Incorporating sparse labels into hidden Markov models using weighted likelihoods improves accuracy and...

Area of Science:

• Ecology
• Behavioral Ecology
• Computational Biology

Background:

• Ecologists utilize hidden Markov models (HMMs) to interpret animal behavior from biologging data.
• Advances in observation technology (e.g., drones) enable direct behavioral labeling, potentially improving HMM accuracy.
• Infrequent or sparse labels often have minimal impact on HMM parameter estimates, limiting decoding accuracy.

Purpose of the Study:

• To develop a novel weighted likelihood approach for HMMs to enhance the influence of sparse observational labels.
• To apply this method to decode the foraging behavior of killer whales (Orcinus orca) using limited observational data.
• To evaluate the effectiveness of the weighted likelihood approach against existing methods and single-frame machine learning techniques.

Main Methods:

• Introduction of a weighted likelihood function to increase the impact of labeled data points within HMMs.
• Development and application of weighted HMMs for analyzing killer whale foraging behavior in British Columbia.
• Utilizing cross-validation and simulation studies for rigorous performance evaluation.

Main Results:

• The weighted likelihood approach significantly improved the accuracy and interpretability of decoded latent processes (animal behaviors).
• The developed HMMs outperformed traditional HMMs and single-frame machine learning methods in decoding accuracy.
• The method demonstrated effectiveness in leveraging sparse observational data to enhance behavioral decoding.

Conclusions:

• The weighted likelihood approach offers a robust solution for improving HMMs when dealing with sparse observational labels in ecological studies.
• This method enhances the ability to accurately decode hidden animal processes, applicable across various scientific disciplines.
• The study provides a valuable tool for researchers using observational data to understand animal behavior.