In this project, I chose to analyze e-commerce fraud data, sourced from Kaggle. The data was synthetically generated using Python. The world has a heavy reliance on e-commerce, and as such it is important to recognize the leading factors that contribute to fraud. There are many factors at play when making an online transaction, such as where you are buying from, the device you are using, and the payment method, among many other things. This dataset contains the following columns: Transaction ID, Customer ID, Transaction Amount, Transaction Date, Payment Method, Product Category, Quantity, Customer Age, Customer Location, Device Used, IP Address, Shipping Address, Billing Address, Is Fraudulent, Account Age Days, and Transaction Hour. The dataset has 1,472,952 rows. As such, I had to trim the data to make it easier to process. One main issue I ran into during this process was running out of memory on my computer to process. Once the data was trimmed, I was left with ID, Transaction Amount, Transaction Date, Payment Method, Product Category, Quantity, Customer Location, Customer Age, Device Used, Is Fraudulent, Account Age Days, and Transaction Hour. This trim also allowed me to focus specifically on these factors for my analysis.
I approached this project with several questions of interest. The driving question I asked is “what factors lead to fraudulent transactions?” Additionally, I created questions related to specific aspects of the data. Are young people more likely to be scammed than older people? IS there a certain type of item that is frequently fraudulently sold? Are certain payment methods more fraud resistant? Does price matter, or are larger purchases fraudulent as frequently as larger purchases?
For my analysis, I relied on four libraries: tidyverse, ggplot2, patchwork, and plotly. Following the cleaning of the data, I started my analysis. First, I plotted transaction amounts over time to determine what day had the highest valued transactions. As seen in figure 1, transaction amounts peak in late January. This is a nice entry point into the analysis. We can select transactions from this specific data range to determine what day this was. I plotted the transactions from January 1, 2024 to Jan 31, 2024, seen in figure 2. According to this graph, the day with the highest valued transactions was January 26, 2024. Let’s look at two factors on this day. First, we will investigate product categories. There are five categories in this dataset: clothing, electronics, health & beauty, home & garden, and toys & games. Figure 3 plots the proportion offraudulent transactions for these five categories. As seen in figure 3, health & beauty had the highest level of fraud on January 26th. I had initially thought that something like clothing or electronics would be higher, given the number of fakes that exist of these products. We can now turn to the payment methods used on January 26th. This may unearth a trend. On January 26th, the most frauded transactions were completed using debit cards, seen in figure 4. This is in line with my expectations, as debit cards are typically less secure than other forms of payment, like PayPal, as they are done on-site and don’t always have strict verification methods in place if transactions are not made through a trusted vendor. To continue, we can count the number of health & beauty transactions that were made with debit cards. My R script counted 826 transactions. We can now calculate the percentage of these that were fraudulent. 4.12% of these specific transactions were fraudulent. This doesn’t seem like a lot, but in the grand scheme of things it is quite large in terms of the number of people getting frauded. 34 people were victims of fraud in this case. Since we have determined a trend, we can now determine if it continues throughout the year, not just on the day with the highest number of transactions. To do this, we can use a similar script to perform the same operations on the entire dataset. Over the course of the data range, there were 73,475 health & beauty transactions made with debit cards. Of those transactions, 3679 were fraudulent, equaling 5.01 %. This is inline with the data from January 26th. I don’t want to conclude that a factor leading to fraud is purchasing health & beauty products, so let’s just focus on debit card purchases. Based on figure 4, we already know that debit cards are the most frequently frauded payment methods. We can perform similar operations to draw more conclusions. Using another R script, we can calculate that 5.02 of all purchases made with debit cards were fraudulent. Of the 368,429 debit card purchases recorded, 18,313 were fraudulent, equaling 4.97%. Again, a small percentage but on a dataset this large that has a massive impact. Thus, it can be concluded that credit cards have the largest rate of fraud across the entire data range. We can now look at some other factors and interesting aspects of this data, such as the most common ages that experience fraud. By filtering the cleaned data to select only Customer_Age and Is_Fraudulent, we can plot these values. Figure 5 is the resultant graph. I was very surprised by this. The data created a bell curve. This was the opposite of what I expected. I had assumed that older people (>50 years) and younger people (<20 years) would have experienced fraud at a higher level than those residing in the middle. This could be due to the data being synthetic, however we can roll with it for further investigation. Let’s combine this age data and the credit card data to see if we can extract another trend. I selected all transactions that were made by people from age 25 to age 45. My R script calculated that 3.54% of debit card transactions made by people in that age range were fraudulent. The percentage of debit card fraud is relatively standard across the board.
This analysis of e-commerce fraud data aimed to identify factors that contribute to fraudulent transactions. By analyzing the data using R, several trends were revealed. My analysis showed that health and beauty transactions made with debit cards were particularly fraud-prone, with 4.12 percent of such transactions made on January 26, 2024 (the date with the highest volume of transactions) being fraudulent. This trend carried through the entire dataset, with 5.01 percent of all health and beauty transactions made with debit cards being fraudulent. Through additional analysis, it was revealed that debit cards were the most frequently targeted payment methods across the board, with 5.02 percent of all debit card transactions being fraudulent. Although these percentages are small, the dataset is very large with 1,472,952 unique transactions making these percentages have a large impact. Contrary to my expectations, the analysis of age vs fraud percentage revealed a bell curve, showing that individuals aged 25-45 experienced higher levels of fraud than both younger and older people. While certain factors such as payment method and product category contributed to fraud percentages, this analysis also opened my eyes to how complex fraud prediction is.