Alpha Miner

What is the Alpha Miner?

The Alpha Miner is one of the earliest process discovery algorithms. It takes an event log and produces a process model, drawn as a Petri net, that shows how the activities in a process relate to each other.

The idea behind it is simple. If you record every step of a process, you can look at which activity tends to come straight after which other activity, and from those local patterns you can rebuild the shape of the whole thing.

The Alpha Miner matters mostly because it showed this was possible at all. It was first described in the early 2000s by Wil van der Aalst and colleagues, and it became the reference point that later algorithms improved on. You will rarely run it on a real business log, but understanding it makes every other discovery algorithm easier to follow.

How the Alpha Miner reads a log

The algorithm starts from the directly-follows relation. If activity B ever happens immediately after activity A, it notes that down as A > B, and it does this for every pair of activities in the log.

From those observations it sorts each pair into a category. If A is followed by B but B is never followed by A, it treats that as a causal step from A to B. If both A > B and B > A occur, it reads the two activities as running in parallel. If neither order ever appears, the two are unrelated in the process.

These categories are collected in a grid called the footprint. The Alpha Miner then uses the footprint to work out where the connecting points go (in a Petri net these are called places) and adds a single start and end point so every case has a clear beginning and end.

A worked example

Suppose a log contains two kinds of cases: A, B, C, D and A, C, B, D. Activities B and C appear in both orders, while A always comes first and D always comes last.

Because B and C swap places, the Alpha Miner concludes they run in parallel. The model it draws splits into two branches after A, lets B and C happen in either order, and joins back together before D.

Now imagine the log only ever contained A, B, C, D. The algorithm would never see C happen before B, so it would miss the parallelism. This is the catch: the log has to contain enough real behaviour for the relations to show up.

Alpha Miner versus Heuristics Miner

The Alpha Miner only cares whether a relation appears, not how often. A single mistyped event or one badly linked case counts for as much as ten thousand normal ones.

The Heuristics Miner was built to fix exactly that. It counts how frequently each relation occurs and uses thresholds to keep the common patterns and drop the rare ones, which makes it far more usable on messy real-world logs.

Alpha Miner versus Inductive Miner

The Inductive Miner is the modern default. It splits the log recursively and guarantees a sound model, one where every case that starts can always finish cleanly. The Alpha Miner offers no such guarantee and can produce a model that deadlocks or leaves activities stranded.

Where the Alpha Miner still earns its place is in teaching. It shows the direct line from raw event data to a formal process model more clearly than any of its successors.

What to watch out for with the Alpha Miner

Short loops trip it up. A step that repeats immediately, such as A, A, or a back-and-forth like A, B, A, breaks the simple relation logic. An extension called Alpha+ handles some of these cases, but not all of them.

It assumes the log is complete. The algorithm treats any order it never saw as impossible, when in reality it might just be a rare but valid variant that did not make it into your sample.

It has no defence against noise. With no frequency threshold, one stray event can add a whole branch to the model. On real logs, teams filter the data first or reach for another algorithm.

The result is not the truth. Any discovered model reflects the specific log and the algorithm's assumptions, not an objective picture of how the business runs.