Auto ML. A different approach to data science

As we can see, we can distinguish three stages:

1. Importing and cleaning.
2. Scanning and modelling.
3. Communication.

Depending on the type of source data and the result we seek to achieve with this data, the modelling process may vary. However, regardless of the model, the data scientist must be able to obtain a clean dataset ready to serve as input to the model. In this post we will focus on the second stage: exploration and modelling. 

Once this clean and error-free data has been obtained (after import and cleaning in step 1), the data scientist must decide which transformations to apply to the data, with the aim of making some data derived from the originals (in conjunction with the originals), the best indicators of the model underlying the dataset. We call these transformations features.

The next step is to divide our dataset into two parts: one part, for example 60% of the total dataset, will serve as the training dataset. The remaining 40% will be reserved for applying our model, once it has been trained. We call this second part the test subset. This process of splitting the source data is done with the intention of assessing the reliability of the model before applying it to new data unknown to the model. An iterative process now unfolds in which the data scientist tests various types of models that he/she believes may work on this dataset. Each time he/she applies a model, he/she observes and measures the mathematical parameters (such as accuracy and reproducibility) that express how well the model is able to reproduce the test data. In addition to testing different types of models, the data scientist may vary the training dataset with new transformations, calculating new and different features, in order to come up with some features that make the model in question fit the data better.

We can imagine that this process, repeated dozens or hundreds of times, is a major consumer of both human and computational resources. The data scientist tries to perform different combinations of algorithms, models, features and percentages of data, based on his or her experience and skill with the tools. However, what if it were a system that would perform all these combinations for us and finally come up with the best combination? Auto ML systems have been created precisely to answer this question.

In my opinion, an Auto ML system or tool is not intended to replace the data scientist, but to complement him or her, helping the data scientist to save a lot of time in the iterative process of trying different techniques and data to reach the best model. Generally speaking, we could say that an Auto ML system has (or should have) the following benefits for the data scientist:

1. Suggest the best Machine Learning techniques and automatically generate optimised models (automatically adjusting parameters), having tested a large number of training and test datasets respectively.
2. Inform the data scientist of those features (remembering that they are transformations of the original data) that have the greatest impact on the final result of the model.
3. Generate visualisations that allow the data scientist to understand the outcome of the process carried out by Auto ML. That is, to teach the Auto ML user the key indicators of the outcome of the process. 
4. Generate an interactive simulation environment that allows users to quickly explore the model to see how it works.

Finally, we mention some of the best-known Auto ML systems and tools, such as H2O.ai, Auto-Sklearn end TPOT.  It should be noted that these three systems cover the entire Machine Learning process that we saw at the beginning. However, there are more solutions and tools that partially cover some of the steps of the complete process. There are also articles comparing the effectiveness of these systems for certain machine learning problems on open and accessible datasets.

In conclusion, these tools provide valuable solutions to common data science problems and have the potential to dramatically improve the productivity of data science teams. However, data science still has a significant art component and not all problems are solved with automation tools. We encourage all algorithm alchemists and data craftsmen to continue to devote time and effort to developing new techniques and algorithms that allow us to turn data into value quickly and effectively.

The aim of this post is to explain to the general public, in a simple and accessible way, how auto ML techniques can simplify the process of advanced data analysis. Sometimes oversimplifications may be used in order not to overcomplicate the content of this post.