data envelopment analysis with principal component ... · data envelopment analysis with principal...
TRANSCRIPT
![Page 1: Data envelopment analysis with principal component ... · Data envelopment analysis with principal component analysis (final presentation) Toni Huuhka 27.8.2020 Instructor: Fabricio](https://reader034.vdocument.in/reader034/viewer/2022051605/600bd75d574bb10a0416f9d2/html5/thumbnails/1.jpg)
Data envelopment analysis with principal
component analysis
(final presentation)Toni Huuhka
27.8.2020
Instructor: Fabricio Oliveira
Supervisor: Fabricio Oliveira
Työn saa tallentaa ja julkistaa Aalto-yliopiston avoimilla verkkosivuilla. Muilta osin kaikki oikeudet pidätetään.
![Page 2: Data envelopment analysis with principal component ... · Data envelopment analysis with principal component analysis (final presentation) Toni Huuhka 27.8.2020 Instructor: Fabricio](https://reader034.vdocument.in/reader034/viewer/2022051605/600bd75d574bb10a0416f9d2/html5/thumbnails/2.jpg)
Background
• The majority of companies are not operating in an ideal
manner
• How could a company improve its operational
efficiency?
1. Rank similar companies and identify the most efficient ones
2. Identify the factors that cause the difference in efficiency
3. Implement these ”best-practices”
![Page 3: Data envelopment analysis with principal component ... · Data envelopment analysis with principal component analysis (final presentation) Toni Huuhka 27.8.2020 Instructor: Fabricio](https://reader034.vdocument.in/reader034/viewer/2022051605/600bd75d574bb10a0416f9d2/html5/thumbnails/3.jpg)
Contents
• Data
• Methods
– Principal Component Analysis
– Data Envelopment Analysis
• Results
• Conclusions
![Page 4: Data envelopment analysis with principal component ... · Data envelopment analysis with principal component analysis (final presentation) Toni Huuhka 27.8.2020 Instructor: Fabricio](https://reader034.vdocument.in/reader034/viewer/2022051605/600bd75d574bb10a0416f9d2/html5/thumbnails/4.jpg)
Data
• 221 active Scandinavian retailers
– No retailers of motor vehicles and motorcycles
– Revenue ≥ 50 million €
• Data retrieved from Orbis Database
• 3 input variables, 2 output variables
– Inputs: Number of Employees (I1), Number of current Directors
& Managers (I2), Number of current Advisors (I3)
– Outputs: Sales (O1), Cost of Employees (O2)
![Page 5: Data envelopment analysis with principal component ... · Data envelopment analysis with principal component analysis (final presentation) Toni Huuhka 27.8.2020 Instructor: Fabricio](https://reader034.vdocument.in/reader034/viewer/2022051605/600bd75d574bb10a0416f9d2/html5/thumbnails/5.jpg)
Data: Correlations
O1 O2 I1 I2 I3
O1 1.0000
O2 0.9573 1.0000
I1 0.9173 0.9676 1.0000
I2 0.4708 0.5245 0.4931 1.0000
I3 0.1576 0.1733 0.1845 0.2295 1.0000
![Page 6: Data envelopment analysis with principal component ... · Data envelopment analysis with principal component analysis (final presentation) Toni Huuhka 27.8.2020 Instructor: Fabricio](https://reader034.vdocument.in/reader034/viewer/2022051605/600bd75d574bb10a0416f9d2/html5/thumbnails/6.jpg)
Methods: Data Envelopment Analysis
• Linear programming
method
• Identifies a ”best-
practice frontier”
• Sensitive to outliers
![Page 7: Data envelopment analysis with principal component ... · Data envelopment analysis with principal component analysis (final presentation) Toni Huuhka 27.8.2020 Instructor: Fabricio](https://reader034.vdocument.in/reader034/viewer/2022051605/600bd75d574bb10a0416f9d2/html5/thumbnails/7.jpg)
Methods: Principal Component Analysis
• A method for creating
uncorrelated basis
vectors
• Commonly used for
variable reduction
• A more robust version
– Robust PCA
![Page 8: Data envelopment analysis with principal component ... · Data envelopment analysis with principal component analysis (final presentation) Toni Huuhka 27.8.2020 Instructor: Fabricio](https://reader034.vdocument.in/reader034/viewer/2022051605/600bd75d574bb10a0416f9d2/html5/thumbnails/8.jpg)
Methods: Principal Component Analysis
• Both DEA and PCA are sensitive to outliers
• The effects of centering and scaling the data before
PCA were studied
• PCA-DEA and RPCA-DEA was performed with every
combination of scaling and centering
A total of six efficiency scores were
computed in addition to the original
three
![Page 9: Data envelopment analysis with principal component ... · Data envelopment analysis with principal component analysis (final presentation) Toni Huuhka 27.8.2020 Instructor: Fabricio](https://reader034.vdocument.in/reader034/viewer/2022051605/600bd75d574bb10a0416f9d2/html5/thumbnails/9.jpg)
Results: PCA and Robust PCAPCA
IPC1 IPC2 IPC3 OPC1 OPC2
𝑹𝟐 0.543 0.289 0.168 0.979 0.021
Cum.
𝑹𝟐0.543 0.832 1.000 0.979 1.000
Robust PCA
IPC1 IPC2 IPC3 OPC1 OPC2
𝑹𝟐 0.534 -0.087 0.129 0.978 -0.460
Cum.
𝑹𝟐0.534 0.447 0.576 0.978 0.519
![Page 10: Data envelopment analysis with principal component ... · Data envelopment analysis with principal component analysis (final presentation) Toni Huuhka 27.8.2020 Instructor: Fabricio](https://reader034.vdocument.in/reader034/viewer/2022051605/600bd75d574bb10a0416f9d2/html5/thumbnails/10.jpg)
Results: DEA, PCA-DEA, RPCA-DEA
DEA PCA-DEA RPCA-DEA
![Page 11: Data envelopment analysis with principal component ... · Data envelopment analysis with principal component analysis (final presentation) Toni Huuhka 27.8.2020 Instructor: Fabricio](https://reader034.vdocument.in/reader034/viewer/2022051605/600bd75d574bb10a0416f9d2/html5/thumbnails/11.jpg)
Results: Non-centered PCA-DEA & RPCA-
DEAPCA-DEA RPCA-DEA
IPC1 IPC2 IPC3 OPC1 OPC2
𝑹𝟐 0.735 0.192 0.073 0.982 0.018
Cum.
𝑹𝟐0.735 0.927 1.000 0.982 1.000
IPC1 IPC2 IPC3 OPC1 OPC2
𝑹𝟐 0.785 0.122 -0.397 0.978 -0.404
Cum.
𝑹𝟐0.785 0.907 0.510 0.978 0.574
![Page 12: Data envelopment analysis with principal component ... · Data envelopment analysis with principal component analysis (final presentation) Toni Huuhka 27.8.2020 Instructor: Fabricio](https://reader034.vdocument.in/reader034/viewer/2022051605/600bd75d574bb10a0416f9d2/html5/thumbnails/12.jpg)
Results: Non-scaled PCA-DEA & RPCA-
DEAPCA-DEA
IPC1 IPC2 IPC3 OPC1 OPC2
𝑹𝟐 1.000 0.000 0.000 0.999 0.001
Cum.
𝑹𝟐1.000 1.000 1.000 0.999 1.000
RPCA-DEA
IPC1 IPC2 IPC3 OPC1 OPC2
𝑹𝟐 1.000 0.000 0.000 0.999 -0.013
Cum.
𝑹𝟐1.000 1.000 1.000 0.999 0.986
![Page 13: Data envelopment analysis with principal component ... · Data envelopment analysis with principal component analysis (final presentation) Toni Huuhka 27.8.2020 Instructor: Fabricio](https://reader034.vdocument.in/reader034/viewer/2022051605/600bd75d574bb10a0416f9d2/html5/thumbnails/13.jpg)
Results: Non-scaled, non-centered PCA-
DEA & RPCA-DEAPCA-DEA
IPC1 IPC2 IPC3 OPC1 OPC2
𝑹𝟐 1.000 0.000 0.000 0.999 0.001
Cum.
𝑹𝟐1.000 1.000 1.000 0.999 1.000
RPCA-DEA
IPC1 IPC2 IPC3 OPC1 OPC2
𝑹𝟐 1.000 0.000 0.000 0.999 -0.014
Cum.
𝑹𝟐1.000 1.000 1.000 0.999 0.985
![Page 14: Data envelopment analysis with principal component ... · Data envelopment analysis with principal component analysis (final presentation) Toni Huuhka 27.8.2020 Instructor: Fabricio](https://reader034.vdocument.in/reader034/viewer/2022051605/600bd75d574bb10a0416f9d2/html5/thumbnails/14.jpg)
Conclusions
• Scaling more important than centering prior to any form
of (R)PCA-DEA
• Robust estimators improve the results of PCA-DEA
• No clear advantage was found between centered and
not-centered data with either PCA-DEA or RPCA-DEA
![Page 15: Data envelopment analysis with principal component ... · Data envelopment analysis with principal component analysis (final presentation) Toni Huuhka 27.8.2020 Instructor: Fabricio](https://reader034.vdocument.in/reader034/viewer/2022051605/600bd75d574bb10a0416f9d2/html5/thumbnails/15.jpg)
Future prospects
• The effects of removing outliers should be studied, the
outliers were not removed in this study
• Different assumptions regarding returs to scale could be
investigated
• Additional research regarding when variable reduction
techniques are necessary to implement