Banking77 Fraud Detection

Application Areas
Banking / Finance
Datatypes
Numeric
Datetime
Text
Location

← Back to Datasets

Midterm Vignette

This dataset is available through HuggingFace’s Datasets portal here, and it has the name “Banking77” because it represents data recorded during testing of a bank website’s chatbot helper, where testers were asked to work through 77 different intents. The intents are encoded numerically within this dataset, and correspond to the following descriptions:

Code 
import pandas as pd
# from great_tables import GT
import itables

intent_df = pd.read_csv('intents.csv')
itables.show(
  intent_df,
  layout={
    "topStart": None,
    # "topEnd": None,
  },
  style="width: 100%",
  autoWidth=True,
)
Loading ITables v2.8.1 from the internet... (need help?)

The bank has hired you as a data scientist (congratulations!) because they would like to improve their fraud detection rate: the rate at which they detect fraudulent usages of the chatbot to (e.g.) “circumvent” reasonable checks that a human might carry out in their heads during an interaction.

Question 1 Starter: OLTP Structure

The entries streaming in the bank’s OLTP system, in this case, represent individual messages sent to the chatbot, like:

{
  'label': 11,
  'text': 'I am still waiting on my card?',
  'received': '2026-10-22_12:00:00'
}

However, note that here (since this is training data) we have the actual “correct” value of label: when future messages (after this test run) stream into your firm’s OLTP database, this value will simply be NA, since the intent is not yet known!

For the purposes of the midterm, however, you can assume that the bank has implemented some simple manual intervention procedure so that some proportion of the messages that are coming in do have intent labels (provided by, e.g., the same firm that first carried out these chatbot tests).

Note the complexities that may be involved in turning this OLTP-format stream of events into an analytics-friendly OLAP infrastructure meeting the bank’s fraud-detection needs. For example:

  • How might you “route” the training data (the messages arriving with label value already filled in) and non-training data (the messages without this value already filled in) differently, to enable continuous updating of a supervised-machine-learning-based fraud detection model?
  • One challenge might come from the fact that message contents might change over time in light of new external contexts: for example, in 2024 Zimbabwe changed the name of its currency from the Zimbabwe dollar to Zimbabwe gold, so that there may be no previously-existing training messages asking about “Zimbabwe gold” labeled with intent 54 (supported_cards_and_currencies).
  • Another challenge might come about in the future, if the bank recognizes new forms of fraud that they would like to detect, perhaps via new intents above the 77 occurring in this dataset. How might you adapt your OLTP \(\leadsto\) OLAP setup to incorporate new additions to or removals from the set of possible intents?