BPI Challenge 2012

 

BPI Challenge 2012 is an anonymized event log capturing a real-world loan application process from a Dutch financial institution, spanning from October 2011 to March 2012.

This dataset is one of the most highly analyzed and benchmarked in the process mining community—featured in hundreds of academic papers—and serves as a gold standard for evaluating and developing process discovery, conformance checking, and performance analysis techniques tf-pm.orgResearchGate. Researchers have leveraged it to explore control-flow discovery, resource behavior, process deviations, and performance insights using tools like ProM and alignment-based diagnostics

Use case:
Process Mining Benchmark for Analyzing and Optimizing Loan Application Workflows

Example observations	• Static case attributes:  – amount_requested, first_withdrawal_amount, salary, number_of_terms, application_type, loan_goal … • Dynamic process attributes:  – current_activity (one-hot of 36 distinct events)  – time_since_case_start, number_manual_steps_so_far  – risk_score_so_far (can be computed from historical decision outcomes) • Calendar context: day_of_week, hour_of_day (SLA compliance)
Example actions	Typical set (≈ 12 high-level choices derived from the 36 raw events): 1 = Send_to_Automatic_Risk_Check 2 = Send_to_Manual_Credit_Scoring 3 = Request_Missing_Documents 4 = Skip_to_Offer_Creation 5 = Decline_Application 6 = Cancel_Case 7 = Send_Contract 8 = Book_Loan … (The legal action mask is inferred from the log’s directly-follows graph.)
Example Reward	+10 if the case is successfully booked and risk ≤ threshold −(lead_time_hours / 24) −5 for each manual step executed −20 if the case is declined after unnecessary manual work Sparse terminal reward can be reshaped into step rewards via potential-based shaping (e.g. negative elapsed time).
Example Transition matrix estimate	• 13 087 cases • 262 200 events ⇒ approx. 262 200 state→next-state transitions (average 20.0 per case). • Empirical transition probabilities P(s′|s,a) can be built directly from the log for offline RL / fitted-Q-evaluation.
This dataset therefore delivers:	– rich, real-world observations (process + business data) – clearly identifiable actions (routing decisions) – measurable rewards (time, cost, risk) and enough transitions to pre-train or evaluate an enterprise workflow-optimising RL agent.