Modeling multiple time series annotations as noisy distortions of the ground truth: An Expectation-Maximization...

Area of Science:

• Behavioral Science
• Psychology
• Data Science

Background:

• Time-continuous human behavior studies often use multiple annotator ratings.
• Current methods like averaging ratings may inaccurately represent the true construct.
• The ground truth in such phenomena is frequently latent and unobservable.

Purpose of the Study:

• To develop a novel method for modeling multiple time series annotations.
• To compute a more accurate ground truth by modeling annotator-specific distortions.
• To predict children's smile confidence ratings in a naturalistic interaction setting.

Main Methods:

• A novel method modeling multiple time series annotations over a continuous variable.
• Conditioning the ground truth on extracted data features.
• Modeling annotator ratings as modifications of the ground truth, incorporating specific distortion tendencies.
• Training the model using an Expectation-Maximization (EM) algorithm.

Main Results:

• The proposed model was evaluated on a study of child-psychologist interaction.
• It predicted confidence ratings of children's smiles.
• Performance was compared against two baselines: framewise mean of ratings and aligned ratings mean.

Conclusions:

• The novel method provides a more accurate estimation of the ground truth compared to standard averaging techniques.
• Modeling annotator distortions improves the reliability of behavioral data analysis.
• This approach enhances the understanding of latent constructs in time-continuous behavioral studies.