נושא הפרוייקט
מספר פרוייקט
מחלקה
שמות סטודנטים
אימייל
שמות מנחים
העברת ידע מרובת משתנים בסדרות זמן
Multivariable Transfer Learning in Timeseries
תקציר בעיברית
תקציר באנגלית
At the heart of this research is the quest to optimize deep learning performance in the context of limited multivariate time series data. Born from the difficulty faced in many real-world applications of deep learning, the scarcity of data often hinders the efficiency and robustness of predictive models. This project employs a transfer learning approach, combining research and experimentation to investigate whether a model, pretrained on synthetic data and then fine-tuned on the original dataset, can surpass the performance of a model trained only on the original data. We employed 14 datasets from the UAE repository, a recognized benchmark for academic research in time series classification. Two types of deep learning architectures – CNN and RNN - were selected and trained with a variety of hyperparameters, serving as our baseline models. These models were then contrasted with counterparts pretrained on synthetic data and fine-tuned with the same hyperparameters. Synthetic data was generated using a GAN trained on different portions of the original training set, to generate synthetic data for the pretraining phase. Upon completion of the experiments, normality and equal variance assumption tests were performed to ensure the validity of our findings. Lastly, statistical tests like the t-test and Welch test were employed to examine the significance of mean differences between the two groups. Our research produced a comprehensive statistical analysis alongside a modular, and generic experimentation code which allows to incorporate more models, hyperparameter search methods, and settings, given sufficient computational resources. All these resources, housed on our GitHub repository, are freely available to other researchers, fostering open collaboration and further advancements in the field. Our research revealed key factors - learning rate, sequence length, number of classes, and features that substantially influence model robustness and performance in terms of improvements achieved using our method. Notably, our two-phase training approach resulted in up to 37.5% improvement in F1 and other metrics across various data aggregations. Remarkable enhancements were observed consistently across certain datasets which seems less complex and may indicate that different and improved data generation techniques that better catch underlying interconnection and complexities should be investigated