How to ‘Netflix’ your HR decisions
Settling on the sofa after a busy day, you turn on the TV. You want to watch something but aren’t sure what. Netflix presents a dozen recommendations. You don’t fancy the first couple, but the third you do like the look of. Click, and away you go.
You may not know this but Netflix has just gained valuable data from your actions. It gave you a list of recommendations, and your eventual click told them what you were ultimately most interested in. Your action will then be used to optimise the recommendations Netflix will make to you and other customers in the future.
This is what data-driven companies do. They self-learn from data at every possible opportunity. And they make a fortune out of it.
So how does this relate to HR? Well, the same Machine Learning (ML) approaches used to recommend films can also be used to suggest HR actions likely to work for your organisation. The potential is vast.
How recommender systems work
This sort of recommender ML involves three key things, related to each other in a closed loop:
- First, a prediction of what is likely to happen in the future.
- Second, an action based on that prediction.
- And finally, learning - a further optimisation of the ML model to optimise the next set of predictions.
This is how it works in the case of Netflix:
How this can work in HR
The same logic can be applied to HR decisions. As an example, let’s look at an organisation wanting to improve retention of its highest-performing employees.
The Learn and Prediction boxes are almost the same as with Netflix. The ML uses the success (or failure) of the previous predictions to get smarter at making a new set of predictions, in this case about what interventions might help us retain our high-performing staff.
The key difference to Netflix is that we need four steps. And in the Action and Outcome boxes we’ll need some human input. That’s because every ML system needs to know whether its predictions worked or not - that’s how it learns. For Netflix that’s simple: it recommends a film; if you click through, it looks like its prediction was a good one; if you don’t, it looks like it wasn’t. It is simple and automated. But in HR it’s a little more complicated. You need a manager to close the feedback loop and allow learning. Let me explain ...
Putting the human back in the loop
Firstly, a line manager or member of HR staff is needed to actually perform an action recommended by the ML. That might be a pay rise, a motivational chat over a nice lunch, an option for flexible working, a promotion, or indeed nothing, if that looks like the best option to achieve your desired outcome.
Secondly, that person will have a crucial role in logging the results of the intervention. That’s because judging the success of the outcome cannot just be based entirely on whether the employee is still with the company six months or a year later.
The manager will need to provide some more meaningful information that can help to improve the next set of predictions. Let me explain why. Choosing a film is straightforward and not much influenced by external factors. Once we’re on Netflix, it’s reasonable to assume we want to watch a film. If we don’t choose one of the recommendations displayed prominently, it’s fair to assume that the reason we haven’t gone with a recommendation is because we don’t like the film.
Losing an employee is different. It’s not binary like the click of a mouse. Perhaps you gave a promotion but the employee left six months later anyway: this doesn’t mean the recommendation was wrong or the outcome was negative. Perhaps that person’s partner got a job at the other end of the country and they moved as a result. Perhaps they got ill. Perhaps the economy boomed and many more options became available. There are many things that influence a human’s job decisions that will never be captured in your HR system.
But a manager can provide the input needed to make the results data meaningful - telling the system, say, to disregard this case, because our imaginary employee left to pursue a different career direction. Human beings have a vital role in producing meaningful results.
However, once we have those results, we can start to do some very powerful analysis with ML. In this example, we can start to produce really robust predictions about which types of action or intervention are likely to help us retain high performers. And that knowledge is gold dust.
Use cases in HR that add serious value
The results of this kind of ML work would help you answer many business-critical HR questions better. They let you answer much more robustly and effectively the question of ‘What works?’. Here are a few examples:
- Pay reviews. Do pay rises work to improve performance or retention? Which types of employees respond best to monetary rewards? Which responds better to something else?
- Retention. How do you spot employees or categories of employees at high risk of going, leaving you with tricky gaps to fill? What indicators best identify a risk of leaving? ML can bring a whole new dimension to this by finding indicators in text data (say employee survey responses) as well as structured numeric data (say a satisfaction score).
- Recruitment. What does a candidate likely to perform very well in the long term look like? Of course this too will need human input: an HR manager or line manager will have to give feedback after 6 months or a year on whether the hire was a success or not.
- Learning and Development. If an employee is struggling, what does the evidence suggest is most likely to bring about better performance?
- Costs of rehire. Predicting the likely financial costs of losing a high-value employee. These are often much greater than might meet the eye. As well as recruitment expenses, there are costs that come with training and loss of productivity as a new person beds in.
Key benefits of recommender ML in HR
- Better decision making. Instead of gut-feelings and subjective choices, ML allows managers to make more objective, data-driven and effective decisions. For companies with hundreds of employees, let alone hundreds of thousands, this is a necessity.
- Constant improvement in decision making. Since the model works as a form of virtuous circle, it constantly improves, adapting to changing patterns. By measuring the success of recommendations, those decisions become part of an updated and enriched set of data, on which the next predictions are based. The accuracy of future predictions is always improving.
- It helps you win the battle for talent. A good recommender system helps to make more informed and reliable HR decisions. Finding the most compatible candidates, advising on the best way to improve their performance, and helping to keep them: all will help a firm achieve an upward trend in the quality of staff performance.
So, is it worth it?
‘Netflixing’ crucial HR decisions, such as how to improve retention of high-performing employees, isn’t a quick bolt-on project. You need to design your HR systems and adapt your business processes in a way that allows you to get the quality data that will produce quality results.
The effectiveness of these types of systems is proven in other arenas. Netflix’s personalised recommender system is responsible for 80% of total stream time, with the company saying its ML models save it $1bn a year (as of 2016).
Of course, transformation projects cost money. But with the average costs of rehire at more than £30,000 the benefits in employee retention alone could be huge. By reducing a 15% turnover rate, the UK’s national average, to 10%, a company of even 1,000 employees could save £1.5m a year. One of 200,000 could save £300m.
As the likes of Netflix have already shown, the potential is vast.