Fuck your machine learning algorithm

In the few months I have between graduation and work, I’ve been volunteering for a high school entrepreneurship program, coaching 16 and 17 year olds on transforming an idea into a business. In the 10 week program, the participants brainstorm business ideas, hash out the details, and pitch to a panel of “investors”.

I participated in this program back when I was in high school, but this time around, I noticed something different among the participants — specifically their brainstorming. When I asked my team to identify a few problems in their lives and list solutions for them, here’s an excerpt of what they came up with:

An app that helps you pick out your outfit for the next day

An app that prevents you from losing one of your Airpods

An app that helps people quit smoking

While it certainly was an amalgamation of perhaps, interesting, ideas there was one shared commonality between them: they all fit the template of

“An app for xyz”

Now, what made this particularly interesting is the fact that when I participated in the program, none of the “envisioned companies” were app based. Perhaps this can be attributed to the fact that the app economy was still in its infancy when I participated. Today, the app economy has more than flourished, and has undoubtedly inserted its way into almost every element of our life.

In response to their ideas, the feedback I gave them was underscoring that an app was a means and platform for a solution, not necessarily a solution itself. I don’t blame the students for having this mindset. In fact, I don’t think this mindset goes away, and it’s heavily evident in entrepreneurship today:

Notice a pattern?

“machine learning for”

In fact, pick an any industry you want, and I guarantee that you’ll find a machine learning startup in that space. Here’s the bone I have to pick.

From my involvements within entrepreneurship, hackathons/business plan competitions, and peers I’ve engaged with in college, I can’t count the permutations of machine learning startups or ideas for machine learning startups I’ve come across:

And here’s an observation with a vast majority of them:

They are solutions chasing a problem

Machine learning is a tool. In fact, it’s a very powerful tool that operates on the principle of making decisions based on incoming data. Undoubtedly, ML has changed the way we analyze, predict, and operate, and will change our lives moving forward.

However, far too often, I see aspiring entrepreneurs starting with machine learning as the basis for their solution — not because machine learning is the best solution, most apt solution, or even the most feasible solution to the problem they are attempting to address, but because it’s the most obvious solution that comes to mind. Machine learning, in essence, uses what is known to answer questions and make predictions about the unknown. It’s a powerful tool with endless potential, and it’s understandable why many have the inclination to gravitate towards ML oriented approaches.

But here’s the reality — I don’t think most college students, especially computer science students, are taught how to properly problem solve. They are taught relational databases, shortest path algorithms, and everything in between, but when are students taught to properly define a problem space? Or how to understand the root causes of a problem and evaluate why this problem hasn’t already been solved before?

Far too often I see these students take a problem they are presented with, throw various tools at the problem and see what sticks. Here’s the challenge with that: you can’t properly solve the problem if you haven’t taken the time understand the full scope, context, and root of the problem. More importantly, it’s a solution chasing a problem when it should be the other way around.

When you have a hammer, everything looks like a nail

As an engineer that graduated from a predominantly engineering school, I think one reason behind this mindset is the fact that we are taught that the most intricate, detailed, and complex solutions are the best solutions. That the higher level of analysis and more heuristics you use, the more valuable your solutions has to be. I learned that this wasn’t the case from firsthand experience.

In my last semester at Georgia Tech, I had to take on a capstone project — find a client, identify a challenge that they’re facing, and solve it for them. For my capstone, my team worked with one of the country’s largest healthcare providers to reduce their no-show rates, i.e patients that book appointments but don’t show up.

After over eight months of work, we presented our client with three solutions, ordered below from most complex to least complex

1. A machine learning model that predicted which patients were likely not to show up and hence should be overbooked with other patients
2. A probabilistic/stochastic model that used simulation to determine the time of optimal overbooking slots
3. A cost benefit analysis of a new transportation system we recommended for implementation

Of the three solutions we developed, the client indicated that they were the most receptive to the third one, the least sophisticated of our solutions.

Now, this wasn’t because our other solutions weren’t functional — in fact, our ML model had a very high accuracy, precision, and recall. The issues they had with that solution in particular was threefold

1. Change management — implementing such a new tool for would require them to revise all of their current call center and scheduling workflows
2. Behavior — Staff would have to retrained and adjust their behavior and line of questioning with patients when scheduling through the use of this tool. Changing the behavior of others is much, much easier said than done
3. Maintenance — The model was trained on the historical data they provided us. In order to ensure that the model remained up to date, it needed to be connected to a back end and the data coming into the model needed to maintained. This was not a skillset that they had at the moment, and was consequently out of their capabilities.

For me, the biggest takeaway from that experience was that the best solution doesn’t always win.

You see this all the time in technology — no matter how good Bing becomes, people won’t stop using Google. No matter how much more ergonomic a DVORAK keyboard is, people won’t stop using a QWERTY keyboard. What wins almost every time is the product that’s the easiest to adopt without users having to change their behavior.

If you start with a machine learning algorithm without properly understanding the problem space you’re working with and fully understanding the nuances of the user, you’ll more than likely end up with the same results I did.

The point of this piece isn’t to undermine machine learning, or the incredible work many machine learning startups are doing, and have done.

It’s to shed light on the notion that we often define problems as being problems only because have the tools and know how to solve them.

A recent report from MMC shed light on the fact nearly half of European AI startups don’t actually have any form of Artificial Intelligence in them, yet they all received funding. This is a very blatant a example of investors buying into the hype of the solution instead of investing in the most effective way to address the problem at hand.

So why does this matter? This sort of public sentiment incentivizes quantity over quality when it comes to developing solutions.

And in a day and age where entrepreneurship is heavily romanticized, that’s the last thing we need.

Share your thoughts! @kamesh__