany, UK, and across the world. Paper further analyzes and suggests which populations are at risk of contracting COVID-19.

On-demand food delivery (OFD) platforms have greatly impacted the food service industry, where OFD recommendation systems play a central role in enhancing user experience and raising revenues. OFD recommendation, compared with existing online e-commerce recommendation systems, needs to put more emphasis on fulfillment time related variables, because the order fulfillment cycle time (OFCT) which refers to the time elapsed between a user placing a food order and receiving the food significantly influences a user's choice from the recommended items. In this paper, we investigate the OFCT related information and propose a Fulfillment-Time-Aware Personalized Ranking (FTAPR) method for recommendation. FTAPR mainly consists of three components. First, Transformers are used to estimate OFCT based on a large amount of user order sequences. Then, the predicted OFCT and other OFCT related features are fused and encoded by a deep & cross network to learn fulfillment time related feature representation. At the last step, the time bias representation from the deep & cross network is integrated into the ranking system to deliver final search results. Extensive offline and online experiments on real-world datasets collected from one of China's largest OFD platforms Ele.me show the superiority of our model, e.g., an online A/B testing shows that FTAPR brings 1.3% and 2.5% gains in CTR and CVR compared with baselines.

Malicious Web crawlers threaten information system due to heavily taking up bandwidth resources and stealing private user data. Ele.me, a prevalent on-demand food delivery platform in China, suffers from the negative impact of crawlers. The crawler detection systems face two major challenges: spatial patterns of the crawler behaviors and limited labeled data for training. In this paper, we present efficient solutions to tackle these challenges. Specifically, we propose a new Attributed Action Net (AANet for short) model to detect Location-Based Services~(LBS) crawlers and a three-stage learning framework to train the model. AANet consists of three different embedding modules, including the action token sequence, temporal-spatial attributes of users, and the context information of the raw data. We have deployed the model at Ele.me, and both offline experiments and online A/B tests show that the proposed method is superior to the state-of-the-art models for sequence data classification on the food delivery platform.

Live-streaming platforms have recently gained significant popularity by attracting an increasing number of young users and have become a very promising form of online shopping. Similar to the traditional online shopping platforms such as Taobao, live-streaming platforms also suffer from online malicious fraudulent behaviors where many transactions are not genuine. The existing anti-fraud models proposed to recognize fraudulent transactions on traditional online shopping platforms are inapplicable on live-streaming platforms. This is mainly because live-streaming platforms are characterized by a unique type of heterogeneous live-streaming networks where multiple heterogeneous types of nodes such as users, live-streamers, and products are connected with multiple different types of edges associated with edge features. In this paper, we propose a new approach based on a heterogeneous graph neural network for LIve-streaming Fraud dEtection (called LIFE). LIFE designs an innovative heterogeneous graph learning model that fully utilizes various heterogeneous information of shopping transactions, users, streamers, and items from a given live-streaming platform. Moreover, a label propagation algorithm is employed within our LIFE framework to handle the limited number of labeled fraudulent transactions for model training. Extensive experimental results on a large-scale Taobao live-streaming platform demonstrate that the proposed method is superior to the baseline models in terms of fraud detection effectiveness on live-streaming platforms. Furthermore, we conduct a case study to show that the proposed method is able to effectively detect fraud communities for live-streaming e-commerce platforms.

Graph classification is an important tool for analyzing data with structure dependency, where subgraphs are often used as features for learning. In reality, the dimension of the subgraphs crucially depends on the threshold setting of the frequency support parameter, and the number may become extremely large. As a result, subgraphs may be incrementally discovered to form a feature stream and require the underlying graph classifier to effectively discover representative subgraph features from the subgraph feature stream. In this paper, we propose a primal-dual incremental subgraph feature selection algorithm (ISF) based on a max-margin graph classifier. The ISF algorithm constructs a sequence of solutions that are both primal and dual feasible. Each primal-dual pair shrinks the dual gap and renders a better solution for the optimal subgraph feature set. To avoid bias of ISF algorithm on short-pattern subgraph features, we present a new incremental subgraph join feature selection algorithm (ISJF) by forcing graph classifiers to join short-pattern subgraphs and generate long-pattern subgraph features. We evaluate the performance of the proposed models on both synthetic networks and real-world social network data sets. Experimental results demonstrate the effectiveness of the proposed methods.

Graph classification is an important tool for analysing structured and semi-structured data, where subgraphs are commonly used as the feature representation. However, the number and size of subgraph features crucially depend on the threshold parameters of frequent subgraph mining algorithms. Any improper setting of the parameters will generate many trivial short-pattern subgraph fragments which dominate the feature space, distort graph classifiers and bury interesting long-pattern subgraphs. In this paper, we propose a new Subgraph Join Feature Selection (SJFS) algorithm. The SJFS algorithm, by forcing graph classifiers to join short-pattern subgraph fragments, can defrag trivial subgraph features and deliver long-pattern interesting subgraphs. Experimental results on both synthetic and real-world social network graph data demonstrate the performance of the proposed method.