Machine learning supervized algorithm to detect fake money. ONCFM is french for Organisation Nationale Contre la Fausse Monnaie.
It could be fastidious for a human being to detect by it-self if it's either a fraud or a conform banknote. Even if credit card payment system is the most used and sophisticated way to pay, especially with e-commerce's rises, still merchants (bookstore, grocery, supermarket...) use traditional old-fashion means of payment : cash. For this reason, I designed and built an artifical intelligence ables to identify the true nature of a traded banknote.
In this section, I will explain what the data looks like before feeding and training the machine learning models. Before starting with the data presentation, there are 2-differents ways of thinking.
- Use of image detection machine learning algorithm to split on fake/true banknote samples.
- Use of banknote dimensional parameters to determine the perfect fake/true banknote profil.
For the purpose of this project, I worked with the banknote dimensional parameters. Indeed, data sample that I used shaped like banknote for rows and 6-dimensional parameters for columns with an is_genuine
extra-column. Feel free to check-out the whole data csv here.
Every american banknote is described as follow :
height_left
: banknote left vertical side in mmheight_right
: banknote right vertical side in mmmargin_low
: distance between US president picture and lower edgemargin_up
: distance between US president picture and upper edgelenght
: banknote horizontal size in mmdiagonal
: banknote diagonal size in mm
For remaning, we used dimensional data featured that allows us to not build an image-based but regression-based artificial intelligence instead. Please take a look at the top of my classification models notebook to get well-informed on logistic regression method. You can go straight to the result and see how well our regression logistic is performed
In addition, I tested clustering approach in order to improve my model forecast accuracy and compare it with the previous methods. Even if both of these methods have good results, logistic regression is the most suitable model for my data.
Comment : Profiling between true and false banknote's nature show us that margin_low
and lenght
are both good features for our models predicting the correct answer.
Accuracy speaking, I let you take a look at logistic regression and clustering models performance with confusion matrix I made to compare these 2-differents artifical intelligence approach.