Real-Time Recommendations in a Multi-Domain Environment

Everyone running large-scale recommenders knows the trade-off: accuracy, speed, and adaptability. Most frameworks pick one or two and ignore the rest, especially when pressed to serve multiple domains (think hotels, shopping malls, events) with diverse, evolving data and rising user demands. This paper argues that’s not just a limitation of existing recommender algorithms—it’s also about the architecture behind them.

We introduced and evaluated a framework (ScaR) that doesn’t just tweak algorithms but rethinks the whole deployment model. Instead of training big models offline (slow) or using single-domain data schemas (inflexible), ScaR is microservice-based and designed for real-time response, letting each domain run its isolated recommender with customized data features—be it items, social signals, or location pings.

Concrete example: with SecondLife and Foursquare datasets (covering item, social, and location interactions), we tested whether blending multiple feature types outperforms item-only recommenders for different use cases. Social signals won out for item-level recommendations, while item and location data could edge ahead for categorical recommendations (like top-level browsing). Mixing feature sources consistently produced more robust recommendations overall, and location data was especially valuable for tackling cold-start users.

On scalability, we ran runtime experiments: 325,000+ simultaneous recommendation requests, scaling from 1 to 4 distributed nodes. A single-node solution buckles fast—runtime grows exponentially as requests rise. Add nodes, and runtimes flatten: horizontal scaling works, but only with an architecture built for it. The microservices in ScaR ensured not just speed but smooth customization per domain, without recalculating recommendations after every tiny data update.

The main takeaway: if real-time, multi-domain recommendations matter, you can’t just bolt new models onto old systems. Architecture must support streaming data, dynamic updates, and independent domain scaling. The detailed evaluations in the paper show where the payoffs—and bottlenecks—really are. If you’re still batch-updating recommenders, it’s time to rethink the stack.

@inproceedings{lacic2016real,
  title = {Real-Time Recommendations in a Multi-Domain Environment},
  author = {Laci\'{c}, E.},
  year = {2016},
  booktitle = {Extended Proceedings of the 27th ACM Conference on Hypertext and Social Media},
  publisher = {ACM},
}