A data-driven approach to addressing unemployment in Jordan
Co-authored by Marieta Fitzcharles, Eva Kaplan, Anthony Pusatory, and Soha Shami
IRC Jordan’s Project Match, which aims to address unemployment in Jordan, uses data-driven algorithms and machine learning to guide our programming and link job applicants with potential employers. Here, we describe the two algorithms at the heart of the program — one which predicts which intervention is best suited to support different profiles of job seekers, and one which provides a probability score on how likely a job seeker is to succeed in any particular job.
Demystifying algorithms
An algorithm is a list of clearly articulated steps required to carry out an operation. In our everyday lives, an analogy would be a cake recipe, a well-defined set of instructions used to produce an intended output, i.e. a cake. In data-driven algorithms, the instructions in the “recipe” are calculations, which use inputs from various data sets as the ingredients. Except in this case, the exact output from the recipe differs based on the “ingredients” (data) that have been used.
Data-driven algorithms are as strong as the data they pull from, and their strength grows as more data is available. In the case of Project Match, as more job seekers, jobs, and successful outcomes are recorded, the algorithms get smarter.
The data is based on surveys conducted both with job seekers and firms. Intake surveys record job seeker preferences, demographic information, and firm profiles and requirements. Follow-up surveys are conducted at one, three, and six months after intake to record details about the job search process, including job matches and retention. As the amount of available information increases during implementation, the effectiveness of both the targeting and the job match algorithms are strengthened.
Precision targeting: Thompson’s Algorithm
Project Match uses an adapted Thompson’s Algorithm to target different job seekers with the most impactful interventions. The Thompson’s Algorithm (or Thompson Selection) identifies the probability that various interventions will lead to a desired outcome; it then uses this probability to continually update who receives what service (called “probability matching”). In other words, as a service shows increasing probability of success, the algorithm automatically recommends more people to receive that service.
In Project Match, the Thompson’s Algorithm is used to identify which of the three services designed to assist job seekers is optimal for different demographics. The services, which have been selected based on particular barriers that job seekers have identified when trying to apply for jobs, include:
- One-time cash assistance (92 USD per job seeker) to offset costs associated with job searching like transportation and childcare.
- Information support in the form of one video on interview techniques to help prepare for interviews and one on the labor law to better understand rights.
- Planning support featuring guidance and a job search planning tool to help job seekers organize their search with clear actions and deadlines.
In the early stages of the project, job seekers are randomly selected to receive one of these three services, and data on job seekers’ profiles and their success in finding jobs is collected. As more data about the success rates of participants from each service is captured, it is used to identify which service is most likely to have the greatest impact for different demographic profiles. For example, if the data shows that Syrian men who have completed secondary education and have a certain amount of work experience are most likely to find a job if they receive information support, they are more likely to be recommended for that service.
This is a key advantage of using the adapted Thompson’s algorithm — it allows the IRC to course correct and prioritize the most effective interventions during implementation rather than redesigning the program based on an after-the-fact evaluation.
The Match Algorithm
In addition to the Thompson Algorithm, the Match Algorithm uses certain data points to match job seeker characteristics and qualifications to business firms’ requirements, which results in a recommendation score for each possible candidate and discards any non-matches. As information about past placements is fed back, the algorithm uses machine learning to update the success criteria to optimize job recommendations — in addition to firm requirements, this could hypothetically include elements like commuting time or informal work experience.
The Match Algorithm is a tool to support the Project Match field team, but the field team’s role remains crucial. The field team knows the job seekers, they know the firms, and they have a great deal of valuable experience. All of this informs their perspective on who will be a good fit for which jobs. The algorithm simply helps by using information about past placements to optimize job recommendations while the field team maintains the critical role of reviewing and finalizing a shortlist of candidates to recommend.
Some thoughts on next steps
In many ways, using algorithms to make our work smarter is low-hanging fruit for the humanitarian sector. But challenges remain. For example, as far as we know, this is the first time a Thompson’s algorithm has been used in the humanitarian sector. Based on our experience so far, we think it’s worth exploring whether more programs could benefit from this type of micro-targeting of services, but it’s not straightforward to implement, especially for field teams who have to actually deliver these services. If we want to make this type of data science more common in the humanitarian sector, more work needs to be done to understand how it impacts our operations, and learn how to make it easier to actually deploy.
Have you used algorithms in your work in the field? Comment below, or drop us a line at airbel@rescue.org. We’d love to learn more.