Recommender systems (RSs) have become a central part of daily digital life, shaping which items, people, and opportunities users are exposed to at scale. Given that personal values are fundamental to individual identity, decision-making, and consumer behavior, and that RSs learn from interaction data that is itself shaped by these values, the question arises whether the recommendations produced by standard RSs already reflect users' personal values. To date, no prior work has empirically investigated this question.

This thesis addresses that gap by examining the extent to which user values are reflected in recommendation outcomes, and whether explicitly incorporating value information can improve this alignment. Using Schwartz's Theory of Basic Human Values as a theoretical framework, we conduct an offline experiment on the Goodreads dataset. We construct value profiles for both users and recommended items using the Personal Values Dictionary, which maps over a thousand English words to their corresponding Schwartz value. These profiles are derived from user reviews and book descriptions respectively, and are used to measure the alignment between a user's personal values and the values embedded in their recommendations.

Our results show that standard RSs exhibit a weak but positive degree of value alignment, suggesting that interaction-based optimization procedures partially capture users' values without explicitly modeling them. Furthermore, we find that explicitly incorporating user value profiles as features within the RS increases this alignment. These findings carry important implications for the design of value-aware recommender systems, and suggest that early integration of value information is a promising direction for future research. ... Read more