Utilizing Online Social Network and Location-Based Data to Recommend Products and Categories in Online Marketplaces

Most online marketplaces lean heavily on transactional data—think purchase histories and product clicks—to power recommendations. But is that the whole story? This paper systematically compares marketplace, social network, and location-based data as sources for product and category recommendations, using a large-scale dataset from SecondLife that uniquely includes all three.

Marketplace data is solid—but social data wins for product predictions. When deploying user-based collaborative filtering, user similarity features extracted from the social network (e.g., groups, interests, interactions) significantly outperformed those derived from pure purchase histories in predicting which specific products users would want. The result is both intuitive and a bit surprising: who you interact with socially is a better predictor of future purchases than what you’ve bought, at least in this data-rich virtual economy.

Category prediction is a trickier beast. Here, marketplace and location-based data become more valuable. Knowing a user’s purchasing domains or favored locations improved the ability to suggest both top-level and detailed product categories. Essentially, product recommendations benefit from social context, category recommendations from behavioral and locational nuance.

A hybrid (weighted sum) approach, drawing from all three data types, proved most robust, increasing coverage and maintaining high accuracy. But even so, no single feature or source dominates across all tasks. The real takeaway is that tailoring your data source and feature selection to the level of recommendation granularity—product vs. category—pays dividends.

Methodologically, the work is grounded: standard collaborative filtering, thoughtfully designed user similarity measures (content- and network-based), rigorous offline evaluation on a dataset of 7,029 users. The inclusion of fine-grained location and social interaction graphs sets this study apart from much of the prior work, which often focuses on only one source or type of similarity.

For anyone building recommender systems in environments where multiple behavioral signals are available, these findings suggest a more nuanced data strategy: social for product, marketplace and location for category, and hybrid approaches when scale and robustness matter.

@incollection{lacic2015utilizing,
  title = {Utilizing online social network and location-based data to recommend products and categories in online marketplaces},
  author = {Laci\'{c}, E. and Kowald, D. and Eberhard, L. and Trattner, C. and Parra, D. and Marinho, L. B.},
  year = {2015},
  booktitle = {Mining, Modeling, and Recommending 'Things' in Social Media},
  pages = {96-115},
  publisher = {Springer International Publishing},
}