A real-estate data analytics company at its core, Proportunity’s mission is to help ease the homeowner’s burden of saving for the entire down payment to help buyers ‘purchase their own castle’. Proportunity uses ML to train models that can accurately forecast house prices region by region and identify those with growing value. Forecasts predict with enough confidence for us to arrange equity loans against these realistic future prices. Depending on the buyers’ target market, we can reduce borrowers’ down payment requirements by 15 percent.
Proportunity’s solution is not a simple analytics problem. Our data scientists deal with a significant number of geo-spatial datasets, and preprocessing these requires a lot of compute power. The ensemble of models we rely on contains both traditional ML models, such as k-d trees and splines, as well as cutting edge predictive auto-encoders. We developed models in the TensorFlow* framework using Python* and the AlexNet* topology. While AlexNet was originally developed for GPUs, our models favor processing on traditional CPUs over GPUs.
In late 2018, to help optimize bottlenecks, our data scientists collaborated with the Intel® AI Builder team of engineers. Intel engineers shared their expertise and helped us achieve significant speed up in certain areas by simply switching to Intel optimized libraries, Intel® Optimization for TensorFlow*, and Intel® Distribution for Python*. When using the Intel® DevCloud, our overall inferencing times improved noticeably, performing more than an order of magnitude faster. We have since built a similar HPC cluster in-house for the data science team.
Using Intel AI technologies enabled us to rapidly prototype ML models without worrying about optimizing for performance at the time. From our perspective, this offered incredible benefits because we were able to push more models in to the ensemble, thus boosting the accuracy of our predictions. Regular, continued interaction with Intel engineers has significantly improved the performance of our models, both in terms of speed and accuracy.
While we continue to optimize our models and the associated inferencing, our next challenge is how to deploy the solution. Data science is compute-intensive, and novel machine learning solutions necessitate fast iterations. To develop and deploy our models, we had three choices for compute platforms:
- Build an in-house cluster with GPUs
- Use a GPU-based cluster at a cloud provider
- Run them on a CPU-only cloud-based solution
Proportunity incubated at the Google campus, and, since our ensemble favors CPUs, it became a natural choice to go for a CPU-only Google Cloud Platform (GCP) cluster. Augmenting the value of already being a Google user, a GCP cluster allowed us to continue to develop on Intel® Xeon® Scalable processors, which are ideally suited for machine learning.
At Proportunity, we are at the cutting edge of property price prediction and we are pushing the limits of spatio-temporal forecasting. We are actively collaborating with Intel engineers to bring advanced computing technologies for AI into our forecasting framework. We will be continuing our benchmarking exercise to the latest 2nd Generation Intel® Xeon® Scalable processors with Intel® Deep Learning Boost.
Using ML and Intel® AI technologies, we can offer U.K. home buyers an alternative to the challenges of purchasing their properties. Instead of the never-ending battle to save enough money for a down payment, hoping to take advantage of the government’s Help to Buy program (which ends in 2021), or shared ownership (a particularly unexciting option), we use advanced machine learning and data analytics technology to find areas that are growing and spot undervalued properties in these areas. We uncover hidden gems within the housing market, so that our customers can be confident of making the right choice on a growth investment, one that is as much an emotional decision as it is a financial one.
Learn more about Proportunity at www.proportunity.co or view our market ready solution here