ocel-mine

Fast object-centric process mining analysis for OCEL 2.0 event logs: per-type trace variants, directly-follows graphs, OC-DFG, and process metrics — in Rust, on top of the ocel crate.

Deterministic computation only: OCEL in, analysis structures (serde-JSON-ready) out. No I/O of its own beyond what ocel provides, no UI, no configuration state — those live in ocel-studio.

Status

Shipped: per-type trace variants, per-type DFG (frequency / distinct objects / gap statistics, start-end counts), the OC-DFG overlay, per-type model discovery — the alpha algorithm (educational; its textbook limits are returned as warnings), the inductive miner (practical; sound by construction, once-per-trace fall-through before the flower, tunable IMf-style noise threshold), and the heuristics miner (noise-robust; dependency graph with tunable thresholds, PM4Py-compatible 5% pre-cleaning, dedicated length-1/length-2 loop measures) — and replay fitness: tree_replay decides exact language membership per variant (the miner's cuts partition the alphabet, so membership is ownership routing, not a token-game approximation), net_replay token-replays alpha nets, and the heuristics net reports how many observed direct successions its kept edges explain.

Discovery honesty notes: alpha cannot model self-loops (a self-looping activity joins no place and its transition fires freely — textbook behavior) and caps at 20 activities. The inductive miner matches PM4Py exactly on structured logs (e.g. the orders type below, at any noise level); on heavily interleaved types (items) the trees agree up to how optional stages nest (ours marks the out-of-stock pair optional per activity, PM4Py per pair — both sound). The noise threshold implements the IMf frequency filter (edges below the fraction of the source's strongest outgoing edge are ignored at every recursion step), not the complete IMf fall-through set. Read fitness together with simplicity: the basic miner fits 100% at noise 0 by construction, and a flower replays anything over its alphabet.

Quickstart

let log = ocel::io::read_path("order-management.sqlite")?;

let report = ocel_mine::variants(&log, "orders");
for v in report.variants.iter().take(5) {
    println!("{:>6}  {}", v.count, v.activities.join(" -> "));
}

let graph = ocel_mine::dfg(&log, "orders");           // nodes + edges
let overlay = ocel_mine::oc_dfg(&log, &["orders", "items"]); // per-type edges, honest totals
println!("{} edges", graph.edges.len() + overlay.edges.len());

Or from the command line:

cargo run --release --example variants -- order-management.sqlite orders
cargo run --release --example dfg -- order-management.sqlite orders

Performance

Official Zenodo Order Management log (21,008 events), Apple Silicon laptop, single run, warm cache:

Replay percentages match pm4py.fitness_token_based_replay exactly on orders and items at noise 0.0 and 0.2 (100 / 100 / 100 / 95.40%) and on the alpha net; validating this uncovered and fixed an alpha bug (independent sets must require a # a, so self-looping activities join no place).

Variant counts, DFG edge frequencies, and start/end counts match PM4Py's flattening exactly on both types. Note when reproducing the PM4Py variants: compute them from the flattened DataFrame with an explicit per-case timestamp sort + groupby — pm4py.get_variants(df) applied directly to a flattened OCEL frame returns scrambled sequences.

Semantics

Object-centric logs punish naive flattening (divergence/convergence). ocel-mine computes per object type: a trace is one object's events ordered by time, a variant is its activity sequence. Cross-type views (OC-DFG) overlay per-type edges with object-count annotations instead of squashing everything into one log. Results are cross-checked against PM4Py on public datasets.

The ocel family

License

MIT