Loading...
Difficulty
Your best score
N/A
The goal of this problem is to identify fraudulent credit card transactions based on a real, anonymized dataset. The problem reflects a real-world scenario where data is heavily imbalanced and fraudulent transactions are rare.
The project uses two CSV files:
train.csv - the training datasettest.csv - the testing dataset| Column | Description |
|---|---|
| id | Unique identifier for each transaction |
| Time | Number of seconds elapsed since the first transaction |
| V1 - V28 | Anonymized features |
| Amount | Transaction value |
| Class | Target variable (0 = legitimate transaction, 1 = fraud) |
Note: The Class column is only available in train.csv.
Important observation: The dataset is severely imbalanced, so accuracy is not a relevant metric for evaluating classification models.
Determine how many fraudulent transactions have an Amount greater than the average Amount of legitimate transactions.
The answer is an integer.
Evaluate how "abnormal" the fraudulent transactions are using the Mahalanobis distance, which considers all numerical variables simultaneously (Amount + V1-V28).
Suggested steps:
Class = 1) from train.csv.For each transaction in test.csv, predict the fraud label:
1 = fraud0 = legitimate transactionPrimary metric: F1-score on class 1.
| F1-score | Score |
|---|---|
| F1 >= 0.85 | 80 points |
| F1 < 0.60 | 0 points |
| Intermediate | Linear scaling between 0 and 80 |
The submission file must be in CSV format with the following columns:
| Column | Description |
|---|---|
| subtaskID | Subtask number (1, 2, or 3) |
| datapointID | Observation identifier |
| answer | The answer or prediction |
Rules:
datapointID = 1 (a single answer).test.csv, with datapointID = the value from the id column.Example:
subtaskID,datapointID,answer1,1,422,1,0.273,227846,13,227847,03,227848,0Note: Solving this problem does not require the use of Transformer-type architectures.