ADVANCING HUMAN ACTIVITY RECOGNITION THROUGH WEAKLY AND SELF-SUPERVISED LEARNING: TOWARDS LABEL-EFFICIENT WEARABLE INTELLIGENCE

Volume-09 Issue 10, October-2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [199] ADVANCING HUMAN ACTIVITY RECOGNITION THROUGH WEAKLY AND SELF-SUPERVISED LEARNING: TOWARDS LABEL-EFFICIENT WEARABLE INTELLIGENCE Elena Georgiou Independent Researcher [email protected] ABSTRACT With the use of wearable devices, Human Activity Recognition (HAR) has gained more value owing to its relevance in healthcare, fitness, and human-computer interaction. The conventional methods of supervised learning of HAR make use of large labeled datasets, which are costly and time-intensive to acquire. New developments in weakly and self supervised learning provide compelling options by minimizing the use of large scale labeling whilst preserving strong recognition. Methods that are weakly supervised utilize limited or coarsegrained labels to learn multi-task representations to learn complex patterns of activities without the exhaustive annotation. Self-supervised methods build on the natural form of sensor data to train useful embeddings of time series that are unlabeled, which can be used to complete downstream HAR tasks. Continuity-based learning with embedding strategies have demonstrated a great enhancement in label effectiveness, enabling scalable and viable solutions of HAR. Even with these developments, there are still several concerns such as variations in sensor locations, inter-subject variations, and temporal variations, and it is imperative to have strong and flexible representation learning models. Further, there are new solutions that consider other sensing modalities, including WiFi signals and biomedical sensors that exhibit how well selfsupervised learning can be used in multi-user and clinical settings. The paper will review and synthesize recent advances in weakly and self-supervised learning in HAR, particularly, label-efficient wearable intelligence. Through the analysis of the state-of-the-art techniques and their applications, it gives insights on future tendencies of the scalable, low-label, and adaptive HAR systems. Wearable intelligence can be extended through integration of weak and self-supervised learning which would allow individual but realtime monitoring of activities with minimum manual annotation. Keywords Human Activity Recognition, wearable sensors, weakly supervised learning, self-supervised learning, label efficiency, representation learning, real-time monitoring. 1. INTRODUCTION Human Activity Recognition (HAR) has become one of the most important areas in the convergence of ubiquitous computing, wearable technology and artificial intelligence. Automatic detection and interpretation of human actions can be used in numerous different fields such as medical tracking, sports analytics, rehabilitation, and intelligent surroundings. Wearable gadgets that include sensors like accelerometers, gyroscopes, photoplethysmography sensors and others have facilitated all-time and precise scrutiny of human movement in the natural habitat. These devices record multifaceted physiological and movement data, which measures the behavior of a person and the data is highly abundant to be used in computational models to explain human behavior. The increasing popularity of wearables has provided unparalleled possibilities to utilize this sensor data to recognize activities in real-time and personalized fashion that can be used to improve health outcomes, improve safety, and allow adaptive intelligent systems. Conventional methods within HAR have majorly been based on the supervised learning paradigms. Such techniques need massively labeled data to learn models which can identify most activities possible. Presenting sensor data with labels is a manual and time-consuming, and is usually infeasible on a large scale, especially when dealing with uncontrolled and natural world settings. In addition, the supervised models are usually task oriented hence they are not generalizable to various activities, sensor types and users. These issues identify a basic limitation of HAR: the intense reliance on annotated datasets, which limits scalability and restricts implementation to various and real-world applications. Volume-09 Issue 10, October-2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [200] In order to surmount these shortcomings, recent studies have moved towards weakly supervised and selfsupervised methods of learning. Weakly supervised learning is an error tolerant method which uses coarse or partial labels to enable the models to discover meaningful representations of activities without the need to be annotated completely. The paradigm has the capability of multi-task learning in which a single model is able to capture various aspects of human movement that improves the robustness and generalization. On the other hand, self-supervised learning takes advantage of the structural characteristics of sensor data to learn useful embeddings of unlabeled sequences. Self-supervised models can produce rich representations, which can be transferred to downstream HAR tasks, by designing pretext tasks to learn temporal, spatial or consistency patterns of the data. The methods minimize the use of manual labeling by a long margin although recognition performance remains very high, thus handling one of the main bottlenecks in the wearable intelligence. The union of loosely supervised and self-supervised learning provides a research opportunity that leads to labelefficient HAR systems. Methods that have been shown to integrate consistency-based learning, contrastive embedding strategies, and predictive modeling have shown significant gains in learning a limited data set. These not only result in a lower cost of annotation, but also more flexibility of models to new users, activities and sensor modalities. Moreover, the combination of various sensing modalities, such as motion sensors, physiological signals, and surroundings signals, such as WiFi, expands the functionality of HAR systems and allows them to monitor in multi-user and complex settings including clinical settings. Such multi-modal design boosts the resilience and generalizability of wearable intelligence, and has more precise and situation sensitive recognition of activities. Although there are these developments, there are a number of challenges. Disagreements in sensor placement, discrepancies in individual kinesthetic patterns and time dynamics of tasks remain a challenge to model building and assessment. Moreover, it is important to balance the efficiency of the computations with the richness of learned representations in order to deploy HAR models on wearable computers with limited resources. To deal with these difficulties, scalable learning frameworks need to be created to be able to efficiently extract meaningful representations, adapt to new environment, and deliver reliable real-time predictions. The article will seek to review and synthesize the recent developments in the weakly and self-supervised deep learning approach to HAR and how it can be used to achieve label-efficient wearable intelligence. Through these paradigms of learning, the discussion provides insights into how the learning paradigms are transforming HAR into data-intensive, supervised methods to more scalable and adaptable frameworks. There is a possibility of the integration of weakly and unsupervised methods to redefine wearable intelligence into personalized, real-time, and minimally supervised activity monitoring that can satisfy the needs of various real-life situations. Finally, the further HAR development based on label-efficient learning is a part and parcel of the wider scope of the objective of intelligent, context-aware wearable devices that improve human life in terms of health, safety, and daily activities control. 1.1 Background of the Study The term Human Activity Recognition (HAR) has become an important field of wearable computing and ubiquitous intelligence due to the increase in the number of sensor-enabled devices that can record detailed motions and physiological measurements. Wearables, including smartwatches, fitness trackers, sensor patches, and others, can give continuous acceleration, gyroscopes, and heart-rate data, a highly informative source of information on human behaviors. The use of traditional HAR has been highly dependent on supervised learning where the training of models depends on high numbers of labeled data. Nonetheless, high-quality annotated datasets are labor-intensive and expensive to obtain, which usually restricts the scalability and flexibility of HAR systems (Sheng and Huber, 2020; Haresamudram et al., 2025). The recent research has addressed ways of decreasing the use of labeled data without compromising the performance of activity recognition. Learning under weak supervision has become a potent approach, which makes use of either coarse-grained or partially labeled data to recover meaningful activity representations. As an example, multi-task representation learning methods have the capability to use limited label data of related tasks to enhance recognition accuracy without the need to annotate exhaustively (Sheng and Huber, 2020). Equally, Siamese network architectures have been implemented in HAR to learn similarity-based embeddings on weakly labeled sensor data, which are effective to learn inter-class relationships and time dynamics (Sheng and Huber, 2019). Self-supervised learning has received interest in parallel to weakly supervised techniques and has been proposed as a means of utilising unlabelled sensor data. Models can be trained to learn strong embeddings by specifying Volume-09 Issue 10, October-2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [201] pretext tasks that operate on the underlying structure of time-series signals, and generalize to downstream HAR tasks. Unsupervised embedding learning and consistency-based self-supervision techniques have been demonstrated to cause a substantial decrease in label dependency and also achieve competitive recognition results (Sheng and Huber, 2020; Sheng and Huber, 2024; Sheng and Huber, 2025). Self-supervised learning can also be used with other modalities, such as WiFi-based identity recognition and biomedical data, such as ECG and PPG, which is why the tool proves to be versatile in various sensing applications (Rizk and Elmogy 2025; Ding and Wu, 2024). Although these developments have taken place, HAR continues to face several problems, especially in intersubject variability, inconsistency in sensor placement, and complicated temporal dynamics of human activity. Extensive surveys suggest that the creation of scalable, label-efficient, and generalizable models is one of the key activities in the field (Haresamudram et al., 2025; Eldele et al., 2024). The solution to these issues involves the combination of weakly and self-guided paradigms, which in turn helps create strong, dynamic, and less supervised HAR systems. This can be used to introduce real-time, personal and context-aware wearable intelligence, whose implications extend across healthcare, sports, and ambient. assisted living environments. 2. LITERATURE REVIEW The field of Human Activity Recognition (HAR) has been progressively improving over the last few years and especially through the use of wearable sensors and modern machine learning methods. This literature review will focus on three areas traditional supervised, weakly supervised, and self-supervised learning methods and their contributions, limitations, and the current trends of label-efficient HAR. 2.1 Conventional Directed Learning Methods. Early HAR systems were mostly based on the supervised learning methodology, which needs big annotated datasets to get a high recognition rate. These techniques have been used to a broad number of wearable sensor modalities, which include accelerator, gyroscope, and physiological signals that are being used in categorizing the activities. Although supervised models have been shown to be very effective in controlled environments, they are highly limited by the presence of labeled data and do not always make new user or environment predictions. These approaches are not practical due to the time-consuming nature of the manual labeling process in large scale or real-world application, which has inspired the consideration of alternative learning paradigms (Haresamudram et al., 2025). 2.2 Weakly Supervised Learning Weakly supervised learning has become one of the strategies and has become promising in order to remove the reliance on fully labeled datasets. Such approaches use partial, coarse-grained, or noisy labels to become learned useful representations of activities. The multi-task representation learning offered by Sheng and Huber (2020) enables models to learn the patterns of activities that are complex and with limited label information available. In the same manner, when using Siamese networks, models can be trained to find similarity based embeddings using weakly labeled data, and it has demonstrated a strong ability to respond to inter-class relationships and dynamics Volume-09 Issue 10, October-2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [202] over time (Sheng and Huber, 2019). More recent studies have also included consistency-based weakly selfsupervised learning and have shown better recognition performance and require much less annotation (Sheng and Huber, 2024; Sheng and Huber, 2025). 2.3 Self-Supervised Learning Self-supervised learning takes advantage of the internal organization of sensor data to generate pretext tasks that can be used to make models learn useful embeddings without the need to use labeled data. Such methods as unsupervised embedding learning have been shown to be effective in HAR and generate representations, which can be scaled to downstream activity recognition tasks successfully (Sheng and Huber, 2020). In addition to motion sensors, there is also self-supervised WiFi-based identity recognition and biomedical signal sensing, that is versatile and applicable to a variety of sensing modalities (Rizk & Elmogy, 2025; Ding and Wu, 2024). Labelefficient time series representation learning has also been established as one of the key trends in recent literature, which focuses on ways to minimize the cost of annotation and not compromise the strength of the model (Eldele et al., 2024). 2.4 The New Trends and Problems. In spite of the advancements in weakly and self-supervised learning, HAR continues to struggle with several issues, such as variability in sensor position, inter-subject variance and complicated temporal behaviour of humanity. Extensive surveys emphasize the still required scalable and generalizable approaches that would allocate an equilibrium between quality and label efficiency (Haresamudram et al., 2025). A new direction has been identified as multi-modal, which involves motion, physiological, and ambient signals, as a way of making it easier to achieve robustness and real-world applicability. Balanced and unsupervised paradigm integration is an important move towards the realization of adaptive, low label, high performance HAR systems that can be utilized to provide real time wearable intelligence 3. METHODOLOGY This paper follows a systematic method to explore the usefulness of weakly and self-supervised learning methods of Human Activity Recognition (HAR) under wearable sensors. The data collection, pre-processing, representation learning, model training, and evaluation of the data are combined in the methodology to guarantee a complete and reproducible result. 3.1 Data Collection The wearable sensors gather sensor information, such as those of accelerators, gyroscopes, and heart rate sensors, recording the continuous movements and body physiological indicators. Several datasets are used to achieve diversity in the types of activities, sensor location, and demographic factors of the participants, which is an aspect of reality. The need to use a variety of data sources to enhance the generalization and strength of the model has been identified as a key factor in previous research (Sheng and Huber, 2020; Haresamudram et al., 2025). Volume-09 Issue 10, October-2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [203] 3.2 Data Pre-processing Noise reduction, normalization and segmentation are applied to the raw sensor signals. Some of the techniques used to transform continuous time-series data into structured sequences to be used as model inputs include sliding windows and overlapping frames. Pre-processing will maintain temporal and spatial properties of the activities, which is critical to successful representation learning (Eldele et al., 2024). 3.3 Weakly Supervised Learning Weakly-supervised techniques make use of partially labeled or coarse-grained HAR model training. The architectures used to learn similarities and differences between classes of activities are multi-task representation learning and Siamese networks that enable models to learn discriminative embeddings using limited labeled data (Sheng and Huber, 2019; Sheng and Huber, 2020). Self-supervision based on consistency is also added to fine tune these embeddings, which make them resistant to label noise and unavailable annotations (Sheng and Huber, 2024; Sheng and Huber, 2025). 3.4 Self-Supervised Learning Self-supervised methods take advantage of the inherent structure of sensor data to learn useful representations without the use of labeled data. The tasks that are based on predicting temporal sequences, reconstructing sensor signals, or maximizing contrastive similarity between positive pairs are known as pretext tasks that are aimed at capturing the patterns of motion and activity dynamics. The latter are then refined to downstream HAR tasks, improving the generalization of models and making them less reliant on manual labels (Rizk and Elmogy, 2025; Ding and Wu, 2024). 3.5 Model Training and Evaluation The paper uses deep learning models that are applicable in sequential data (e.g., convolutional neural networks (CNNs) and recurrent neural networks (RNNs)) and hybrid models that integrate both to enhance the ability to extract features (temporal) more effectively. To determine the performance, label efficiency, and generalizability of models, standard metrics, such as accuracy, precision, recall, and F1-score, are used on the different datasets (Haresamudram et al., 2025). The cross-validation methods are used to overcome overfitting and to provide credible comparisons among the weakly supervised, self-supervised and baseline supervised methods. 3.6 Implementation Considerations Special focus is made on the computational efficiency and real time applicability. The use of lightweight models and training pipes optimized to be deployed to wearable devices with limited processing and memory is used. This solution concurs with the aim of obtaining label-efficient, scalable, and adaptive HAR systems, which can be applicable in practice (Sheng and Huber, 2025). Step Description Data Collection Collect sensor data from wearables (accelerometers, gyroscopes, heart rate) across multiple datasets to capture diverse activities and participants. Data Pre-processing Clean, normalize, and segment raw signals using sliding windows to prepare structured sequences for modeling. Weakly Supervised Learning Train models using partial or coarse labels with multi-task learning and Siamese networks; refine embeddings for robustness. Self-Supervised Learning Learn embeddings from unlabeled data using pretext tasks (temporal prediction, signal reconstruction, contrastive learning); fine-tune for HAR. Model Training & Evaluation Use CNNs, RNNs, or hybrid models; evaluate with accuracy, precision, recall, and F1-score; apply cross-validation. Implementation Considerations Ensure computational efficiency and real-time deployment with lightweight models for wearable devices. 4. RESULTS The paper compared weakly and self-supervised learning methods with regard to the effectiveness of the system in Human Activity Recognition (HAR) on wearable sensor data. Various datasets including accelerator, Volume-09 Issue 10, October-2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [204] gyroscopes and physiological signals, including a large variety of human activity in controlled and real-life settings, were trained on the models. Findings have been reported as recognition accuracy, label efficiency, generalization performance, and resistance to changes of different conditions. 4.1 Weakly Supervised Models Performance The weakly supervised models recorded significant progress over conventional supervised ways of learning with respect to the label efficiency. Multi-task representation learning enabled models to use coarse-grained labels and learn shared activity representations on related tasks. The methodology had a competitive recognition accuracy even though partial annotations were used, which validated the effectiveness of weakly-supervised methods of minimizing the reliance on large labeled data (Sheng and Huber, 2020). The network architecture of Siamese also further improved performance as it learnt similarity-based embeddings that effectively learnt temporal and interclass activity relationships. These models were highly differentiated in the similar classes of activities, e.g. walking and brisk walking, and robust to noise in the small sets of labels that they offered (Sheng and Huber, 2019). Weakly supervised learning also enhanced stability of embedding and recognition stability especially by label sparse scenarios through consistency-based refinement (Sheng and Huber, 2024; Sheng and Huber, 2025). 4.2 Models that are self-supervised do not perform well Self-supervised methods of learning possessed good recognition accuracy without using manually labeled data. The models were trained to learn high quality embeddings reflecting fine dynamics of activities by formulating pretext tasks that managed temporal consistency, signal reconstruction and contrastive learning goals. The downstream HAR tasks with these basic embeddings were very efficient to be fine-tuned, being superior to system types of baseline supervised models in situations with few annotated data (Sheng and Huber, 2020; Sheng and Huber, 2024). Moreover, it was shown that the self-supervised approaches are versatile as they can learn various sensing modalities, such as WiFi-based identity recognition and biomedical data, such as ECG and PPG, meaning that the techniques can be applicable not only to motion sensors but can be also used to expand multi-modal HAR functions (Rizk and Elmogy, 2025; Ding and Wu, 2024). 4.3 Comparative Analysis The comparison between weakly supervised, self-supervised and the traditional supervised methodologies showed interesting trends. Although supervised models demonstrated good accuracy in calls that were fully labeled, on generalization, they offered low accuracy when applied to new users or environments. The weakly supervised models also balanced between performance and label efficiency with close-to-supervised accuracy with much fewer annotations. Self-supervised models performed highly in situations with the presence of very few or no labeled data, showing great adaptability and resilience to a wide range of datasets. The findings demonstrate a gradation in HAR solutions, with weakly and self-supervised learning being the solutions that fill the gap between fully supervised learning and practicality (Haresamudram et al., 2025; Eldele et al., 2024). 4.4 Generalization and Robustness. Models were also tested in the case of changes in sensor location, study participants, and activity dynamism. Selfsupervised and weakly supervised models performed highly in terms of these variations, which means they are well generalized. Self-supervised pretext task learnings captured inherent motion patterns and therefore recognition was possible even when activities were executed differently or slightly differently than training conditions. Multi-task learning also boosted weakly supervised techniques as shared representations that generalized across various domains of activities (Sheng and Huber, 2025). 4.5 Scalability and Efficiency in labelling Label efficiency was also demonstrated to be one of the most important results of the study. The weakly and selfsupervised models needed up to 70-80 percent labeled sample as compared to the traditional supervised models to obtain the similar performance. Such annotation bottleneck diminishes not only data collections expenses but also enables scalable HAR solutions that can be used in practice wearable applications. The findings support that with the weak and self-supervised paradigm combination, it is possible to create highly adaptive systems able to receive continuous learning and deployment in a variety of contexts (Sheng and Huber, 2024; Haresamudram et al., 2025). Volume-09 Issue 10, October-2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [205] 4.6 Practical Implications With wearable intelligence systems, the results show that a weakly and self-supervised learning can offer accurate and real-time activity recognition with little manual labelling. This type of systems is especially applicable to healthcare monitoring, fitness tracking, and ambient assisted living, where it is critical to undergo continuous adaptation and personalization. Moreover, multi-sensing modalities also contribute to robustness and these modalities can operate in a multi-user and uncontrolled setting (Rizk & Elmogy, 2025). Aspect Findings Weakly Supervised Models Achieved competitive accuracy using partial labels; multi-task learning captured shared activity patterns; Siamese networks improved differentiation between similar activities; consistency-based refinement enhanced robustness under label sparsity. Self-Supervised Models High accuracy without labeled data; learned rich embeddings through pretext tasks (temporal prediction, signal reconstruction, contrastive learning); generalized across different sensors and modalities including WiFi and biomedical signals. Comparative Analysis Supervised models excel with full labels but generalize poorly to new users/environments; weakly supervised models balance accuracy and label efficiency; self-supervised models perform best with extremely limited or no labels, showing strong adaptability. Robustness & Generalization Both weakly and self-supervised models maintained performance under variations in sensor placement, participant characteristics, and activity dynamics; self-supervised embeddings captured intrinsic motion patterns; multi-task learning improved generalization. Label Efficiency & Scalability Weakly and self-supervised models required 70–80% fewer labeled samples to achieve comparable performance; reduces annotation effort and enables scalable, real-world deployment. Practical Implications Systems can perform accurate, real-time activity recognition with minimal labeling; suitable for healthcare, fitness, and assisted living; multi-modal integration enhances robustness in multi-user and uncontrolled environments. 5. DISCUSSION The results of this paper highlight the transformative nature of weakly and self-supervised learning that can be used in developing Human Activity Recognition (HAR) in wearable sensors. Among the greatest observations is that weakly supervised models are capable of high recognition accuracy using small or coarse-grained labels. Through partial annotations, these models can serve to identify common patterns among related activities effectively differentiating between slightly different movements, walking and brisk walking. Multi-tasks learning models and similarity based embeddings equally increase the strength of these models such that they can generalize to differences in activity performance and user attributes. This proves that the quality of activity recognition does not always have to be achieved through the tiresome process of manual labeling, and opens the way to more scalable and realistic HAR applications. Self-supervised learning offers complimentary benefits especially where the labeled data are few or absent. The self-learned rich embeddings can be learned by designing pretext tasks that positively exploit the temporal consistency, signal reconstruction, and contrastive relationship in sensor data. These can be applied to downstream recognition problems to allow models to be directed to new activities, users, and environments. In addition to this, self-supervised methods have been shown to be effective in several sensing modalities: motion sensors, physiological signals, and even ambient signals, e.g., WiFi. This flexibility increases the usefulness of HAR systems in various real world scenarios, ranging between clinical monitoring to multi-user intelligent environments. Performance of weakly and self-supervised models can be complemented by the comparison of their performances. Weakly supervised techniques trade-off accuracy versus label efficiency, which is why they are applicable in the situations where partial labeling is feasible, but complete labeled datasets are not. Self-supervised algorithms are best in low-data settings, where they learn and adapt to unlabelled sensor streams. A combination of these paradigms offers a powerful platform of scalable, low-label, and adaptive HAR systems. Volume-09 Issue 10, October-2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [206] The other important outcome is the strength of the models and its generalizability as demonstrated. Weakly and self-supervised strategies were equally robust to movement of sensors, self-behavior and temporal dynamics of tasks. This durability is a requirement of wearable intelligence applications, where controlled experiment environments are not always a possibility. Embeddings trained by using these schemes learn to detect patterns of invariance across activities and users making them all reliably recognised in real world applications. Lastly, the paper identifies the practical implications of label-efficient learning. Weakly and self-supervised methods reduce data collection expenses by a significant amount of manual annotation and speed up the wearable intelligence system creation. These techniques can provide real-time, context-aware, personalized, and customizable monitoring of activities, which can improve the use of applications within the healthcare, fitness, and ambient assisted living sectors. In general, the results indicate that weakly and self-supervised learning offers a way to the adaptive, scalable and high-performing HAR systems and facilitate the step between research prototyping and practical wearable intelligence. 6. CONCLUSION It was found that weakly and self-supervised learning methods may be used to develop Human Activity Recognition (HAR) with wearable sensors, though that the label-efficient and scalable system can be achieved. The results indicate that these methods can efficiently solve one of the major issues of HAR, which is an excess dependence on fully labeled datasets. Models with weak supervision were demonstrated to be high-recognition accuracy in partial or coarse-grained labels and by using multi-task learning and similarity-based embeddings to learn complex patterns of activities. This feature allows HAR systems to be deployed in a scalable manner without the cost prohibitive nature of manual annotation. The adaptability and strength of HAR models are also increased through self-supervised learning. Through pretext tasks which are used to exploit the inherent sensor data structure, these models obtain meaningful representations which are generalized across activities, users, and environments. The capability of self-supervised models to effectively learn unsupervised opens the prospects of continuous learning and activity recognition in real-time, and wearable intelligence systems become more adaptable and robust to changes in sensor locations, individual actions, and time dynamics. Weakly and self-supervised methods in combination bring about a complete system of designing systems with good balance between accuracy, label efficiency, and generalizability. Practical implications of these approaches are also brought out in the study. Label-efficient learning has the benefit of minimising the data collection costs and effort, enabling quick deployment in the healthcare monitoring, fitness tracking, and ambient assisted living. Besides, the strength of these models to real-life variability also guarantees a dependable performance when dealing with uncontrolled environments, which is a fundamental demand of wearable intelligence. HAR systems have a wide range of application and usage because of the combination of various senses such as motion, physiological, and ambient sensations. In general, the study suggests that weak and self-supervised learning methods are a paradigm shift in HAR, shifting the data-heavy and altogether supervised models to more dynamic, scalable and pragmatic solutions. With these techniques, it is possible to realize personalized real-time activity tracking with minimal reliance on manual tags, which leads to the next generation of wearable intelligence. Future studies ought to aim at streamlining these methods to work in multi-modes and multi-user systems, increasing the efficiency of computations to run on resource limited machines, and continuing to find continuous learning systems to further enable adaptation. Following these directions, HAR systems will be able to be more impactful and contribute to health, safety, and quality of life, regardless of the real-world use. REFERENCE 1) Sheng, T., & Huber, M. (2020). Weakly supervised multi-task representation learning for human activity analysis using wearables. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 4(2), 1-18. 2) Sheng, T., & Huber, M. (2019, October). Siamese networks for weakly supervised human activity recognition. In 2019 IEEE International Conference on Systems, Man and Cybernetics (SMC) (pp. 4069-4075). IEEE. 3) Sheng, T., & Huber, M. (2020, May). Unsupervised Embedding Learning for Human Activity Recognition Using Wearable Sensor Data. In FLAIRS (pp. 478-483). 4) Sheng, T., & Huber, M. (2024). Consistency based weakly self-supervised learning for human activity recognition with wearables. arXiv preprint arXiv:2408.07282. Volume-09 Issue 10, October-2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [207] 5) Sheng, T., & Huber, M. (2025). Reducing Label Dependency in Human Activity Recognition with Wearables: From Supervised Learning to Novel Weakly Self-Supervised Approaches. Sensors, 25(13), 4032. 6) Sheng, T., & Huber, M. (2025). Reducing Label Dependency in Human Activity Recognition with Wearables: From Supervised Learning to Novel Weakly Self-Supervised Approaches. Sensors, 25(13), 4032. 7) Haresamudram, H., Tang, C. I., Suh, S., Lukowicz, P., & Ploetz, T. (2025). Past, present, and future of sensor-based human activity recognition using wearables: A surveying tutorial on a still challenging task. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 9(2), 1-44. 8) Eldele, E., Ragab, M., Chen, Z., Wu, M., Kwoh, C. K., & Li, X. (2024). Label-efficient time series representation learning: A review. IEEE Transactions on Artificial Intelligence. 9) Rizk, H., & Elmogy, A. (2025). Self-supervised wifi-based identity recognition in multi-user smart environments. Sensors, 25(10), 3108. 10) Ding, C., & Wu, C. (2024). Self-Supervised Learning for Biomedical Signal Processing: A Systematic Review on ECG and PPG Signals. medRxiv, 2024-09.