Back

Finance & Insurance, Security

Fraud Detection

fraud.jpg

Financial fraud amounts for considerable money.

Hackers and crooks around the world are always looking for new ways of committing financial fraud every day.

Relying exclusively on rule-based, conventionally programmed systems for detecting financial fraud does not provide the appropriate response time.

This is where Machine Learning brings a unique solution for this type of problem.

This case study is about building a predictive model to detect fraud.

 

Problems to solve

How to detect a fraud before it’s too late?

In a case of fraudulent transactions, how to investigate without upsetting innocent people ?

How to predict weak signals within a large amount of data ?

Can machine learning help in these matters and how accurate predictive models can be to detect frauds ?

 

Solution

Like humans, machines can learn to make predictions by analyzing past information (historical data). Machines can quickly identify patterns in a set of data and produce a mathematical formula (model) using the variables from this historical data.

In order to demonstrate the performance of MyDataModels solution for this type of problem we choose a specific case study.

 

Case Study

This case study is based on real data from a public dataset* which can be found from Kaggle or UCI *.

https://www.kaggle.com/mlg-ulb/creditcardfraud

The objective in this case study is to build a model to detect fraudulent transactions

* see detailed information on this dataset at Dataset information section

The screenshot below shows an extract of the public dataset.

Each line is a transaction and each Column (aka feature) is a variable which can be used in

dataset.png

The dataset contains transactions made by credit cards in September 2013 by european cardholders. This dataset presents transactions that occurred in two days, where we have 492 frauds out of 284,807 transactions, therefore this dataset is highly unbalanced.

It contains only numerical data which are the result of a PCA transformation. Unfortunately, due to confidentiality issues, we cannot provide the original features and more background information about the data. Features V1, V2, ... V28 are the principal components obtained with PCA, the only features which have not been transformed with PCA are 'Time' and 'Amount'. Feature 'Time' contains the time in seconds between each transaction and the first transaction in the dataset. The feature 'Amount' is the transaction amount. Feature 'Class' is the response variable and it takes value 1 in case of fraud and 0 otherwise.

Since we want to prove the performance of MyDataModels’ technology on Small Data, we reduced the original dataset from 285.000 transactions to 6.026 and frauds from 492 to 147.


Dataset information

  • Task: Binary Classification
  • Number of features: 32
  • Size of data: 6026 samples
  • Weight: Positive class (fraud) 8%, Negative class: 92%
  • Target variable: (Class) Is it a fraudulent transaction ? (yes or no).
  • Score: True Positive Rate

Given the unbalance of the dataset the best way to evaluate the performance of the model is not to look at the usual accuracy but the True Positive Rate (TPR) since fraud is the positive class.


Used features

  1. Random
  2. Time : Number of seconds elapsed between this transaction and the first transaction in the dataset
  3. V1-V28 : may be result of a PCA Dimensionality reduction to protect user identities and sensitive features(v1-v28)
  4. Amount : The amount of transaction
  5. Class : 1 for fraudulent transaction, 0 otherwise

 

Results

To create predictive models MyDataModels has a performant solution called TADA.

In order to build a model, professional need to :

  1. upload the historical dataset into TADA
  2. set what they want to predict (here “class”)
  3. select the other variables to use (all the other columns)

Below here are the statistics of a model obtained with TADA in 10 minutes.

metrics.png

ACC = Accuracy
TPR = True Positive Rate
TNR = True Negative Rate
MCC = Matthew’s Correlation Coefficient


Confusion matrix
confusion_matrix.png

How to use this model?

Once a model is built from historical data, professionals can easily make predictions by using information from the transactions.

To use a model, domain experts need to create a new dataset with:
As rows, all the transactions they want a prediction for.
As columns, all the variables used in the model.

Then they need to import this file into TADA (see screen below) and click on “Generate score”.

TADA.png

A new file will be generated with a new columns giving the predictions of the type of transactions : 1 (fraud) or 0 (Regular)

 

Benefits of TADA

Most professionals from Banking & Insurance industries can benefit from predictive models but are not data scientists and may not have the required skills in machine learning nor coding to build them. Even if the data handled by these professionals could be considered as Big Data (transactions for example) the risks they want to predicts are Small Data, frauds in this specific use case.

In this case, historical data contains a limited number of frauds compared to the number of transactions. Traditional machine learning tools work well with Big Data but do not perform as well for prediction of Small Data within Big Data (imbalanced dataset).

MyDataModels allows domain experts to build automatically predictive models from Small Data. No training is required and raw data can be used, no need to normalize it, handle outliers, no feature engineering is required. Thanks to this limited data preparation and in few clicks the above results from this specific dataset were obtained in 10 minutes using a standard laptop.

MyDataModels brings a self-service solution for those who have Small Data and no data scientists.

 

Conclusion

According to a report published by Nilson, in 2017 the worldwide losses in card fraud related cases reached 22.8 billion dollars and is expected to reach 32.96 billion dollars by 2021.

Frauds are costly but it is too expensive and inefficient to investigate every transaction. Investigating innocent customers can be a bad customer experience, leading to greater customer churn. By accurately predicting which transactions are likely fraudulent, banks can significantly reduce these illegal transactions while providing card holders an excellent customer experience.

In this fraud detection use case, the results obtained from MyDataModels’ predictive models are satisfying with a 99.3% True Positive rate.

Regardless the industry, fraudulent activity is a high-cost threat that can compromise the integrity of companies as well as hurt bottom line. The use of fraud detection analytics using an automatic machine learning solution like TADA enables companies to discover fraudulent activity before it’s too late.

 

Dataset information

The dataset contains transactions made by credit cards in September 2013 by european cardholders. This dataset presents transactions that occurred in two days, where we have 492 frauds out of 284,807 transactions, therefore this dataset is highly unbalanced.

It contains only numerical input variables which are the result of a PCA transformation. Unfortunately, due to confidentiality issues, we cannot provide the original features and more background information about the data. Features V1, V2, ... V28 are the principal components obtained with PCA, the only features which have not been transformed with PCA are 'Time' and 'Amount'. Feature 'Time' contains the seconds elapsed between each transaction and the first transaction in the dataset. The feature 'Amount' is the transaction Amount, this feature can be used for example-dependant cost-senstive learning. Feature 'Class' is the response variable and it takes value 1 in case of fraud and 0 otherwise.

Since we want to prove the performance of MyDataModels’ technology on Small Data, we reduced the original dataset from 285.000 transactions to 6.000 and frauds from 492 to 147.

Dataset information

  • Task: Binary Classification
  • Number of features: 32
  • Size of data: 6026 samples
  • Weight: Positive class (fraud) 8%, Negative class: 92%
  • Target variable: (Class) Is it a fraudulent transaction ? (yes or no).
  • Score: True Positive Rate

Given the unbalance of the dataset the best way to evaluate the performance of the model is not to look at the usual accuracy but the True Positive Rate (TPR) since fraud is the positive class.

 

Get started for free

Detailed informations

General

Artificial intelligence: Theories and techniques aiming to simulate intelligence (human, animal or other).

Binary Classification: It is the problem type when you are trying to predict one of two states, e.g. yes/no, true/ false, A/B, 0/1, red/green, etc. This kind of analysis requires that the Goal variable type is of type CLASS. Binary Classification analysis also requires that there be only 2 different values in the Goal column. Otherwise, it is not a binary problem (two choices and no more).

Convolutional Neural Network: This type of network is dedicated to object recognition. They are generally composed of several layers of convolutions + pooling followed by one or more FC layers. A convolutional layer can be seen as a filter. Thus, the first layer of a CNN make it possible to filter the corners, curves and segments and the following ones, more and more complex forms.

Data Mining: Field of data science aimed at extracting knowledge and / or information from a body of data.

Deep Learning: Deep Learning is a category of so-called "layered" machine learning algorithms. A deep learning algorithm is a neural network with a large number of layers. The main interest of these networks is their ability to learn models from raw data, thus reducing pre-processing (often important in the case of classical algorithms).

Fully Convolutional Networks: An FCN is a CNN with the last FC layers removed. This type of network is currently not used much but can be very useful if it is succeeded by an RNN network allowing integration of the time dimension in a visual recognition analysis.

GRU (Gated Recurrent Unit): A GRU network is a simplified LSTM invented recently (2014) and allowing better predictions and easier parameterization.

LSTM (Long Short-Term Memory): An LSTM is an RNN to which a system has been added to control access to memory cells. We speak of "Gated Activation Function". LSTMs perform better than conventional RNNs.

Machine learning : Subfield of Artificial Intelligence (AI), Machine Learning is the scientific study of algorithms and statistical models that provides systems the ability to learn and improve any specific tasks without explicit programming.

Multi Classification: Classification when there is more than two classes in the goal variable, e.g. A/B/C/D, red/orange/green, etc.

Multilayer perceptron: This is a classic neural network. Generally, all the neurons of a layer are connected to all the neurons of the next layer. We are talking about Fully Connected (FC) layers.

RCNN (Regional CNN): This type of network compensates for the shortcomings of a classic CNN and answers the question: what to do when an image contains several objects to recognize? An RCNN makes it possible to extract several labels (each associated with a bounding box) of an image.

Regression: Set of statistical processes to predict a specific number or value. Regression analysis requires the type of Goal variable to be numeric (INTEGER or DOUBLE).

Reinforcement learning: Reinforcement learning is about supervised learning. It involves using new predicted data to improve the learning model (calculated upstream).

RNN (Recurrent Neural Networks): Recurrent networks are a set of networks integrating the temporal dimension. Thus, from one prediction to another, information is shared. These networks are mainly used for the recognition of activities or actions via video or other sensors.

Semi supervised learning: Semi-supervised learning is a special case of supervised learning. Semi-supervised learning is when training data is incomplete. The interest is to learn a model with little labeled data.

Stratified sampling: It is a method of sampling such that the distribution of goal observations in each stratum of the sample is the same as the distribution of goal observations in the population. TADA uses this method to shuffle the data set from binary and multi classification projects.

Simple random sampling: It is a method of sampling in which each observation is equally likely to be chosen randomly. TADA uses this method to shuffle the data set from regression projects.

Supervised learning: Sub-domain of machine learning, supervised learning aims to generalize and extract rules from labeled data. All this in order to make predictions (to predict the label associated with a data without label).

Transfer learning: Brought up to date by deep learning, transfer learning consists of reusing pre-learned learning models in order not to reinvent the wheel at each learning.

Unsupervised learning: Sub-domain of machine learning, unsupervised learning aims to group data that are similar and divide/separate different data. We talk about minimizing intra-class variance and maximizing inter-class variance.


Metrics

Binary

ACC (Accuracy): Percentage of samples in the test set correctly classified by the model.

Actual Negative: Number of samples of negative case in the raw source data subset.

Actual Positive: Number of samples of positive case in the raw source data subset.

AUC: Area Under the Curve (AUC) of the Receiver Operating characteristic (ROC) curve. It is in the interval [0;1]. A perfect predictive model gives an AUC score of 1. A predictive model which makes random guesses has an AUC score of 0.5.

F1 score: Single value metric that gives an indication of a Binary Classification model's efficiency at predicting both True and False predictions. It is computed using the harmonic mean of PPV and TPR.

False Negative: Number of positive class samples in the source data subset that were incorrectly predicted as negative.

False Positive: Number of negative class samples in the source data subset that were incorrectly predicted as positive.

MCC (Matthews Correlation Coefficient): Single value metric that gives an indication of a Binary Classification model's efficacy at predicting both classes. This value ranges between -1 to +1 with +1 being a perfect classifier.

PPV (Positive Predictive Value/Precision): Number of a model's True Positive predictions divided by the number of (True Positives + False Positives) in the test set.

Predicted Positive: Number of samples in the source data subset predicted as the positive case by the model.

Predicted Negative: Number of samples in the source data subset predicted as the negative case by the model.

True Positive: Number of positive class samples in the source data subset accurately predicted by the model.

True Negative: Number of negative class samples in the source data subset accurately predicted by the model.

TPR (True Positive Rate/Sensitivity/Recall): Ratio of True Positive predictions to actual positives with respect to the test set. It is calculated by dividing the true positive count by the actual positive count.

TNR (True Negative Rate/Specificity): Ratio of True Negative predictions to actual negatives with respect to the test set. It is calculated by dividing the True Negative count by the actual negative count.

 

Multi classification

ACC (Accuracy): Ratio of the correctly classified samples over all the samples.

Actual Total: Total number of samples in the source data subset that were of the given class.

Cohen’s Kappa (K): Coefficient that measures inter-rater agreement for categorical items, it tells how much better a classifier is performing over the performance of a classifier that simply guesses at random according to the frequency of each class. It is in the interval [-1:1]. A coefficient of +1 represents a perfect prediction, 0 no better than random prediction and −1 indicates total disagreement.

False Negative: Number of positive class samples in the source data subset that were incorrectly predicted as negative.

False Positive: Number of negative class samples in the source data subset that were incorrectly predicted as positive.

Macro-PPV (Positive Predictive Value/Precision): The mean of the computed PPV within each class (independently of the other classes). Each PPV is the number of True Positive (TP) predictions divided by the total number of positive predictions (TP+FP, with FP for False Positive) within each class. PPV is in the interval [0;1]. The higher this value, the better the confidence that positive results are true.

Macro-TPR (True Positive Rate/Recall): The mean of the computed TPR within each class (independently of the other classes). Each TPR is the proportion of samples predicted Truly Positive (TP) out of all the samples that actually are positive (TP+FN, with FN for False Negative). TPR is in the interval [0;1]. The higher this value, the fewer actual samples of positive class are labeled as negative.

Macro F1 score: Harmonic mean of macro-average PPV and TPR. F1 Score is in the interval [0;1]. The F1 Score can be interpreted as a weighted average of the PPV and TPR values. It reaches its best value at 1 and worst value at 0.

MCC (Matthews Correlation Coefficient): Represents the multi class confusion matrix with a single value. Precision and recall for all the classes are computed and averaged into a single real number within the interval [-1;1]. However, in the multiclass case, its minimum value lies between -1 (total disagreement between prediction and truth) and 0 (no better than random) depending on the data distribution.

Predicted Total: Total number of samples in the source data subset that were predicted of the given class.

True Positive: Number of positive class samples in the source data subset accurately predicted by the model.

True Negative: Number of negative class samples in the source data subset accurately predicted by the model.

 

Regression

MAE (Mean Absolute Error): represents the average magnitude of the errors in a set of predictions, without considering their direction. It’s the average over the test sample of the absolute differences between prediction and actual observation where all individual differences have equal weight. MAE is in the intervall [0;+∞]. A coefficient of 0 represents a perfect prediction, the higher this value is the more error (relative error) the model have.

MAPE (Mean Absolute Percentage Error): MAPE is computed as the average of the absolute values of the deviations of the predicted versus actual values.

Max-Error: Maximum Error. The application considers here the magnitude (absolute error when identifying the maximum error. Thus -1.5 would be consider the maximum error over +1.3. The sign of the error however is still reported in this column in case it has domain significance for the user.

R2 (R Squared): also known as the Coefficient of Determination. The application computes the R2 statistic as 1 - (SSres / SStot) where SSres is the residual sum of squares and SStot is the total sum of squares.

RMSE: Root Mean Square Error against the Dataset partition selected. RMSE is computed as the square root of the mean of the squared deviations of the predicted from actual values.

SD-ERROR (Standard Deviation Error): Standard statistical measure used to quantify the amount of variation of a set of data values.