Structured data question answering (QA), including table QA, Knowledge Graph (KG) QA, and temporal KG QA, is a pivotal research area. Advances in large language models (LLMs) have driven significant progress in unified structural QA frameworks like TrustUQA. However, these frameworks face challenges when applied to small-scale LLMs since small-scale LLMs are prone to errors in generating structured queries. To improve the structured data QA ability of small-scale LLMs, we propose a self-correction distillation (SCD) method. In SCD, an error prompt mechanism (EPM) is designed to detect errors and provide customized error messages during inference, and a two-stage distillation strategy is designed to transfer large-scale LLMs' query-generation and error-correction capabilities to small-scale LLM. Experiments across 5 benchmarks with 3 structured data types demonstrate that our SCD achieves the best performance and superior generalization on small-scale LLM (8B) compared to other distillation methods, and closely approaches the performance of GPT4 on some datasets. Furthermore, large-scale LLMs equipped with EPM surpass the state-of-the-art results on most datasets.

Tao Li, Wenshuo Ge, Zhichao Wang, Zihao Cui, Yong Ma, Yingying Gao, Chao Deng, Shilei Zhang, Junlan Feng. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The majority of data in businesses and industries is stored in tables, databases, and data warehouses. Reasoning with table-structured data poses significant challenges for large language models (LLMs) due to its hidden semantics, inherent complexity, and structured nature. One of these challenges is lacking an effective evaluation benchmark fairly reflecting the performances of LLMs on broad table reasoning abilities. In this paper, we fill in this gap by presenting a comprehensive table reasoning benchmark, TReB. Firstly, we propose a taxonomy to systematically measure both shallow table understanding abilities and deep table reasoning abilities, covering a total of 26 sub-tasks. We then construct a high quality dataset through a dedicated data processing and synthesis procedure. Based on these well-constructed samples, we design an evaluation framework to robustly measure table reasoning capabilities with three distinct inference modes. Experimental results with our data and framework reveal that existing LLMs still have significant room for improvement in addressing the complex and real world table related tasks. Both the dataset and evaluation framework are publicly available, with the dataset hosted on huggingface.co/datasets/JT-LM/JIUTIAN-TReB, and the framework on github.com/JT-LM/jiutian-treb.

Large Language Models (LLMs) are often English-centric due to the disproportionate distribution of languages in their pre-training data. Enhancing non-English language capabilities through post-pretraining often results in catastrophic forgetting of high-resource languages. Previous methods either achieve good expansion with severe forgetting or slight forgetting with poor expansion, indicating the challenge of balancing language expansion while preventing forgetting. In this paper, we propose a method called MoE-LPR (Mixture-of-Experts with Language Priors Routing) to alleviate this problem. MoE-LPR employs a two-stage training approach to enhance the multilingual capability. First, the model is post-pretrained into a Mixture-of-Experts(MoE) architecture by upcycling, where all the original parameters are frozen and new experts are added. In this stage, we focus improving the ability on expanded languages, without using any original language data. Then, the model reviews the knowledge of the original languages with replay data amounting to less than 1% of post-pretraining, where we incorporate language priors routing to better recover the abilities of the original languages. Evaluations on multiple benchmarks show that MoE-LPR outperforms other post-pretraining methods. Freezing original parameters preserves original language knowledge while adding new experts preserves the learning ability. Reviewing with LPR enables effective utilization of multilingual knowledge within the parameters. Additionally, the MoE architecture maintains the same inference overhead while increasing total model parameters. Extensive experiments demonstrate MoE-LPR’s effectiveness in improving expanded languages and preserving original language proficiency with superior scalability.

Cross-lingual context retrieval (extracting contextual information in one language based on requests in another) is a fundamental aspect of cross-lingual alignment, but the performance and mechanism of it for large language models (LLMs) remains unclear.In this paper, we evaluate the cross-lingual context retrieval of over 40 LLMs across 12 languages, using cross-lingual machine reading comprehension (xMRC) as a representative scenario.Our results show that post-trained open LLMs show strong cross-lingual context retrieval ability, comparable to closed-source LLMs such as GPT-4o, and their estimated oracle performances greatly improve after posttraining.Our mechanism analysis shows that the cross-lingual context retrieval process can be divided into two main phases: question encoding and answer retrieval, which are formed in pre-training and post-training respectively.The phasing stability correlates with xMRC performance, and the xMRC bottleneck lies at the last model layers in the second phase, where the effect of post-training can be evidently observed.Our results also indicate that largerscale pretraining cannot improve the xMRC performance.Instead, larger LLMs need further multilingual post-training to fully unlock their cross-lingual context retrieval potential.

Transformer has become one of the most popular architectures for multivariate time series (MTS) forecasting. However, existing Transformer-based methods still lack consideration of cross-time-and-channel dependency modeling, which is important to MTS forecasting. In addition, existing methods either completely ignore the channel dependency for robustness or model channel dependency at the sacrifice of robustness. How to achieve a balance between robustness and information capacity has not been investigated. To address these problems, a Transformer-based method, InjectTST, is proposed in this paper. Instead of designing a channel-dependent model directly, we retain the channel-independent backbone and restrainedly inject global information into individual channels. A joint temporal-channel attention scheme is proposed to better model the global information and a token dropout mechanism is designed to improve the model robustness and calculation efficiency. Through a Transformer-based injection module, the independent channels get improved by concentrating on useful global information. Experiments indicate that InjectTST can achieve stable improvement compared with state-of-the-art methods.

Large-scale pre-trained language models (PLMs) require significant computational resources to train from scratch on large volumes of data. But in the real world, emerging data from diverse sources may not be initially available for pre-training. Recent studies on lifelong learning have tried to solve this problem by exploring the use of model growth techniques to effectively incorporate new knowledge without the need for complete re-training. However, model growth approaches utilized have issues with growth operators that do not ensure strict function preservation or growth schedules that only include a few growth dimensions, reducing lifelong learning's effect. Furthermore, existing approaches often assume that emerging data has the same distribution as pre-training data, causing catastrophic forgetting of previously acquired knowledge. To address the aforementioned issues, we introduce LOIRE, a framework for lifelong learning that enables PLMs to effectively grow their capacity using incremental data. LOIRE employs growth operators for all feasible dimensions and a growth schedule to generate the optimal expansion sequence in the field of lifelong learning. Specifically, we present a novel plug-in layer growth operator with residual connections that skip the newly added layer during initial training while ensuring function preservation. We additionally propose an iterative distillation strategy for LOIRE that allows an intermediate model in the growth stages to switch between being a student and a teacher, reducing catastrophic forgetting during growth. Experiments show that LOIRE can reduce computational expenses by an average of 29.22\% while retaining equivalent or better downstream performance.

The recent boom of Transformer-based models have enhanced state-of-the-art results of multivariate time series (MTS) forecasting. However, MTS forecasting remains a challenging problem, primarily because of the intricate temporal patterns and obscured spatial correlations. Existing models are not only computationally expensive in modeling long-term temporal dependencies, but more importantly, fail to adequately capture the interrelations among variables. To address these problems, we propose STNet, a Spatial-Temporal Transformer Network with self-supervised pre-training scheme for MTS forecasting. In STNet, the MTS are formalized as a data-driven graph structure, which is learned through the training process to extract the latent patterns within the spatial dependencies of the data. Then we encode the structural information of the graph into the spatial Transformer encoder to help STNet better model refined spatial dependencies. A patch-level Transformer encoder is implemented to efficiently enhance locality and comprehensively capture semantic information pertaining to temporal dependencies. Moreover, the pre-training model generates rich contextual information, which consistently leads to reliable outcomes in transfer performance for downstream tasks. Extensive experiments conducted on five real-world benchmark datasets show the proposed STNet improves the prediction accuracy by 5.0%-25.2% compared to previous state-of-the-arts.

Large-scale mobile traffic prediction is critical for intelligent telecommunication networks, but remains challenging due to complex spatial-temporal patterns and network heterogeneity. Existing transfer learning-based prediction methods, tailored for grid data, inadequately handle graph-based mobile traffic data and ignore network heterogeneity. In this paper, we propose a Knowledge-enhanced Hierarchical Graph Transfer Learning (KHGTL) framework integrating graph-level and node-level knowledge transfer, for large-scale mobile traffic prediction. We also present a Spatial-Temporal Adaptive Graph Convolutional Network to capture the intricate spatial-temporal correlations. Experiments on real datasets show KHGTL reduces MAE by 44.7% compared to baselines, outperforms full supervision with 80% faster training, addressing scalability and efficiency challenges in large-scale traffic prediction.

Peng Hu, Sizhe Liu, Changjiang Gao, Xin Huang, Xue Han, Junlan Feng, Chao Deng, Shujian Huang. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Zhe Yang, Yi Huang, Yaqin Chen, XiaotingWu XiaotingWu, Junlan Feng, Chao Deng. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

This paper presents a comprehensive exploration of the phenomenon of data redundancy in video understanding, with the aim to improve computational efficiency. Our investigation commences with an examination of spatial redundancy, which refers to the observation that the most informative region in each video frame usually corresponds to a small image patch, whose shape, size and location shift smoothly across frames. Motivated by this phenomenon, we formulate the patch localization problem as a dynamic decision task, and introduce a spatially adaptive video recognition approach, termed AdaFocus. In specific, a lightweight encoder is first employed to quickly process the full video sequence, whose features are then utilized by a policy network to identify the most task-relevant regions. Subsequently, the selected patches are inferred by a high-capacity deep network for the final prediction. The complete model can be trained conveniently in an end-to-end manner. During inference, once the informative patch sequence has been generated, the bulk of computation can be executed in parallel, rendering it efficient on modern GPU devices. Furthermore, we demonstrate that AdaFocus can be easily extended by further considering the temporal and sample- wise redundancies, i.e., allocating the majority of computation to the most task-relevant video frames, and minimizing the computation spent on relatively "easier" videos. Our resulting algorithm, Uni-AdaFocus, establishes a comprehensive framework that seamlessly integrates spatial, temporal, and sample- wise dynamic computation, while it preserves the merits of AdaFocus in terms of efficient end-to-end training and hardware friendliness. In addition, Uni-AdaFocus is general and flexible as it is compatible with off-the-shelf backbone models (e.g., TSM and X3D), which can be readily deployed as our feature extractor, yielding a significantly improved computational efficiency. Empirically, extensive experiments based on seven widely-used benchmark datasets (i.e., ActivityNet, FCVID, Mini-Kinetics, Something-Something V1&V2, Jester, and Kinetics-400) and three real-world application scenarios (i.e., fine-grained diving action classification, Alzheimer's and Parkinson's diseases diagnosis with brain magnetic resonance images (MRI), and violence recognition for online videos) substantiate that Uni-AdaFocus is considerably more efficient than the competitive baselines. Code and pre-trained models are available at https://github.com/blackfeather-wang/AdaFocus and https://github.com/LeapLabTHU/AdaFocusV2.

App usage prediction aims to predict the next app most likely to be used based on historical behaviors, which is beneficial for smartphone system optimization, such as system resource management, battery energy optimization, and user experience enhancement. Existing studies have treated it as a simple time series prediction problem and overlooked the sessionization characteristic of mobile app usage, i.e., neglecting the intent context in which the user interacts with apps. In this paper, we explore the context of user intents and incorporate app sessionization features into prediction models to improve prediction accuracy. Specifically, we first extract the semantic meaning of spatio-temporal contextual information of app usage by constructing an urban knowledge graph. Second, we devise a hypergraph-based embedding model to extract the hyper-relations of intra-session apps. Third, we utilize a self-attention mechanism to fuse intra-session apps’ representations and combine spatio-temporal contextual embedding to form the session representation. We further leverage a transformer for inter-session intent transition modeling to extract users’ dynamic intent (i.e., the semantic meaning of sessions) for app usage. Finally, we jointly fuse dynamic intent and recently used app features using the MLP model for the prediction. The novelty of our method is that we are the first to leverage dynamic hypergraphs for modeling sessionization features, and we model both inter-session and intra-session relations. We evaluate our model based on two real-world datasets collected in Shanghai and Nanchang. In terms of prediction accuracy, mean reciprocal rank, and normalized discounted cumulative gain, our proposed framework outperforms state-of-the-art baselines by more than 30% in the Shanghai dataset and 20% in the Nanchang dataset, respectively.

By integrating machine learning with edge devices, we can augment the capabilities of edge devices, such as IoT devices, household appliances, and wearable technologies. These edge devices generally operate on microcontrollers with inherently limited resources, such as constrained RAM capacity and limited computational power. Nonetheless, they often process data in a high-velocity stream fashion, exemplified by sequences of activities and statuses monitored by advanced industrial sensors. In practical scenarios, models must be interpretable to facilitate troubleshooting and behavior understanding. Implementing machine learning models on edge devices is valuable and challenging, striking a balance between model efficacy and resource constraint. To address this challenge, we introduce our novel Onfesk, which combines online learning algorithms with an innovative interpretable kernel. Specifically, our Onfesk trains an online classifier over the kernel's feature sketches. Benefiting from our specially designed modules, the kernel's feature sketches can be efficiently produced, and the memory requirements of the classifier can be significantly reduced. As a result, Onfesk delivers effective and efficient performance in environments with limited resources without compromising on model interpretability. Extensive experiments with diverse real-world datasets have shown that Onfesk outperforms state-of-the-art methods, achieving up to a 7.4% improvement in accuracy within identical memory constraints.

Shimao Zhang, Changjiang Gao, Wenhao Zhu, Jiajun Chen, Xin Huang, Xue Han, Junlan Feng, Chao Deng, Shujian Huang. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Active learning (AL) aims to reduce labeling costs by selecting the most valuable samples to annotate from a set of unlabeled data. However, recognizing these samples is particularly challenging in multi-label text classification tasks due to the high dimensionality but sparseness of label spaces. Existing AL techniques either fail to sufficiently capture label correlations, resulting in label imbalance in the selected samples, or suffer significant computing costs when analyzing the informative potential of unlabeled samples across all labels. Facing these challenges, we propose an efficient two-stage sample acquisition strategy for multi-label active learning, called ALMuLa-mix. For saving the computational costs, ALMuLa-mix first attempts to identify unlabeled samples with novel features by employing a time-efficient feature-mixing method in conjunction with label correlations. Regarding the label imbalance, ALMuLa-mix then leverages the minority class in the labeled set to select a small batch of candidate unlabeled samples with greater inter-class diversity from the candidates with novel features. Experimental results on publicly available datasets show that ALMuLa-mix is superior to other strong baselines for handling multi-label text classification tasks.

To improve the robustness of pre-trained language models (PLMs), previous studies have focused more on how to efficiently obtain adversarial samples with similar semantics, but less attention has been paid to the perturbed samples that change the gold label. Therefore, to fully perceive the effects of these different types of small perturbations on robustness, we propose a RObust Self-supervised leArning (ROSA) method, which incorporates different types of perturbed samples and the robustness improvements into a unified framework. Subsequently, to implement ROSA, a perturbed sample generation strategy supported by the large language models (LLMs) is proposed, which adaptively controls the generation process based on the fine-grained similarity information among the training samples. The experimental results demonstrate the remarkable performance of our ROSA.

Fine-grained action recognition aims to classify complex coarse actions into finer ones. It suffers due to varying temporal scales and subtle differences between categories, which brings significant challenges to the temporal perception capabilities of current models. In this paper, we proposed a novel Temporal Cues Transformer (TCT) to guide the recognition of fine-grained action by exploiting detailed temporal cues in video duration and action sequence. The proposed TCT consists of a duration-aware encoder and a hierarchical sequence aggregation decoder. In the encoder, we extract duration-aware representations from the video and its duration. In the decoder, we reinforce the learning of action sequence by first searching the fine-grained elements of action with a hierarchical element query module, then aggregating the elements with a sequence aggregation module to predict the action category. Extensive experiments on widely used Diving48 and FineGym datasets demonstrate the superiority of the proposed method compared to the state-of-the-art methods.

In recent years, vision transformers have been introduced into face recognition and analysis and have achieved performance breakthroughs. However, most previous methods generally train a single model or an ensemble of models to perform the desired task, which ignores the synergy among different tasks and fails to achieve improved prediction accuracy, increased data efficiency, and reduced training time. This paper presents a multi-purpose algorithm for simultaneous face recognition, facial expression recognition, age estimation, and face attribute estimation (40 attributes including gender) based on a single Swin Transformer. Our design, the SwinFace, consists of a single shared backbone together with a subnet for each set of related tasks. To address the conflicts among multiple tasks and meet the different demands of tasks, a Multi-Level Channel Attention (MLCA) module is integrated into each task-specific analysis subnet, which can adaptively select the features from optimal levels and channels to perform the desired tasks. Extensive experiments show that the proposed model has a better understanding of the face and achieves excellent performance for all tasks. Especially, it achieves 90.97% accuracy on RAF-DB and 0.22 ϵ-error on CLAP2015, which are state-of-the-art results on facial expression recognition and age estimation respectively. The code and models will be made publicly available at https://github.com/lxq1000/SwinFace.

Fine-grained image recognition is a longstanding computer vision challenge that focuses on differentiating objects belonging to multiple subordinate categories within the same meta-category. Since images belonging to the same meta-category usually share similar visual appearances, mining discriminative visual cues is the key to distinguishing fine-grained categories. Although commonly used image-level data augmentation techniques have achieved great success in generic image classification problems, they are rarely applied in fine-grained scenarios, because their random editing-region behavior is prone to destroy the discriminative visual cues residing in the subtle regions. In this paper, we propose diversifying the training data at the feature-level to alleviate the discriminative region loss problem. Specifically, we produce diversified augmented samples by translating image features along semantically meaningful directions. The semantic directions are estimated with a covariance prediction network, which predicts a sample-wise covariance matrix to adapt to the large intra-class variation inherent in fine-grained images. Furthermore, the covariance prediction network is jointly optimized with the classification network in a meta-learning manner to alleviate the degenerate solution problem. Experiments on four competitive fine-grained recognition benchmarks (CUB-200-2011, Stanford Cars, FGVC Aircrafts, NABirds) demonstrate that our method significantly improves the generalization performance on several popular classification networks (e.g., ResNets, DenseNets, EfficientNets, RegNets and ViT). Combined with a recently proposed method, our semantic data augmentation approach achieves state-of-the-art performance on the CUB-200-2011 dataset. Source code is available at https://github.com/LeapLabTHU/LearnableISDA.

Understanding and accurately predicting cellular traffic data is vital for communication operators and device users, as it facilitates efficient resource allocation and ensures superior service quality. However, large-scale cellular traffic data forecasting remains challenging due to intricate temporal variations and complex spatial relationships. This article proposes a Knowledge Graph Driven Decomposition Approach (KGDA) for precise cellular traffic prediction. The KGDA breaks down the impact of static environmental factors and dynamic autocorrelations of cellular traffic time series, enabling the capture of overall traffic changes and understanding of traffic dependence on past values. Specifically, we propose an urban knowledge graph to capture the static environmental context of base stations, mapping these entities into the same latent space while retaining static environmental knowledge. The cellular traffic is divided into a regular pattern and fluctuating residual components, with the KGDA comprising four modules: a Knowledge Graph Representation Learning model, a traffic regular pattern prediction module, a traffic residual dynamic prediction module, and an attentional fusion module. The first leverages graph neural networks to extract spatial contexts and predict regular patterns, the second utilizes the Bi-directional Long Short-Term Memory (Bi-LSTM) model to capture autocorrelations of traffic time series, and the final module integrates the patterns and residuals to produce the final prediction result. Comprehensive experiments demonstrate that our proposed model outperforms state-of-the-art models by more than 10% in forecasting cellular traffic.

With the rapid development of smart TV, TV recommendation is attracting more and more users. TV users usually distribute in multiple regions with different cultures and hence have diverse TV program preferences. From the perspective of engineering practice and performance improvement, it's very essential to model users from multiple regions with one single model. In previous work, Multi-gate Mixture-of-Expert (MMoE) has been widely adopted in multi-task and multi-domain recommendation scenarios. In practice, however, we first observe the embeddings generated by experts tend to be homogeneous which may result in high semantic similarities among embeddings that reduce the capability of Multi-gate Mixture-of-Expert (MMoE) model. Secondly, we also find there are lots of commonalities and differences between multiple regions regarding user preferences. Therefore, it's meaningful to model the complicated relationships between regions. In this paper, we first introduce contrastive learning to overcome the expert representation degeneration problem. The embeddings of two augmented samples generated by the same experts are pushed closer to enhance the alignment, and the embeddings of the same samples generated by different experts are pushed away in vector space to improve uniformity. Then we propose a Graph-based Gating Mechanism to empower typical Multi-gate Mixture-of-Experts. Graph-based MMoE is able to recognize the commonalities and differences among multiple regions by introducing a Graph Neural Network (GNN) with region similarity prior. We name our model Multi-gate Mixture-of-Contrastive-Experts model with Graph-based Gating Mechanism (MMoCEG). Extensive offline experiments and online A/B tests on a commercial TV service provider over 100 million users and 2.3 million items demonstrate the efficacy of MMoCEG compared to the existing models.

The optimization and management of transportation play an important role in the Intelligent Transportation System (ITS), and the prediction of traffic flow data is the basis of it. Recently, graph convolution network (GCN) is widely applied to extract the features of non-Euclidean data, some models usually use the predefined distance-based adjacency matrix for GCN to extract node features. However, the distance graph can only capture part of the spatial correlation and limit the effective learning of those models. In fact, there are close similarities between the remote intersection with similar time sequence, they may be living quarters or industrial zones in the city. Besides, some previous models capture incomplete temporal correlation, which will affect the long-term prediction task to a certain extent. To address these limitations, we propose a novel Spatial-Temporal Residual Multi-Graph Convolution Network (ST-RMGCN) for traffic flow forecasting. Specifically, our model utilizes SimHash to calculate the time sequence similarity between nodes and build a semantic graph which is be instrumental in capturing remote spatial correlation. Moreover, multi-graph convolution network also extracts the features of adjacent nodes through distance graph. In addition, we design a novel time embedding method which contains the exact time information of the time step, and it contributes to the attention mechanism and Gate Recurrent Unit (GRU) to capture the local and global time information. Extensive experiments on six real-world datasets show that our proposed model performs better than the state-of-the-art methods.

Fine-grained address entity recognition (FGAER) from multi-turn spoken dialogues is particularly challenging.The major reason lies in that a full address is often formed through a conversation process.Different parts of an address are distributed through multiple turns of a dialogue with spoken noises.It is nontrivial to extract by turn and combine them.This challenge has not been well emphasized by main-stream entity extraction algorithms.To address this issue, we propose in this paper a logic-guided fine-grained address recognition method (Log-FGAER), where we formulate the address hierarchy relationship as the logic rule and softly apply it in a probabilistic manner to improve the accuracy of FGAER.In addition, we provide an ontology-based data augmentation methodology that employs ChatGPT to augment a spoken dialogue dataset with labeled address entities.Experiments are conducted using datasets generated by the proposed data augmentation technique and derived from real-world scenarios.The results of the experiment demonstrate the efficacy of our proposal.Where are you in Suzhou city, Jiangsu Prov? 您现在在江苏 苏州市 的哪里？ I'm in Wujiang District, Shengze Town in Wujiang District.

Accurately predicting base station traffic volumes and understanding mobile traffic patterns is essential for smart city development, enabling efficient resource allocation and ensuring high-quality communication services. However, existing works have limitations in capturing spatial information, though the surrounding environment plays a critical role in mobile traffic prediction. In this paper, we utilize a spatial knowledge graph to represent spatial information and add important urban components to augment it making it a more effective tool for capturing environmental information. we further propose a multi-relational knowledge graph convolutional network model for mobile traffic prediction, which consists of three parts. The environmental context modelling captures spatial information from the augmented spatial knowledge graph using tucker decomposition and relational graph convolutional network. The semantic relationship modelling extracts semantic relationships between base stations and employs transformer and causal convolution to capture temporal features. The inter-attentional fusion modelling utilizes the self-attention mechanism to further capture base station relationships and predict future traffic volumes. Extensive experiments demonstrate that our proposed model significantly outperforms the state-of-the-art models by over 10% in mobile traffic prediction. The code is available at https://github.com/tsinghua-fiblab/Mobile-Traffic-Prediction-sigspatial23

We tend to assume most available speech corpora we use are either completely clean or completely noised. However, the reality is most of them are a mix of both. In this paper, we propose a semi-supervised speech enhancement framework to enhance such typical speech datasets. This framework includes an estimator to measure the speech purity. Utterances with high speech purity are considered clean, otherwise noised. For clean speech utterances, we follow the supervised learning mechanism to train a deep learning speech enhancement model. For noised speech, we update the model in an unsupervised manner. Hence, we design our training loss as a combination of the supervised loss and unsupervised loss. We refer to this framework as SemiEnhance. Experimental results show that SemiEnhance substantially improves the speech quality, and achieves new state-of-the-art results on benchmark datasets: 2022 DNS Challenge and NoiseX-92.

Previous contrastive learning methods for sentence representations often focus on insensitive transformations to produce positive pairs, but neglect the role of sensitive transformations that are harmful to semantic representations. Therefore, we propose an Equivariant Self-Contrastive Learning (ESCL) method to make full use of sensitive transformations, which encourages the learned representations to be sensitive to certain types of transformations with an additional equivariant learning task. Meanwhile, in order to improve practicability and generality, ESCL simplifies the implementations of traditional equivariant contrastive methods to share model parameters from the perspective of multi-task learning. We evaluate our ESCL on semantic textual similarity tasks. The proposed method achieves better results while using fewer learning parameters compared to previous methods.

Due to the growth of densification and multi-band services in emerging cellular networks, the next generation NodeB will confront with a plethora of handover-related Configuration and Optimization Parameters (COPs). Adjusting the handover-related COPs will directly impact the networks Key Performance Indicators (KPIs). Thus, accurately estimating the effect of the handover-related COPs adjustment on KPI values and modeling the COP-KPI relationship can provide guidelines for intelligent network management. Existing studies have applied machine learning to model the intricate relationship between COP and KPI, but these methods merely exploit their correlations. In this paper, considering that there exist the causal effects between COPs and KPIs that go beyond the correlation analysis of them, we propose the causal inference aided handover parameter adjusting effect estimation and design the Causal-Stable Decomposition Network (CSDnet). Specifically, we propose the causal-stable deconstructor module based on self-attention mechanism to characterize the complex causality-stability of KPI induced by the COP adjustment. To ensure the training, we propose the gradient matching loss and the consistency constraint loss. Then, we design the stable decoder and the causal decoder that incorporate the causal information into COP-KPI modeling. Extensive experiments on the COP-KPI dataset collected from the real-world cellular networks verify the effectiveness of CSDnet.

Cellular traffic prediction is an indispensable part for intelligent telecommunication networks. Nevertheless, due to the frequent user mobility and complex network scheduling mechanisms, cellular traffic often inherits complicated spatial-temporal patterns, making the prediction incredibly challenging. Although recent advanced algorithms such as graph-based prediction approaches have been proposed, they frequently model spatial dependencies based on static or dynamic graphs and neglect the coexisting multiple spatial correlations induced by traffic generation. Meanwhile, some works lack the consideration of the diverse cellular traffic patterns, result in suboptimal prediction results. In this paper, we propose a novel deep learning network architecture, Adaptive Hybrid Spatial-Temporal Graph Neural Network (AHSTGNN), to tackle the cellular traffic prediction problem. First, we apply adaptive hybrid graph learning to learn the compound spatial correlations among cell towers. Second, we implement a Temporal Convolution Module with multi-periodic temporal data input to capture the nonlinear temporal dependencies. In addition, we introduce an extra Spatial-Temporal Adaptive Module to conquer the heterogeneity lying in cell towers. Our experiments on two real-world cellular traffic datasets show AHSTGNN outperforms the state-of-the-art by a significant margin, illustrating the superior scalability of our method for spatial-temporal cellular traffic prediction.

Early exiting has become a promising approach to improving the inference efficiency of deep networks. By structuring models with multiple classifiers (exits), predictions for "easy" samples can be generated at earlier exits, negating the need for executing deeper layers. Current multi-exit networks typically implement linear classifiers at intermediate layers, compelling low-level features to encapsulate high-level semantics. This sub-optimal design invariably undermines the performance of later exits. In this paper, we propose Dynamic Perceiver (Dyn-Perceiver) to decouple the feature extraction procedure and the early classification task with a novel dual-branch architecture. A feature branch serves to extract image features, while a classification branch processes a latent code assigned for classification tasks. Bi-directional cross-attention layers are established to progressively fuse the information of both branches. Early exits are placed exclusively within the classification branch, thus eliminating the need for linear separability in low-level features. Dyn-Perceiver constitutes a versatile and adaptable framework that can be built upon various architectures. Experiments on image classification, action recognition, and object detection demonstrate that our method significantly improves the inference efficiency of different backbones, outperforming numerous competitive approaches across a broad range of computational budgets. Evaluation on both CPU and GPU platforms substantiate the superior practical efficiency of Dyn-Perceiver. Code is available at https://www.github.com/LeapLabTHU/Dynamic_Perceiver.

Item representation learning (IRL) plays an essential role in recommender systems, especially for sequential recommendation. Traditional sequential recommendation models usually utilize ID embeddings to represent items, which are not shared across different domains and lack the transferable ability. Recent studies use pre-trained language models (PLM) for item text embeddings (text-based IRL) that are universally applicable across domains. However, the existing text-based IRL is unaware of the important collaborative filtering (CF) information. In this paper, we propose CoWPiRec, an approach of Collaborative Word-based Pre-trained item representation for Recommendation. To effectively incorporate CF information into text-based IRL, we convert the item-level interaction data to a word graph containing word-level collaborations. Subsequently, we design a novel pre-training task to align the word-level semantic-and CF-related item representation. Extensive experimental results on multiple public datasets demonstrate that compared to state-of-the-art transferable sequential recommenders, CoWPiRec achieves significantly better performances in both fine-tuning and zero-shot settings for cross-scenario recommendation and effectively alleviates the cold-start issue. The code is available at: https://github.com/ysh-1998/CoWPiRec.

Traffic forecasting has numerous application scenarios, but it is also a challenging task considering complex spatial and temporal correlations. Deep learning has shown great potential in recent years. Graph convolutional networks, recurrent neural networks and other deep learning architectures have been used in traffic forecasting. However, most existing methods capture spatial and temporal correlations sequentially or synchronously and couple spatial and temporal representations, which may introduce spurious dependencies. The heterogeneity of spatio-temporal data also increases the difficulty of modeling and makes it harder to obtain accurate predictions. In this paper, we propose a novel multi-branch spatial-temporal decoupling neural network (MBSTDN) for traffic forecasting. Specifically, we independently map hidden features into spatial and temporal representations and process them through corresponding branches to produce partial predictions. The partial predictions are then aggregated into final predictions by residual links. At the same time, we reconstruct input signals through a separate branch, introducing layer-by-layer decomposition based on residual learning to deal with the intricate dependencies. Furthermore, we design an embedding layer utilizing multi-dimensional information to reduce the influence of spatio-temporal heterogeneity on predictions. Experiments on four public traffic datasets demonstrate the effectiveness of MBSTDN which performs better than the state-of-the-art baselines.

With the rapid development of the cellular network, network planning is increasingly important. Generating large-scale urban cellular traffic contributes to network planning via simulating the behaviors of the planned network. Existing methods fail in simulating the long-term temporal behaviors of cellular traffic while cannot model the influences of the urban environment on the cellular networks. We propose a knowledge-enhanced GAN with multi-periodic patterns to generate large-scale cellular traffic based on the urban environment. First, we design a GAN model to simulate the multi-periodic patterns and long-term aperiodic temporal dynamics of cellular traffic via learning the daily patterns, weekly patterns, and residual traffic between long-term traffic and periodic patterns step by step. Then, we leverage urban knowledge to enhance traffic generation via constructing a knowledge graph containing multiple factors affecting cellular traffic in the surrounding urban environment. Finally, we evaluate our model on a real cellular traffic dataset. Our proposed model outperforms three state-of-art generation models by over 32.77%, and the urban knowledge enhancement improves the performance of our model by 4.71%. Moreover, our model achieves good generalization and robustness in generating traffic for urban cellular networks without training data in the surrounding areas.

The emergence of the metaverse is being widely viewed as a revolutionary technology owing to a myriad of factors, particularly the potential to increase the accessibility of learning for students with disabilities. However, not much is yet known about the views and expectations of disabled students in this regard. The fact that the metaverse is still in its nascent stage exemplifies the need for such timely discourse. To bridge this important gap, we conducted a series of semi-structured interviews with 56 university students with disabilities in the United States and Hong Kong to understand their views and expectations concerning the future of metaverse-driven education. We have distilled student expectations into five thematic categories, referred to as the REEPS framework: Recognition, Empowerment, Engagement, Privacy, and Safety. Additionally, we have summarized the main design considerations in eight concise points. This paper is aimed at helping technology developers and policymakers plan ahead of time and improving the experiences of students with disabilities.

Calibration in recommender systems ensures that the user’s interests distribution over groups of items is reflected with their corresponding proportions in the recommendation, which has gained increasing attention recently. For example, a user who watched 80 entertainment videos and 20 knowledge videos is expected to receive recommendations comprising about 80% entertainment and 20% knowledge videos as well. However, with the increasing calls for responsible recommendation, it has become inadequate to just match users’ historical behaviors especially when items are grouped by their qualities, which could result in undesired effects at the system level (e.g., overwhelming clickbaits). In this paper, we envision the two-sided calibration task that not only matches the users’ past interests distribution (user-level calibration) but also guarantees an overall target exposure distribution of different item groups (system-level calibration). The target group exposure distribution can be explicitly pursued by users, platform owners, and even the law (e.g., the platform owners expect about 50% knowledge video recommendation on the whole). To support this scenario, we propose a post-processing method named PCT. PCT first solves personalized calibration targets that minimize the changes in users’ historical interest distributions while ensuring the overall target group exposure distribution. Then, PCT reranks the original recommendation lists according to personalized calibration targets to generate both relevant and two-sided calibrated recommendations. Extensive experiments demonstrate the superior performance of the proposed method compared to calibrated and fairness-aware recommendation approaches.

This paper is concerned with TV recommendation, where one major challenge is the coupling behavior issue that the behaviors of multiple users are coupled together and not directly distinguishable because the users share the same account. Unable to identify the current watching user and use the coupling behaviors directly could lead to sub-optimal recommendation results due to the noise introduced by the behaviors of other users. Most existing methods deal with this issue either by unsupervised clustering algorithms or depending on latent user representation learning with strong assumptions. However, they neglect to sophisticatedly model the current session behaviors, which carry the information of user identification. Another critical limitation of the existing models is the lack of supervision signal on distinguishing behaviors because they solely depend on the final click label, which is insufficient to provide effective supervision. To address the above problems, we propose the Coupling Sequence Model (COSMO) for TV recommendation. In COSMO, we design a session-aware co-attention mechanism that uses both the candidate item and session behaviors as the query to attend to the historical behaviors in a fine-grained manner. Furthermore, we propose to use the data of accounts with multiple devices (e.g., families with various TV sets), which means the behaviors of one account are generated on different devices. We regard the device information as weak supervision and propose a novel pair-wise attention loss for learning to distinguish the coupling behaviors. Extensive offline experiments and online A/B tests over a commercial TV service provider demonstrate the efficacy of COSMO compared to the existing models.

Topology impacts important network performance metrics, including link utilization, throughput and latency, and is of central importance to network operators. However, due to the combinatorial nature of network topology, it is extremely difficult to obtain an optimal solution, especially since topology planning in networks also often comes with management-specific constraints. As a result, local optimization with hand-tuned heuristic methods from human experts is often adopted in practice. Yet, heuristic methods cannot cover the global topology design space while taking into account constraints, and cannot guarantee to find good solutions. In this paper, we propose a novel deep reinforcement learning (DRL) algorithm for graph searching, called DRL-GS, for network topology optimization. DRL-GS consists of three novel components, including a verifier to validate the correctness of a generated network topology, a graph neural network (GNN) to efficiently approximate topology rating, and a DRL agent to conduct a topology search. DRL-GS can efficiently search over relatively large topology space and output topology with satisfactory performance. We conduct a case study based on a real-world network scenario, and our experimental results demonstrate the superior performance of DRL-GS in terms of both efficiency and performance.

Graphs are widely used to represent the relations among entities. When one owns the complete data, an entire graph can be easily built, therefore performing analysis on the graph is straightforward. However, in many scenarios, it is impractical to centralize the data due to data privacy concerns. An organization or party only keeps a part of the whole graph data, i.e., graph data is isolated from different parties. Recently, Federated Learning (FL) has been proposed to solve the data isolation issue, mainly for Euclidean data. It is still a challenge to apply FL on graph data because graphs contain topological information which is notorious for its non-IID nature and is hard to partition. In this work, we propose a novel FL framework for graph data, FedCog, to efficiently handle coupled graphs that are a kind of distributed graph data, but widely exist in a variety of real-world applications such as mobile carriers’ communication networks and banks’ transaction networks. We theoretically prove the correctness and security of FedCog. Experimental results demonstrate that our method FedCog significantly outperforms traditional FL methods on graphs. Remarkably, our FedCog improves the accuracy of node classification tasks by up to 14.7%.

Sequential recommendation aims at predicting the next item that the user may be interested in given the historical interaction sequence. Typical neural models derive a single history embedding to represent the user's interests. Moving one step forward, recent studies point out that multiple sequence embeddings can help to better capture multi-faceted user interests. However, when ranking candidate items, these methods usually adopt the greedy inference strategy. This approach uses the best matching interest for each candidate item to calculate the ranking score, neglecting the target interest distribution in different contexts, which might lead to incompatibility with the current user intent. In this paper, we propose to enhance multi-interest recommendation by predicting the target user interest with a separate interest predictor and a specifically designed distillation loss. The proposed framework consists of two modules: the 1) multi-interest extractor to generate multiple embeddings regarding different user interests; and the 2) target-interest predictor to predict the interest distribution in the current context, which will be further utilized to dynamically aggregate multi-interest embeddings. To provide explicit supervision signals to the target-interest predictor, we devise a target-interest distillation loss that uses the similarity between the target item and multi-interest embeddings as the soft label of the target interest. This helps the target-interest predictor to accurately predict the user interest at the inference stage and enhances its generalization ability. Extensive experiments on three real-world datasets show the effectiveness and flexibility of the proposed framework.

With the widespread use of mobile devices in recent years, accurate prediction of base station traffic is vital for maintaining a good quality of mobile network services. In this paper, we propose an ensemble learning framework to predict the cellular traffic of base stations. Specifically, we introduce the Granger causality test to find the causal relationship in the base stations and model the spatial relationship between them simultaneously. We also employ a temporal convolutional network (TCN) to extract the sequential temporal features of base station traffic. Aiming at modelling the long-tail characteristics of the traffic distribution of base stations, we use the technique of redundant encoding to refine the prediction task to learn the base station traffic of different scales combined with ensemble learning. Extensive experimental results demonstrate that our method can predict base station traffic precisely and outperforms the best baseline by nearly 13% on average in terms of NMSE and NRMSE.

Spoken language understanding (SLU) treats automatic speech recognition (ASR) and natural language understanding (NLU) as a unified task and usually suffers from data scarcity.We exploit an ASR and NLU joint training method based on meta auxiliary learning to improve the performance of low-resource SLU task by only taking advantage of abundant manual transcriptions of speech data.One obvious advantage of such method is that it provides a flexible framework to implement a lowresource SLU training task without requiring access to any further semantic annotations.In particular, a NLU model is taken as label generation network to predict intent and slot tags from texts; a multi-task network trains ASR task and SLU task synchronously from speech; and the predictions of label generation network are delivered to the multi-task network as semantic targets.The efficiency of the proposed algorithm is demonstrated with experiments on the public CATSLU dataset, which produces more suitable ASR hypotheses for the downstream NLU task.

Understanding mobile data traffic and forecasting future traffic trend is beneficial to wireless carriers and service providers who need to perform resource allocation and energy saving management. However, predicting wireless traffic accurately at large-scale and fine-granularity is particularly challenging due to the following two factors: the spatial correlations between the network units (i.e., a cell tower or an access point) introduced by user arbitrary movements, and the time-evolving nature of user movements which frequently changes with time. In this paper, we use a time-evolving graph to formulate the time-evolving nature of user movements, and propose a model Graph-based Temporal Convolutional Network (GTCN) to predict the future traffic of each network unit in a wireless network. GTCN can bring significant benefits to two aspects. (1) GTCN can effectively learn intra- and inter-time spatial correlations between network units in a time-evolving graph through a node aggregation method. (2) GTCN can efficiently model the temporal dynamics of the mobile traffic trend from different network units through a temporal convolutional layer. Experimental results on two real-world datasets demonstrate the efficiency and efficacy of our method. Compared with state-of-the-art methods, the improvement of the prediction performance of our GTCN is 3.2 to 10.2 percent for different prediction horizons. GTCN also achieves 8.4× faster on prediction time.

Mobile network traffic forecasting is one of the key functions in daily network operation. A commercial mobile network is large, heterogeneous, complex and dynamic. These intrinsic features make mobile network traffic forecasting far from being solved even with recent advanced algorithms such as graph convolutional network-based prediction approaches and various attention mechanisms, which have been proved successful in vehicle traffic forecasting. In this paper, we cast the problem as a spatial-temporal sequence prediction task. We propose a novel deep learning network architecture, Adaptive Multi-receptive Field Spatial-Temporal Graph Convolutional Networks (AMF-STGCN), to model the traffic dynamics of mobile base stations. AMF-STGCN extends GCN by (1) jointly modeling the complex spatial-temporal dependencies in mobile networks, (2) applying attention mechanisms to capture various Receptive Fields of heterogeneous base stations, and (3) introducing an extra decoder based on a fully connected deep network to conquer the error propagation challenge with multi-step forecasting. Experiments on four real-world datasets from two different domains consistently show AMF-STGCN outperforms the state-of-the-art methods.

Kernel density estimation method is a powerful tool and is widely used in many important real-world applications such as anomaly detection and statistical learning. Unfortunately, current kernel methods suffer from high computational or space costs when dealing with large-scale, high-dimensional datasets, especially when the datasets of interest are given in a stream fashion. Although there are sketch methods designed for kernel density estimation over data streams, they still suffer from high computational costs. To address this problem, in this paper, we propose a novel Rotation Kernel. The Rotation Kernel is based on a Rotation Hash method and is much faster to compute. To achieve memory-efficient kernel density estimation over data streams, we design a method, RKD-Sketch, which compresses high dimensional data streams into a small array of integer counters. We conduct extensive experiments on both synthetic and real-world datasets, and experimental results demonstrate that our RKD-Sketch saves up to 216 times computational resources and up to 104 times space resources than state-of-the-arts. Furthermore, we apply our Rotation Kernel in active learning. Results show that our method achieves up to 256 times speedup and saves up to 13 times space to achieve the same accuracy as the baseline methods.

Submodular optimization lies at the core of many data mining and machine learning applications such as data summarization and subset selection. For data streams where elements arrive one at a time, streaming submodular optimization (SSO) algorithms are desired. Existing SSO solutions are mainly designed for insertion-only streams where elements in the stream all participate in the analysis, or sliding-window streams where only the most recent data participates in the analysis. SSO for insertion-only streams does not sufficiently emphasize recent data. SSO for sliding-window streams abruptly forgets all past data. In this work, we propose a new SSO problem, i.e., temporal biased streaming submodular optimization (TBSSO), which embraces the special settings of all previous studies. TBSSO leverages a temporal bias function to force each element in the stream to participate in the analysis with a probability decreasing over time and hence elements in the stream are forgotten gradually. We design novel streaming algorithms to solve the TBSSO problem with provable approximation guarantees. Experiments show that our algorithm can find high quality solutions and improve the efficiency to about one order of magnitude faster than the baseline method.

As an efficient tool for approximate similarity computation and search, Locality Sensitive Hashing (LSH) has been widely used in many research areas including databases, data mining, information retrieval, and machine learning. Classical LSH methods typically require to perform hundreds or even thousands of hashing operations when computing the LSH sketch for each input item (e.g., a set or a vector); however, this complexity is still too expensive and even impractical for applications requiring processing data in real-time. To address this issue, several fast methods such as OPH and BCWS have been proposed to efficiently compute the LSH sketches; however, these methods may generate many sketches with empty bins, which may introduce large errors for similarity estimation and also limit their usage for fast similarity search. To solve this issue, we propose a novel densification method, i.e., BiDens. Compared with existing densification methods, our BiDens is more efficient to fill a sketch's empty bins with values of its non-empty bins in either the forward or backward directions. Furthermore, it also densifies empty bins to satisfy the densification principle (i.e., the LSH property). Theoretical analysis and experimental results on similarity estimation, fast similarity search, and kernel linearization using real-world datasets demonstrate that our BiDens is up to 106 times faster than state-of-the-art methods while achieving the same or even better accuracy.