Predictive Process Monitoring (PPM) aims at predicting at runtime and as early as possible the future states of a process execution. Common tasks include predicting the next event, the time to completion of a trace, and outcomes. Existing approaches typically consider an event from the perspective of the executed activities along with their timestamps and case identifiers. This leads to the disadvantage that in real-life settings, there is much more information recorded in the event log that is not captured or completely ignored when performing prediction tasks. We introduce GNN4PPM, an approach that predicts all next events along with their complete data payload at once. We represent event information in a heterogeneous knowledge graph that captures the event log as an RDF semantics, and train the embeddings with a Relational Graph Convolutional Network (R-GCN). Our approach is promising in comparison to existing solutions, and experiments with state-of-the-art solutions prove the accuracy and applicability of GNN4PPM in complex settings.
Figures & tables
Figure 1: Pipeline of GNN4PPM
Figure 2: Knowledge graph representation of the entity relations of an event log.
Dataset
#Traces
#Events
#Variants
#Case attr.
#Event attr.
Triples
Entities
Predicates
BPIC12A
13,087
60,849
17
1
3
597,071
144,870
10
BPIC12W
9,658
170,107
2921
1
3
1,554,389
360,122
10
BPIC12WC
9,658
72,413
2963
1
3
685,039
164,741
10
BPIC13O
819
2,351
108
3
6
32,925
5,985
14
BPIC20P
10,500
56,437
896
8
5
254,401
59,081
16
BPIC20R
6,886
36,796
76
3
4
368,206
110,655
7
Table 1: Structural characteristics of the event logs and constructed knowledge graph information.
Approach
BPIC12A
BPIC12W
BPIC12WC
BPIC13O
BPIC20P
BPIC20R
SEPHIGRAPH [ 13 ]
67.8
78.6
46.6
20.2
73.8
76.5
LUPIN [ 33 ]
19.3
35.7
21.2
24.4
39.7
51.1
PROPHET [ 34 ]
76.7
89.2
80.3
62.0
83.8
87.8
BEST [ 37 ]
76.1
69.0
40.0
48.8
86.9
89.0
CAMARGO [ 7 ]
75.9
76.4
68.9
77.2
84.4
91.9
EVERMANN [ 15 ]
75.8
75.4
67.5
65.1
88.9
90.5
Table 2: Accuracy of the Next Activity Prediction Task compared to different approaches. Best result per dataset highlighted in bold; second-best is underlined.
Dataset
activity
resourcecountry
org_resource
lifecycle
org_group
org_role
timestamp (MAE)
BPIC12A
88.75
64.20
100
–
–
–
0.3184
BPIC12W
80.34
56.76
83.46
–
–
–
0.5476
BPIC12WC
83.19
14.83
100
–
–
–
0.7667
BPIC13O
78.15
63.16
67.04
97.78
87.14
91.85
0.7100
BPIC20P
80.44
99.38
–
–
88.67
–
0.3785
BPIC20R
87.79
99.58
–
–
90.53
–
0.6400
Table 3: Multi-target prediction results across datasets given as accuracy for event attributes and as normalized MAE for timestamp.
Figure 4: Influence of each event attribute on activity prediction accuracy when removed from the model