GRAPH NEURAL NETWORKS IN INTERMITTENT TIME-SERIES FORECASTING

• Main Author: Belmiro Tirta Kusuma, Mikhael
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/80063
• Institution: Institut Teknologi Bandung

Intermittent time-series forecasting is a subproblem of time-series forecasting, where in intermittent time-series forecasting there is more considerable amount of zeroes present in the data as opposed to regular time-series forecasting. This causes method usually used in time-series forecasting like ARIMA cannot be used directly due to stationarity assumption ARIMA holds. Therefore, there are several models that specifically tackle intermittent time-series forecasting that have been introduced. There exists two family of models that are predominant in intermittent time-series forecasting. The two family of models are demand arrival time, which uses information regarding when a non-zero value will be present and aggregatedisaggregate, which aggregates data in the temporal dimension to obtain a smoother form of the data. Up until now there is no model that directly combines the ideas of those two methods, furthermore the research in deep-learning regarding intermittent time-series is not yet as well-researched as its regular timeseries forecasting counterpart. Therefore in this thesis, a new model which imitates those two ideas and implements it using deep – learning methods using graph neural networks components is introduced. The model is called intermittent time graph neural forecast or ITNGF. ITNGF achieves the best result in terms of ???????????????? and ???????????????????? in three different benchmark datasets if compared to methods that are frequently used in intermittent time-series forecasting. In addition to that, ITNGF tries to bridge the gap between GNN and intermittent time-series forecasting since ITNGF is the first GNN model which is used for intermittent time-series forecasting.