Behavioral Patterns of Users in Digital Content Marketplaces: A Steam Dataset Study

Abstract

Digital content marketplaces generate large volumes of data reflecting collective user preferences. However, explicit user-level interaction logs are often unavailable due to privacy and accessibility constraints. This paper investigates how such aggregated signals can be leveraged as proxies for user preference modeling in recommendation tasks.

Content attributes (genres, tags, and textual descriptions), together with popularity indicators (ratings and ownership statistics), are treated as representations of collective demand. A structured Bronze-Silver-Gold preprocessing pipeline is applied to ensure data consistency and analytical reliability. Three recommendation strategies are evaluated: content-based similarity, popularity-based ranking, and a hybrid model combining both signals.

Unlike traditional recommender systems that rely on user interaction data, the proposed framework focuses on item-to-item recommendation under data-constrained conditions. Evaluation is performed using semantic consistency metrics (Tag-Precision@10 and Tag-Jaccard@10), which capture thematic alignment rather than actual user satisfaction.

Experimental results show that the hybrid model achieves high semantic consistency (Tag-Precision@10 = 0.945) while providing a better balance between relevance and popularity compared to baseline approaches. The findings suggest that aggregated marketplace data can support effective and scalable recommendation, offering a practical solution for platforms where user-level data is unavailable.