Anomaly Detection in Walking Data Using Isolation Forest: An Unsupervised Learning Approach

Abstract

Detecting anomalies in walking data is crucial for ensuring data quality in wearable devices and understanding irregular physical activity patterns. Traditional methods often rely on labeled data, which is scarce in real-world applications. This study presents an unsupervised learning approach using Isolation Forest to detect anomalies in walking datasets. The data, comprising features such as step count, distance, and time, was preprocessed and analyzed to identify patterns and deviations. Isolation Forest was employed due to its efficiency in handling high-dimensional data and its ability to separate anomalies without prior labeling. The model successfully detected 5 anomalous data points out of the dataset, with anomaly scores ranging from -0.15 to 0.2. These outliers corresponded to extreme walking patterns, such as unusually high step counts with disproportionate time and distance. Visualization of anomaly scores and statistical evaluations validated the model's effectiveness, showing clear distinctions between normal and abnormal data. The proposed approach highlights the potential of Isolation Forest in improving data quality and enabling real-time anomaly detection in fitness tracking applications. This work contributes to the broader field of unsupervised anomaly detection by demonstrating a scalable and effective method for handling real-world activity data.