BPI Challenge 2017 is part of the renowned Business Process Intelligence (BPI) Challenge series and provides a comprehensive event log for the loan application process from a Dutch financial institution. It includes all loan applications submitted via an online system in 2016 and tracks their evolution up to February 1, 2017, 15:11
The dataset is modeled as a multi-dimensional event graph and is made available in graph database formats—specifically Neo4j dump and GraphML—allowing for advanced analysis of entities (such as applications, offers, workflows, users) and their interactions, captured through labeled relationships and event nodes
Use case:
Next-Best-Action Advisor for Optimizing Loan Application Decisions
| Example Observations (state vector available at decision time) | • Static case attributes – amount_requested, product_type, first_time_customer, income_band, etc. • Dynamic attributes – – current_activity (one-hot over 26 activities) – elapsed_time_since_start, waiting_time_since_last_event – #offers_generated_so_far, #previous_rejections, credit_score_updates – resource_group_in_charge (front-office, back-office, risk, …) |
| Example Actions (discrete, 8 controllable choices) | 1. Trigger risk assessment (“O_SELECTED”) 2. Ask for additional documents (“W_Afhandelen_leads”) 3. Generate provisional offer (“A_PREACCEPTED”) 4. Escalate to senior underwriter (“O_SENT_BACK”) 5. Skip to final approval (“A_APPROVED”) 6. Reject application (“A_DECLINED”) 7. Cancel application (“A_CANCELLED”) 8. Do nothing (let process follow historical path) |
| Example Reward | +10 if case eventually reaches “A_ACCEPTED”, -10 if “A_DECLINED” or “A_CANCELLED”, -0.05 per extra day of throughput time, -1 each time the file is sent back (“O_SENT_BACK”) or re-opened. A dense intermediate reward can be shaped by giving ‑0.01 per hour of idleness between events. Transition estimation Each event→event pair represents an environment transition. Total transitions ≈ (#events – #cases) ≈ 1 204 267 – 31 509 ≈ 1 172 758. Average episode length ≈ 38 events (1 204 267 / 31 509). With the 8-action space and a state space of ~1000 unique encodings (after categorical hashing), the empirical transition matrix is sparse but fully observable from the log, enabling off-line RL (e.g., Fitted Q-Iteration, BCQ, CWPDIS). |