Deciding on a method

From Machine Learning
Jump to: navigation, search

There is no one size fits all when it comes to algorithms in machine learning. The best method for a given problem is highly dependent on the desired goal and the type and quality of available data.

Which algorithm generates the best model for this problem? It is important to consider accuracy and latency.

Accuracy involves how the algorithm handles the data, whether it is tall and skinny or short and fat, if it is messy with outliers and if it has many complex relationships.

Latency is concerned with how long it will take for the model to train and how long it will take to make predictions following training.

As well as considering what kind of data you have, what is the goal of the machine learning model? The types of machine learning page has a good mind map showing what area of machine learning different applications require.

Choosing a method.png

Read this for a clear overview of what to watch for when preparing data and an indication of how to choose the method. Microsoft Azure has a nice table of comparison for different methods as well.

Baseline Results

Once you have chosen a method, or shortlisted a few, it is possible to conduct tests and analysis to help decide which method suits the data best, or whether the method will be detailed or accurate enough for your needs.

Developing a baseline of performance of the dataset is important when developing a machine learning algorithm. A baseline prediction algorithm allows for comparison of different algorithms to determine which one will be best to develop further into a complex model[1]. It is a very simple algorithm that should not take long to implement.

At its most basic (for simple classification or regression problems), a baseline algorithm could use simple maths and statistics to determine performance metrics.

For more complex problems, two common approaches are random prediction and zero-rule.

References