Target inductive methods for zero-shot regression

• DOI: 10.1016/j.ins.2022.03.075
• Publication Date: 2022-03-28T15:26:55Z
• Authors: Miriam Fdez-Díaz; José Ramón Quevedo; Elena Montañés
• Keywords: Zero (linguistics)

This research arises from the need to predict amount of air pollutants in meteorological stations. Air pollution depends on location stations (weather conditions and activities surroundings). Frequently, surrounding information is not considered learning process. known beforehand absence unobserved weather remains constant for same station. Considering as side facilitates generalization predicting new stations, leading a zero-shot regression scenario. Available methods typically lean towards classificat are easily extensible regression. paper proposes two The first method similarity based approach that learns models features aggregates them using information. However, potential knowledge feature may be lost aggregation. second overcomes this drawback by replacing aggregation procedure correspondence between feature-induced models, instead. Both proposals compared with baseline artificial datasets, UCI repository communities crime pollutants. approaches outperform method, but parameter manifests its superiority over method.