Business Intelligence and Process Modelling

Business Intelligence and Process Modelling

F.W. Takes Universiteit Leiden

Where are we?

Business Intelligence: anything that aims at providing actionable information that can be used to support business decision making

Business Analysis Business Analytics

Visual Analytics Descriptive Analytics Predictive Analytics

Network Intelligence: Network Science in a BI context

Process Modelling

Data

→

Network Science (recap)

Network science: analyzing “big” structured data consisting of

objects connected via certain relationships, in short: networks

Interest from: mathematics, computer science, physics, biology, public administration, social sciences, . . .

Notation (recap)

Small World Networks (recap)

1 Sparse networks density

2 Fat-tailed power-law degree distribution degree

3 Giant component components

4 Low pairwise node-to-node distances distance

Many real-world networks: communication networks, citation

networks, collaboration networks (Erd¨os, Kevin Bacon), protein

interaction networks, information networks (Wikipedia), webgraphs, financial networks (Bitcoin) . . .

Topics

Graph Representation and Structure Paths and Distances

Graph Evolution, Link Prediction ←

Spidering and Sampling

Centrality

Visualization Algorithms and Tools Graph Compression

Community Detection

Contagion, Gossipping and Virality Privacy, Anonymity and Ethics

Network evolution

Graphsevolve over time

Social networks: users join the network and create new friendships Webgraphs: new pages and links to pages appear on the internet Scientific networks: new papers are being co-authored and new citations are made in these papers

Interesting: small world properties emerge and are preserved during evolution!

Evolving graphs

Graph Gt = (Vt,Et)

Time window 0≤t ≤T −1

Usually at t = 0, either

V0=∅ and a new edge may bring new nodes, or

V0=VT−1 and only edges are added at each timestamp

Timestamp on node v ∈V:

t(v)∈[0;T −1]

Timestamp on edgee ∈E:

t(e)∈[0;T −1], or as common input format:

e = (u,v,t_(u,v)) with u,v ∈V andt_(u,v) ∈[0,T −1]

LIACS collaboration network (v2012)

Two schools

Synthetic graphs model-driven

Model or algorithm to generate graphs from scratch

Tune parameters to obtain a graph similar to an observed network Statistical analysis

Real-world graphs data-driven

Obtain data from an actual network

Compute and derive properties and determine similarity with other networks

Apple collaboration network

http://www.kenedict.com/apples-internal-innovation-network-unraveled/

Link prediction

Link prediction problem: given a networkGt = (Vt,Et), denoting

the network at time t, predict the newly formed links in the evolved

network Gt0 = (V_t0,E_t0) at timet0 >t, i.e., predict the contents of Et0\E_t.

Applicable to weighted and unweighted, directed and undirected networks

Supervised learning problem

Features based on the structure of the network Train on first 95%, test on last 5% (randomized) Validate result using AUROC

J.E. van Engelen, H.D. Boekhout and F.W. Takes, Explainable and Efficient Link Prediction in Real-World Networks (working paper), 2016.

Link prediction

Link prediction problem: given a networkGt = (Vt,Et), denoting

the network at time t, predict the newly formed links in the evolved

network Gt0 = (V_t0,E_t0) at timet0 >t, i.e., predict the contents of Et0\E_t.

Applicable to weighted and unweighted, directed and undirected networks

Supervised learning problem

Features based on the structure of the network Train on first 95%, test on last 5% (randomized) Validate result using AUROC

J.E. van Engelen, H.D. Boekhout and F.W. Takes, Explainable and Efficient Link Prediction in Real-World Networks (working paper), 2016.

Feature set goals

efficient in terms of time complexity;

accurate in its future link predictions;

explainable in its performance based on simple features;

consistent in its accuracy relative to larger feature sets across networks;

Link prediction features

Compute features for each possible future edge (i,j)∈/Et

Node features: degree, volume (total weight)

Neighborhood features: neighbor count, common neighbor count, transitive common neighborhood, Jaccard coefficient, preferential attachment, and others

Path features: shortest path length, number of shortest paths, restricted Katz measure, and others

Efficient Feature Set (EFS)

Large number of features Black box type of approach

Cover individual, local and global properties Explainable result

Node features

Feature Variant Complexity EFS Degree (source) - O(1) X Degree (source) din O(1)

Degree (source) dout O(1) Degree (target) - O(1) X Degree (target) din O(1)

Degree (target) dout O(1)

Volume (source) - O(m/n) Volume (source) din O(m/n) X

Volume (source) dout O(m/n) X Volume (target) - O(m/n) Volume (target) din O(m/n) X

Volume (target) dout O(m/n) X

Neighbourhood features

Total neighbours - O(m/n) Total neighbours Γin O(m/n)

Total neighbours Γout O(m/n) Common neighbours - O(m/n) X Common neighbours Γin O(m/n)

Common neighbours Γout O(m/n)

Transitive comm. neigh. - O(m/n) Jaccard Coeff. - O(m/n) X Jaccard Coeff. Γin O(m/n)

Jaccard Coeff. Γout O(m/n)

Transitive Jacc. Coeff. - O(m/n) X Adamic/Adar - O(m/n) Preferential attachment - O(1) Preferential attachment Γin O(1)

Preferential attachment Γout O(1)

Opposite direction link - O(1) X

Path features

Shortest path length - O(m+n) X Num. shortest paths `max= 3 O(m+n)

Restricted Katz measure `max= 3, O(m+n) β= 0.05

Datasets

Table : Characteristics of network data sets used for testing

Data set Nodes Links CC Type Dist 3N

digg 30,398 86,404 0.01 + D 4.68 45% fb-links 63,731 817,035 0.22 - U 4.31 88% fb-wall 46,952 274,086 0.11 + D 5.71 61% infectious 410 2,765 0.46 + U 3.57 83% liacs 1,036 4,650 0.84 + U 3.86 100% lkml-reply 27,927 242,976 0.30 + D 5.19 99% slashdot 51,083 131,175 0.02 + D 4.59 75% topology 34,761 107,720 0.29 + U 3.78 97% ucsocial 1,899 20,296 0.11 + D 3.07 99% wikipedia 100,312 746, 114 0.21 - D 3.83 89%

Experiments

Large candidate set of size (|V| × |V −1|)− |E|

Restrict based on maximum distance a new edge bridges Class imbalance

Randomly leave out edges in training to get to 9 : 1 ratio Measure result using AUROC

Determine difference between All features, Node features, Neighborhood Features and EFS

Results

Features digg fb-links fb-wall infectious liacs lkml slashdot topology ucsocial wikipedia All 0.830 0.933 0.887 0.967 0.997 0.975 0.928 0.967 0.913 0.970 Node 0.827 0.700 0.710 0.955 0.969 0.971 0.922 0.949 0.911 0.941 Neighbourhood 0.761 0.911 0.866 0.794 0.986 0.974 0.920 0.961 0.920 0.926 Path 0.632 0.897 0.819 0.579 0.979 0.925 0.777 0.940 0.673 0.827 EFS 0.825 0.930 0.876 0.958 0.995 0.973 0.921 0.965 0.910 0.967 EFS Performance 99.4% 99.6% 98.8% 99.1% 99.8% 99.8% 99.2% 99.8% 99.7% 99.7%

Table : AUROC for each network and each set of features. EFS Performance lists performance of EFS relative to All features.

Conclusions

Network science treats data as an annotated set of objects and relationships

The structureof the network provides new insights in the data

Centrality measuresare able to identify prominent actors in the network solely based on its structure

Community detection algorithms reveal groups and clusters based on the network structure

Process Modelling

Recap

Business Intelligence

Process Modelling

Business process modelling Modelling languages Process discovery

Business Process Management (recap)

Process: a set of related actions and transactions to achieve a certain objective

Business process: a sequence of activities aimed at producing something of value for the business (Morgan02)

Management processes Operational processes Supporting processes

Business Process Management: the discipline that combines knowledge from information technology and knowledge from management sciences and applies this to operational business processes (v.d. Aalst)

Extension of WorkFlow Management (WFM)

Business Process Modelling (recap)

Business Process Model: abstract representation of business processes, functionality is:

Descriptive: what is actually happening? Prescriptive: what should be happening?

Explanatory: why is the process designed this way?

In practice: formalizeand visualizebusiness processes

Process Discovery: derive the process from a description of activities

Process Mining: the task of converting eventdata into process models (discovery, conformance, enhancement)

Why Model Processes? (recap)

Process Mining (recap)

Business Process. . . Intelligence?

M. Castellanos et al.,Business process intelligence,Handbook

Process Modelling

Informal models: used for discussion and documentation (process descriptions)

Formal models: used for analysis or enactment

Petri Nets— today PN Business Process Model Notation — later BPMN

Petri Nets

Event logs (1)

Case ID Event ID dd-mm-yyyy:hh.mm Activity Resource Costs

1 35654423 30-12-2010:11.02 register request Pete 50 1 35654424 31-12-2010:10.06 examine thoroughly Sue 400 1 35654425 05-01-2011:15.12 check ticket Mike 100 1 35654426 06-01-2011:11.18 decide Sara 200 1 35654427 07-01-2011:14.24 reject request Pete 200 2 35654483 30-12-2010:11.32 register request Mike 50 2 35654485 30-12-2010:12.12 check ticket Mike 100 2 35654487 30-12-2010:14.16 examine casually Sean 400 2 35654488 05-01-2011:11.22 decide Sara 200 2 35654489 08-01-2011:12.05 pay compensation Ellen 200 3 35654521 30-12-2010:14.32 register request Pete 50 3 35654522 30-12-2010:15.06 examine casually Mike 400 3 35654524 30-12-2010:16.34 check ticket Ellen 100 3 35654525 06-01-2011:09.18 decide Sara 200 3 35654526 06-01-2011:12.18 reinitiate request Sara 200 3 35654527 06-01-2011:13.06 examine thoroughly Sean 400 3 35654530 08-01-2011:11.43 check ticket Pete 100 3 35654531 09-01-2011:09.55 decide Sara 200 3 35654533 15-01-2011:10.45 pay compensation Ellen 200 4 35654641 06-01-2011:15.02 register request Pete 50 4 35654643 07-01-2011:12.06 check ticket Mike 100 4 35654644 08-01-2011:14.43 examine thoroughly Sean 400 4 35654645 09-01-2011:12.02 decide Sara 200 4 35654647 12-01-2011:15.44 reject request Ellen 200 . . .

Event logs (2)

Case ID Event ID dd-mm-yyyy:hh.mm Activity Resource Costs

. . .

5 35654711 06-01-2011:09.02 register request Ellen 50 5 35654712 07-01-2011:10.16 examine casually Mike 400 5 35654714 08-01-2011:11.22 check ticket Pete 100 5 35654715 10-01-2011:13.28 decide Sara 200 5 35654716 11-01-2011:16.18 reinitiate request Sara 200 5 35654718 14-01-2011:14.33 check ticket Ellen 100 5 35654719 16-01-2011:15.50 examine casually Mike 400 5 35654720 19-01-2011:11.18 decide Sara 200 5 35654721 20-01-2011:12.48 reinitiate request Sara 200 5 35654722 21-01-2011:09.06 examine casually Sue 400 5 35654724 21-01-2011:11.34 check ticket Pete 100 5 35654725 23-01-2011:13.12 decide Sara 200 5 35654726 24-01-2011:14.56 reject request Mike 200 6 35654871 06-01-2011:15.02 register request Mike 50 6 35654873 06-01-2011:16.06 examine casually Ellen 400 6 35654874 07-01-2011:16.22 check ticket Mike 100 6 35654875 07-01-2011:16.52 decide Sara 200 6 35654877 16-01-2011:11.47 pay compensation Mike 200

Table : Event logs of a support desk handling customer compensations

Simplified event log

Table : Simplified event log of a support desk handling customer compensations (a = register request, b = examine thoroughly, c = examine

casually, d = check ticket, e = decide, f = reinitiate request, g = pay compensation, h = reject request)

In short: {ha,b,d,e,hi,ha,d,c,e,gi,ha,c,d,e,f,b,d,e,gi,

Simplified event log

Table : Simplified event log of a support desk handling customer compensations (a = register request, b = examine thoroughly, c = examine

casually, d = check ticket, e = decide, f = reinitiate request, g = pay compensation, h = reject request)

In short: {ha,b,d,e,hi,ha,d,c,e,gi,ha,c,d,e,f,b,d,e,gi,

ha,d,b,e,hi,ha,c,d,e,f,d,c,e,f,c,d,e,hi,ha,c,d,e,gi}

Example (1)

Example (2)

Figure : Petri net based on event log{ha,b,d,e,hi,ha,d,b,e,hi}

Play out

Replay

Connecting models to real events is crucial Possible uses

Conformance checking Repairing models

Extending the model with frequencies and temporal information Constructing predictive models

Operational support (prediction, recommendation, etc.)

Automata (remember?)

Finite automaton FA= (Q,Σ,qo,A, δ)

Q is a finite set of states

Σ is a finite alphabet of input symbols qo∈Q is the initial state

A⊆Q is the set of accepting states

δ:Q×Σ→Q is the transition function

Figure : Deterministic Finite Automaton for the functionx mod 3

Automata (remember?)

Finite automaton FA= (Q,Σ,qo,A, δ)

Q is a finite set of states

Σ is a finite alphabet of input symbols qo∈Q is the initial state

A⊆Q is the set of accepting states

δ:Q×Σ→Q is the transition function

Petri Nets

Petri netN= (P,T,F)

P is a finite set of places

T is a finite set of transitions

F ⊆(P ×T)∪(T ×P) is a finite set

of directedarcscalled the flow

relation

Labeled Petri Nets

Petri netN= (P,T,F,A, `)

P is a finite set of places

T is a finite set of transitions

F ⊆(P ×T)∪(T ×P) is a finite set

of directedarcscalled the flow

relation

Ais a set of activity labels

Enabling

A transition is enabled if each of its input places contains at least

one token

Firing

An enabled transition can fire (i.e., it occurs),consuming a token

fromeach input place andproducing a token for each output

Petri Nets

Connections are directed

No connections between two places or two transitions Places may hold zero or more tokens

At most one arc between nodes (for now)

Firing is atomic

Multiple transitions may be enabled, but only one fires at a time During execution, the number of tokens may vary if there are transitions for which the number of input places is not equal to the number of output places

The network is static

Example (1)

Petri net for atraffic light

States: red, orange and green

Transitions from red to green, green to orange, and orange to red

Example (1)

Petri net for atraffic light

States: red, orange and green

Transitions from red to green, green to orange, and orange to red

Example (1)

Petri net for atraffic light

States: red, orange and green

Transitions from red to green, green to orange, and orange to red

Example (1)

Petri net for atraffic light

States: red, orange and green

Transitions from red to green, green to orange, and orange to red

Example (2)

Petri net for

Example (2)

Petri net for

2 traffic lights

Example (3)

Petri net for

Lab session

Continue with Assignment 2

Do the pandas, scikit-learn and Algorithmia tutorials Create features

Machine learning

Implement (a small part of) your data mining algorithm on

Algorithmia, and add it to your dashboard Write the (scientific!) report for the assignment Start reading relevant book chapters . . .

Credits

Lecture based on slides belonging to the course book

W. van der Aalst,Process Mining: Discovery, Conformance and