Lei Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Lei Zhang has a total of 152 co-authors. Here are the top co-authors by number of shared publications:
- Hao Wu
School of Information Science and Engineering, Yunnan University, Kunming, China
54 shared publications - Aiguo Song
Southeast University
51 shared publications - Wenbo Huang
Southeast University, Nanjing, China
18 shared publications - Dongzhou Cheng
School of Cyber Science and Engineering, Southeast University, Naning, Jiang Su, China
17 shared publications - Jun He
School of Electronic and Information Engineering, Nanjing University of Information Science and Technology, Nanjing, China
17 shared publications Hao Wu
School of Information Science and Engineering, Yunnan University, Kunming, China
11 shared publications
Latest Publications
- Open Access
- 13 November 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Information Systems
- Vol. 44  (8) , 1-33
FAESRec: Frequency Adaptive Encoder for Sequential Recommendation
Sequential recommender systems predominantly focus on time-domain modeling, overlooking rich seasonal and periodic structures revealed in the frequency domain. While recent frequency-aware methods leverage Fourier transforms, their reliance on fixed filters introduces two critical limitations: (i) inadequate sequence encoding due to indiscriminate spectral suppression, causing information degradation, and (ii) semantic-degrading contrastive augmentations that compromise critical frequency components essential for behavioral semantics. To address these challenges, we propose FAESRec, a Frequency Adaptive Encoder with three synergistic innovations: First, a Frequency Adaptive Mixture-of-Experts (FAMoE) layer decomposes sequences into non-overlapping spectral bands with end-to-end learned boundaries, enabling dynamic weighting of heterogeneous frequency drivers. Second, a Filter-Enhanced module (FAFe) employs learnable global/local filters with adaptive thresholds to denoise spectral components while preserving energy-concentrated bands. The filtered spectral features are projected back to the time domain and encoded by a Mamba block. Third, a Frequency Semantic-aware Contrastive Learner (FSaCL) synthesizes semantically coherent views through probabilistic frequency masking, theoretically guaranteeing mutual information conservation between original and augmented sequences. Extensive experiments on six benchmarks from Amazon, Movielens, and LastFM demonstrate FAESRec’s consistent superiority over 13 SOTA baselines. It achieves significant improvements of 2.74–67.69% HR@1 and 3.95–20.38% NDCG@5, particularly under extreme sparsity. Theoretical analysis proves that FSaCL preserves minimal sufficient semantics, establishing a new paradigm for frequency-aware sequential recommendation. Code is available at: https://github.com/0-1user/FAESRec .
- 13 November 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Information Systems
- Vol. 44  (8) , 1-33
Open AccessFAESRec: Frequency Adaptive Encoder for Sequential Recommendation
Sequential recommender systems predominantly focus on time-domain modeling, overlooking rich seasonal and periodic structures revealed in the frequency domain. While recent frequency-aware methods leverage Fourier transforms, their reliance on fixed filters introduces two critical limitations: (i) inadequate sequence encoding due to indiscriminate spectral suppression, causing information degradation, and (ii) semantic-degrading contrastive augmentations that compromise critical frequency components essential for behavioral semantics. To address these challenges, we propose FAESRec, a Frequency Adaptive Encoder with three synergistic innovations: First, a Frequency Adaptive Mixture-of-Experts (FAMoE) layer decomposes sequences into non-overlapping spectral bands with end-to-end learned boundaries, enabling dynamic weighting of heterogeneous frequency drivers. Second, a Filter-Enhanced module (FAFe) employs learnable global/local filters with adaptive thresholds to denoise spectral components while preserving energy-concentrated bands. The filtered spectral features are projected back to the time domain and encoded by a Mamba block. Third, a Frequency Semantic-aware Contrastive Learner (FSaCL) synthesizes semantically coherent views through probabilistic frequency masking, theoretically guaranteeing mutual information conservation between original and augmented sequences. Extensive experiments on six benchmarks from Amazon, Movielens, and LastFM demonstrate FAESRec’s consistent superiority over 13 SOTA baselines. It achieves significant improvements of 2.74–67.69% HR@1 and 3.95–20.38% NDCG@5, particularly under extreme sparsity. Theoretical analysis proves that FSaCL preserves minimal sufficient semantics, establishing a new paradigm for frequency-aware sequential recommendation. Code is available at: https://github.com/0-1user/FAESRec .
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13  (20) , 47603-47620
Unsupervised Sensor-Based Human Activity Recognition via Contrastive Learning of Representation
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13  (20) , 47603-47620
Unsupervised Sensor-Based Human Activity Recognition via Contrastive Learning of Representation
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13  (13) , 28866-28882
Rep-MMB: Bridging Mobile CNN and Transformer for Sensor-Based Human Activity Recognition
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13  (13) , 28866-28882
Rep-MMB: Bridging Mobile CNN and Transformer for Sensor-Based Human Activity Recognition
- Top Cited Papers
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 25  (7) , 11227-11243
Sensor-Prompt Tuning: Aligning Time Series Foundational Models With Motion Sensors for Few-Shot Activity Recognition
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 25  (7) , 11227-11243
Top Cited PapersSensor-Prompt Tuning: Aligning Time Series Foundational Models With Motion Sensors for Few-Shot Activity Recognition
- Open Access
- 15 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
- Vol. 10  (2) , 1-33
Domain Generalization for Sensor-Based Activity Recognition via Decomposed Feature Representation and Semantic Augmentation
Domain generalization (DG) has recently offered a promising solution to tackle distribution-shift problem in sensor-based human activity recognition (HAR) scenario. Particularly, the sensor data augmentation has gained a lot of attention, which aims to generate more diverse representations to learn model invariance across different domains. However, most existing methods seek to directly perform augmentation in the sensor input space, which result in a limited variety of augmented features. Intuitively, performing augmentation in feature space would be more versatile and diverse, which has been seldom considered in sensor-based cross-domain activity recognition. In this paper, we propose a new feature space augmentation method by decomposing the obtained features into four main components, i.e., category-invariant domain-invariant, category-invariant domain-specific, category-specific domain-invariant, and category-specific domain-specific, which enable us to increase sensor sample diversity while emphasizing invariance learning to enhance domain generalization for activity recognition. Extensive experiments on four public HAR benchmarks including DSADS, USC-HAD, PAMAP2 and UCI-HAR demonstrate that our proposed feature space augmentation approach is able to achieve state-of-the-art performance under Cross-person, Cross-dataset, Cross-position and One-person-to-another settings. Detailed ablation and visualizations analyses verify its effectiveness, efficiency, and universality under cross-domain activity recognition scenario.
- 15 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
- Vol. 10  (2) , 1-33
Open AccessDomain Generalization for Sensor-Based Activity Recognition via Decomposed Feature Representation and Semantic Augmentation
Domain generalization (DG) has recently offered a promising solution to tackle distribution-shift problem in sensor-based human activity recognition (HAR) scenario. Particularly, the sensor data augmentation has gained a lot of attention, which aims to generate more diverse representations to learn model invariance across different domains. However, most existing methods seek to directly perform augmentation in the sensor input space, which result in a limited variety of augmented features. Intuitively, performing augmentation in feature space would be more versatile and diverse, which has been seldom considered in sensor-based cross-domain activity recognition. In this paper, we propose a new feature space augmentation method by decomposing the obtained features into four main components, i.e., category-invariant domain-invariant, category-invariant domain-specific, category-specific domain-invariant, and category-specific domain-specific, which enable us to increase sensor sample diversity while emphasizing invariance learning to enhance domain generalization for activity recognition. Extensive experiments on four public HAR benchmarks including DSADS, USC-HAD, PAMAP2 and UCI-HAR demonstrate that our proposed feature space augmentation approach is able to achieve state-of-the-art performance under Cross-person, Cross-dataset, Cross-position and One-person-to-another settings. Detailed ablation and visualizations analyses verify its effectiveness, efficiency, and universality under cross-domain activity recognition scenario.
- 8 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. PP  (15209210) , 1-12
Training-Free Zero-Shot Temporal Action Detection with Vision-Language Models
Existing zero-shot temporal action detection (ZSTAD) methods predominantly use fully supervised or unsupervised strategies to recognize unseen activities. However, these training-based methods are susceptible to domain shifts and entail high computational costs. Unlike previous works, we present FreeZAD, an training-free method for ZSTAD, which leverages image-pretrained vision-language models (VLMs) to directly classify and localize unseen activities within untrimmed videos. We mitigate the need for explicit temporal modeling and reliance on pseudo-label quality by designing the Logarithmic decay weighted Outer-Inner-Contrastive Score (LogOIC) and frequency-based actionness calibration. Furthermore, we introduce a test-time adaptation (TTA) strategy using Prototype-Centric Sampling (PCS) to expand FreeZAD, enabling VLMs to adapt more effectively for ZSTAD. Extensive experiments on the THUMOS14 and ActivityNet-1.3 datasets demonstrate that our training-free method outperforms state-of-the-art unsupervised methods while requiring only 1/13 of the runtime, and the adaptation-enhanced extension further narrows the gap with fully supervised approaches. The code are available at: https://github.com/Chaolei98/FreeZAD.
- 8 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. PP  (15209210) , 1-12
Training-Free Zero-Shot Temporal Action Detection with Vision-Language Models
Existing zero-shot temporal action detection (ZSTAD) methods predominantly use fully supervised or unsupervised strategies to recognize unseen activities. However, these training-based methods are susceptible to domain shifts and entail high computational costs. Unlike previous works, we present FreeZAD, an training-free method for ZSTAD, which leverages image-pretrained vision-language models (VLMs) to directly classify and localize unseen activities within untrimmed videos. We mitigate the need for explicit temporal modeling and reliance on pseudo-label quality by designing the Logarithmic decay weighted Outer-Inner-Contrastive Score (LogOIC) and frequency-based actionness calibration. Furthermore, we introduce a test-time adaptation (TTA) strategy using Prototype-Centric Sampling (PCS) to expand FreeZAD, enabling VLMs to adapt more effectively for ZSTAD. Extensive experiments on the THUMOS14 and ActivityNet-1.3 datasets demonstrate that our training-free method outperforms state-of-the-art unsupervised methods while requiring only 1/13 of the runtime, and the adaptation-enhanced extension further narrows the gap with fully supervised approaches. The code are available at: https://github.com/Chaolei98/FreeZAD.
- 19 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. PP  (15) , 1
TSA-Former: Linear Transformer with Taylor Series Attention for Sensor-Based Human Activity Recognition
Transformer models have demonstrated superior capability in capturing long-range temporal dependencies crucial for Sensor-Based Human Activity Recognition (HAR). However, the quadratic computational complexity inherent to the Softmax-Attention mechanism significantly impedes their deployment on resource-constrained wearable devices and real-time streaming tasks. To address this, we propose a novel Linear Transformer with Taylor Series Attention specifically tailored for the HAR domain, named TSA-Former. It leverages the first-order Taylor expansion to approximate the Softmax-Attention and utilizes the norm-preserving mapping to approximate the high-order non-linear information, resulting in a linear computational complexity. In addition, TSA-Former integrates a multi-branch architecture featuring multi-scale patch embedding, which enables the model to dynamically capture multi-scale temporal features while minimizing overhead. Experimental results across four public HAR benchmarks, namely UniMiB-SHAR, UCI-HAR, WISDM, and OPPORTUNITY, demonstrate that TSA-Former achieves state-of-the-art (SOTA) accuracy and efficiency, outperforming conventional Transformers and existing linear-attention models. Deployment experiments conducted on the Raspberry Pi 5 platform further validate the model’s superior low-latency and minimal power consumption profile, confirming its robust suitability for real-world embedded HAR applications. Code will be released.
- 19 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. PP  (15) , 1
TSA-Former: Linear Transformer with Taylor Series Attention for Sensor-Based Human Activity Recognition
Transformer models have demonstrated superior capability in capturing long-range temporal dependencies crucial for Sensor-Based Human Activity Recognition (HAR). However, the quadratic computational complexity inherent to the Softmax-Attention mechanism significantly impedes their deployment on resource-constrained wearable devices and real-time streaming tasks. To address this, we propose a novel Linear Transformer with Taylor Series Attention specifically tailored for the HAR domain, named TSA-Former. It leverages the first-order Taylor expansion to approximate the Softmax-Attention and utilizes the norm-preserving mapping to approximate the high-order non-linear information, resulting in a linear computational complexity. In addition, TSA-Former integrates a multi-branch architecture featuring multi-scale patch embedding, which enables the model to dynamically capture multi-scale temporal features while minimizing overhead. Experimental results across four public HAR benchmarks, namely UniMiB-SHAR, UCI-HAR, WISDM, and OPPORTUNITY, demonstrate that TSA-Former achieves state-of-the-art (SOTA) accuracy and efficiency, outperforming conventional Transformers and existing linear-attention models. Deployment experiments conducted on the Raspberry Pi 5 platform further validate the model’s superior low-latency and minimal power consumption profile, confirming its robust suitability for real-world embedded HAR applications. Code will be released.
- 5 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13  (7) , 14923-14939
Beyond 1 × 1 Convolutions: A Dynamic Select-and-Fuse Channel Sampling Strategy for On-Device Human Activity Recognition
The proliferation of low-cost, portable sensors has made wearable human activity recognition (HAR) a cornerstone for real-time health monitoring and behavior analysis. However, deploying accurate yet lightweight deep learning models on resource-constrained wearable devices poses a significant challenge for on-device activity recognition. While channel pruning is a common solution to accelerate deep convolutional neural networks (CNNs), existing works often require specialized implementations or pretrained models, which potentially degrade performance by simply removing an entire channel, limiting their ability to handle complex multimodal sensor inputs. Moreover, lightweight CNN design, particularly the heavy use of $1\times 1$ convolution layers for channel squeezing, remains inefficient for sensor-based HAR, which consumes resources without expanding the receptive field due to their pointwise nature. To address these issues, we propose a novel dynamic channel sampling module, select-and-fuse (SaF), specifically designed for sensor-based HAR. SaF divides channels into subsets and performs a dynamic, input-dependent selection from them, with the picking decision being made per time step based on the input sensor signal activations, allowing for fine-grained feature adaptation to multimodal sensor signals. While integrated into compact backbones, SaF significantly reduces model size and inference latency while maintaining high accuracy. Extensive evaluations on public UCI-HAR, OPPORTUNITY, WISDM, and UniMiB-SHAR benchmarks confirm a favorable performance–cost tradeoff. Crucially, we measure actual inference latency on a Raspberry Pi, proving its practicality for resource-constrained HAR applications. The code will be released.
- 5 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13  (7) , 14923-14939
Beyond 1 × 1 Convolutions: A Dynamic Select-and-Fuse Channel Sampling Strategy for On-Device Human Activity Recognition
The proliferation of low-cost, portable sensors has made wearable human activity recognition (HAR) a cornerstone for real-time health monitoring and behavior analysis. However, deploying accurate yet lightweight deep learning models on resource-constrained wearable devices poses a significant challenge for on-device activity recognition. While channel pruning is a common solution to accelerate deep convolutional neural networks (CNNs), existing works often require specialized implementations or pretrained models, which potentially degrade performance by simply removing an entire channel, limiting their ability to handle complex multimodal sensor inputs. Moreover, lightweight CNN design, particularly the heavy use of $1\times 1$ convolution layers for channel squeezing, remains inefficient for sensor-based HAR, which consumes resources without expanding the receptive field due to their pointwise nature. To address these issues, we propose a novel dynamic channel sampling module, select-and-fuse (SaF), specifically designed for sensor-based HAR. SaF divides channels into subsets and performs a dynamic, input-dependent selection from them, with the picking decision being made per time step based on the input sensor signal activations, allowing for fine-grained feature adaptation to multimodal sensor signals. While integrated into compact backbones, SaF significantly reduces model size and inference latency while maintaining high accuracy. Extensive evaluations on public UCI-HAR, OPPORTUNITY, WISDM, and UniMiB-SHAR benchmarks confirm a favorable performance–cost tradeoff. Crucially, we measure actual inference latency on a Raspberry Pi, proving its practicality for resource-constrained HAR applications. The code will be released.
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. PP, 1-14
Cross-Modal Feature Adapter for Few-Shot Human Activity Recognition
Recent years have witnessed outstanding success of deep learning in sensor-based human activity recognition (HAR), spanning a wide range of real-world applications like healthcare management, fitness tracking, and fall detection. However, sensor data annotation scarcity still remains a main challenge unresolved, which requires human annotators to take a long-term and tedious observation to segment and timestamp sensor samples meticulously, hampering the wide use of deep learning models, especially in few-shot HAR scenario. To handle such issue, this paper introduces a cross-modal data augmentation, by exploiting activity label text as key words to search for activity-related images to construct an augmented dataset. On this basis, a new cross-modal feature adapter is designed, which fine-tunes a pre-trained CLIP image encoder to effectively align image-sensor pairs. Through a learnable residual ratio, it may adaptively blend the old knowledge inherited from the original zero-shot CLIP with the new knowledge adapted from few-shot training samples, which makes training converges faster whilst forming a streamlined time series sensor encoder. Extensive experiments and ablation studies are performed on three public HAR benchmarks. The experimental results demonstrate that the proposed method outperforms existing state-of-the-art HAR baselines under all few-shot scenarios. A practical on-device inference latency is provided.
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. PP, 1-14
Cross-Modal Feature Adapter for Few-Shot Human Activity Recognition
Recent years have witnessed outstanding success of deep learning in sensor-based human activity recognition (HAR), spanning a wide range of real-world applications like healthcare management, fitness tracking, and fall detection. However, sensor data annotation scarcity still remains a main challenge unresolved, which requires human annotators to take a long-term and tedious observation to segment and timestamp sensor samples meticulously, hampering the wide use of deep learning models, especially in few-shot HAR scenario. To handle such issue, this paper introduces a cross-modal data augmentation, by exploiting activity label text as key words to search for activity-related images to construct an augmented dataset. On this basis, a new cross-modal feature adapter is designed, which fine-tunes a pre-trained CLIP image encoder to effectively align image-sensor pairs. Through a learnable residual ratio, it may adaptively blend the old knowledge inherited from the original zero-shot CLIP with the new knowledge adapted from few-shot training samples, which makes training converges faster whilst forming a streamlined time series sensor encoder. Extensive experiments and ablation studies are performed on three public HAR benchmarks. The experimental results demonstrate that the proposed method outperforms existing state-of-the-art HAR baselines under all few-shot scenarios. A practical on-device inference latency is provided.