Elevating SIoT Service Recommendations: A Context-Aware Framework

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Aug 22 2023

The concept of the Social Internet of Things (SIoT) allows interconnected smart devices to share data, presenting possibilities for tailored service recommendations. However, existing research frequently overlooks vital factors essential for precise SIoT recommendations. To bridge this gap, researchers recently tackled these issues in a paper submitted to the arXiv* server, in which they incorporated device relationships and contextual information into their recommendations.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.

Background

The rapid growth of interconnected devices and technologies has led to the emergence of the transformative Internet of Things (IoT). This paradigm involves objects embedded with sensors and software, enabling autonomous data exchange. A new paradigm, the SIoT, extends connectivity to encompass social interactions and human-centric services. SIoT transitions from a device-centric to a user-centric environment, allowing devices to form social connections. This empowerment enables devices to share information, collaborate, and interact socially, ultimately revolutionizing the process of service discovery and composition.

The diverse range of SIoT services underscores the need for effective recommendation systems. Although certain models attempt to address this need, they encounter challenges when dealing with the dynamic context of SIoT. Existing solutions such as the social correlation group-based recommender system and the time-aware smart object recommendation model fall short of considering diverse content, user preferences, and context. While a framework based on graphs for service recommendations incorporates social relationships, it lacks accuracy due to incomplete data.

Exploring SIoT service recommendations

The rise of the SIoT has introduced interconnected devices and data generation, leading to novel services. Service recommendation systems are essential for providing tailored user experiences. Existing research addresses service recommendations within the SIoT context, including models utilizing social relationships among device owners. However, these models lack consideration of user preferences and context. Another approach integrates time and social similarity, employing a latent probabilistic model to capture user preferences and social ties.

Nevertheless, this approach may struggle with incomplete usage data. Graph-based frameworks incorporate social ties and user preferences but face challenges related to scalability and handling new users or objects. While a recommender system considers social dynamics, it might overlook preferences and context. The integration of the social dimension in IoT enhances trust computing.

Traditional recommendation models encounter difficulties with the diverse data in the SIoT. Pursuing multi-modal service recommendation involves leveraging various modalities for accuracy, employing approaches ranging from graph-based to time-aware models. To bridge these gaps, this study proposes a context-aware service recommendation system that captures latent interactions to enhance SIoT service recommendations.

A context-aware framework

The proposed framework aims to enhance device-service recommendation by utilizing a context-aware representation learning model. For estimating rating scores of device-service pairs, the framework utilizes information from two sources: service reviews and device-service interactions. This framework comprises two feature learning components: engagement-based feature learning and review-based feature learning. The architecture begins with convolutional layers processing review collections and extracting contextual features. A selective layer then prioritizes relevant aspects, followed by an abstraction layer that extracts semantic features using mean-pooling convolutional neural networks.

Additionally, engagement-based feature learning employs separate latent vectors for devices and services. The combination process integrates context-aware latent features from both components dynamically using a dynamic weighting scheme based on the relative prediction scores of review-based and engagement-based components. Factorization machine (FM) and linear regression techniques are adopted to obtain rating scores using the extracted latent feature vector.

Evaluation methodology in SIoT service recommendations

Given the scarcity of publicly available SIoT datasets, the Amazon review dataset was utilized to evaluate the proposed service recommendation approach. Despite its lack of SIoT specificity, the dataset comprising diverse product reviews proved valuable for assessing recommendation methods. This dataset includes various user reviews and ratings, facilitating the evaluation of the proposed methods against existing approaches.

 A comprehensive evaluation methodology was introduced to assess the effectiveness of the proposed recommendation approach. Baseline metrics such as recall, precision, root mean squared error (RMSE), normalized discounted cumulative gain (NDCG), and mean absolute error (MAE) were employed. These metrics provide insights into recommendation accuracy, relevance, and ranking.

The proposed framework's performance was compared with graph-based service recommendation, matrix factorization, and other models. The study addressed three research questions: the impact of combining feature learning approaches, the influence of hyperparameters, and the effectiveness of the selective layer in feature learning. The new framework, combining review-based and engagement-based learning, consistently outperformed baseline methods regarding recall and precision. FM outperformed linear regression regarding MAE values for electronics, cell phones, and appliances. FM excels at handling high-dimensional data, capturing complex interactions, and outperforming linear regression's linear assumption. The significance of the selective layer was evident in identifying relevant semantic information for effective service recommendation in SIoT.

Conclusion

In summary, researchers employed the Amazon review dataset to evaluate the proposed service recommendation approach. Although this dataset lacks SIoT focus, it provides diverse reviews for evaluating recommendation methods.

The study addressed three key aspects: the impact of combining feature learning, the influence of hyperparameters, and the effectiveness of the selective layer. The proposed framework, which combines review-based and engagement-based learning, consistently outperformed baseline methods in terms of recall and precision. Notably, the FM approach demonstrated superiority. The importance of the selective layer in identifying relevant semantic information for effective SIoT service recommendations was evident.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.