prasunshrestha / recommendation_system_for_donors

This project is to build a recommender system for donors of DonorChoose.org platform to increase repeat donations

Social-good-recommender-system

This project is to build a recommender system for donors of DonorChoose.org platform to increase repeat donations

PROJECT OVERVIEW

DonorsChoose.org (DCo) is a fund-raising platform that empowers public school teachers across the nation to request funding support for educational needs through its online platform. Since DCo’s establishment in 2000, 4.6 million donors and partners have funded more than 1.7 million projects. There is a need to increase repeat donation, through targeted marketing to donors with higher likelihood of re-contribution. Thus, we worked on DonorsChoose.org data challenge to build a recommendation system that profiles donors with higher probability of repeating donations and suggested projects that are likely to be supported by the prior donors. And this recommendation system could be expanded to any other fundraising platform beyond DCo. In addition, we can also leverage DCo’s data to address their additional challenges. Teachers have many projects in parallel that require fundings. A classifier that could predict the likelihood of a project being funded will benefit the teachers, as they then can propose the projects which are most likely to be funded. As a result, our project also built a classifier to predict the probability of a project's funding success. (More on my resume)

DATA DESCRIPTION

The data is available at Kaggle’s website, which is where DCo hosted their data challenge as well. It includes all the projects from January 2013 to May 2018, with a total number of 4.6 million observations. Available are six data files: (1) Projects.csv, (2) Schools.csv, (3) Teachers.csv, (4) Resources.csv, (5) Donations.csv, and (6) Donors.csv. Each file includes pertinent details - for example, Projects.csv includes project type, cost, short description, and project’s funding status.

PROJECT GOALS

Increase recurring donations among prior donors via donor profiling and the recommendation system. Build a classifier that will predict the likelihood of a project being funded.

DELIVERABLES

A detailed profiling of donors and projects based on the past data by clustering. The clustering will be used for User-item mapping in the recommendation system Visualizations for understanding the projects, stakeholders and donations made till date A recommender system suggesting projects with funding needs to prior donors of DCo (fund-raising platform). A classifier to predict the likelihood of a project getting funded. The deliverables will be modular and standardized so that any fundraising platform can use the application with few modifications like feature engineering and use it with minimal effort.

HYPOTHESIS

Donors are more likely to donate to projects that match their geographic region, such as zip code or urban/rural. Projects of educational supplies type, such as stationeries, are more likely to get donor support than other types. If a donor makes a recurring donation, s/he is more likely to donate to a similar project.

TECHNICAL APPROACHES

The dataset contains both temporal (5 years) and spatial data (donors and teachers located across cities in the US). Our train, test, validation splits, and k-fold cross validations will take the temporal and spatial nature of the data into consideration. We intend to use unsupervised learning algorithms, such as k-means clustering and collaborative filtering, to cluster different types of donors. The different unsupervised learning algorithms will then be compared for performance. The cluster will also help us understand a donor’s likely project preferences, which we could leverage to market targeted projects to them. We will use collaborative filtering as we are building the recommendation system for existing donors. Both memory based and model based approaches will be tried to get the best model. User-item interaction (donor-project) will be the primary method for recommendation systems. User-user and item-item methods will also be tried out. We will build a classification algorithm, such as logistic regression, decision tree, and SVM, to predict whether a project will be funded. We have historical data that shows us whether a project is funded or not. Thus, we can apply supervised learning models to discern the patterns and build a classifier.