Por: Pedro Castellares Torres, Machine Learning en el Massachusetts Institute Technology. 

Abstract

The greatest value a company can obtain by using data mining or using deeper information analysis is related to predicting certain scenarios with greater accuracy, but one of the most common mistakes for organizations that contemplate implementing Machine Learning or Data Mining, has to do with two extremes. One involves thinking that: “We're not sure if this is really going to work” and the other that “it is going to work perfectly.”

We must keep in mind that the world of Machine Learning is probabilistic, not deterministic. So, it is important to understand that we are changing from a paradigm F(Xi) = C+aXi + bXi + ... +zXi, to another where an algorithm is trained. In other words, Machine Learning focuses on looking for patterns in order to make predictions. 

The algorithms used depend largely on the type of data being analyzed and the result we are trying to predict or analyze. Therefore, each process is different, where we have some data to analyze and a deep exploration of that data is required. In that sense, you are not the one who decides which algorithm to use, but it is the data that determines the algorithm that produces the results based on the existing data.

To estimate the reagent dosage and improve recovery in any flotation process, it is possible to apply the Regression Decision Tree Algorithm by Supervised Learning, since it is a simple way of representation to find homogeneous groups according to a certain response variable. 

This technique allows to represent graphically a series of rules about the decision to be taken according to a main characteristic defined by the algorithm (primary node) and can be applied for the following variables in a flotation process, such as:

ν Primary collector dosage.

ν Secondary collector dosage.

ν Dosage of primary foaming agent.

ν Dosage of secondary foaming agent.

ν Lime dosage.

ν Dispersant dosage.

ν Solids % in Ro Scv pulp.

ν P80 to the flotation circuit.

ν Ore type, etc.

This algorithm can be applied in different software, such as Python, Studio R, MatLab, C/C++, Cart Regression, Xlstat, etc.