User choice, hidden gems, and the quadratic horizon

Area of Science:

• Computer Science
• Machine Learning
• Information Retrieval

Background:

• Content platforms utilize machine learning models to recommend items from vast catalogs to users.
• These models are optimized to predict user preferences and curate small sets of items for engagement.

Purpose of the Study:

• To investigate the limitations of current recommendation models in identifying highly preferred items within large item sets.
• To demonstrate the existence of "hidden gems"—items favored by many users but not surfaced by platforms.

Main Methods:

• Theoretical analysis of recommendation model prediction horizons.
• Empirical studies across diverse datasets to validate findings.
• Examination of model performance with complete user preference information within item sets.

Main Results:

• Recommendation models have a prediction horizon limited to k^2 items, even with complete preference data for k items.
• Examples show models failing to identify universally favored items with better than random chance.
• "Hidden gems" are shown to exist in real-world datasets of moderate size.

Conclusions:

• Current recommendation platforms exhibit a significant gap in identifying widely popular but not yet surfaced items.
• The "quadratic horizon" limitation restricts models from discovering "hidden gems" in extensive catalogs.
• Further research is needed to address these limitations and improve content discovery.