Identification of human behavior in accelerometer data from interactive toys by applying AutoML

Background:

Human behavior is closely tied to our identities, cultures, and illnesses, and is therefore highly relevant to social, commercial, and medical studies. Analyzing interactions between people or between people and items is a method for studying behavior. In this work, we analyze pre-recorded accelerometer data from interactions with embedded sensors to classify 8,946 behavior records from five classes: drop, hit, pickup, shake, and throw.

Methods:

We evaluated multiple machine learning algorithms—Bayes Network, Multinomial Logistic Regression, Multi-layer Perceptron, Naïve Bayes, and Repeated Incremental Pruning to Produce Error Reduction (RIPPER). Also, an AutoML approach was applied for automated model and hyperparameter selection.

Results:

AutoML outperformed traditional classifiers, achieving a precision of 94.4% and a receiver operating characteristic (ROC) area of 0.992 were obtained.

Conclusion:

These findings confirm AutoML’s effectiveness in accurately identifying human behaviors from accelerometer data in interactive toys.