If the company is going to show the workflow of the machine learning model to non-technical audience, decision tree may be a good choice. I tried decision tree model first because it's a very interpretable model and it's very transparent. The deeper the tree, the more complex the rules and fitter the model. It’s more efficient than GridSearchCV and its chance of finding the optimal parameter are comparatively similar.ĭecision tree: The decision true are generally in form of if-then-else statements. This allows me to explicitly control the number of parameter combinations that are attempted. It helped me select models by setting up a grid of hyperparameter values and selecting random combinations to train the model and score. RandomizedSearchCV: Random Search is a hyperparameter tuning algorithm. With cross-validation, I can achieve a generalized model. The accuracies and errors are then averaged of the algorithm. It continued the process by changing the testing part in each iteration and training the algorithm over the other parts. The algorithm helped us reserve 1 part for testing and train the algorithm over the 2 parts. The K-Fold Cross Validation works by first dividing the dataset into k-subsets. After the oversampling process, the data is reconstructed.Ĭross validation: cross validation is an algorithm that can ensure that each and every instance of the dataset will be trained and tested. SMOTE generates the virtual training records by randomly selecting one or more of the k-nearest neighbors for each example in the minority class.
SMOTE: The target variable stating whether a transaction is fraud is highly imbalanced. Methods I used in the model building part include: OpenTime: The time between the account open date and the transaction date.ĬhangeAddTime: The time between the last of address change date and the transaction date. Most of the time, fraud transactions happen in a merchant where the customer hasn't used before.ĮxpTime: The time between the transaction date and the expiration date. Whether first time purchasing merchant: Whether it’s the first time the user purchase in this merchant. If they use their cards in some other countries, it's possible that this transaction would be fraud. most of the time, people will use their cards in the country they acquired cards. SameCountry: Whether the location where the card is acquired is the same as where the card is used. If the entered CVV is incorrect, it's highly possible that this transaction would be fraud.
PURCHASE MULTISWIPE VERIFICATION
Typically, the card's verification value adds extra security to transaction. RightCVV: Whether the entered CVV is correct. I built different functions to identify them and convert them to be dummy variables. Another is the multi-swipe, where a vendor accidentally charges a customer's card multiple times within a short time span. Generally speaking, there’re 2 types of duplicated transactions, one is the reversed transaction, where a purchase is followed by a reversal. This is the most important part of this project, I focused on these features:ĭuplicated transactions: there’re many duplicated transactions in the raw data set and I want to detect them and label them. The raw data is being kept here within the repo EDAīasically, I made histograms, kernel density estimate plots and cumulative distribution plots to show the distribution of some of my predictor variables. Make Plots for Exploratory Data Analysis.This project used data from credit card transactions data loosely resembles real transactional data from a bank, and build a model to predict future fraud. As stated by ‘Javelin Strategy & Research’, more than 20 percent of customers choose to change their banks after experiencing fraud. Fraud can also affect customer satisfaction and customer retention. Only in one year, Fraudsters stole about $6 billion from banks last year. The financial services industry is suffering from fraud-related losses and damages.
Build predictive models to determine whether a given transaction will be fraudulent or not using transaction data.
0 kommentar(er)
