Human Resource Analytics: is machine learning the answer to removing biases from the hiring process?

Human Resource Analytics is one of the more exciting applications of ‘analytics’/’Big Data’/’Data Science’. A very straight forward application is to predict who is likely to leave within a certain time frame, and also, to some degree, what can be done to retain the employee if the organisation wishes to do so. (Not all employees who want to leave are worth keeping). I’ve built this type of model and they work quite well, and I only used the employee’s behaviour at work.

This can be even more interesting in the context of “Big Data”, with the increasing use of social networking and online presence, most of us leave traces for more behavioural clues to be harvested. Recently there has been a trend to use machine learning more in Human Resource, especially in terms of hiring. It is sold as finally unbiased hiring. But how accurate is that? 

Quite a few start-ups are promising unbiased hiring, getting you the best candidate by using proprietary algorithms, very often machine learning algorithms.

Basically, in machine learning, the machine is fed a whole bunch of data, in the context of HR this is likely to be CVs of applicants, any test answers, interview notes, and in the case of supervised algorithms, whether the person was hired or not (and may be the subsequent performance of the people hired). The machine then trawls through the data, learns, and when fed a new bunch of candidates will pick a list of top candidates based on what it has learnt.

My first reaction to this is that, if the data you feed into the algorithm is the data collected by your own organisation within the hiring period, then the machine is likely to replicate what the organisation has been doing. Let’s say that most of your hiring managers come from a specific school, they might be biased in favour of that school, and hire people accordingly. The machine will, in an unbiased way, learn this behaviour and apply it. Training a machine on biased data, is likely to lead to the machine churning out biased outcomes.

Of course there are a few ways round this. 

One way is not to limit the training dataset to a single organisation, but to take a broader view of the market, basically trying to remove or at least dilute the bias. May be your organisation is populated by graduates from school A, your competitor by those from School B, so if you train your algorithm on the combined dataset, you are likely to be reducing the bias. This is the idea behind credit bureau scores, and is likely what quite a few of the start-ups in the HR space are trying to do. 

In this case, the organisation-specific biases are likely to be mitigated, as long as the training dataset is diverse enough. However, unless you have the universe, you still run the risk of being biased. Ideally you should weigh your training set; it’s like stratified sampling in a way, but how will you know how to stratify? You’d need to know the biases.

Another approach is to repair the training dataset, for example: http://www.datasociety.net/output/hiring-by-algorithm/ Basically, the output of the algorithm has to be analysed, and any biases uncovered before implementation, and the training dataset modified to remove that bias. It is not as easy as it seems. For example, imagine if the true reason why people are rejected from the hiring process is not directly apparent in the data, let's say due to criminal record. If some section of society, say people with blue hair are over represented in the population of people with criminal record, then it is possible that the algorithm will be less likely to recommend people with blue hair for hire based on the fact that a smaller proportion of them have been hired in the past, not recognising the real reason. Basically blue hair becomes sort of a proxy for criminal record. 

To go a level deeper, even an algorithm trained on the whole population is unlikely to be free of societal biases https://www.theguardian.com/technology/2016/aug/03/algorithm-racist-human-employers-work . For example if women have not been employed in C-level positions, it is likely that an algorithm would pick that up and not recommend women for these positions, or show people online whom they know to be women, lower paying jobs: http://www.andrew.cmu.edu/user/danupam/dtd-pets15.pdf  . 

To make matters worse, the algorithms often are black boxes. Unless the results of the algorithm are analysed and an attempt is made to find the possible drivers by supervised techniques, it would be almost impossible to uncover the actual drivers. (There are vendors who do that, but clients need to bear in mind that these are not necessarily the real drivers, and to me, it kind of defeats the purpose) 

As an example, imagine a story about a person committing crimes, getting caught by the police, and getting sent to jail is told to a classroom. Everyone agrees that the outcome (ending up in jail) is undesirable. However, what was learnt from the story: do not commit crimes, or do not get caught? A teacher can find out by simply asking the students, but a black box is unlikely to answer.

This is what is called algorithmic transparency, one of the solutions that the ford foundation posits to solving the biases in algorithms. https://www.fordfoundation.org/ideas/equals-change-blog/posts/can-computers-be-racist-big-data-inequality-and-discrimination/ but by definition, black boxes are not transparent.

For those of you who are interested in reading more about the topic I’d recommend the following paper by Barocas and Selbst: Big Data’s disparate impact http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2477899 . For an overview of a method to uncover the hidden biases and try to remove them from the training dataset bearing in mind the trade-offs http://arxiv.org/pdf/1412.3756v3.pdf. Basically it takes a lot of effort and if the actual predictors used are not visible, even more so.

In summary, I’d say that achieving total impartiality in a Human Resource function such as hiring by using machine learning is not as easy as it seems. Humans are generally biased, even societies have biases, and this bias often taints the machines too, especially via the data the algorithms are trained on. And if the algorithms are black-boxes, it makes it that much harder to remove the biases.