Ask HN: What market demands can I solve with machine learning?

14 points by curiously ↗ HN
so while I'm learning machine learning I want to pursue some problems that are in demand by the market.

One example I can think of is a sentiment analysis on social media to measure some degree of product satisfaction. However, at one point can you confidently state that you've found the appropriate product satisfaction? What if there were errors along the way (NLP not working properly or false positives)? How would you consolidate such things to a client? Social media related machine learning seems dime a dozen anyhow.

I guess I am trying to employ the same philosophy to when I was learning how to program. Create something that is in demand by the market to best learn the trade. I want to do the same thing for machine learning, computer vision, but am not able to visualize what potential problems that are in demand by the market are worth pursuing.

So, what immediate problems do businesses or industries face that I can solve with machine learning?

6 comments

[ 27.2 ms ] story [ 928 ms ] thread
Machine learning is a hot topic today, like AI in general.

Facebook and Google use Machine learning algorithms to increase Ad revenue and serve more relevant information to the user.

Machine learning is also used in health care to predict Emergency Room wait times for example.

Many banks and big online retailers use Machine learning for fraud detection.

And of course Machine Learning is used in the stock market.

this is a very broad, top down look at the problems that are being solved but I want to know a bottom up, an specific detail or the problems that is good to tackle.

>ML in stock market

I know this is used here but what exactly? Mine tweets to predict stock prices with a contrarian strategy? Computer vision on stock charts to apply technical analysis trading system? Mine historical volatility of options on futures?

It would be great if there was some simple business or SaaS I can execute using machine learning.

But say you have a website now with 'We use ML to solve X". Then the next bigger challenge, convincing a business to fork over their data for the sake of machine learning without previous experience. Or do you have to work for free in the beginning to build credentials? 'Hey if you hand over your patient data at your ER, we'll figure the waiting time out, never done this before with another ER so we'll do it for free?'

I guess I'm interested in the applying entrepreneurial and business approach to solve an actual real world problem that is in demand using ML.

Anything that is manual, tedious and error-prone. Or requires quick reaction time.

A "silly" example might be clinical diagnostic decision support. Instead of charts with history, lab results and meds to visually wade through and diagnose, there could be a simple UI. "Simple" means simply that instead of potentially 1000 fields that might be read or selectively updated, there would be 5 pre-selected by the computer. This could then go from medical unstructured "narrative" to automated ICD-10 coding. There would be lots of NLP, semantic analysis of large corpii, formal ontologies and RDF-encoded KB tangles to work though.

> Computer vision on stock charts to apply technical analysis trading system?

Why would you use computer vision? Just get a DVD full of the intra-day tick data for the instruments of interest going back 20-or-so years and find correlations.

And, I'm sure you know this, but a reader wandering through here may not: many of the kaggle competitions can be solved with ML techniques and many-or-most of the solution types would be valuable to the market or to large (hopefully funded) research problems. https://www.kaggle.com/

How to determine what restaurants a person will like based on their past ratings. This is not trivial; it is difficult to say what are the features which determine why a person likes a restaurant, so I'm thinking this would be a problem that could be tackled with deep learning i.e., automatic learning of features rather than hand-crafting in advance.

This is similar to the problem of movie/book recommendations, which has not been solved to my satisfaction. I have yet to see any good recommendation systems out there which go beyond recommending movies based on simple things like 'same actor was in this movie', 'people who watched that watched this' etc.

that's a great point, however the fact that those haven't been solved yet suggests that the problem is actually very difficult. These type of problems I would shy away from because the results wouldn't be consistent.

For example, catching a fraudster based on their actions for your website is easy because you can create models with scope limited to your own website with few constraints vs. catching any fraud on any website because they all have different processes, scopes and constraints wildly fluctuate.

The problem with recommendations in general is that they are subjective to people's feelings, often times at a particular point of time.

This is applicable to movies and restaurants. Sometimes I'm in a mood for an action movie but that doesn't mean I want to be seeing those all the time around. Similarly, there is simply time when I'm in a mood for mexican or Indian and not Italian where I may have been going to for past month.

Unless we manage to capture the inner motivations/feelings (maybe based on our online behavioural patterns?) our recommendations will just be artificially built mashed potatoes crap.