UNIFIED REPRESENTATION LEARNING FRAMEWORKS FOR WEARABLEBASED HUMAN ACTIVITY RECOGNITION: FROM UNSUPERVISED EMBEDDING TO MULTI-TASK ADAPTATION

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/) [208] UNIFIED REPRESENTATION LEARNING FRAMEWORKS FOR WEARABLEBASED HUMAN ACTIVITY RECOGNITION: FROM UNSUPERVISED EMBEDDING TO MULTI-TASK ADAPTATION Elena Georgiou Independent Researcher [email protected] ABSTRACT Wearable-based Human Activity Recognition (HAR) has become an important element in ubiquitous computing, healthcare monitoring and intelligent human-computer interaction. Nevertheless, current HAR systems face challenges in terms of generality across different sensor types, users, and activity areas and are mostly because of the lack of labelled data and the over reliance on task-specific learning paradigms. In order to overcome the limitations, recent scholarship has been directed towards unified representation learning frameworks that make use of unsupervised, weakly supervised, and multitask adaptation strategies that learn generalisable feature embeddings from multimodal wearable sensor data. This manuscript forms a detailed study on these unified frameworks that describes how these frameworks have evolved from unsupervised embedding learning (Sheng & Huber, 2020) and weakly supervised Siamese networks (Sheng & Huber, 2019) to state-of-the-art consistencybased weakly self-supervised approaches (Sheng & Huber, 2024; Sheng & Huber, 2025). It explores how the representation learning methods of contrastive learning, self-supervised pre-training and multitask architectures (Samyoun et al., 2022) can improve cross-domain generalisation and reduce the dependence on large amounts of annotation. Moreover, the review brings together the results of new emergent models such as multimodal and transformer-based models such as Harformer (Wang, Mo, & Zhu, 2025) and CrossHAR (Hong et al., 2024) pointing to the promise of hierarchical and channel-separated embeddings for domain-invariant representations. The synthesis of the confluence of unsupervised embedding, weakly supervised adaptation and multitask learning has been identified as the cornerstone of next-generation unified frameworks for scalable, adaptive HAR in a variety of contexts. The article concludes with an overview of outstanding challenges related to modality imbalance, temporal dynamics alignment, and interpretability, as well as some recommendations for future research in the area of universal representation learning for wearable-based HAR systems. Keywords Human Activity Recognition (HAR), Wearable Sensors, Unsupervised Learning, Multi-Task Adaptation, Representation Learning, Weakly Supervised Learning, Self-Supervised Frameworks, Transfer Learning 1. INTRODUCTION Human activity recognition (HAR) using wearable sensor technology has become a central topic in the fields of ubiquitous computing, artificial intelligence and mobile health systems. By utilizing wearable devices such as smartwatch, fitness tracker, inertial measurement unit (IMU), etc. Wearable-based HAR provides continuous context-aware monitoring of humans' motion, posture, and physiological parameters. Such capabilities are essential in a host of applications such as personalized healthcare, monitoring the elderly, rehabilitation and sports analytics (Ni et al., 2024). However, current HAR systems are facing deep challenges with respect to scalability, adaptivity and robustness, especially in case models are deployed among sensor setups or different user groups. However, the large number of labeled datasets is limiting its applicability, as the manual annotation of sensor signals is time consuming, expensive, and often impossible in a real-world environment (Sheng & Huber, 2025). Conventional supervised learning in HAR is based on labeled activity segments and hand-designed feature engineering, which generally cannot capture the nonlinear and context-aware dynamics of human movement (Cai et al., 2025). Further, the domain gaps caused by the differences in device places, sampling rates, and noise levels prevent cross-user and cross-device generalization (Hong et al., 2024). These shortcomings have motivated the evolution of unified representation learning frameworks that aim at uniting supervised, unsupervised and multitask paradigms into one unifying framework. A class of general discriminative latent representation learning 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/) [209] methods that learn latent representations which are not only able to be used for activity classification, but are also transferable among datasets and robust to sensor differences (Sheng & Huber, 2020; Samyoun et al., 2022). 1.1 Supervised Paradigms to Representation Learning The path of HAR research has been characterized by a progressive movement from supervised pipelines through traditional representations to more and more autonomous representation learning pipelines. While supervised deep learning techniques have shown amazing performance on benchmark data sets, they are bound by the fact that they require large amounts of labeled data. In view of this, unsupervised and self-supervised learning schemes have been developed to learn effective representations from unlabeled data by exploiting the structure of intrinsic signals as well as the contextual dependencies. Sheng and Huber (2020) have pioneered an unsupervised learning framework for wearable based HAR using reconstruction based objectives and context prediction to capture temporal dependencies and transitions of motion without any human annotation. This study showed that unsupervised embeddings could be used as transferable task substrates with a substantially reduced amount of labelled data needed for subsequent task fine-tuning. Similarly, Siamese and contrastive architectures Sheng & Huber (2019) used similarity constraints between activity segments to impose inter-class separability and intra-class compactness, thus paving the way for selfsupervised learning in HAR and highlighting the usefulness of pretext tasks, e.g. temporal order prediction and sensor signal augmentation, for robust representation learning. Based on this, Sheng and Huber (2024, 2025) proposed weakly self-supervised and consistency-based methods to further alleviate the dependence of the labels. Their models involved pseudolabelling, temporal consistency regularization and cross modality alignment that refines latent feature spaces in various sensors and activities. These hybrid frameworks proved that combining unsupervised pretraining with limited weak supervision was capable of achieving performance on par with, if not better than, fully supervised models. The evolution of wearable-based HAR demonstrates a key transition from annotated data-dependent methods to unsupervised representation learning techniques that can automatically extract meaningful relations from raw sensor data. 1.2 Development of Multi-Task and Domain Generalizable Frameworks While unsupervised and weakly supervised learning solve data sparsity, multi-task representation learning (MTRL) takes this work further by allowing models to learn representations that are shared across related tasks. Rather than separating activity recognition from other motion-related goals, MTRL combines several learning objectives - namely, activity classification, intensity estimation and affect recognition - in a single architecture (Samyoun et al., 2022). The M3Sense framework is a typical example of this concept, with the idea of using affect-agnostic multitask learning with multimodal wearable sensors. This architecture enables the sharing of features among the tasks and it improves generalization and also increases the interpretability of the learned embeddings. By aligning multiple goals in a single model, MTRL avoids overfitting and encourages robustness in general activity settings. Recent research has also added cross dataset generalization and transformer based architectures to obtain domain independent performance. Hong et al. (2024) CrossHAR: Hierarchical Self-Supervised Pretraining through MultiLevel Attention Mechanisms Hong, et al., in Proceedings of the 35th International Conference on World Wide Web, WWW 2024, pp. 1-3. Likewise, Wang, Mo, and Zhu (2025) also developed Harformer, which is a channelseparated transformer model that targets unsupervised HAR embedding learning. Harformer captures temporal and inter-channel dependency in sensor streams, which results in transferability of heterogeneous data sources. These innovations are examples of the convergence of unsupervised learning of embedding, self-supervised pretraining, and multitask adaptation in unified frameworks. Nevertheless, there are still many open research issues. Modality imbalance, where some sensors exert an inexplicably larger influence on the learning process of a model, is also a major obstacle (Ni et al., 2024). In multimodal signals, temporal synchronization is usually inaccurate, which makes joint representation learning complicated (Cai et al., 2025). In addition, latent embeddings are still limited in their interpretability which hinders transparency and interpretability of HAR applications in health-critical applications. Solving these problems requires novel methods that incorporate temporal alignment strategies, cross-model fusion strategies and explainable AI strategies within cohesive representation strategies. 2. LITERATURE REVIEW 2.1 Overview of Human Activity Recognition using Wearable 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/) [210] Wearable based Human Activity Recognition (HAR) has quickly become a multidisciplinary field combining machine learning, signal processing and pervasive sensing technologies. Its aim is to automatically recognize physical and contextual human activities through analysis of data collected from body-worn sensors such as accelerometers, gyroscopes, magnetometers, and a number of physiological devices. Investigations in HAR were mostly based on handcrafted features and traditional classifiers (e.g. decision trees and support vector machines). However, these approaches were limited in scalability and adaptability since they required manual feature engineering and were complicated by sensor heterogeneity and inter-user variability (Ni et al., 2024). The development of deep learning models led to convolutional and recurrent neural networks that took over the area in the following order: They automatically learn discriminative features directly from raw sensor streams. Despite their success, deep supervised models still rely on large labeled datasets, and it's been shown that they still have limited transferability when transferred to different users and devices (Cai et al., 2025). Consequently, representation learning, the paradigm that aims at finding latent embeddings that are able to generalise across activities, domains, and modalities, has become increasingly popular among researchers. Representation learning allows for the models to learn semantic signal representations in an unsupervised way, which benefits and improves the generalisation and reduces annotation cost (Sheng & Huber, 2020; Hong et al., 2024). 2.2 Unsupervised and Weakly Supervised Representation Learning. In particular, recent advances in unsupervised and weakly supervised learning methods have played a major role in reshaping the HAR paradigms. Sheng and Huber (2020) proposed an unsupervised learning of embedding using reconstruction and context prediction, to extract meaningful representations from raw sensor data. By this approach, they showed that unsupervised pre-training of a model can help to transfer learned embeddings to supervised downstream classification tasks and hence reduce annotation needs. Extending this work, Sheng and Huber (2019) used Siamese network architectures in the weakly supervised HAR, based on pairwise similarity constraints for discriminating between similar and dissimilar activity patterns. This approach supported greater inter-class separability in the domain where the comprehensive annotation was not feasible. In 2024, Sheng and Huber further developed these ideas by introducing a weakly-supervised consistency-based learning framework which focuses on temporal stability and cross-sensor alignment to derive consistent representations under different sensing conditions (Sheng & Huber, 2024). Based on this, Sheng and Huber (2025) proposed a joint weakly self-supervised model to alleviate the dependence of the labels by combining pseudolabeling and consistency regularization. Taken together, these contributions demonstrate a paradigm shift towards label-efficient and generalisable HAR systems. In addition to these developments, multimodal methods have achieved popularity as well. Samyoun, Islam, Iqbal, & Stankovic (2022) introduced M3Sense, which is a multitask representation learning model for multimodal wearable data (e.g., motion and physiological signals). This results in a framework that is independent of affective states and means that the model is dependent only on cues related to physical activity. Their results showed that multi-task representation learning significantly enhances performance and generalization especially in the complex activity setting. 2.3 Towards Generalizable and Multi-Task HAR Architectures Aiming to achieve domain generalization in HAR, multi-task learning and self-supervised learning architectures have been employed in this research. In their detailed survey, Cai, Guo, Salim, and Hong (2025) pointed out that the future HAR systems need to cross the domain shifts by learning universal and transferable embeddings. To this end, a few models have been developed that use cross-dataset training and hierarchical self-supervised pretraining. Hong et al. CrossHAR, a generalisation framework across datasets through hierarchical temporal attention and self-supervised objectives. The model obtained state-of-the-art results in the cross-domain evaluation which demonstrates the potential of hierarchical embeddings for universal HAR. Similarly, Harformer has been introduced by Wang, Mo and Zhu (2025) to create an unsupervised transformer based model to capture channel separated embeddings for multimodal sensor fusion. Harformer applies attention mechanisms to learn dependencies across the channels and temporal relationships, which results in better generalization across the heterogeneous wearable devices. Along with weakly supervised and multitask paradigms, these models are a convergence towards consolidated representation learning paradigms that can support robust and scalable HAR. Table 1 below gives an overview of major studies in the evolution of unified frameworks for wearable-based HAR based on learning strategies, modalities, and key contributions. 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/) [211] Table 1. Summary of Key Studies on Representation Learning Frameworks for Wearable-Based HAR Study Learning Type Core Methodology Key Contribution Sheng & Huber (2019) Weakly supervised Siamese network for pairwise similarity learning Improved inter-class separability without full labels Sheng & Huber (2020) Unsupervised Contextual prediction and reconstruction-based embedding Learned transferable representations from unlabeled data Sheng & Huber (2024) Weakly selfsupervised Consistency regularization and temporal alignment Enhanced robustness under heterogeneous sensors Sheng & Huber (2025) Weakly selfsupervised Pseudo-labeling and hybrid objective optimization Reduced dependence on annotated datasets Samyoun et al. (2022) Multi-task learning M3Sense multimodal representation learning Joint learning across multiple sensor tasks Hong et al. (2024) Self-supervised pretraining Hierarchical attention and crossdataset generalization Domain-invariant representation learning Wang et al. (2025) Unsupervised transformer Channel-separated embedding (Harformer) Improved temporal and inter-channel representation Cai et al. (2025) Survey Generalizable HAR frameworks Synthesized trends and open challenges in HAR research Ni et al. (2024) Survey Multimodal sensor-based HAR Provided comprehensive review of multimodal HAR methods 2.4 Trends of Literature Summarized An accelerated path to unified and generalizable representation learning for wearable-based human activity recognition (HAR) is revealed from the literature explored. Although early models focused on unsupervised embedding learning and weak supervision, recent work has integrated self-supervised and multi-task architectures to achieve domain-invariant performance (Sheng & Huber, 2024; Wang et al., 2025). The coming together of these methodological paradigms points to a shift in focus from the traditional single task recognition systems to a holistic system that can adapt to users, environments and modalities. However, to achieve such an approach, there are still challenges in terms of the trade-off between performance and interpretability. Most deep representation learning frameworks are black box systems that do not provide much insight into physiological or biomechanical correlates for learned embeddings. Moreover, the complications associated with the synchronization of the sensor data, missing data and the imbalance between modalities would be introduced when multimodal data integration occur (Ni et al., 2024). Addressing these issues will be critical in developing reliable, transparent and ethically deployable HAR systems that will be able to support large-scale applications in real-world settings of health monitoring and ubiquitous computing (Cai et al., 2025). 3. METHODOLOGY AND CONCEPTUAL FRAMEWORK 3.1 Research Design Overview This study adopts a conceptual-analytical research design that synthesizes and integrates recent advancements in unified representation learning frameworks for wearable-based Human Activity Recognition (HAR). The methodology follows a structured process involving (1) systematic review of foundational and emerging approaches, (2) comparative analysis of representation learning paradigms, and (3) formulation of a conceptual unified framework integrating unsupervised, weakly supervised, and multi-task adaptation mechanisms. The proposed conceptual model builds upon existing works by Sheng and Huber (2019, 2020, 2024, 2025), Samyoun et al. (2022), and other contemporary researchers who emphasize autonomous and scalable HAR architectures. These studies collectively highlight the evolution from independent learning paradigms toward 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/) [212] unified representation frameworks, where multiple learning objectives interact within a shared latent space. The overarching goal is to achieve robustness, adaptability, and domain generalization across heterogeneous wearable sensors (Hong et al., 2024; Wang et al., 2025). The methodology integrates three primary components: 1. Unsupervised Embedding Learning – for autonomous feature extraction from unlabeled sensor data. 2. Weakly Supervised and Consistency-Based Adaptation – for leveraging limited annotations and enforcing representation stability. 3. Multi-Task Learning (MTL) Integration – for optimizing shared embeddings across multiple correlated HAR tasks. Each component contributes to the formation of a unified representation learning pipeline, enabling knowledge transfer and reducing the dependency on large-scale labeled datasets. 3.2 Data Sources and Feature Representation The conceptual framework assumes the use of multimodal wearable sensor data, typically collected from accelerometers, gyroscopes, and physiological sensors such as heart rate or electrodermal activity (EDA). According to Ni et al. (2024), multimodal sensor fusion enhances the semantic richness of HAR data and enables more robust feature representations. The data are represented as time-series signals X={x1,x2,...,xT}X = \{x_1, x_2, ..., x_T\}X={x1 ,x2 ,...,xT }, where each xt ∈ Rdx_t \in \mathbb{R}^dxt ∈Rd corresponds to the multichannel sensor readings at time ttt. Following the works of Sheng and Huber (2020), the raw sensor data undergo a preprocessing pipeline that includes normalization, sliding window segmentation, and augmentation (e.g., jittering or rotation) to simulate real-world variability. These operations ensure that the learned embeddings are invariant to sensor noise and orientation differences. Feature extraction occurs through deep encoders such as convolutional neural networks (CNNs) or transformer-based encoders, which transform the raw sequences into latent embeddings ztz_tzt . The representation learning objective can be expressed as: where Lunsup\mathcal{L}_{unsup}Lunsup represents the unsupervised embedding objective (e.g., contrastive or reconstruction loss), Lweak\mathcal{L}_{weak}Lweak captures weakly supervised or consistency-based constraints, and Lmtl\mathcal{L}_{mtl}Lmtl denotes the multi-task learning loss. The hyperparameters λ1\lambda_1λ1 and λ2\lambda_2λ2 balance the contributions of each learning component (Sheng & Huber, 2024). 3.3 Conceptual Framework for Unified Representation Learning The proposed conceptual framework (Figure 1) integrates unsupervised embedding, weak supervision, and multitask adaptation into a cohesive architecture for wearable-based HAR. 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/) [213] Figure 1. Conceptual Framework for Unified Representation Learning in Wearable-Based HAR 3.4 Algorithmic Workflow The unified learning framework follows a three-phase workflow: 1. Unsupervised Pretraining Phase: a. Employs autoencoding, contrastive learning, or transformer-based self-attention mechanisms to learn unsupervised embeddings (Wang et al., 2025). b. The embeddings capture both temporal and inter-sensor relationships without requiring labels. 2. Weakly Supervised Adaptation Phase: a. Introduces limited label information to align embeddings using consistency-based or pseudolabeling strategies (Sheng & Huber, 2024). b. Employs cross-sensor alignment and temporal consistency to reduce modality bias. 3. Multi-Task Fine-Tuning Phase: a. Jointly optimizes multiple related tasks such as recognition, segmentation, and intensity estimation (Samyoun et al., 2022). b. Promotes feature sharing and enhances the interpretability of the learned embeddings. Table 2. Key Components of the Unified Representation Learning Framework Framework Component Learning Objective Representative Techniques Core References Unsupervised Embedding Learn general features from unlabeled data Autoencoders, Contrastive Learning, Transformers Sheng & Huber (2020); Wang et al. (2025) Weakly Supervised Adaptation Refine embeddings with limited labels Pseudo-labeling, Temporal Consistency Sheng & Huber (2024, 2025) Multi-Task Learning Share knowledge across related HAR tasks Joint optimization, Shared latent spaces Samyoun et al. (2022); Hong et al. (2024) 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/) [214] Cross-Domain Generalization Adapt to new devices/users Hierarchical Self-Supervision, CrossHAR Framework Hong et al. (2024); Cai et al. (2025) 3.5 Implementation Issues The adoption of the integrated framework requires consideration of some pivotal factors. Primarily, the data diversity is indispensable; training datasets should cover a broad range of users, activities, and sensor configurations in order to enable as robust as possible generalisation (Ni et al., 2024). Secondly, the choice of the model architecture is a critical factor for determining the representational capacity. Transformer based architectures, such as Harformer (Wang et al., 2025), demonstrate the potential to capture long-range dependencies, while CNN based encoders allow for computational efficiency, which proves to be a particularly salient attribute for resource constrained devices. Thirdly, in order to increase training efficiency, curriculum learning or progressive task scheduling can be used in which an unsupervised pre-training phase can be seamlessly transitioned to a multi-task training phase (Sheng & Huber, 2025). Finally, the evaluation metrics should include both the recognition quality and the representation quality, and the evaluation criteria such as cross-domain accuracy, separability of embeddings and task transferability (Cai et al., 2025) should be employed. 3.6 Summary of methodology approach In summary, the proposed methodology integrates state-of-the-art representation learning paradigms in a unified architecture for wearable-based human activity recognition (HAR). By making use of unsupervised pre-training, weakly supervised refinement, and multi-task adaptation, the framework provides a scalable and data-efficient foundation for next generation HAR systems. This integrative approach is in line with current research trends towards domain-invariant and label-efficient learning, which eventually leads to universal HAR models that can be learnt continuously across different users, devices, and environmental conditions (Sheng & Huber, 2024; Hong et al., 2024; Wang et al., 2025). RESULTS 4.1 Description of Analytical Findings The theoretical integration of existing frameworks shows that the joint representation learning has the significant performance and generalization enhancement in wearable-based HAR. Compared with typical supervised models, unified models have better cross-domain robustness, lower dependence on labelled data, and better adaptability to unobserved sensor modalities (Sheng & Huber, 2025; Hong et al., 2024). Multiple benchmarks have been used to demonstrate that models combining unsupervised embedding, weak supervision, and multi-task learning are all uniformly better than models based only on any one of these paradigms. The use of unsupervised pretraining training (Sheng & Huber, 2020) produces basic embeddings that contain temporal and contextual information without being explicitly labeled. Further adaptation by consistency-based adaptation (Sheng & Huber, 2024) results in better accuracy of alignment across different wearable devices. Moreover, the multi-task adaptation (Samyoun et al., 2022) can easily share the parameters and help the model to learn the task-invariant representations which are most important to cross-user generalization and scalability. In addition to this, the findings collectively indicate that such unifying models are not only label-efficient but also architecturally generalizable in that they can be implemented in various types of models - like CNNs, LSTMs, and transformers - without any performance loss (Wang et al., 2025). 4.2 Comparative Analysis of Performance Comparative performance studies of representative HAR models show that when consistent unified learning principles are applied, the recognition accuracy and transferability of the HAR models are consistently improved. Studies in weakly self-supervised paradigms or multi-task paradigms report performance gains of 8-15% over traditional supervised baselines for cross-domain evaluations (Sheng & Huber, 2025; Hong et al., 2024). Likewise, Harformer (Wang et al., 2025) which is a channel-separated unsupervised embedding achieves remarkable advancements in temporal modeling as well as robustness to noisy data. Fig. 2 (below) summarises these results conceptually by depicting the interaction between model complexity, label dependence and generalisation performance. As the supervised models are moved to unified models, label dependency decreases and cross-domain generalization significantly improves. 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/) [215] Figure 2. Conceptual Performance Comparison Across Learning Paradigms in HAR This conceptual curve highlights that as frameworks integrate unsupervised and multi-task objectives, their capacity for domain-invariant feature extraction improves, confirming the superiority of unified learning strategies in wearable-based HAR (Sheng & Huber, 2024; Samyoun et al., 2022). 4.3 Implications of Results The synthesized results have a number of important implications for future HAR development: Data Efficiency: Unified frameworks allow for substantial reductions in labeled data needs through the use of unsupervised pretraining and pseudo-labeling (Sheng & Huber, 2025). This technology has a significant application, especially in health care and ambient intelligence applications, where the availability of annotated data is limited. Cross-Domain Generalization: Using hierarchical self-supervised and consistency-based learning, models are able to produce stable performance across users, devices, and sensor modalities (Hong et al., 2024). This generalization ability makes it practical to use in the real world. Task Adaptability: Multi-task architectures are flexible as HAR models simultaneously support various recognition tasks, such as activity classification, movement intensity estimation, and behavioral context inference (Samyoun et al., 2022). Model Scalability: Transformer based architectures such as Harformer (Wang et al., 2025) exhibit scalability and ability to better manage long term dependencies making them good candidates for the future unified HAR frameworks. In conclusion, the analyzed results strongly support the progress towards monolithic representation learning as the basis for the next generation of wearable intelligence. This approach is in line with the more general goal of 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/) [216] achieving label-efficient, adaptable, and generalizable HAR systems that can operate well in diverse and dynamically changing human environments. 5. DISCUSSION 5.1 Fusion of Unsupervised Embedding and Weakly Supervised Learning to Robust HAR. The confluence of unsupervised and weakly supervised representation learning is a crucial transition in wearable based Human Activity Recognition (HAR). Although traditional supervised methods give high performance in well-controlled environments, they are fundamentally limited by their dependence on annotated data and low ability to generalize to unexpected conditions. More recent researches have proved that unsupervised embedding methods can be used to recover latent temporal and spatial structures from sensor streams without manual annotation (Sheng & Huber, 2020). These embeddings are used as low-level representations of cross-modal correlations, thus allowing the models to generalize not only out of domain of a particular dataset. Weakly supervised techniques are further used to add to this capacity by using pseudo labeling and pairwise contrastive objectives, guiding the embedding space to be semantically consistent between similar activity patterns (Sheng & Huber, 2019). Such methods reduce the reliance on explicit labels and maintain high levels of discriminative performance - a property that is of great importance in areas where the collection of annotated data is either very expensive or very impractical. These principles have recently been extended to consistency based self-supervised learning proposed by Sheng and Huber (2024, 2025). In this kind of paradigm, the model is trained on the requirement of having invariant representations during temporal shifts, sensor specific noise, or user variability. This consistency objective causes the network to focus on stable features that are relevant for the activity, and reject variations that are considered to be a nuisance. Empirical data collected from weakly self-supervised HAR systems is in agreement that such techniques provide significant improvements in cross-user and cross-device generalisation. As a result, unsupervised and weakly supervised paradigms of representation that are combined into unified representation paradigms represent an important step toward autonomous and adaptive activity recognition systems. Council of Europe, 2004). 5.2 Multitask adaptation as a Bridge to Generalised Representation A key feature of unified frameworks is the fact that they allow for multi-task adaptation, i.e. joint optimisation across correlated recognition tasks. Rather than learning representations for one single classification task, models can be trained for a set of auxiliary tasks, such as gesture segmentation, motion intensity estimation and temporal prediction, to promote the emergence of shared representations that can be effectively generalised (Samyoun et al., 2022). This approach reduces the risk of overfitting to specific activity labels and encourages a good parameter sharing between tasks. It is a natural fit between multi-task learning and representation unification, since both of them consider the learning of features that extend semantically over different contexts. As an illustration, when applied to multimodal wearable data, the M3Sense framework shows that multi-task objectives can significantly improve HAR performance. Through the simultaneous training of activity and context prediction, this framework obtains better affect-agnostic generalization, which demonstrates the worthiness of latent disentanglement. In addition, the recent CrossHAR (Hong et al., 2024) model demonstrates that hierarchical pre-training when combined with multi-task objectives can enable cross-dataset transfer learning achieving strong transferability across sensor modalities and environments. The embedding representations learned in these systems are stacking representations from low-ranking motion patterns to high-ranking semantic behaviors, which are invariant across domains. By incorporating multi-task learning into unified frameworks researchers are starting to close the gap between representation learning which is generally applicable and the actual implementation in heterogeneous and diverse user environments. However, issues remain in the design of multiple tasks to reduce interference while ensuring that common features are used for all objectives. Modular architecture design, dynamic loss balancing and adaptive weighting techniques have been proposed to eliminate the negative transfer effects (Ni et al., 2024). Despite these challenges, multi-task adaptation is still a promising direction for scalable contextual human activity recognition (HAR) solutions that are applicable to dynamic and heterogeneous human activity scenarios. 5.3 Moving towards Universal and Interpretable Unified Models While the unified representation learning has produced good empirical results, there are still questions left regarding the interpretability, the transparency and the universality. In the case of wearable based HAR, interpretability is not only a theoretical matter of preference but an operational necessity - especially in the health, eldercare and clinical monitoring domains where the system's decisions have a direct impact on human well-being.