By: Martín Rago
In this 3rd Industrial Revolution, the radical change is that we now have algorithms capable of learning by themselves.
We are participating in and witnessing the 3rd industrial revolution. In the 1st one, man's strength, his manual abilities, were replaced. With the extraction of fossil minerals, human power was artificially replaced. The 2nd was the development of production lines and today we have cars with 250 horsepower, electric pumps that draw water on farms to irrigate crops.
In this 3rd industrial revolution we are giving intelligence to that car to drive itself, to that water pump to draw the necessary amount of water taking into account the reserves, the weather, the time of the year and the time of planting.
It was Arthur Samuel, the creator of Machine Learning, who in 1956 wanted his computer to beat him at checkers, and after hours and hours of trying to teach it the best moves, he decided to make the computer play against itself so that it would learn by itself. In 1962 his computer won the Connecticut championship.
The beginning of artificial intelligence started with expert knowledge, where we have to tell the computer what to do in detail. In this 3rd industrial revolution, the radical change is that we now have algorithms capable of learning on their own, either with a training dataset or cold.
Algorithms that classify/predict without prior knowledge are called "Unsupervised" and those that require training are called "Supervised", where the training data consists of pairs of objects (usually vectors): one component of the pair is the input data and the other, the desired results. The output of the function can be a numerical value (as in regression problems) or a class label (fraud / normal transaction / suspicious transaction).
An important feature of supervised algorithms is "overfitting". This occurs when the algorithm is overtrained in the training set and loses the ability to generalise, i.e. it becomes very specialised in what is known in the training set and performs very poorly for any situation not foreseen in the training set.
If a model is over-fit it has poor predictive performance, as it overreacts to small fluctuations in the training data. Models have to generalise from the data presented to previously unseen situations.
These technologies are part of our DNA as a company and have been present since the gestation of Monitor Plus, and over the last 17 years we have incorporated both supervised and unsupervised learning techniques.
The Machine Learning algorithms we have developed and incorporated into MONITOR PLUS are: Naive Bayes Classifier, Dynamic Scoring, Online Data Mining, Adaptive Rules, K-Means Clustering, Deep Learning Neural Networks and Logistic Regression.
In conclusion, there is the premise that if one classifier algorithm is good... possibly using several is even better.
In this sense, these techniques are assembled to generate a single final classifier that provides the best performance in detection levels as well as in the ratio of False Positives.
