Leakage Prediction in Machine Learning Models When Using Data from Sports Wearable Sensors

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Data leakage is a critical issue in machine learning, undermining AI model validity and reliability.
• It occurs when training data inadvertently contains information about the target variable, leading to poor generalization.
• Preventing data leakage is essential for developing robust and accurate forecasting models.

Purpose of the Study:

• To present an innovative system for predicting data leakage in machine learning models.
• To enhance the reliability and generalizability of artificial intelligence models.
• To address concerns regarding adversarial attacks and data integrity in AI.

Main Methods:

• Utilizes Bayesian inference to calculate the reverse probability of unseen variables.
• Establishes statistical conclusions about relevant correlated variables.
• Calculates a lower limit on the marginal likelihood of observed variables, indicating potential data coupling.

Main Results:

• A higher marginal probability for a set of variables indicates a better data fit and a greater likelihood of data leakage.
• The proposed system effectively identifies potential data leaks within machine learning models.
• Methodology validated on a specialized dataset from sports wearable sensors.

Conclusions:

• The Bayesian inference system offers a robust approach to detecting and preventing data leakage.
• Accurate data leak prediction is crucial for trustworthy and generalizable AI.
• This method contributes to the reliability of machine learning applications, particularly in sensor data analysis.