Webscale Recommender Systems
Problem being addressed
Deep learning systems have made significant progress over the last couple of years. But can they handle data on the order of magnitude of the web?
In this research, a deep learning recommendation system for graph structured data is developed. This involves the creation of an efficient algorithm called PinSage which uses a combination of random walks and graph convolutions. PinSage is then deployed on Pinterest web scale data that involves a network data set of 7,5 billion examples and 3 billion nodes. It is found that PinSage generates recommendation benchmarks that are much better than any other deep learning or graph based architecture that was considered state-of-the-art at the time of publication.
Advantages of this solution
This is a highly scalable algorithm that outperforms the best benchmarks at the time for making recommendations to users on Pinterest.
Solution originally applied in these industries
Possible New Application of the Work
Chemical and Materials Industry
Chemical and material manufacture can sometimes require large scale simulations of molecular structure. This deep learning approach and its ability to scale up could be of use for the materials industry.
Real Estate Industry
Real estate data is quite extensive, and real estate recommendation engines often serve up information to users. Perhaps this algorithm could be adapted to the real-estate industry as a way of making recommendations to prospective customers. One possibility is, for example, to embed it into AirBNB type businesses.
Source URL: #############