Evaluation Techniques for Long Short-Term Memory Models: Overfitting Analysis and Handling Missing Values

Long Short-Term Memory models have demonstrated their effectiveness across various domains, providing solutions to numerous problems. In time-series data originating from process monitoring, Long Short-Term Memory models are particularly useful due to their capability to capture and represent spatial and temporal dependencies within the data. However, when it comes to modeling nonlinear systems described by differential equations, the relationship between input and output variables cannot be established solely by considering the current inputs. A well-known technique called Teacher Forcing enables the inclusion of previous output true values as additional inputs to Recurrent Neural Networks. However, applying this technique poses a significant challenge: the resulting models might tend to excessively rely on the previous true value, leading to predictions that replicate the previous value at each time step. As a solution, this paper presents a novel methodology to assess whether these models suffer from such overfitting to the previous true value. Additionally, this paper introduces a method to evaluate the robustness of these models against missing values, which may occur due to unforeseen events such as communication faults or erroneous sensor readings. The experimental results, conducted on the Tennessee Eastman Process Dataset, demonstrate the effectiveness of our proposed solutions in terms of overfitting tests and handling missing values.