Monday 12 January 2015

The Mis-Match of Algorithmic Recruitment

It's the not so distant future.

A mobile app linked to a wrist mounted wearable wakes you, at precisely the right moment.  It monitors your sleep patterns and pulse rate and greets you each morning with a chipper "Go get 'em!".  You dress and get ready to leave the house, the fridge has emailed to remind you that you'll need to buy milk on your return.  You lock the door behind you with a swipe of your cell phone, keys are no more.  Outside, you step into a self driving car and take a different route to the usual commute - the car knew about the traffic before you did.  You arrive at work and boxes are moved into the previously vacant office next to yours.  You weren't aware of a new co-worker. There were no interviews. They were algorithmically selected from the passive talent pool.  Kept warm on a diet of Pinterest photos of the office and Youtube videos of kittens selected to be the most humanising for the Mega Corp you happen to work in...



As far as predictions of the future go the vision I offer above is hardly advanced.  The technology exists for the wearables, the Internet of Things and the self driving cars, it's just that last part that seems incongruent.

In the growing adoption of technology for HR departments seeking to differentiate their sourcing efforts, the idea of algorithmic matching is seen to be the magic bullet in the "War for Talent".  Beyond the clichéd war metaphors and gullibility of HR Tech buyers is the future of recruitment to be left to the robots?

Technology has made the discipline of talent acquisition better.  We've moved far beyond the data entry and green screen databases of a decade ago.  As a modern workforce migrates to online services so their digital footprint increases making them all the more easy for the new breed of sourcers to find.  Now the future, according to some, looks set to be the automated addition of new workers and a touted increase in the skill of selection.  I'm no Luddite but I can't help thinking this is a version of a technological utopianism whose primary supporters are those that seek to benefit financially from the adoption of the technology in question.

So many of the products available that claim to have solved matching are the same providers who don't recognise some of the fatal flaws that their products exacerbate. The primary example of this is the reliance on the quality of data on both sides necessary for a match.  The majority of matching systems are parsing CV's and then matching against a job description analysed in the same way.  This is exactly the limited key word matching that these systems say is so weak.  Even when other data are relied upon to beef up the input, suggestions of LinkedIn profiles and even LinkedIn endorsements are laughable. Especially in the case of unverifiable LinkedIn endorsements like mine for "Midwifery" and "Cheese Making".  Of course I'm totally brilliant at both of these things...

Even the more advanced of the matching algorithms that incorporate some elements of semantic search (context of search, location, intent, variation of words, synonyms, generalised and specialised queries, concept matching and natural language processing) are constrained both by the data the candidates provide and the job description or criteria the employer matches against.  Anyone who works in recruiting will be able to quickly see that both of these sources of data are flawed and subject to constant change.  Data in both these areas can be knowingly falsified, incomplete and always out of date.

This data is inherently flawed because people themselves are inherently flawed.  Candidates will always seek to portray themselves in the best light, hiring managers will always add some extra "nice to haves" or even make the work of two people into one mythical job description.  A matching algorithm is forced to make sense of too many moving parts and results will suffer.

In moving towards this style of recommendation the people in the processes are reduced to the status of commodities.  Subtle nuance is lost and the chance for innovation curtailed by inelastic parameters.  People are not a product.  When Amazon presents you with a book based on your buying preferences it has only to reckon with your fickle, transient tastes.  A book doesn't reject you because it feels it's too far to get to your house, or because the other books on the shelf don't feel your reputation is strong enough, a book doesn't want to work from is own home or have a counter offer from a series of rival readers...people do.

Recruiting is a constant stream of edge cases.  Whilst a matching engine might work for less complex roles at large numbers, it won't help you compete in winning that all important "War for Talent" you were so desperately spending your way out of.  The current level of technology is no match for the ability of a good recruiter.  This is not an indictment of the technology, it's an acknowledgement of the greater problem that exists in the institutionally flawed HR departments and Recruiting processes the world over.  Using a tool like this to gain another datapoint to inform decision making is a valid use - it's the shame of HR Tech that every new tool is paraded as "the answer".  If the industry could wean itself off it's obsession with the novel and shiny we might be able to tackle some of these issues at the root cause and realise that the skills we learnt whilst toiling at our green screens might not be entirely redundant.

3 comments:

Vincent L said...

Well, I can't agree this time. Algorithmic recruitment doesn't work not because "The current level of technology is no match for the ability of a good recruiter" but because the dataset is too small. If there were a few billion people a day getting hired and "let go" (or "objectively rated in their job" - lol), you could just let the machine learning algo spin and find a new job :-)

Unknown said...

The dataset is both too small and a large amount of HR "data" is in fact not a real metric at all. Any performance review is always subjective and when translated into a number is, apparently, instant science!

Building the dataset would be almost impossible too, what employers would want to share all their employee data with each other and even if they did there's still the privacy issues that would bring with it...

The Algos might be good enough but the only people who think it's the ultimate answer are the one's selling it!!

Unknown said...

I've always been very suspicious of solutions that boastfully quote figures of how effective their matching tools are. As you mention, the data they're analysing isn't even close to being sufficient to give you a genuine insight into how suitable someone will be for your company.

The only way I see matching being effective is if there was infinitely more data freely available on each candidate - forget LinkedIn bio, think LinkedIn book :) Then we might be onto something...