Tailoring Recommendations for a Multi-Domain Environment

Most recommender systems are tuned for a single type of data—think movies or music, but not both. This paper targets a gap: what does it actually take to build a recommender platform that handles multiple, heterogeneous domains without bespoke solutions for each? We share practical guidance drawn from real-world deployments (like e-commerce, hotels, scientific talks, and news) on architectures, configuration, and the subtle challenges of service isolation, data heterogeneity, and customization.

Our proposed architecture adopts a microservices pattern, splitting core functions (data handling, recommendation logic, request balancing, customization, and evaluation) into standalone modules. Key insight: each domain gets its own configuration profile—the so-called recommender profile—where algorithm parameters, similarity metrics, and filtering logic are tailored to the needs of the domain (for example, the features and actions that matter in music are very different from those in movies or hotel bookings). These profiles propagate throughout the system so new domains or requirements can be handled with minimal friction.

For data storage at the item level, we adopt a schema-less approach (with Apache Solr) to ingest heterogeneous data models, making it trivial to add new types of interactions or entities as domains expand. Distributed orchestration (using ZooKeeper) allows horizontal scaling and robust fault tolerance, a must when one domain suddenly surges in traffic or a component fails.

We illustrate the impact with experiments on LastFM, Foursquare, and MovieLens—covering tens of millions of interactions and both implicit and explicit feedback. Results highlight how domain-specific configuration (like tuning neighborhood size in collaborative filtering) can make a measurable difference; no one-size-fits-all setup exists. A hybrid recommendation approach, blending collaborative and popularity-based methods with domain-tuned parameters, consistently provided robust coverage and accuracy. For those designing recommendation infrastructure across organizational silos, this work offers a concrete blueprint and experiential findings rather than abstract generalizations.

@inproceedings{lacic2017tailoring,
  title = {Tailoring Recommendations for a Multi-Domain Environment},
  author = {Laci\'{c}, E. and Kowald, D. and Lex, E.},
  year = {2017},
  booktitle = {Proceedings of the Workshop on Intelligent Recommender Systems by Knowledge Transfer & Learning at the 11th ACM Conference on Recommender Systems},
}