Before embarking on the journey of collecting and cleansing data, and training and deploying the models, Product Managers (PM) must frame the problem as precisely as possible.
Neal Lathia in his article “Machine Learning for Product Managers” has made the effort to describe the six types of problems that nearly every AI product should focus on.
The six types of problems
1) Our product helps users find the right information – you’re solving the ranking problem
The common search engines are the best examples here. Whenever you hit the search button, the Google’s and the Yahoo’s try to rank the results according to your preferences.
2) Our product shows users things they could be interested in (without actively searching) – you’re solving the recommendation problem
Companies who's software platforms use these algorithms often want to keep you engaged. This engagement helps to generate more data and revenue. Sites using it are Netflix or social platforms such as Twitter and Facebook.
3) Our product helps to understand what kind of thing something is – you’re solving the classification problem
One of the best examples is spam detection. Most of the email providers like Gmail classify every incoming mail and flag it either as spam or not spam.
Another example is face detection that detects faces in photos and tags them appropriately (e.g. Facebook). This use case, however, has been highly criticized in the past months because face detection is often biased, and when used for e.g. crime prevention can cause severe consequences.
4) Our product predicts a numerical value of something – then you’re solving the regression problem
5) Our product is putting similar things together – you’re solving the clustering problem
The algorithms used for this problem group similar objects together based on a set of features. When you buy stuff on Amazon, you’ll most likely see this algorithm in action (“Other Customers bought also” or similar).
Similar action happens on music streaming apps (Spotify).
6) Our product finds uncommon things – you’re solving the anomaly detection problem
It’s about uncovering trends. Platforms such as Twitter (“Trends for you”) use it. But also financial services might use anomaly detection algorithms for fraud prevention (e.g. credit cards)
The ranking, recommendation, classification, and regression problems are examples of supervised learning (predicting outcomes). The clustering and anomaly detection belong to the unsupervised learning domain (identifying patterns).
By identifying the problem, the next stages of developing an AI product will be much easier.
If you’re a buyer of such products, you might be interested in this article.