Creative Data Mining : Documentation of the teaching results from the Spring Semester 2017

Documentation of the teaching results from the spring semester 2017 

Creative Data Mining 

Danielle Griego, Dani Zünd and Prof. Dr. Gerhard Schmitt

Chair of Information Architecture 


Documentation of teaching results 

Danielle Griego, Dani Zünd, and Gerhard Schmitt 


Teaching 

Danielle Griego, Dani Zünd and Gerhard Schmitt 

Syllabi 

http://www.ia.arch.ethz.ch/category/teaching/fs2017-creative-data-mining/ 

Seminar 


Students 

Biyu Wang, Jiani Liu, Jack Kenning, Hong Liang, Gong Chen, Zhonghau Dai 

Published by 

Swiss Federal Institute of Technology in Zurich (ETHZ) 

Department of Architecture 

Institute of Technology in Architecture 


Wolfgang-Pauli-Strasse 27, HIT H 31.6 

8093 Zurich 

Switzerland 

Zurich, August 2017 

Layout 

Brigitte M. Clements 

Contact 

grigod@arch.ethz.ch | http://www.ia.arch.ethz.ch/griegod/ 

Cover picture: 

Front side: Labelled Frames. Jack Kenning

Course Description and Program 

Creative Data Mining: Uncover and Evaluate 

Mondays 10:00 - 12:00 

051-0726-17U | 2 ECTS* 


Uncover and Evaluate 

The participants of this course learn how to collect, process, 

analyze and interpret real spatial and temporal data in order to 

by using actual data from a recent study and analysing it with 

about the urban environment and test the stated hypothesis 

students with a skill-set to further support their design and 

decision making processes. 

The course focuses on creating deeper insights to critically 

evaluate design alternatives for urban planning projects. Students 

will work with time-series and geo-referenced data including 

temperature, relative humidity, illuminance, noise, people density, 

and dust particulate matter. Subjective impression survey data 

will also be integrated into the student projects to further explore 

20.02.2017 













Course Introduction 

Introduction to Python & Programming I 

Introduction to Python & Programming II 

Data Processing 

Seminar week (No lecture) 

Intro to time-series data analysis 

Time series data analysis ctd. & Machine learning 

Machine learning ctd. 

Holiday (No lecture) 

Programming tutorial applications 

Holiday (No lecture) 

Q&A Feedback Workshop I 

Final iA critique 

Combined critique with the other iA courses 

(13:00 - 18:00) 

experiences. Non-architectural skills the participants can develop 

during this course are 1) an introduction to programming 2) how 

clustering methods like PCA or K-Means could be applied in an 

architectural context. 

Requirement 

Former knowledge of any digital tool or coding language is 

most welcome but NOT required. You only need to provide a 

reasonable amount of motivation and of course a notebook. 

Where 

HIT H 34.1 (Video Wall) 

Supervision 

Danielle Griego 

Daniel Zünd 

Artem Chirkin 

griego@arch.ethz.ch 

zuend@arch.ethz.ch 

chirkin@arch.ethz.ch 

* Total 60 h = 2 ECTS 

Ungraded Semester Performance 

The most recent outline will be found on www.ia.arch.ethz.ch 

Prof. Dr. Gerhard Schmitt 


Information Science Lab 

Wolfgang-Pauli-Strasse 27, 8093 Zurich 

www.ia.arch.ethz.ch

Content 

Does Distance Affect our Feelings? 

Student: Biyu Wang & Jiani Liu 

Urban Perception: A deep learning approach 

Student: Jack Kenning 

Street Cross Section 

Student: Hong Liang 

Security Analysis 

Student: Gong Chen & Zhonghau Dai 

p.9 

p.32 

p.49 

p.103

Objectivity, Subjectivity, Colour 

Student: Andrea Panzeri

Motivation 

As two geography students, Tobler’s first law of geography is one of the most important rules we 

learned about spatial relationship: 

Everything is related to everything else, but near things are more related than distant things. 

This applies to many issues in real world. For example, if you map population density of a country, it 

is highly possible that a dense city is surrounded also by dense cities. We want to examine if this also 

applies to the cognition world. 

In the ESUM (Analyzing trade-offs between Energy and Social performance of Urban Morphologies) 

project, 32 participants were asked to go through a path with 14 checkpoints and give feedback 

about their perception. The checkpoints were specifically selected so that they differ from each other. 

However, we wonder if the participants could feel the difference as planned instead of relating their 

current perception with their impression from last several checkpoints. 

Hypothesis & Research Question 

In our project, we investigated if distance affects people’s feeling. Based on Tobler’s first law of geography, 

we assume that people’s feelings towards a specific place are related to feelings of places nearby. Thus, 

people’s attitudes towards near space are similar. 

The concept of distance we used here is not Euclidian distance, but the length of walking distance in 

the experiment. 

Approach & Methods 

Similarity is something hard to measure precisely. We used clustering as a reference: objects in one 

cluster are more similar than those in two different clusters. 

We used data from the ESUM project and divided them into 3 groups: survey results which reflect 

people’s subjective perception, biofeedback data which measures people’s physical reaction, and 

environment data which reflects the objective difference. Without offering geometry data, we checked 

if the cluster distribution conforms to certain spatial patterns. 

10 

New Methods in Creative Data Mining | Final project documentation

We We did did clustering for for each each group group participant by by participant rather rather than than aggregating 32 32 people’s data 

together, data together, mainly because mainly cognition because cognition among people among as people well as as dynamic well as features dynamic of environment features of are 

different. environment For example, are different. some people For example, might find some it noisy, people but might others find might it noisy, not evaluate but others sounds might with not same 

decibel evaluate value. sounds Some with participants same decibel walked value. in the Some morning, participants while others walked might in do the the morning, same task while in the 

afternoon with significant higher temperature. We also standardized data, since different variable varies 

others might do the same task in the afternoon with significant higher temperature. We also 

in different scale. 

standardized data, since different variable varies in different scale. 

For 

For 

survey 

survey 

results, 

results, 

which 

which 

are small 

are small 

in sample 

in sample 

size, K-means 

size, K-means 

clustering 

clustering 

was used 

was 

and 

used 

14 checkpoints 

and 14 

were clustered into three classes for each participant. To compare the result, we used a metric to 

checkpoints were clustered into three classes for each participant. To compare the result, we 

compare if two checkpoints are more similar than another two: count of participants to recognize each 

two 

used 

checkpoints 

a metric 

into 

to compare 

one cluster. 

if 

For 

two 

all 

checkpoints 

91 (14*13/2) 

are 

checkpoint 

more similar 

pairs, 

than 

the higher 

another 

the 

two: 

count 

count 

is (maximal 

of 

32, participants minimal 0), to the recognize more similar each they two are. checkpoints We checked into if one checkpoints cluster. For closer all 91 to (14*13/2) each other checkpoint show higher 

similarity. pairs, the higher the count is (maximal 32, minimal 0), the more similar they are. We checked if 

checkpoints closer to each other show higher similarity. 

For biofeedback data and environment data, which are almost continuous along the path with 666 

points, For biofeedback we used DBSCAN data and with environment same parameters data, which so the are count almost of clusters continuous would along not be the predetermined 

path with 

and 666 the points, clusters we are used comparable DBSCAN between with same different parameters dataset. so We the mapped count the of results clusters and would compared not be how 

these predetermined two clustering and pattern the clusters differ from are each comparable other for between the same different participant. dataset. We mapped the 

results and compared how these two clustering pattern differ from each other for the same 

participant. 

Results & Discussion 

Subjective feeling 


In the 

Subjective 

first impression, 

feeling 

it looks like the first several checkpoints are more into one cluster, and the last 

several 

In the 

in 

first 

another, 

impression, 

while checkpoints 

it looks like 

in 



first 

middle 

several 

show 

checkpoints 

more uncertainty. 

are more 

See 

into 

example 

one cluster, 

of participant 

and 

No.12 in figure 1. Though, there are also a few exceptions. 

the last several in another, while checkpoints in the middle show more uncertainty. See 

example of participant No.12 in figure 1. Though, there are also a few exceptions. 

Figure 1 Cluster result of participant No.12 

Then we calculated the similarity parameters explained before and the results are shown in 

figure 2. 14 checkpoints were place on a circle and the width of the chords connecting two 

Then checkpoints we calculated represents the similarity the similarity parameters metric explained of these before two. Note and that the results checkpoint are shown 1 and 14 in figure are the 2. 14 

checkpoints were place on a circle and the width of the chords connecting two checkpoints represents 

the similarity metric of these two. Note that checkpoint 1 and 14 are the starting and ending points 

Creative Data Mining | Spring 2017 | Final Projects 

which is the furthest pair although they were placed close to each other. Similarly, 1 and 13, 2 and 14 

are also very far away. Then it shows a tendency that checkpoints closer to each other give out a higher 

New Methods in Creative Data Mining | Final project documentation 

11

similarity in general. Take checkpoint 11 for example, it’s clear that chords by two sides are much wider 

than starting chords and in ending the middle. points which is the furthest pair although they were placed close to each 

other. Similarly, 1 and 13, 2 and 14 are also very far away. Then it shows a tendency that 

checkpoints closer to each other give out a higher similarity in general. Take checkpoint 11 for 

starting and ending points which is the furthest pair although they were placed close to each 

example, it’s clear that chords by two sides are much wider than chords in the middle. 

starting other. Similarly, and ending 1 and points 13, which 2 and is 14 the are furthest also very pair far although away. Then they were it shows placed a tendency close to each that 

other. checkpoints Similarly, closer 1 to and each 13, other 2 and give 14 out are a also higher very similarity far away. in Then general. it shows Take checkpoint a tendency 11 that for 

checkpoints example, it’s closer clear that to each chords other by two give sides out a are higher much similarity wider than in general. chords in Take the middle. checkpoint 11 for 

example, it’s clear that chords by two sides are much wider than chords in the middle. 

Figure 2: Similarity of subjective feelings 

Following the Moran’s I index measuring two-dimensional autocorrelation, we also calculated 

Following the overall the autocorrelation Moran’s I index for measuring the Figure similarity 2: Similarity two-dimensional metric of subjective using following autocorrelation, feelings formula: we also calculated the overall 

autocorrelation for the similarity Figure metric 2: using Similarity following of subjective feelings 

Following the Moran’s I index measuring 14 two-dimensional D formula: 

∑ 

ij autocorrelation, we also calculated 

j=1 C ij ∗ exp⁡(− ⁡) 

Following the overall the autocorrelation Moran’s I index for the measuring similarity two-dimensional metric using ∑ j D ij following autocorrelation, formula: we also calculated 

I 

the overall autocorrelation for the i = 

similarity ∑14 

⁡⁡, j ≠ i 

j=1 

metric C ij /13using following formula: 

14 

D 

∑ 

ij 

j=1 C ij ∗ exp⁡(− 

14 

D 

⁡) 

∑ D 

Cij is the similar metric and Dij is the ∑walking ij 

j=1 C ij ∗ 

distance. 

exp⁡(− j It actually ij 

⁡) 

uses distance-weighted average 

I i = 

∑14 

∑ j D 

⁡⁡, j ≠ i 

of similarity metric divided by ij 

I 

arithmetic 

i = 

average. j=1 C ij /13 When similar 

∑14 

⁡⁡, j ≠ 

metric 

i 

is evenly distributed, the 

Ii would equal 1. Ii larger than 1 illustrates j=1that C ij /13 in general closer points have higher similar 

Cij is the similar metric and Dij is the walking distance. It actually uses distance-weighted average 

metric. The result shows that for most checkpoints it confirms our assumption but only check 

Cij Cij of is is similarity the similar metric metric divided and by Dij is arithmetic is the the walking average. distance. When It actually similar It actually uses metric uses evenly distance-weighted distributed, average the average of 

point 7 and 8. 

similarity of Ii would similarity metric equal metric divided 1. Ii divided larger by arithmetic than by arithmetic 1 illustrates average. that When When in general similar closer metric points is is evenly have distributed, higher similar 

the Ii would 

equal 

Ii metric. Table would 1. 1 Overall Ii The equal larger result autocorrelation 1. than 

Ii shows larger 1 illustrates that Index than for 1 each most that illustrates in checkpoints general that closer it general confirms points closer our have assumption points higher have similar but higher only metric. similar check The result 

shows metric. point 

that 

7 and The for 

8. result most shows checkpoints that for it most confirms checkpoints our assumption it confirms but our only assumption check point but 7 only and check 8. 

Check Point 1 2 3 4 5 6 7 

point 7 and 8. 

Table Overall 1 Overall Correlation autocorrelation 1.87 Index for 1.99 each check point 1.33 1.35 1.33 1.12 0.49 

Table Check 1 Overall Point autocorrelation 8 Index for each 9 check point 10 11 12 13 14 

Check Overall Point Correlation 0.81 1 1.07 2 1.15 3 1.40 4 1.47 5 1.47 6 1.15 7 

Check Overall Point Correlation 1.87 1 1.99 2 1.333 1.35 4 1.33 5 1.12 6 0.49 7 

Overall Check Point Correlation 1.878 1.999 1.33 10 1.35 11 1.33 12 1.12 13 0.49 14 

Check Overall One more Point Correlation thing that worth 0.818 our attention: 1.07 9 checkpoint 1.15 10 9 1.40 and 11 14 are 1.47 actually 12 the 1.47 13 same places 1.15 14 (see 

Overall figure 1 Correlation on the map). However, 0.81 the 1.07 similar metric 1.15 is only 1.40 24, meaning 1.47 that only 1.47 2/3 respondents 1.15 

feel it similarly when they come back to the exact same place several minutes later. So there 

One more thing that worth our attention: checkpoint 9 and 14 are actually the same places (see 

must be something affecting people’s feelings except the environment of the place itself. 

One figure more 1 on thing the map). that worth However, our attention: the similar checkpoint metric is only 9 and 24, 14 meaning are actually that only the same 2/3 respondents places (see 

figure feel it 1 similarly on the map). when However, they come the back similar to the metric exact is only same 24, place meaning several that minutes only 2/3 later. respondents So there 

One 

feel 

more 

must Creative it be similarly 

thing that 

Data something Mining when 

worth 

affecting | Spring they 

our 

come 

attention: 

2017 people’s | back Final feelings Projects to 

checkpoint 

the exact except same 

9 and 

the environment place 

14 are 

several 

actually 

of minutes 

the same 

the place later. 

places 

itself. So there 

(see figure 

1 on the map). However, the similar metric is only 24, meaning that only 2/3 respondents feel it similarly 

must be something affecting people’s feelings except the environment of the place itself. 

when they come back to the exact same place several minutes later. So there must be something 

affecting Creative people’s Data Mining feelings | Spring except 2017 | the Final environment Projects of the place itself. 


12 



Physical reaction and environment difference 

Generally Speaking, biofeedback data follows specific spatial pattern, as the closer points tend to 

be in the same cluster as well as there are less clusters for each participant, even though we do not 

take spatial information into consideration during clustering; while participants’ environment data is 

opposite, which seems to be more randomly distributed. It makes sense since environment data such 

as the level Physical of sound, reaction dust, and etc. environment can be suddenly difference high when a truck pass by and can go back to normal 

level immediately after it runs away. Therefor the result can be interpreted as people’s feelings are 

relatively Generally stable Speaking, and change biofeedback gradually data in spite follows of the specific sudden spatial change pattern, in environment. 

as the closer points tend 

to be in the same cluster as well as there are less clusters for each participant, even though we 

Furthermore, do not take the spatial comparison information between into consideration participants’ during biofeedback clustering; data while and participants’ environment data can 

be summarized environment as data three is opposite, groups, namely, which seems biofeedback to be more data randomly fluctuates distributed. more than It makes environment sense data, 

biofeedback since environment data and environment data such as data the level conform of sound, to similar dust, etc. pattern, can be and suddenly biofeedback high when data a truck fluctuates 

less than pass environment by and can go data. back As to we normal expected, level immediately the first two after pattern it runs exist away. but Therefor only minority. the result Interesting can 

is that, be even interpreted these as cases, people’s actually feelings biofeedback are relatively data stable still and conforms change gradually to specific in spite spatial of the pattern. The 

possible 

sudden 

reason 

change 

is probably 

in environment. 

the too stable environment rather than sudden changes in biofeedback. 

Furthermore, the comparison between participants’ biofeedback data and environment data 

can be summarized as three groups, namely, biofeedback data fluctuates more than 

environment data, biofeedback data and environment data conform to similar pattern, and 

biofeedback data fluctuates less than environment data. As we expected, the first two pattern 

exist but only for minority. Interesting is that, even in these cases, actually biofeedback data still 

conforms to specific spatial pattern. The possible reason is probably the too stable environment 

rather than sudden changes in biofeedback. 


13

The pattern “environment data fluctuates more” is dominant, as 22 out of 29 participants’ show 

this pattern, which proves our hypothesis more or less that biofeedback data possess spatial 

The 

similarity. 

pattern 

As 

“environment 

some examples 

data 

shown 

fluctuates 

below, 

more” 

closer 

is dominant, 

space is highly 

as 22 

likely 

out 

to 


be 

29 

classified 

participants’ 

in the 

show 

this pattern, which proves our hypothesis more or less that biofeedback data possess spatial 

same cluster even though the corresponding environment does not show spatial correlation. 

similarity. As some examples shown below, closer space is highly likely to be classified in the 

The Therefore, pattern “environment we can confirm data that fluctuates people react more” to environment is dominant, as changes 22 out in of a 29 more participants’ smooth way. 

same cluster even though the corresponding environment does not show spatial correlation. show 

this Besides, pattern, there which are proves always our more hypothesis outliers, more which or is less represented that biofeedback as black point data possess along path spatial in the 

similarity. Therefore, graph, in As environment we some can examples confirm data that which shown people also below, proves react closer its to space environment instability. is highly changes likely to be in a classified more smooth in the way. 

same Besides, cluster there even are though always the more corresponding outliers, which environment is represented does not as show black spatial point correlation. along path in the 

Therefore, graph, in environment we can confirm data that which people also react proves to environment its instability. changes in a more smooth way. 

Besides, there are always more outliers, which is represented as black point along path in the 

graph, in environment data which also proves its instability. 

Conclusions 

Through our three groups of clustering, we concluded that in general, people’s feelings towards 

a specific place are related to feelings of places nearby. 

14 

Conclusions Creative Data Mining | Spring 2017 | Final Projects 

Through our New three Methods groups in Creative of clustering, Data Mining we concluded | Final that project in general, documentation people’s feelings towards 

a specific place are related to feelings of places nearby.

Conclusions 

Through our three groups of clustering, we concluded that in general, people’s feelings towards a 

specific place are related to feelings of places nearby. People tend to feel that close places are similar 

both physically, which is proved by biofeedback data, and mentally, which is proved by subjective 

survey data. Not only your brain feels that way, so do other parts of your body. 

However, the conclusion does not apply to all participants in the experiment. The validity is also limited 

by the sample size. We only confirmed that based on the ESUM project data, most people’s feelings 

are also affected by distance besides the actual environment. Further study on larger sample is needed 

to validate our conclusion. 

References 

Tobler, W. R. (1970). A computer movie simulating urban growth in the Detroit region. Economic 

geography, 46(sup1), 234-240. 

Scikit-learn(2011): Machine Learning in Python, Pedregosa et al., JMLR 12, pp. 2825-2830 


15

Urban Perception: A deep learning approach 

Student: Jack Kenning

mation process and ends with a visualization step, making the results visible. From 

Data Set (Energy and Social Performance of Urban Morphologies) collected b 

tion Architecture Motivation chair of ETH Zürich, there were survey responses from specific path-p 

ty of Zürich. The survey asked respondents at each path-point a series of questions, su 

The creative data mining course leads students through the process of data mining with a specific questionbased 

is this approach. location?" The collection, or "How cleaning, ordered selection do of you relevant find data this is location?". followed by a One transformation popular app 

eautiful 

process and ends with a visualization step, making the results visible. From the ESUM Data Set (Energy 

mpare the subjective survey responses to an objective measure that could guide a 

and Social Performance of Urban Morphologies) collected by the Information Architecture chair of ETH 

rban Zürich, design there process. were survey responses from specific path-points in the city of Zürich. The survey asked 

respondents at each path-point a series of questions, such as: “How beautiful is this location?” or 

“How ordered do you find this location?”. One popular approach is to compare the subjective survey 

responses to an objective measure that could guide a more global urban design process. 

sis & Research Questions 

Hypothesis & Research Questions 

an element that seemed to be under evaluated was the presence of vegetation and 

e can 

One 

compare 

urban element 

this 

that 

to 

seemed 

how beautiful, 

to under 

interesting 

evaluated was 

or 


ordered 

presence of 

that 

vegetation 

people 

and 

found 

green 

a sp 

areas. We can compare this to how beautiful, interesting or ordered that people found a specific point. 

“Trees make an urban situation more beautiful.” Do they also make it more interesting and/or more 

ordered? 

ake an urban situation more beautiful." 

also make it more interesting and/or more ordered? 


The ESUM Project Data was a valuable source, but it was necessary to collect other data to evaluate 

the presence of vegetation in an urban context. 

ch & Methods 

M Project Data was a valuable source, but it was necessary to collect other data to 

e the presence of vegetation in an urban context. 

Figure 1: Data Sources 

Data Mining | Spring 2017 | Final Projects 

18 


The presence of vegetation is not something that is easily measured, but a logical approach is to use 

The 

an 

presence 

image-based 

of vegetation 

method. 

is 

This 

not something 

also makes 

that 

sense 

is easily 

as the 

measured, 

research 

but 

question 

a logical 

directly 

approach 

relates 

is to use 

to an 

an 

image-based method. This also makes sense as the research question directly relates to an individual’s 

individual's perception. However, it is not possible to simply identify vegetation by the color of a 

perception. However, it is not possible to simply identify vegetation by the color of a pixel for example, 

because pixel The presence for many example, things of vegetation because in an urban is many not context something things are green an that urban is and easily context this measured, would are result green but in a and a logical large this approach error. would result is to in use a 

large an image-based error. method. This also makes sense as the research question directly relates to an 

Deep 

Deep individual's 

learning 

learning perception. 

based convolution 

based on However, convolution it 

neural 

is neural not 

networks 

possible 

makes 

networks to makes simply 

it possible 

it identify 

to 

possible vegetation 

quite easily 

to quite easily by 

accomplish 

the accomplish color of 

this 

a 

task. Through an established framework, we can train the framework to recognize certain features in 

this pixel task. for example, Through because an established many things framework, in an urban we can context train the are green framework and this to recognize would result certain 

a 

an image - in this case vegetation (trees, bushes, grass, etc.). 

features large error. in an image - in this case vegetation (trees, bushes, grass, etc.). 

Caffe is a framework developed by the Berkeley Vision and Learning Centre and was used in this 

Caffe Deep is learning a framework based on developed convolution by the neural Berkeley networks Vision makes and Learning it possible Centre to quite and easily was used accomplish 

example. The advantage is not requiring much coding knowledge to implement the framework. in Below this 

is example. this 

the proposed 

task. The Through 

workflow. advantage an established not requiring framework, much we coding can train knowledge the framework to implement to recognize the framework. certain 

Below features is the in an proposed image - in workflow. this case vegetation (trees, bushes, grass, etc.). 

Caffe is a framework developed by the Berkeley Vision and Learning Centre and was used in this 

example. The advantage is not requiring much coding knowledge to implement the framework. 

Below is the proposed workflow. 

The extractor can learn to identify certain features Figure 2: corresponding Workflow to what we are trying to identify from 

the training set that we will use and then the classifier makes it possible to identify vegetation on any 

The extractor can learn to identify certain features corresponding to what we are trying to identify 

urban image. These can both be set up in Caffe. 

from the training set that we will use and then the classifier makes it possible to identify vegetation 

on any urban image. These can both be set Figure up in 2: Caffe. Workflow 

The extractor can learn to identify certain features corresponding to what we are trying to identify 

from the training set that we will use and then the classifier makes it possible to identify vegetation 

on any urban image. These can both be set up in Caffe. 

Figure 3: Caffe setup 


Figure 3: Caffe setup 



19

In this case, the project relied largely on SegNet, a modified version of Caffe, developed by Cambridge 

University. Hand-labeled images are used to generate a predictive model that can segment an image 

into different urban elements such as buildings, roads, pavements, trees, vehicles etc. There are 12 

different In this categories case, the that project SegNet relied can largely be trained on to SegNet, identify a from modified the labeled version set, these of Caffe, were developed the categories by 

used for the analysis, more specifically the tree elements. 

Cambridge University. Hand-labeled images are used to generate a predictive model that can 

segment an image into different urban elements such as buildings, roads, pavements, trees, vehicles 

etc. There are 12 different categories that SegNet can be trained to identify from the labeled set, 

these were the categories used for the analysis, more specifically the tree elements. 

Figure 4: SegNet, Cambridge University 

The pre-trained model runs the classification pixel by pixel based on a multitude of input factors, 

but ultimately outputs a segmented image that can then be used as an objective source of urban 

The 

data. 

pre-trained model runs the classification pixel by pixel based on a multitude of input factors, but 

ultimately outputs a segmented image that can then be used as an objective source of urban data. 

20 


Creative Data Mining | Spring 2017 | Final Projects



In order Results to take & Discussion into account peripheral vision as someone walks through a neighborhood, we ran the 

segmentation In order to on take three into images, account taken peripheral at eye-level vision from as someone a vertical walks position. through a neighborhood, we ran 

the In order segmentation to take into on account three images, peripheral taken vision at eye-level as someone from walks a vertical through position. a neighborhood, we ran 

the segmentation on three images, taken at eye-level from a vertical position. 

Figure 5: Segmented images from Zürich survey path 

Figure 5: Segmented images from Zürich survey path 

After this segmentation was done, it was relatively simple to read the number of pixels of each color 

with After a python this segmentation script and establish was done, the percentage it was relatively of image simple corresponding to read the number to a given of pixels category. of each The color 

first 

After with assumption a this python segmentation 

was script to plot and was 

the establish done, 

percentage the it was percentage of 

relatively 

vegetation of simple image to different 

to corresponding read the 

survey 

number 

responses to of a given pixels 

and category. of 

to 

each color The 

analyze with first the assumption a python output: script was and to establish plot the percentage the percentage of vegetation of image corresponding to different survey to a given responses category. and The to 

first analyze assumption the output: was to plot the percentage of vegetation to different survey responses and to 

analyze the output: 

Figure 6: Survey Responses plotted against the segmented tree percentage 

Figure 6: Survey Responses plotted against the segmented tree percentage 


Creative Data Mining New | Spring Methods 2017 in | Creative Final Projects Data Mining | Final project documentation 

21

The The output output had had a a general coherence but but a lot lot of of specific cases which didn't didn’t fit to the trend of a place 

being more beautiful with a larger presence of vegetation. For this reason, it made more sense to move 

being more beautiful with a larger presence of vegetation. For this reason, it made more sense to 

back 

The 

towards 

move output back had 

a more 

towards a general 

general 

a more coherence 

analysis 

general but 

where 

analysis a lot 

all 

where of 


specific 

different 

all the cases 

categories 

different which didn't 

could 

categories fit 

be 

to the 

seen. 

could trend 

The 

be seen. of 

expected 

a place 

output would be to see an increase in the “vegetation” part of the image in places that the survey 

The 

expected being more output beautiful would with a to larger see an presence increase of in vegetation. the "vegetation" For this part reason, of the it made image more in places sense that to 

respondents found more beautiful. As the graph below shows, that was not necessarily the case. 

the move survey back respondents towards a more found general more analysis beautiful. where As the all graph the different below shows, categories that was could not be necessarily seen. The 

the expected case. output would be to see an increase in the "vegetation" part of the image in places that 

the survey respondents found more beautiful. As the graph below shows, that was not necessarily 

the case. 

Figure 7: Relative importance of urban elements plotted to survey beauty index 

The stacked surfaces 

Figure 

represent 

7: Relative importance 

the relative 

of urban 

presence 

elements 

of the 

plotted 

street, 

to survey 

the buildings, 

beauty index 

the trees and the 

The sky, stacked but it doesn't surfaces appear represent that the any relative of these presence have a direct of the influence street, the on buildings, how beautiful the trees people and found the sky, a 

but The given it doesn’t stacked location. appear surfaces To be that sure represent any of the of these the output, relative have the a presence same direct graph influence of the was street, on generated how the beautiful buildings, relative people to the the trees found questions and a given the of 

location. sky, "Like-Dislike", but To it doesn't be sure "Interest-Boring", appear of the output, that any "Order-Disorder". 

the of these same have graph a direct was generated influence relative on how to beautiful the questions people of found “Like- a 

Dislike”, given location. “Interest-Boring”, To be sure “Order-Disorder”. 

of the output, the same graph was generated relative to the questions of 

"Like-Dislike", "Interest-Boring", "Order-Disorder". 

Figure 8: Relative importance for different survey questions 

Figure 8: Relative importance for different survey questions 


22 

Creative Data New Mining Methods | Spring Creative 2017 Data | Final Mining Projects | Final project documentation

Conclusions 

The hypothesis and research questions seem to indicate that the factors leading to a place being 

described as “beautiful”, “ordered”, or “likeable” are more complex than just the presence of vegetation 

in a given place. Multiple factors may have influenced these results, such as the specific location of 

the photos compared to the position of the survey respondents which were not always very precise. 

There were also minor (and major) modifications to the urban environment with one tree having been 

chopped down between the survey and the photo analysis. The sample-size may also have induced 

error in these findings, as only around 30 survey responses were used. With a larger data set it may be 

possible to extract more meaningful results. 

It seems reasonable to imagine that the presence of trees makes a place-point more beautiful but this 

does not directly influence how beautiful a place is. The characteristic appears to be a lot more complex 

that solely quantitative. However, the first data used from the segmented image was an average of the 

number of pixels representing vegetation in the photo. This was considered quite a poor measurement 

in relation to the other information contained in a segmented image, so in the second visualisation the 

relative importance of segments is taken into account. For further analysis, maybe an indicator such as 

the distribution across the image can be looked at. 

References 

Caffe - Berkeley Vision and Learning Centre http://caffe.berkeleyvision.org/ 

SegNet - Cambridge University http://mi.eng.cam.ac.uk/projects/segnet 

NVIDIA Caffe Course http://on-demand.gputechconf.com/gtc/2015/webinar/deep-learning-course/ 

gettingstarted-with-caffe.pdf 

ESUM Project - ETH Zürich http://www.ia.arch.ethz.ch/esum/ 


23

Street Cross Section 

Student: Hong Liang

Motivation 

Street cross section design is an import part urban planning. In most cases, it depends on the road 

grade, and the property and function of the road. Once the road grade and capacity are confirmed, 

the type of street cross section is roughly decided. Street cross section offers functional services to 

pedestrian, on the other hand, different fraction of various street cross section elements may create 

different atmosphere, which can also effect pedestrian’s feeling and behavior. But how does it work? 

In this project, with the ESUM data set and some suitable analysis methods, the relationship between 

street cross section and pedestrian’s reaction will be explored and discussed. 

Hypothesis & Research Question(s) 

The main question of this project is to explore the effect of street cross section on pedestrian. The 

elements of street cross section mainly include drive lane, parking lane, bike lane and side walk. So the 

more specific question is: which element plays a more important roll on effecting pedestrian’s feeling? 

For example, parking lane could be a negative factor. More wide the parking lane is, it could make 

pedestrian feel more insecure and chaotic. 


Data Collection and Preparation 

In order to describe the street cross section along the path, I measured every street cross section 

element on 14 street points on google map, and compared them with the photos. The collected data 

and street cross section diagrams are as follow: 

26 


Table 1 street cross section data of 14 points 

point street width sidewalk width sidewalk rate parking lane width parking lane rate trees crossing way 

1 12 3 0.2500 4.2 0.3500 1 none 

2 12 3 0.2500 4.2 0.3500 1 none 

3 4.2 1.2 0.2857 0 0.0000 1 none 

4 12.6 3.6 0.2857 0 0.0000 1 none 

5 12.4 4.3 0.3468 0 0.0000 0 zebra crossing 

6 13.5 4.4 0.3259 2.1 0.1556 0 zebra crossing 

7 11.3 3.6 0.3186 4.2 0.3717 1 none 

8 16.8 3 0.1786 0 0.0000 1 zebra crossing 

9 13.8 3.6 0.2609 4.2 0.3043 0 none 

10 11.7 3.6 0.3077 2.1 0.1795 1 zebra crossing 

11 18.5 4 0.2162 0 0.0000 0 zebra crossing 

12 19.3 3.6 0.1865 0 0.0000 0 zebra crossing 

13 15 3.6 0.2400 2.1 0.1400 0 zebra crossing 

14 13.8 3.6 0.2609 4.2 0.3043 0 none 

Figure 1 street cross section diagram of 14 points 

As shown in the diagram above, due to the different street grade and function, the street cross 

section elements at 14 points are different. In particular, street width and parking lane are the 

As primary shown in differences the diagram among above, 14 due points. to the At different the same street time, grade side walk and function, is the most the important street cross part section of 

elements street for at 14 pedestrian. points are Therefore, different. In I particular, chose street street width width (width and parking between lane two are buildings), the primary parking differences 

among 14 points. At the same time, side walk is the most important part of street for pedestrian. 

lane fraction (parking lane width / street width) and side walk fraction (side walk width / street 

Therefore, I chose street width (width between two buildings), parking lane fraction (parking lane width 

/ street 

width) 

width) 

to describe 

and side 


walk 

property 

fraction 


(side 

street 

walk 

cross 

width 

section. 

/ street 

Some 

width) 

binary 

to describe 

variables, 


such 

property 

as whether 

of street 

cross there section. are trees Some of zebra binary crossings variables, are such also as taken whether into there consideration. are trees of zebra crossings are also taken 

into consideration. 

To describe pedestrian’s feeling, I chose 3 questions from the survey result of ESUM: like of 

To dislike, describe ordered pedestrian’s of chaotic, feeling, secure I chose or insecure. 3 questions ESUM from experiment the survey was result taken of by ESUM: 37 participants. like of dislike, 

ordered In order of to chaotic, compare secure the results or insecure. with the ESUM street experiment cross section’s was taken property by 37 at 14 participants. points more In easily, order to 

compare first, I calculate the results the with average the street value cross for 3 section’s questions property at 14 points. at 14 points more easily, first, I calculate the 

average value for 3 questions at 14 points. 

Creative Data Mining | Spring 2017 | Final Projects 2 


27

Figure 2 survey result 

As shown in the diagrams above, the results of 3 survey questions present similar tendency. It 

As shown can be in also the concluded diagrams from above, the the analysis, results that of 3 the survey deviation questions at some present points similar (e.g. point tendency. 7) is It can 

be also relatively concluded small, from which the means analysis, the that average the deviation value is more at some credible, points while (e.g. point the feelings 7) is relatively of small, 

which pedestrian means the at average point 5 and value 6 are is more more credible, disperse. while the feelings of pedestrian at point 5 and 6 are 

more disperse. 

Data Analysis Method 

Data Analysis Method 

Data analysis is divided into 4 steps. 

Data analysis 1. Average is divided value into 4 steps. 

In first step, I compare the calculated 3 average values of survey results with 3 street cross 

section properties (street width, parking lane fraction and side walk fraction) in turn, to get 

1. Average value 

the first impression of the relation ship between them. 

In first step, 2. Cluster I compare the calculated 3 average values of survey results with 3 street cross section 

properties (street width, parking lane fraction and side walk fraction) in turn, to get the first impression 

of the relation I choose ship k-mean between as them. cluster method to analyze the relationship between 3 average 

values of survey results with 3 street cross section properties. In this step, some semiquantitative 

results can be 2. Cluster 

concluded. 

I choose 3. k-mean Multi-variable as the linear cluster regression method to analyze the relationship between 3 average values of 

survey results with 3 street cross section properties. In this step, some semiquantitative results can be 

concluded. In multi-variable regression, two more binary variables (trees and zebra crossing) are also 

taken into consideration. For each survey result, we will get a formula consisting of 5 street 

3. Multi-variable cross section linear variables. regression And according to its coefficient, we can get which variable’s effect is 

more significate. 

In multi-variable regression, two more binary variables (trees and zebra crossing) are also taken into 

consideration. 4. Simple For each linear survey regression result, we will get a formula consisting of 5 street cross section variables. 

And according to its coefficient, we can get which variable’s effect is more significate. 

Once we get the most import factor for each survey result, we can do the simple linear 

4. Simple regression linear to regression test its significance and credibility again. 

Once we get the most import factor for each survey result, we can do the simple linear regression to 

test its significance and credibility again. 


28 




1. Average value 

Figure 3 comparison between survey result and street cross section property 

According to the diagrams, it is more or less clear that there is correspondent relationship between side 

walk fraction and pedestrian’s feeling in all 3 survey results (like, ordered and secure). The biggest 

difference occurs at point 5. But point 5 is an intersection, and the street cross data can only describe on 

According to the diagrams, it is more or less clear that there is correspondent relationship between 

street. That could be the reason to explain that big gap. Additionally, the deviation of survey results at 

side walk fraction and pedestrian’s feeling in all 3 survey results (like, ordered and secure). The biggest 

point 5 is relatively big, it could also result in this difference. 

difference occurs at point 5. But point 5 is an intersection, and the street cross data can only describe 

on street. That could be the reason to explain that big gap. Additionally, the deviation of survey results 

at point 5 is relatively big, it could also result in this difference. 



29

2. Cluster 

Figure 4 cluster analysis 

Due to the limitation of the amount of data set, the results of cluster analysis are not very 

significant. But we can still get the similar semi-quantitative result that low side walk fraction 

will result in more negative feeling, while the pedestrian is more likely to have a positive 

Figure 

evaluation 

4 cluster 

for those 

analysis 

points which have high side walk fraction. 

Due In additional, to the limitation there is of also the amount some obvious of data cluster set, the between results of parking cluster lane analysis fraction are not and very secure. significant. It 

But 

seems 

we 

that 

can 

higher 

still get 

parking 

the similar 

lane 

semi-quantitative 

fraction leads to 

result 

an insecure 

that low 

feeling 

side walk 

for pedestrian. 

fraction will result in more 

negative feeling, while the pedestrian is more likely to have a positive evaluation for those points 

which If the experiment have high side could walk have fraction. more participants and cover more streets, cluster analysis will 

have more credible conclusion. 

In additional, there is also some obvious cluster between parking lane fraction and secure. It seems 

that higher parking lane fraction leads to an insecure feeling for pedestrian. If the experiment could 

have more participants and cover more streets, cluster analysis will have more credible conclusion. 


30 


3. Multi-variable linear regression 

Like = 

– 0.046 * street width 

+ 0.312 * sidewalk fraction 

– 1.350 * parking lane fraction 

+ 0.259 * trees 

– 0.600 * zebra crossing 

Figure 5 multi-variable linear regression for 'like or dislike' 

Ordered = 

+ 0.046 * street width 



+ 0.383 * trees 


Figure 6 multi-variable linear regression for ‘ordered or chaotic’ 



31

Secure = 



Secure = 

0.720 parking lane fraction 


0.208 trees 


0.521 zebra crossing 


+ 0.208 * trees 


Figure 7 multi-variable linear regression for ‘secure or insecure’ 

Table 2 coefficients of multi-variable linear regression 

Figure 7 multi-variable linear regression for ‘secure or insecure’ 

street width sidewalk fraction parking lane fraction trees zebra crossing 

like -0.046 0.312 -1.350 0.259 -0.600 

Table 2 coefficients of multi-variable linear regression 

ordered 0.046 5.898 -0.293 0.383 -0.503 

secure street width -0.008 sidewalk fraction 2.269 parking lane fraction -0.720 trees0.208 zebra crossing -0.521 

From 

like -0.046 0.312 -1.350 0.259 -0.600 

the table and diagrams, we can get the conclusion that side walk fraction and trees have 

and 

ordered 0.046 5.898 -0.293 0.383 -0.503 

positive effect on pedestrian, because all the coefficients are positive. On the other hand, 

parking 

secure -0.008 2.269 -0.720 0.208 -0.521 

lane fraction and zebra crossing effect pedestrian in a more negative way. 

From the table and diagrams, we can get the conclusion that side walk fraction and trees have and 

positive From the effect table on and pedestrian, diagrams, because we can get all the coefficients conclusion are that positive. side walk On fraction the other and hand, trees parking have 

lane and positive fraction and effect zebra on pedestrian, crossing effect because pedestrian all the in coefficients a more negative are positive. way. On the other hand, 

parking 

4. Simple 

lane fraction 

linear regression 

and zebra crossing effect pedestrian in a more negative way. 

4. Simple linear regression 

Figure 8 simple linear regression 

Figure 8 simple linear regression 


32 


From last analysis, we find the most important factor for each question, and test it with simple linear 

regression. According to the diagrams above, the negative effect of parking lane fraction on ‘like or 

dislike’ is not very significant. But the positive effect of side walk fraction is still credible. 

Conclusions 

After all these analysis, we can now answer the questions at the beginning. The different elements 

of the street cross section effect pedestrian in a different way. We can find that parking lane fraction 

has a negative effect on pedestrian, while side walk fraction affect pedestrian in a more significantly 

positive way. 

But to get a more meaningful and credible results, we still need more data. Additionally, the binary 

variable should be avoided in multi-variable regression. For example, ‘trees’ variable could be 

replaced by green area. Or there could be some other methods more suitable for binary variable 

analysis. That’s could be the further step of this project. 

References 

Marc Schlossberg, John Rowell, Davd Amos, and Kelly Sanford, Rethinking Streets: 

An Evidence-Based Guide to 25 Complete Street Transformations, 2013. 


33

34 



Student: Gong Chen & Zhonghau Dai 


35

Motivation 


Security level is one important index to evaluate urban planning. Security level can be treated as a 

citizens’ comprehensive subjective reflection on urban surroundings. Therefore, based on the project 

ESUM- Name Analyzing trade-offs between Energy and Social performance of Urban Morphologies, 

various datasets are available. Among there datasets, survey questions are of most interest, which 

covers twelve questions (Figure 1), including security, noise, beauty evaluation and etc. Gong Hence, Chen 

there should be a best combination of these parameters to provide us a high urban security. 

Zhonghao Dai 

Motivation 

Security level is one important index to evaluate urban planning. Security level can be treated as a 

citizens’ comprehensive subjective reflection on urban surroundings. Therefore, based on the project 

ESUM- Analyzing trade-offs between Energy and Social performance of Urban Morphologies, various 

datasets are available. Among there datasets, survey questions are of most interest, which covers twelve 

questions (Figure 1), including security, noise, beauty evaluation and etc. Hence, there should be a best 

combination of these parameters to provide us a high urban security. 

Figure 1: Survey questions 

Hypothesis & Research Question(s) 

Hypothesis 

Following are questions 

& Research 

conveyed 

Question(s) 

during ESUM project, dislike/like, chaotic/ordered, noisy/quiet, 

private/public, boring/interesting, crowded/empty, insecure/secure, ugly/beautiful, narrow/spacious, 

Following 

enclosed/open, 

are questions 

dark/light, 

conveyed 

unfamiliar/familiar, 

during ESUM 

among 

project, 

which insecure/secure 

dislike/like, chaotic/ordered, 

is response variable 

noisy/quiet, 

private/public, focused during boring/interesting, this project, all these crowded/empty, variables are evaluated insecure/secure, -2, -1, 0, 1, ugly/beautiful, 2. narrow/spacious, 

enclosed/open, dark/light, unfamiliar/familiar, among which insecure/secure is response variable 

focused during this project, all these variables are evaluated in -2, -1, 0, 1, 2. 

Based on intuition, when a place is quite, it will make people feel secure while when it goes noisy, 

Based people on tend intuition, to feel when upset and a place become is quite, feeling it unsecure. will make In people this case, feel one secure hypothesis while is made when that it goes noisy, 

people secure/insecure tend to feel has upset a negative and (because become the feeling high level unsecure. in questionnaire In this case, means one quiet) hypothesis relationship is made with that 

secure/insecure noisy/quiet a has wide a range. negative And, (because by searching the high such level relationship, in questionnaire a best range means of noise quiet) are relationship intended to be with 

noisy/quiet found. in a wide range. And, by searching such relationship, a best range of noise are intended to 

be found. 


36 


Besides, another hypothesis made in project is that people’s evaluation truly reflects noise level in 

decibel. 


Since dataset is continuous and labeled, regression method is applied in project. Main procedures 

carried in project are listed following: 

1. Besides, Decided another weather hypothesis to use made noise in in project decibel is level that people’s or in evaluation evaluation level truly in survey reflects questions, noise level i.e. in does 

people’s decibel. evaluation on noise level truly corresponds to decibel level 

2. Searching potential variables related with security, both from visualization of data and linear 

regression method, i.e. single variable liner regression 


3. Since Do multivariable dataset is continuous linear regression and labeled, to regression acquire a method simple is linear applied regression in project. model Main to procedures quantify security 

level carried in project are listed following: 

4. Check 1. Decided if the above weather linear to use model noise can in decibel simplified level or further in evaluation to get level the most in survey crucial questions, parameters i.e. does 

people’s evaluation on noise level truly corresponds to decibel level 

2. Searching potential variables related with security, both from visualization of data and linear 

regression method, i.e. single variable liner regression 

3. Do multivariable linear regression to acquire a simple linear regression model to quantify 


security level 

4. Check if the above linear model can be simplified further to get the most crucial parameters 

1. Noise in decibel level versus evaluation level Boxplot of decibel level and evaluation level are 

both shown below (Figure 2 and Figure 3). Although boxplot provides more information e.g. outlier, 

percentile, it is hart to tell relationship. 


1. Noise in decibel level versus evaluation level 

Boxplot of decibel level and evaluation level are both shown below (Figure 2 and Figure 3). 

Although boxplot provides more information e.g. outlier, percentile, it is hart to tell relationship. 

Figure 2: Decibel level 

Figure 3: Noise level 

Then two boxplots are shown in Figure 4 in the same axis, and mean values of these two data are 

shown in Figure 5. Important thing here is that in question survey, -2 means noisy and 2 means 

Then two 

quite, 

boxplots 

so noise 

are 

level 

shown 

is replaced 

in Figure 

with 

4 in 

its 


opposite 

same 

number. 

axis, and 

Basically, 

mean values 

from Figure 

of these 

4 & 5, 

two data are 

shown in 

evaluation 

Figure 5. Important 

on noisy/quite 

thing 

level 

here 

corresponds 

is that in question 

to decibel 

survey, 

level obtained 

-2 means 

by sensor. 

noisy and 2 means quite, 

so noise level is replaced with its opposite number. Basically, from Figure 4 & 5, evaluation on noisy/ 

quite level corresponds to decibel level obtained by sensor. 


37

Figure 4: Decibel versus noise level 

2. Potential variable searching 

2.1 Visualization 

2. Potential variable searching 

Figure 5: Mean value 

Following figures show mean value of 11 questions mentioned above versus security evaluation 

2.1 Visualization 

level at 14 path points (Figure 6), from which the most related variables are selected: dislike/like, 

Following figures show mean value of 11 questions mentioned above versus security evaluation level 

at 14 path points (Figure 6), from which the most related variables are selected: dislike/like, chaotic 

ordered, Creative noisy/quiet, Data Mining | boring/interesting, Spring 2017 | Final Projects crowded/empty, ugly/beautiful, which means these variables 3 

determines insecure/secure level significantly. 

38 


chaotic/ordered, noisy/quiet, boring/interesting, crowded/empty, ugly/beautiful, which means 

these variables determines insecure/secure level significantly. 

Figure 6: Mean value of each path point 

Also, scatter plot methods can be employed here to show linear relationship between these 

variables and security more clearly, for example shown in figure 7. 

Also, scatter plot methods can be employed here to show linear relationship between these variables 

and security more clearly, for example shown in figure 7. 



39

Figure 7: Scatter plot 

2.2 Single variable linear regression 

2.2 Single Apart variable from mean linear value visualization regression of 14 path points, single variable linear regression is also 

applied here to find the coefficient between each of these 11 survey questions and security 

Apart from level. mean In value these visualization case, all datasets of 14 are path employed points, instead single of variable mean values linear only regression in section is also 2.1. applied The 

here to find results the coefficient together with between coefficient each and of R-square these 11 values survey are questions shown in Figure and security 8. level. In these 

case, all datasets are employed instead of mean values only in section 2.1. The results together with 

coefficient The and same R-square result values is given: are dislike/like, shown in Figure chaotic/ordered, 8. noisy/quiet, boring/interesting, 

crowded/empty, ugly/beautiful, these 6 variables are most likely to determine insecure/secure 

The same result 

level. 

is given: dislike/like, chaotic/ordered, noisy/quiet, boring/interesting, crowded/empty, 

ugly/beautiful, these 6 variables are most likely to determine insecure/secure level. 

Notes: Since there are only 5*5 combinations in each subfigure, overlaps are inevitable, thus 

Notes: Since there are only 5*5 combinations in each subfigure, overlaps are inevitable, thus bigger the 

bigger the circle is, more overlaps there are. 

circle is, more overlaps there are. 

3. Multivariable linear regression 

First, linear regression with 6 dependent variables and security level as response variable is 

accomplished with its corresponding model: 

3. Multivariable linear regression 

First, linear 

Security 

regression 

level 

with 6 dependent variables and security level as response variable is accomplished 

with its corresponding model: 

= 0.53 + 0.0685*noise + 0.2211* order + 0.1121*beauty + 0.0975*density+ 

0.1670*like+ 0.1473* interesting (1) 

Security level 

= 0.53 + 

Note: 

0.0685*noise 

Abbreviation 

+ 0.2211* 

is used 

order 

here: 

+ 

like 

0.1121*beauty 

= dislike/like, 

+ 

order 

0.0975*density+ 

= chaotic/ordered, noise = 

0.1670*like+ 0.1473* interesting (1) 

noisy/quiet, interesting = boring/interesting, density = crowded/empty, beauty = ugly/beautiful 

In this case, the R-squared value is 0.323, which is satisfying since when do linear regression of 

Note: Abbreviation is used here: like = dislike/like, order = chaotic/ordered, noise = 

all other parameters i.e. all questions except security level, the value is 0.339. The accuracy is 

noisy/quiet, interesting = boring/interesting, density = crowded/empty, beauty = ugly/beautiful. 

not destroyed while 5 parameters are dropped, making the model much simple and clear. 

In this case, the R-squared value is 0.323, which is satisfying since when do linear regression of all other 

parameters i.e. all questions except security level, the value is 0.339. The accuracy is not destroyed 

while 5 parameters are dropped, making the model much simple and clear. 


40 




41

Figure 8: Single variable linear regression 

From this equation, From this noise equation, level does noise not level have does much not to have do much with security to do with compared security compared with resting with parameters, resting 

in next step, parameters, feasibility of in model next step, simplification feasibility of will model be checked. simplification will be checked. 

4. Model simplification 

4. Model simplification 

In this section, the model is check to see if the above linear model can be simplified further to get 

the most crucial parameters, basically, R-squared mean is applied to check the accuracy of 

In this section, model. the model In this case, is check the noise to see is dropped, if the above and the linear new model can is followed: be simplified further to get the 

most crucial parameters, basically, R-squared mean is applied to check the accuracy of model. In this 

case, the noise Security is dropped, level and the new model is followed: 

= 0.49 + 0.2285* order + 0.1312*beauty + 0.1209*density+ 0.1312*like+ 0.1469* 

interesting (2) 

Security level 

Meanwhile, R value drops only by 0.002 i.e. 0.625%, so this model is still effective. 

= 0.49 + 0.2285* order + 0.1312*beauty + 0.1209*density+ 0.1312*like+ 0.1469* interesting (2) 

Meanwhile, Conclusions R value drops only by 0.002 i.e. 0.625%, so this model is still effective. 

According to equation (1), the security level does have relation with noise, but a positive coefficient is 

shown in equation (2) which means people feels more secure when it is quiet. So a wrong hypothesis is 

made although security and noise level are correlated, but the relationship is negative. However, 

Conclusions 

According to equation (1), the security level does have relation with noise, but a positive coefficient is 

shown in equation (2) which means people feels more secure when it is quiet. So a wrong hypothesis 


is made although security and noise level are correlated, but the relationship is negative. However, 

compared with other five parameters in equation (2), noise level is negligible and order level, i.e. the 

place is chaotic or ordered, is dominates in security level. If security level is intended to be increased, 

one should make the place more ordered. 

42 



43

Chair of 

Information 

Architecture

Creative Data Mining : Documentation of the teaching results from the Spring Semester 2017

Create successful ePaper yourself

Delete template?

Save as template?