This master's thesis addresses the problem of analyzing smart home sensor data and predicting user activities based on event sequences. Due to the large volume of heterogeneous data, class imbalance, and the presence of noise, processing data from smart environments represents an important research challenge in the fields of human activity recognition and intelligent living systems.
As part of this work, a modular software framework for processing the CASAS/Kyoto dataset was developed. The system includes procedures for importing, validating, and normalizing sensor data, empirical sensor classification, binary value mapping, and event segmentation using temporal and spatio-temporal methods. Several machine learning classification models were implemented and evaluated for activity prediction, including a naive Bayes classifier as the primary baseline model, a Support Vector Machine (SVM), a Random Forest classifier, and a Hidden Markov Model (HMM). Laplace smoothing was applied during the training of the naive Bayes model to address rare and previously unseen events. The performance of the individual models was evaluated using the classification metrics precision, recall, F1-score, and AUC.
The evaluation results demonstrate that all considered models are capable of successfully recognizing user activities from sensor events. The differences between the individual algorithms were relatively small, making it difficult to draw reliable conclusions regarding the clear superiority of any single approach. The analysis of micro- and macro-averaged metrics confirms the significant class imbalance present in the dataset, which substantially affects classification performance. The results validate the usefulness of the proposed architecture for analyzing behavioral patterns in smart environments and provide a solid foundation for the further development of advanced machine learning methods and sequential activity modeling.
|