The Price of Data Processing: Impact on Predictive Model Performance

Abstract

The study investigates how data processing affects the performance of a predictive model trained on processed (vs. unprocessed) data. By unprocessed data, we refer to data that includes information on the influence of each mixture of transmissions from the input neurons. By processed data, we refer to data where the transmission from each input neuron is treated independently, and its co-activity behaviors with other neurons are processed into a single signal. It is intuitive that predicting the output may be less accurate when the input data undergo such processing. However, it is not immediately clear what factors determine the degree of accuracy loss.

Employing the simplest structure for the described data processing, namely two input neurons and one output neuron, we built predictive models that can forecast the activity of the output neuron based on the states of the input neurons, using a synthetic processed (vs. unprocessed) data set of historical system states. It has been discovered that a significant decrease in accuracy occurs when the conditional output firing probabilities are high for opposite inputs. The prediction of significant effects of data processing was demonstrated by utilizing a real data set of breast cancer. The study emphasizes the significance of evaluating the degree of impact caused by data processing, with possible uses for predictive models and applications, such as explainable AI.