Cutting Edge Computing

But For the Want of An Algorithm November 21, 2021

I’ll bet you never heard that statement before.  Algorithms are the basis of machine learning and decision-making today.  Recently, for the want of an accurate and representative algorithm, Zillow is laying off a quarter of their staff and exiting its home-buying and selling business.

To be fair, it’s not all about the machine learning algorithm(s) that it used to determine what homes to buy and what to pay for them.  The algorithm was the starting point for an in-person inspection and negotiation by actual people.  But the algorithm overpriced homes across the country, resulting in Zillow losing millions of dollars and not able to sell many homes at the list price.

But it does illustrate overt dependence on algorithms to represent a particular problem domain.  And home prices are particularly susceptible to local conditions.  One algorithm does not fit all.

There are a number of ways that algorithms can be biased.  They may not fully represent the problem domain, or the domain may have changed since the algorithm was developed.  In this case, probably both of these contributed to the problem.

So there are dangers in using algorithms to represent domains and scenarios, dangers that can cost a great deal of money.  This illustrates the importance of testing; not just testing, but the testing and evaluation of algorithms.  And this seems to be a step that Zillow didn’t take.