After some analyses, we realized that there might be unexpected issues with our on-demand job marketplace, Biscate. Despite the number of requests being four times superior to the number of workers, we found that many of the registered workers had never received a request (around 54%); and even the ones that had received one had to wait an excessive amount of time (50% of workers with requests waited over 7 months before receiving their first request). These surprising results led to the following research question: “How are the requests distributed among the workers on the platform?”. Through an analysis in Python, we created three graphs that help us understand the root of the above-mentioned issues.

The first graph shows what percentage of total requests were owned by each percent of the top workers of the platform. It reveals a shocking reality: the current request distribution system is highly unequal. As we can see, the top 1% of workers receive 28% of all the requests on the platform. Then, the percentages of the total requests possessed by the following groups of top users keep decreasing drastically. This decrease goes on until the 45% group, then the following groups have all received 0% of total requests, no request at all.

 

 

In this second analysis, we wanted to discover the average number of requests per worker inside of each group, hence the below graph. Again, the results are conspicuous, there is a high level of inequality in the request distribution. The workers in the top 1% have received on average 120 requests each. Thus, we see that the top 1% is receiving an extreme number of requests compared to the global mean of 4 requests per worker and the median of 0 requests per user. This median of 0 requests per user indicates that over 50% of all the workers never received a request.

 

 

The two previous analyses have the same insight implied: the current request distribution system is highly unequal as only a small portion of the workers benefit from the Biscate platform. It is then clear that we need to develop a fairer request distribution system. The first step of this process is to be able to measure the level of inequality of the system and monitor its evolution after the implementation of adjustments. Thus, we decided to adopt the Gini coefficient, an economics concept used to measure inequality. The Gini coefficient is represented by the area between the cumulative share of requests and the Perfect Equality. For our system, the Gini coefficient is 0.84, close to the Perfect Inequality level (value of 1). Also, we can also see that 54% of all the workers have 0% of the total requests, while 87% of all the workers only receive 20% of all the requests.

 

 

To better understand the reason behind this unequal distribution of opportunities (requests) between all workers registered on Biscate it's important to analyze which design features led to these results, and we believe that the culprit is the way clients request the worker’s contact. On Biscate a client selects a location and profession (e.g. Plumber in Maputo), and is given a list of all workers registered in those two categories. Furthermore, this list is organized by number of requests and ratings, meaning that the workers with most requests, and therefore, ratings, are on top of the list - they become super workers that receive most of the requests - while new workers with no requests end up in the bottom. To further complicate this, USSD, the technology that powers the platform has a character limitation per each menu it displays, so only 8 workers are available on each menu displayed, and clients have to navigate to the second, third, and so forth menus in order to find more workers. These extra steps inevitably lead to a decline in requests for users further down the list, leading to the unequal distribution of opportunities (requests). This is similar to what happens on Google, where page 1 captures a whopping 95% of all traffic, while page 2 is jokingly referred to by SEO professionals as “the best place to hide a dead body”.

This analysis perfectly illustrates the importance of data deep dives for improving the social impact of services like Biscate. Prior to D4WN, the Biscate team was just looking at the global mean of requests per user as a key performance indicator, and was quite pleased that the average worker received 4 requests from the platform. The problem here is that averages can hide what you need to know, especially when used in a KPI, and this analysis highlighted how design decisions can lead to unintended consequences - in this case, they created a biased platform. We are now busy developing a matchmaking algorithm that will distribute opportunities more evenly across all workers, an attempt to remove selection bias and striving for perfect equality in opportunity distribution. Furthermore, the Gini coefficient will now become a KPI of Biscate to be monitored through time, because social impact is only achieved if everyone has an equal opportunity in the system. This does not mean that highly rated workers will be penalized, it will simply provide more opportunities to new workers to provide their services and improve their ratings. In the end, it will be up to the clients to choose which variables are most important in the hiring process, and not for the platform to condition this choice.

Coming soon.