Problem Types

There are many ways to hierarchically group Machine Learning problems, but lets start with either: Insights vs accuracy.

​

Insight Gaining

Many machine learning models are designed to be easily interpretable. In those cases, you fit a Machine Learning model to data with hopes to analyze what the model learned to gain insights. These problems usually start with business people wanting to understand a dataset. The project is identified for Machine Learning because standard statistical techniques cannot deliver the type of insights they are looking for.

 

We, the MLEs, then design a machine learning model that is capable of surfacing the insight types they are looking for. We can either use an off the appropriate shelf package or implement a custom model if necessary using a customization tool set such as Tensorflow or PyTorch. Common questions consist of identifying common user personas or quantifying significance of positive or negative effects on some outcome. Common tools consist of clustering, custom regressions, or decision trees.

 

Building Accurate Models for Production

The converse is building a Machine Learning model to solve a task. These problems start with a business problem such as counting people in a video, routing users from natural language messages, or recommending products to users. The MLE will usually frame the problem as ingesting data objects and outputting decisions, and a loss function to evaluate how accurate their decision is. The MLE then uses an optimization package to minimize their loss function to solve the task better. 

 

The first significant time commitment for the MLE involves experimentally customizing the model representation to fit the dataset. If there are users, allow them to be represented individually. If there is a geo component, represent it. If there is a temporal component, try to represent it. A good MLE will know that standard academic models are mostly ideas that can be combined for specific problems, such as convolution connections to handle spatial relations within their recurrent models that handle sequences.