Frame Your Use Case Before You Aim Machine Learning
With the mainstreaming of Machine Learning in many contexts, there is such a bewildering array of algorithms and approaches that it is becoming difficult to know where to start when faced with a particular business problem. In our experience, framing your use case is an important trick that can help streamline the search for a solution. Almost all unstructured and structured data analysis problems can be framed as one of the following:
- Value estimation
- Classification
- Anomaly/Outlier Detection
- Recommendations
Ask yourself what questions the ML model is expected to answer to deduce which of the above use case types patterns your problem. Once categorized, the next step is try to one or more of the popular algorithms for each use case as depicted below. Build intuition about the problem and data using these algorithms. It is quite possible that you achieve sufficient accuracy just with these. As the table below shows, there is significant overlap between algorithms used in different use cases - you might find one that always works for you! In our experience, advanced techniques like Neural Networks need to be explored only when these entry-level approaches fail.
