DeepLSF: Fusing Knowledge and Data for Time Series ForecastingMuhammad Ali Chattha; Ludger van Elst; Muhammad Imran Malik; Andreas Dengel; Sheraz Ahmed
In: IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), Vol. NA, Page NA, IEEE, 2023.
Data-driven Deep-Neural-Networks (DNNs) have proven to be highly effective in many domains, but there is growing recognition that relying solely on data to find solutions can be sub-optimal since it neglects an important domain i.e. knowledge. Consequently, hybrid schemes utilizing both knowledge-driven and data-driven approaches are becoming mainstream. In the context of time series forecasting, these hybrid schemes are often based on ensemble methods, where predictions are averaged regardless of the accuracy of constituent models. In contrary, we believe that knowledge-driven and data-driven approaches have unique strengths and an ideal fusion scheme should dynamically adapt to these differences. Inspired by this, we introduce a novel fusion framework, Deep Latent Space Fusion (DeepLSF), that learns a joint transfer function from features drawn from both knowledge-driven approaches and historical data. This leads to a better fusion mechanism where information is shared in a complementary manner. Our tests on 5 real-world forecasting datasets demonstrate that even with off-the-shelf knowledge-driven and data-driven models, DeepLSF outperforms the State-of-the-Art by an average of 22%, highlighting effectiveness of the proposed knowledge fusion scheme. Leveraging knowledge to enhance the performance of DNNs can be useful in a wide range of real-world applications and improve its generalization capabilities.