Comparing Long Short-Term Memory and Graph Convolutional Network Models for Human Activity Recognition using WISDM Dataset
| Author | Ebrahim, Majeed Mohamed Hasan | |
|---|---|---|
| metadata.dc.contributor.department | University of Bahrain, College of Science, Department of Postgraduate Programs | |
| metadata.dc.date.accessioned | 2026-09-20T15:44:52Z | |
| Publication date | 2023 | |
| Degree level | Master | |
| Abstract | Abstract: The development of technology in computer hardware has made observing and analyzing daily performed activities an easy task. Therefore, human activity recognition systems have benefited from the embedded sensors in smart devices in several fields, such as healthcare, security, and fitness. Furthermore, these sensors can generate sequenced data with temporal and spatial relationships. Therefore, Artificial Intelligence techniques, such as machine learning and deep learning, have been utilized for data processing and activity classification. While traditional machine learning methods have performed well, they require manual feature extraction from the raw data. Meanwhile, deep learning methods can extract the required features respecting spatial and temporal dependencies. Among those methods, The Long Short-Term Memory (LSTM) method is proven efficient in dealing with time-series data. In addition, the Graph Convolutional Network (GCN) has the ability to extract spatiotemporal relations using the graph data structure. However, as per author knowledge, the previous studies have yet to compare the performance of both methods, so this research aims to compare both methods using the WISDM dataset. The dataset contains data from accelerometers and gyroscope sensors embedded in smartphones and smartwatches for 18 performed activities. For each smart device, the data of two sensors have been combined to form two sub-datasets which are the preprocessed and prepared datasets for feeding models using the overlapped sliding windows. Hence, both methods have been applied to the resulting sub-datasets. As a result, for the smartphone sub-dataset, the GCN model outperformed the LSTM model by around 4%, scoring 94.4% for accuracy and F1-score. In addition, the GCN model performed better than some previous LSTM models considering a subset of activities. On the other hand, the smartwatch sub-dataset gave a slight advantage to the GCN model over the LSTM model by nearly 1%, as it scored 90.4% for accuracy and F1-score. However, the GCN method can be considered a rival to the LSTM method in such data type, which should be applied to other datasets in future works for generalization. | |
| metadata.dc.identifier.other | https://digitalrepository.uob.edu.bh/id/007dc6d2-03c8-4856-afd3-303f939e81ec | |
| metadata.dc.identifier.other | Drupal node 2382 | |
| metadata.dc.identifier.other | https://digitalrepository.uob.edu.bh/en/dar/comparing-long-short-term-memory-and-graph-convolutional-network-models-human-activity | |
| metadata.dc.identifier.uri | http://165.227.56.122/handle/123456789/6221 | |
| Language | en | |
| Link | https://digitalrepository.uob.edu.bh/en/_flysystem/fedora/2025-10/001106_comparing_long_short-term_memory_and_graph_convolutional_network_models_for_human_activity_recognition.pdf | |
| Subject | Long-term memory -- Dissertations | |
| Subject | Short-term memory -- Dissertations | |
| Subject | Memory -- Data processing -- Dissertations | |
| Title | Comparing Long Short-Term Memory and Graph Convolutional Network Models for Human Activity Recognition using WISDM Dataset | |
| Type | Master's Thesis | |
| College | College of Science | |
| Department | Postgraduate Programs |
Files
Original bundle
1 - 1 of 1
Loading...
- Name:
- node-2382-1-001106_comparing_long_short-term_memory_and_graph_convolutional_network_models_for_human_activity_recognition.pdf
- Size:
- 44.72 MB
- Format:
- Adobe Portable Document Format