This project explores the application of five different machine learning algorithms - Logistic Regression, Support Vector Classifier (SVC), Random Forest Classifier,K-Nearest Neighbors Classifier (KNN), and Gaussian Naive Bayes (GNB) all within the context of credit card fraud detection. What sets our project apart is that we didn’t merely compare these models in their standard configurations. Instead, we initiated a correlation check for each feature with our target variable. This methodology enabled us to observe the influence of correlation on the performance of these models.
This project serves as a comprehensive guide to understanding the role of feature selection in machine learning model performance, particularly in the context of credit card fraud detection.
1-Enormous Data: The models need to process millions of transactions daily, necessitating high computational efficiency.
2-Imbalanced Data: The vast majority of transactions are non-fraudulent, resulting in a highly imbalanced dataset.
In the context of our project, we encountered the common issue of imbalanced datasets. This is a scenario where the classes in the target variable are not represented equally. In our case, the majority of credit card transactions were non-fraudulent, leading to a highly imbalanced dataset.
After conducting research, we decided to use undersampling to address this issue. Undersampling is a technique where you randomly overlook some of the observations from the majority class in order to balance the class distribution.
Initially, we were concerned that undersampling might hurt the model’s performance. This is a valid concern as undersampling can lead to loss of information, given that it involves removing instances from the dataset. However, our results showed that undersampling actually improved the performance of our models.
At first, we applied these models directly to our dataset, without considering the correlation between the features and the target variable. Then, we incorporated a correlation check before building the models. This step involved calculating the correlation of each feature with the target variable, which added an extra layer of depth to our model comparison. This correlation check was ensured by having a look at the heatmap or sample correlation matrix. We then understood the prevalence of certain variables when put into perspective in what we referred to as the ‘grand scheme’.
It goes without saying that we cleaned the data, i.e., dropped the duplicate values, filled the missing data and of course, we standardized it using the standard scaler. This was not arbitrary because we checked every variable to see the distribution of the values. Most variables in this dataset were pre-standardized and had their names written off. The distribution turned out to be in a Gaussian pattern, and therefore we came to the conclusion that this is almost definitely the work of a standard scaler. We left a sample for the distribution checking.
Through our project, we discovered that the performance of the machine learning models was quite similar when using all variables versus a select few. This observation led us to an important insight: working with a handful of relevant variables can be just as effective, if not more so, than working with a larger set.
This finding is significant for several reasons:
Using fewer variables reduces the computational complexity of the models, leading to faster training times and less resource usage.
Models with fewer variables are easier to understand and interpret.
By limiting the number of variables, we reduce the risk of overfitting, where the model learns the training data too well and performs poorly on unseen data.
In retrospect, this may seem self-evident, but the process of reaching this conclusion through our own research proved to be a valuable learning experience. It highlighted the importance of feature selection in machine learning and challenged the assumption that more data invariably leads to better results. This insight is a key takeaway from our project, serving as a reminder of the critical importance of understanding the data and the problem at hand.