[1Q4-GS-10] AI application:

[1Q4-GS-10-03] Leveraging Data from Vast Unexplored Seas: Positive Unlabeled Learning for Refining Prediction Area in Good Fishing Ground Prediction

〇Haruki Konii¹, Teppei Nakano¹, Yasumasa Miyazawa², Tetsuji Ogawa¹ (1. Waseda University, 2. Japan Agency for Marine-Earth Science and Technology)

Keywords:Machine Learning, Semi-supervised Learning, PU Learning, Fishing Industry, Fishing Ground Prediction

This paper proposes a Positive Unlabeled (PU) learning approach to refine good fishing ground prediction.
PU learning, a semi-supervised method, is suitable for scenarios with limited positive examples and abundant unlabeled data, such as fishing ground prediction from unexplored sea areas.
Conventional methods often result in overly broad or restrictive predictions due to the data scarcity.
To address this problem, we utilize PU learning to identify negative examples from unlabeled data and refine the prediction area.
Specifically, we train a model to predict fishing duration as a surrogate indicator of good fishing grounds.
Areas with short predicted durations are treated as negative examples, enabling a binary classification framework to improve prediction accuracy.
To our best knowledge, this is the first application of PU learning in this domain.
Experiments with bullet tuna trolling data validate the effectiveness of our approach.