[Case Study] Optimizing the application screening process with NLP and machine learning

DonorChoose.org is a non-profit crowdfund platform for teachers to request resources for their classrooms. Founded in 2000 by a public school teacher and largely operated by volunteers, the website faced a scaling issue as the number of proposals grew exponentially every year. The solution was to use an algorithm to pre-screen applications and provide a quality score indicating the likelihood of acceptance. With the latter, high quality proposals can be expedited through the process, and good proposals needing a bit of help can receive special attention from volunteers. The data included the paragraph responses to the application prompts and meta data (e.g., demographic information, grade level, subject area, resource requests, cost estimates, etc.). Exploratory data analysis was first performed to assess data quality and investigate its features. While using an ensemble of many learners by stacking individual classifiers from logistic regression to neural networks produces the best result, performance improvements would need to be weighed against other practical factors, such as the need to productionalize and deploy the system. Three potential solutions are proposed along with their pros and cons.

Read more (soon)