Training and Prediction Data Discrepancies: Challenges of Text Classification with Noisy, Historical Data
This addresses a practical issue for industry practitioners using accumulated historical data for text classification, but it is incremental as it builds on existing work in noisy data handling.
The paper tackles the problem of evaluating text classification models when training data is noisy and historical, while prediction data is cleaner and different, showing that performance metrics on noisy data can still reflect future model performance.
Industry datasets used for text classification are rarely created for that purpose. In most cases, the data and target predictions are a by-product of accumulated historical data, typically fraught with noise, present in both the text-based document, as well as in the targeted labels. In this work, we address the question of how well performance metrics computed on noisy, historical data reflect the performance on the intended future machine learning model input. The results demonstrate the utility of dirty training datasets used to build prediction models for cleaner (and different) prediction inputs.