Creating a Real-Time Recommendation Engine using Modified K-Means Clustering and Remote Sensing Signature Matching Algorithms


Built on Google App Engine (GAE), RealMassive encountered challenges while attempting to scale its recommendation engine to match its nationwide, multi-market expansion. To address this problem, we borrowed a conceptual model from spectral data processing to transform our domain-specific problem into one that the GAE’s search engine could solve. Rather than using a more traditional heuristics-based relevancy ranking, we filtered and scored results using a modified version of a spectral angle. While this approach seems to have little in common with providing a recommendation based on similarity, there are important parallels: filtering to reduce the search space; independent variables that can be resampled into a signature; a signature library to identify meaningful similarities; and an algorithm that lends itself to an accurate but flexible definition of similarity. We implemented this as a web service that provides recommendations in sub-second time. The RealMassive platform currently covers over 4.5 billion square feet of commercial real estate inventory and is expanding quickly.

Keywords:algorithmsclusteringrecommendation engineremote sensing