The before
The issue with standard, ‘dumb’ job ads is that their effectiveness is severely limited. The traditional approach is to show a visitor all the available ads, allowing them to search for any keywords they might be interested in.
This means, on the one hand that the same, possibly irrelevant ads get shown to all visitors, even if there is no chance they might be interested – for example, quality assurance people working at startups probably won’t be very interested in ads for .NET backend developers at multinationals. On the other hand, if a visitor isn’t willing to search for any and all keywords they might be interested in, then they may miss on valuable opportunities.
We wanted to change that.
How we helped them
Analyzing Codecamp’s business needs, we proposed implementing a recommender system that takes the information from each and every job ad and matches it with the information gathered about the visitors. Useful job ad data includes information about the company, such as size and domain of activity, but also about the technologies they require, the level of seniority, etc. Similarly, data pertaining to visitors includes the resentations they were interested in, the ones that they gave feedback on, the topics of said presentations, etc.
Based on this information, we designed a content-based recommender system able to learn the similarities present in the data and recommend to each visitor the job ads they are most likely to be interested in. We also designed the data augmentations needed to support building the recommender system, such as additional information collected for each job ad, but also the information collected for each visitor, with data often implied from their recorded behavior on the Codecamp website.
Challenges in implementing Machine Learning
Analyzing Codecamp’s business needs, we proposed implementing a recommender system that takes the information from each and every job ad and matches it with the information gathered about the visitors. Useful job ad data includes information about the company, such as size and domain of activity, but also about the technologies they require, the level of seniority, etc. Similarly, data pertaining to visitors includes the resentations they were interested in, the ones that they gave feedback on, the topics of said presentations, etc.
Based on this information, we designed a content-based recommender system able to learn the similarities present in the data and recommend to each visitor the job ads they are most likely to be interested in. We also designed the data augmentations needed to support building the recommender system, such as additional information collected for each job ad, but also the information collected for each visitor, with data often implied from their recorded behavior on the Codecamp website.
The after
Putting the recommender system in use means that the job ads present on the Codecamp website are matched to each visitor, and only the ones with the highest chance of a match are displayed on the home page. By only showing a limited number of ads that a visitor might be interested in, we decrease their potential “decision paralysis” in the face of many options, effectively increasing visitor-ad engagement.
It should be noted that this system is currently under development, with the expectation that it will go live at the end of 2018.
