tum data innovation lab · 2018-04-10 · team celonis tum data innovation lab 2018 10 undesired...
TRANSCRIPT
Team CelonisTUM Data Innovation Lab 2018
1
Master in Data
Engineering and Analytics
Margarita Ageeva
Master in Management
Keesiu Wong
Master in Mathematical
Finance and Actuarial Science
Olha Tupko
Master in Computational
Science and Engineering
Sebastian Roßner
TEAM CELONISTUM Data Innovation Lab
Master in Electrical and
Computer Engineering
Master in Management
Ahmed Ayadi
Team CelonisTUM Data Innovation Lab 2018
2
AGENDA
STRATEGIC GOAL
PHASE I: DATA ANALYSIS
PART 1: EXISTING USE CASES
PART 2: NEW USE CASES
PHASE II: MACHINE LEARNING
STRATEGIC GOAL
Team CelonisTUM Data Innovation Lab 2018
3
STRATEGIC GOALStrategic positioning of the project within the Ansoff Matrix
MARKET DEVELOPMENT
MARKET PENETRATION
DIVERSIFICATION
PRODUCT DEVELOPMENT
Existing Products
Exi
stin
g M
ark
ets
New
Mark
ets
New Products
Phase I Part 1: Existing Use Cases
Phase I Part 2: New Use Cases
Team CelonisTUM Data Innovation Lab 2018
4
AGENDA
STRATEGIC GOAL
PHASE I: DATA ANALYSIS
PART 1: EXISTING USE CASES
PART 2: NEW USE CASES
PHASE II: MACHINE LEARNING
PHASE I: DATA ANALYSIS
Team CelonisTUM Data Innovation Lab 2018
5
PHASE I: DATA ANALYSIS
PART II: NEW USE CASESPART I: EXISTING USE CASES
Team CelonisTUM Data Innovation Lab 2018
6
PHASE I: DATA ANALYSIS
PART I: EXISTING USE CASES
Team CelonisTUM Data Innovation Lab 2018
7
7
EXISTING USE CASES11 new Analyses in the Content Store
2016-12-01 CREATE PURCHASE ORDER #12342016-06-23 START PRODUCTION #56782016-07-14 RECEIVE PAYMENT #12342016-07-14 SEND EMAIL #9012
EVENT LOG
7
AI-POWERED ROOT
CAUSE ANALYSIS &
IMPROVEMENT
VISUALIZATION OF THE
ACTUAL PROCESSES
Team CelonisTUM Data Innovation Lab 2018
8
PHASE I: DATA ANALYSIS
PART II: NEW USE CASESPART I: EXISTING USE CASES
Team CelonisTUM Data Innovation Lab 2018
9
PHASE I: DATA ANALYSIS
PART II: NEW USE CASES
Team CelonisTUM Data Innovation Lab 2018
10
UNDESIRED ACTIVITIES
Analyze where
undesired activities
happen – and why.
EFFICIENCY ANALYSIS
See exactly which
process steps can be
improved – for sure.
COMPLEXITY ANALYSIS
Get a feeling for the
complexity for your
processes – quantified.
THREE NEW USE CASESfor any process, for any ERP platform
Team CelonisTUM Data Innovation Lab 2018
11
COMPLEXITY ANALYSISGet a feeling for the complexity of your processes – quantified.
Team CelonisTUM Data Innovation Lab 2018
12
Perfect world
COMPLEXITY ANALYSISGet a feeling for the complexity of your processes – quantified.
Team CelonisTUM Data Innovation Lab 2018
13
Perfect world
COMPLEXITY ANALYSISGet a feeling for the complexity of your processes – quantified.
Team CelonisTUM Data Innovation Lab 2018
14
Perfect world Expectation
COMPLEXITY ANALYSISGet a feeling for the complexity of your processes – quantified.
Team CelonisTUM Data Innovation Lab 2018
15
Perfect world Expectation
COMPLEXITY ANALYSISGet a feeling for the complexity of your processes – quantified.
Team CelonisTUM Data Innovation Lab 2018
16
Perfect world Expectation Reality
COMPLEXITY ANALYSISGet a feeling for the complexity of your processes – quantified.
Team CelonisTUM Data Innovation Lab 2018
17
Perfect world Expectation Reality
COMPLEXITY ANALYSISGet a feeling for the complexity of your processes – quantified.
Team CelonisTUM Data Innovation Lab 2018
18
AGENDA
STRATEGIC GOAL
PHASE I: DATA ANALYSIS
PART 1: EXISTING USE CASES
PART 2: NEW USE CASES
PHASE II: MACHINE LEARNINGPHASE II: MACHINE LEARNING
Team CelonisTUM Data Innovation Lab 2018
19
ROBLEMwhere machine learning empowers process mining
Team CelonisTUM Data Innovation Lab 2018
20
ARIMA ARD Regression LSTM
Classical Statistics Advanced Machine LearningBayesian Trade-off
Team CelonisTUM Data Innovation Lab 2018
21
ARIMAAutoRegressive Integrated Moving Average
• Time series specific
• Removes polynomial trend
• Easy to implement
• Computationally fast
Advantages
• Works with stationary
time series only
• Cannot deal with trend
other then polynomial
Limitations
Team CelonisTUM Data Innovation Lab 2018
22
ARDAutomatic Relevance Determination regression
• Automatic feature selection
• Confidence intervals
• Easy to implement
• Computationally fast
Advantages
• Not times series specific
• Not robust to
distributional changes
Limitations
Team CelonisTUM Data Innovation Lab 2018
23
LSTMLong Short-Term Memory recurrent neural network
http://prog3.com/article/2015-11-25/2826323
• Sequential data specific
• Captures long and short-
term dependencies
• Can handle complex
data structures
Advantages
• Gives point estimates
• Requires large datasets
• Computationally expensive
Limitations
Team CelonisTUM Data Innovation Lab 2018
24
RESULTSin terms of Mean Absolute Error
Before Truncation
ARD LSTM LSTMARIMA ARIMA ARD
105.4
120.5
53.1
105.4
57.858.2
After Truncation
Team CelonisTUM Data Innovation Lab 2018
25
MACHINE LEARNING
Team CelonisTUM Data Innovation Lab 2018
26
2
Logistic
Regression
3
K-Nearest
Neighbors
4
Decision
Trees
5
Random
Forests
1
Naïve
Bayes
6
Feedforward
Neural
Networks
Team CelonisTUM Data Innovation Lab 2018
27
𝑝
𝑥𝑆
𝑥𝐷
512 256 128 64 1
ReLu
Sigmoid
PCA
Team CelonisTUM Data Innovation Lab 2018
28
Random Forests
63
74
79
85
89
86
84
86
81
85
78
Neural Networks
85
88
89
85
92
93
90
89
84
83
88
1
2
3
4
5
6
7
8
9
10
Overall
Logistic Regression
14
37
41
50
63
60
68
72
66
74
45
KNN
59
64
69
75
83
77
81
80
78
78
65
Decision Trees
60
70
75
81
85
79
79
80
79
81
74
Naïve Bayes
38
40
42
46
46
47
55
52
50
59
43
Team CelonisTUM Data Innovation Lab 2018
29
Random Forests
63
74
79
85
89
86
84
86
81
85
78
Neural Networks
85
88
89
85
92
93
90
89
84
83
88
1
2
3
4
5
6
7
8
9
10
Overall
Logistic Regression
14
37
41
50
63
60
68
72
66
74
45
KNN
59
64
69
75
83
77
81
80
78
78
65
Decision Trees
60
70
75
81
85
79
79
80
79
81
74
Naïve Bayes
38
40
42
46
46
47
55
52
50
59
43
Team CelonisTUM Data Innovation Lab 2018
30
CONCLUSION
MARKET DEVELOPMENT
MARKET PENETRATION
DIVERSIFICATION
PRODUCT DEVELOPMENT
Existing Products
Exi
stin
g M
ark
ets
New
Mark
ets
New Products
Phase I Part 1: Existing Use Cases
Phase I Part 2: New Use Cases
Team CelonisTUM Data Innovation Lab 2018
31