In April I experimented with a neural network that could dream up resumes of imaginary people. I used TensorFlow, the deep learning framework from Google, along with thousands of anonymized CVs. The results were intriguing. It’s time to take this idea one step further.
Suppose we take a large number of candidates that applied to jobs. We put their resumes in two buckets – those that made it to interview, and those that didn’t. We train a neural network on this data. Then we give it a new resume that it’s never seen before. Could it predict whether this new candidate will make it to interview?
The Neural Network
I dusted off deep_orange, the Recurrent Neural Network that I created in the Spring. Rather than looking at the complete resume, I decided to look at each individual bullet-point on a candidate’s resume. Like:
- Managed a project that moved widgets from A to B
- Designed new websites that doubled our customers
These bullets typically don’t include specifics like the actual company, just a description of responsibilities. It would be fascinating if deep_orange could make a reasonable prediction based only on a bullet-point of text.
We actually have 140,000 bullet-points like this in untapt. I fed this data into deep_orange and left it running on my MacBook Pro overnight. The fan was still whistling loudly the next morning.
When it finally completed a few hours later, deep_orange was ready to do some predicting. It assumes each bullet-point is taken from a job application for a software developer role in FinTech, which is a typical scenario for the untapt platform.
Let’s start with something extreme:
Fair enough! Now for something more relevant:
Seems reasonable. Next I tried a really substantive bullet-point:
That’s a strong opinion! But deep_orange has some views on technology:
When we tested deep_orange on about 1,000 bullet-points that it hadn’t seen before, it predicted the correct outcome 87% of the time.
We’ll be exploring whether we could incorporate deep learning techniques like deep_orange into untapt to increase our accuracy. We’re also considering whether these techniques could provide instant feedback while a candidate writes their resume.
This has potential to be a powerful addition to our algorithm.
Stay tuned – there’s lots more to come.