Making Recommendations Explainable

Learn how to transform traditional recommendation systems into explainable ones that provide transparency for both users and developers in this 16-minute conference talk. Explore the limitations of classic recommender pipelines that rely on single embeddings and understand why they fail to provide meaningful explanations to users. Discover a new approach that stores positive publication IDs and builds explainability indices using relevancy labels and models. Master techniques for ordering candidates based on attractiveness using CTR predictive models, and understand how to generate clear explanation text that users can actually comprehend. Examine methods for keeping positive examples manageable through diversity and ranking strategies, while implementing smarter candidate quotas using bandit-style allocation. Gain insights into debugging recommendation systems more effectively and improving overall relevance through transparency, covering everything from embeddings basics and collaborative filtering to advanced candidate selection with KNN/HNSW algorithms and business logic integration.