Do the topic preferences of social media users differ between producing and consuming content? And if so, do recommenders currently account for such differences? This paper acknowledges that the difference is real and measurable, and exploits that finding to report for the first time (as the study claims) how to build a specialized recommender that produces better results when evaluated with ranking metrics.
The experimental set is from Google Plus user data collected during May and June 2014. This data is preprocessed to extract higher-level semantic concepts, taking advantage of the Google Knowledge Graph. Processing occurs using cascading engines that assess users’ interests in different topics (roughly, the higher-level concepts) using matrix factorization, specializing for behaviors (for example, publish or consume), and capturing interest for combinations of the above. The resulting improvement, evaluated with metrics of recall, normalized discounted cumulative gain, and average percentile metrics, is reportedly impressive: usually 30 percent better (or more) than standard (that is, topic-unaware) results. Interestingly, the work does not report any prediction metrics (for example, mean square error); however, it could be argued that in this context prediction has little meaning.
Preprocessing raw data into semantic topics using Google’s in-house ontology will be laborious to replicate for an independent confirmation, even if essential to the core impact of this approach. The reported improvement, in any case, suggests that user behavior specialized by multiple circumstances, creating multiple customized profiles, is key to computing better recommendations.