Data analysis and impact measurement
A partnership recommender for social entrepreneurs
I developed a recommender to help an impact investor and incubator decide what support to offer the social entrepreneurs in its programs.
It used five years of program and partnership data to suggest funding, mentorship, workshops, and other support.
How the model workedModel overview
Five years of program data
Social entrepreneurs’ needs, partner expertise, and previous support
Find similar organizations
Collaborative filtering using a nearest-neighbor model
Suggested types of support
- Funding
- Mentorship
- Workshops
- Other programming
- My role
- Developing the recommender model
- Data
- Five years of programming, partnerships, and organization profiles
- Use
- Used by the organization; details of its use and results aren’t available
Matching support to an organization’s needs
The organization worked as an impact investor, incubator, and accelerator. It offered funding, direct investment, connections to mentors, and workshops.
I built a tool to help identify which kinds of support might be useful for each social entrepreneur, using information from the organization’s previous programs and partnerships.
Using past partnerships to make recommendations
I used collaborative filtering with a K-nearest-neighbors model. It compared organizations using their needs and characteristics, the support offered, and the expertise of their partners.
- Organization profiles
- The type of work, background, and needs of the social entrepreneurs
- Programs and partners
- The type and size of the support offered, and the partners’ expertise
- Recorded outcomes
- Whether a previous partnership had been described as transformational for the organization
Defining a successful partnership
The model relied on a label for whether a partnership had been transformational. Looking back, I questioned whether that was a useful target, especially when the effects of an investment or partnership might take years to become clear.
I would spend more time defining the outcome before building the model: what progress should look like, when it could reasonably be measured, and what the available data could tell us.
Combinations of support
The model recommended individual types of programming. Another question was whether a combination of support would be more useful—for example, how mentorship and funding worked together.
After this project, I began working with multi-armed bandit methods and considered whether that approach could be useful here. It was a possible direction for further work, not part of the recommender I delivered.
Outcome
A recommender used by the organization
The organization used the tool. I don’t have details of how it was used or evidence of its effect on partnership decisions, so I can describe the model and its limitations but can’t quantify its impact.