Knowledge Forcing: Fusing Knowledge-Driven Approaches with LSTM for Time Series ForecastingMuhammad Ali Chattha; Muhammad Imran Malik; Andreas Dengel; Sheraz Ahmed
In: Springer (Hrsg.). Knowledge Forcing: Fusing Knowledge-Driven Approaches with LSTM for Time Series Forecasting. International Conference on Artificial Neural Networks (ICANN), September 26-30, Crete, Greece, Vol. 14259, Springer, 2023.
Long Short-Term Memory (LSTM) typically relies solely on historical data for training and although, they excel at modelling sequential series and finding hidden patterns in the data, they are unable to utilize expert knowledge. Knowledge-driven systems (KDS), on the other hand, rely on domain knowledge and consist of rules explicitly defined by human experts. Both LSTM and KDS offer unique advantages, hence relying on a single approach can be suboptimal. However, currently there is a lacking of frameworks that can concurrently utilize explicit information in the KDS and hidden features in the data. In this paper, we propose a novel fusion mechanism, knowledge-forced LSTM (KF-LSTM), that combines knowledge-driven approaches with LSTM for time series forecasting. KF-LSTM employs LSTM in an encoderdecoder setting, where the decoder utilizes KDS predictions in a residual connection. This enables the decoder to utilize sequential relations in the historical data passed on by the encoder as well as information present in KDS in a complementary manner. We tested KF-LSTM on 4 realworld datasets in a multi-horizon forecasting setting. Even with utilizing relatively shallow single layered LSTM, KF-LSTM achieves State-of-theArt (SotA) performance on almost all of the datasets, highlighting the information fusion capabilities of the framework. On average, knowledge forcing improves over previous SotA by 20%.