The failure of Predictive Algorithms

The failure of Predictive Algorithms

Most of the Big Data & AI Companies are based on the assumption that predictive analytics will be able to solve problems - which means they are inherently assuming that modelling works or works well enough for a company or a type of business problem. However poking the validity of such assumptions is what is really required , the very short answer to this is it is a very narrow view - most algorithms work narrowly in a way the company who built them or the one who is deploying them wants them to work but probably fails in a lot of other ways - even though it may be irrelevant to that company but is some way may be important to the targets of the algorithm or some other side effects of that algorithm.

A real question is that if a model works for one - does it work for others / all ? An example: Facebook news feed algorithm works for them - helps them make a lot of money - because it is optimized to its definition of success - but the definition of success but certainly when the definition of success changes as per a company it may not work .

A predictive Algorithm is an algorithm is one which predicts success( i.e a condition to be true) - it is trained on historical data (e.g 10 yrs, 20 yrs etc) looks for statistical pattern - so that it is able to learn from things which have happened in the past will predict will happen in future. So to be build a predictive model what is needed is past data and this definition of success .

Taking some real world examples - its no brainer . Let us take some day to day examples like

  1. Getting dressed in the morning - looking at memories in the head - what i wore and was i comfortable- is that a definition of success
  2. Social structure - cooking dinner for family --- looking at the memories - like this person likes carrots but when they are raw ? this person likes bread more than pasta or do kids like vegetables - is that a definition of success ?

So the question is the question is that are the predictive algorithms optimized to the success defined by their owners, builders or deployers - it is all about the power dynamic . When we are scoring people who are getting scored may not agree with the definition of the success may not agree to it but they don't get to vote - however the scorers are the ones who say what i mean by a good score ( eg - Insurance , credit score , hiring for job , sentencing to prison) . The algorithms actually are not standard but they are a trade secret - but people generally think algorithm is a a data / number crunching by a computer which is statistical so it would be always true - but this black box approach may not always be true for all business scenarios.

There are also such of the reasons why algorithms mess up

  1. Bad data - like Garbage in Grbage out
  2. definition of success - if the definition of success is not in line with the business problem even though the data is good



要查看或添加评论,请登录

Ramanpreet Singh Dandona (PMP?)的更多文章

社区洞察

其他会员也浏览了