Alberto Michelini

Erik Otović, Marko Njirjak, Dario Jozinović et al.

In practice, it is very demanding and sometimes impossible to collect datasets of tagged data large enough to successfully train a machine learning model, and one possible solution to this problem is transfer learning. This study aims to assess how transferable are the features between different domains of time series data and under which conditions. The effects of transfer learning are observed in terms of predictive performance of the models and their convergence rate during training. In our experiment, we use reduced data sets of 1,500 and 9,000 data instances to mimic real world conditions. Using the same scaled-down datasets, we trained two sets of machine learning models: those that were trained with transfer learning and those that were trained from scratch. Four machine learning models were used for the experiment. Transfer of knowledge was performed within the same domain of application (seismology), as well as between mutually different domains of application (seismology, speech, medicine, finance). We observe the predictive performance of the models and the convergence rate during the training. In order to confirm the validity of the obtained results, we repeated the experiments seven times and applied statistical tests to confirm the significance of the results. The general conclusion of our study is that transfer learning is very likely to either increase or not negatively affect the predictive performance of the model or its convergence rate. The collected data is analysed in more details to determine which source and target domains are compatible for transfer of knowledge. We also analyse the effect of target dataset size and the selection of model and its hyperparameters on the effects of transfer learning.

Stefan Bloemheuvel, Jurgen van den Hoogen, Dario Jozinović et al.

Machine learning, with its advances in deep learning has shown great potential in analyzing time series. In many scenarios, however, additional information that can potentially improve the predictions is available. This is crucial for data that arise from e.g., sensor networks that contain information about sensor locations. Then, such spatial information can be exploited by modeling it via graph structures, along with the sequential (time series) information. Recent advances in adapting deep learning to graphs have shown potential in various tasks. However, these methods have not been adapted for time series tasks to a great extent. Most attempts have essentially consolidated around time series forecasting with small sequence lengths. Generally, these architectures are not well suited for regression or classification tasks where the value to be predicted is not strictly depending on the most recent values, but rather on the whole length of the time series. We propose TISER-GCN, a novel graph neural network architecture for processing, in particular, these long time series in a multivariate regression task. Our proposed model is tested on two seismic datasets containing earthquake waveforms, where the goal is to predict maximum intensity measurements of ground shaking at each seismic station. Our findings demonstrate promising results of our approach -- with an average MSE reduction of 16.3% - compared to the best performing baselines. In addition, our approach matches the baseline scores by needing only half the input size. The results are discussed in depth with an additional ablation study.