Sustainable techniques to improve Data Quality for training image-based explanatory models for Recommender Systems

Data quality matters more than data quantity for explainable AI: The research demonstrates that improving the quality of training data through smart selection and augmentation techniques can boost performance by 5% in ranking metrics without requiring expensive data collection or larger models, challenging the assumption that "more data" is always necessary.

Responsible AI requires considering environmental sustainability: The paper explicitly addresses the carbon footprint of training AI systems by showing that sustainable enrichment strategies (data selection and augmentation) can achieve comparable results to environmentally costly approaches like massive data gathering or model enlargement.

Labeling noise is a significant but addressable problem: The study identifies that visual explanation systems for recommender systems suffer from unreliable training labels, and proposes using Positive-Unlabelled Learning to identify and filter out unreliable negative examples—a practical solution for real-world datasets where perfect labeling is often impossible.

Generative AI can improve data quality responsibly: Text-to-image generation and transform-based data augmentation offer sustainable ways to expand training datasets without collecting new real-world data, potentially reducing the privacy and environmental costs associated with large-scale data gathering.

Transparency in recommender systems requires rethinking training approaches: Visual explanations help users understand why products are recommended, but achieving this transparency sustainably demands moving beyond conventional "bigger is better" training paradigms toward more thoughtful data curation strategies.