What are the best practices for evaluating machine learning model performance?

Another point: Are we solving the right problem and getting really important outcome which is non-obvious? Or difficult for human being to perform the task? Or saving time and cost? Sometimes even using AI / ML we get some obvious results although driven by data, but can’t help the end client or consumer. For example: Say, someone crafts a classification algorithm to find the most important feature (variable) of cardiac arrest (heart attack), and then after taking hundreds of parameters with thousands of records, the result says: “high cholesterol levels”. Now, any cardiologist would be interested in this ML model and purchase the product? The answer is “no”, because the result is obvious and no one needs an ML model for this. Think…

Machine Learning in general is a problem-solving technique. The first you should do before applying any type of problem-solving method is identifying what the problem is, why is it a problem and learn as a much as you can about that problem. Ask yourself questions like: Why is this problem? What type of solution are we looking for this problem? Is this problem a numeric problem? Is the problem associated with determining the correct class for each instance? Having a properly defined problem and the boundaries of what you expect to achieve by solving it, will give you the information you need to properly evaluate what methods to use for ML, as well as what type of metrics to look into to assess if your solution has been effective.

One of the main reasons to thoroughly study the problem you're addressing before applying AI or ML is that you need some basic expertise in the area to build your databases and train your models effectively. Choosing the right data for your problem is crucial. Each descriptor you include in your final model adds a dimension of complexity to both the model and its interpretation. Ideally, models should be built with as few descriptors as necessary. This approach is usually feasible only when you have sufficient expertise in the relevant field, enabling you to pre-select the descriptors that are most informative about your problem. Remember, having the right data is fundamental!

Choosing the right metric is super important. In the context of classification problem where class probabilities are important, one should consider log-loss and brier score instead of roc-auc, f1, etc. Otherwise, class probability scores maybe quite off.

Depending on the type of problem you are working on, you will be examining different types of metrics. For regression problems, you will likely focus on an R2-related metric, which often shows how well your model properly describes the variance in your data. If you are working on a classification task, all you need is the Confusion Matrix for your model. From there, you can derive other metrics such as F1, ROC, Precision, Recall, MCC, and Kappa scores. These metrics are used to assess the overall effectiveness of your model and for making comparisons between different models. Please keep in mind that good model metrics may not be very meaningful if statistical validation is lacking.

Imagine you have 10 fruits, with 5 apples and 5 oranges, ordered all the apples on the left in a line directly followed by the oranges. This is your data. Lets say we wanted to train our predictive model with 7 fruits and test it on the remaining 3. If we just took data from left to right until we reached 7 fruits, our training dataset would have 5 apples and 2 oranges. The model we train based on this will learn that there are way many more apples than oranges, eventhough we know for a fact this is not true, and the ratio is 50/50. This error is called selection bias and to avoid it you may want to use techniques like Cross-Validation, where you reshuffle your dataset a few times and train various models and take the average of them.

The idea for a ML learning model is to help us humans learn or solve problems in a better way. We must be able to learn from our models. This is done by carefully selecting the descriptors to be used, applying the proper algorithms for your problem, and thinking about your models results in the context of what you already know about the problem and what you would have expected for a valid result to look like, among other considerations. As with any knowledge, it is vital to also develop a proper storytelling for non-technical people to also be able to understand and learn from your results.

What Thomas Husslein suggested is extremely important and is often neglected by both newbies and experts in the field. It's common to think that certain data work is not complex and that each step taken is obvious. However, this is hardly ever the case. Most of the time, you will need your model to be replicated, evaluated, updated, and maintained. For all these instances, it is critical to have (at least) a clear and concise workflow. This should explain where the data is coming from and any initial transformations you may have done to the data. As with any scientific process, keeping a clean track of tasks and results is a key aspect of all work. Such practices ensure accuracy and reliability.