Using Data Mining to Assess Environmental ... - ACS Publications

Subscriber access provided by Kaohsiung Medical University

Environmental Modeling

Using Data Mining to Assess Environmental Impacts of Household Consumption Behaviors Andreas Froemelt, David J. Dürrenmatt, and Stefanie Hellweg Environ. Sci. Technol., Just Accepted Manuscript • DOI: 10.1021/acs.est.8b01452 • Publication Date (Web): 22 Jun 2018 Downloaded from http://pubs.acs.org on June 24, 2018

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 38

Environmental Science & Technology

1

Using Data Mining to Assess Environmental

2

Impacts of Household Consumption Behaviors

3

Andreas Froemelt*, David J. Dürrenmatt† and Stefanie Hellweg

4

Chair of Ecological Systems Design, Institute of Environmental Engineering, ETH Zurich, John-

5

von-Neumann-Weg 9, 8093 Zurich, Switzerland

6

KEYWORDS Household Consumption, Household Behavior, Life Cycle Assessment, Hybrid

7

LCA, Data Mining, Machine Learning, Self-Organizing Map, Kohonen Map, Clustering

8

9

ABSTRACT

10

Household consumption is a main driver of economy and might be regarded as ultimately

11

responsible for environmental impacts occurring over the life cycle of products and services.

12

Given that purchase decisions are made on household levels and are highly behavior-driven, the

13

derivation of targeted environmental measures requires an understanding of household behavior

14

patterns and the resulting environmental impacts. To provide an appropriate basis in support of

15

effective environmental policymaking, we propose a new approach to capture the variability of

16

lifestyle-induced

17

consumption patterns are derived in a two-tiered clustering that applies a Ward-clustering on top

18

of a pre-conditioning self-organizing map. The environmental impacts associated with specific

19

archetypical behavior are then assessed in a hybrid life cycle assessment framework. The

environmental

impacts.

Lifestyle-archetypes

representing

prevailing

ACS Paragon Plus Environment

1


Page 2 of 38

20

application of this approach to the Swiss Household Budget Survey reveals a global picture of

21

consumption that is in line with previous studies, but also demonstrates that different archetypes

22

can be found within similar socio-economic household types. The appearance of archetypes

23

diverging from general macro-trends indicates that the proposed approach might be useful for an

24

enhanced understanding of consumption patterns and for the future support of policymakers in

25

devising effective environmental measures targeting specific consumer groups.

26

1 Introduction

27

Households are major drivers of the economy. Their consumption behavior triggers a

28

multitude of economic activities along the supply chain of each product and service, which

29

subsequently involves the use of resources and the release of emissions. Household consumption

30

is estimated to be responsible for 65% of global greenhouse gas emissions and 50% to 80% of

31

total land, material, and water use1. The United Nation’s Sustainable Development Goal 12

32

(“Sustainable Consumption and Production”)2 demonstrates a large consensus that today’s

33

consumption patterns are unsustainable and changes in consumer behavior are urgently needed3–

34

11

35

complex economic, social, technological and cultural systems. In addition to informing

36

households about their environmental impacts, policymakers should frame an enabling

37

environment for individuals to change towards more sustainable lifestyles4,6,8,12.

. However, changing household consumption behavior is challenging4,6,12, as it is embedded in

38

Several studies quantified environmental impacts induced by household consumption (see

39

e.g.7,8,13–15 for reviews). While many studies focus on a national average household and on

40

identifying environmental priorities of different consumption areas1,3,7,13,14,16,17, several

41

researchers acknowledge the importance of investigating the environmental consequences of


2

Page 3 of 38


42

different household groups (e.g.5,7,10,11,18–25). Being highly influenced by socio-cultural and

43

economic factors, as well as driven by individual preferences, household behavior and

44

consequential purchase decisions are diverse, and “one-size-fits-all”-recommendations are likely

45

to fail5,11,26. Therefore, understanding the variability of consumption patterns and associated

46

environmental impacts is required for devising targeted environmental policies.

47

Studies attempting to capture this lifestyle-induced variability usually combine household

48

budget surveys (also called consumer expenditure surveys) with environmentally-extended

49

input-output-models (EEIOM) and then assess the environmental impacts of different socio-

50

economic cohorts3,5,9–11,17,20,22,23,26, or fit regression models with socio-economic characteristics

51

as explanatory variables24,27. The findings of both (sometimes combined) approaches are very

52

insightful, especially if the applied regression models aim at explaining the drivers of

53

environmental consequences10,16,20,22,23,25,26. However, Girod and De Haan19 found that there

54

might still be significant variability of behavior within investigated household types. This raises

55

the question if the use of household segments that are pre-defined solely based on socio-

56

economic characteristics might prevent recognition of important behavioral patterns by assuming

57

all households within a segment behaving similarly. How could the support for policymaking be

58

improved, especially in view of recent calls to increase the involvement of behavioral economics

59

and psychology when deriving environmental measures4,12,28? Building upon the mentioned

60

lifestyle-studies, we propose a new approach: Instead of pre-defined household segments, we

61

suggest deriving clusters of households which are not only based on socio-economic aspects but

62

also on real observed behavior. The proposed two-stage clustering allows for studying behavior-

63

associated environmental impacts in the context of total consumption and is simultaneously able

64

to capture non-linear effects. This approach thus allows for investigating the nature and


3


Page 4 of 38

65

implications of different household behavior in detail. The emergence of archetypes might then

66

form a new information basis to derive environmental policies tailored to actual consumption

67

patterns.

68

The goal of this article is twofold: First, we will demonstrate the clustering of household

69

behavior by applying our approach to the Swiss Household Budget Survey (HBS). Second, this

70

application will result in ready-to-use consumption archetypes with associated environmental

71

modeling for Switzerland. While the transferability of the Swiss archetypes to other countries is

72

unclear, the proposed methods can definitely be applied to different expenditure surveys.

73

2 Methodology

74

Grouping households based on their characteristics and on their consumption behavior

75

represents the core of our approach to studying household environmental impacts. Utilizing

76

groups of households is also important because of the so-called “infrequency of purchase

77

problem”22,25 that is encountered when working with expenditure surveys in which households

78

participate only during a limited time period. To obtain a representative picture of a certain

79

population group’s consumption behavior, in which infrequent or season-specific purchases

80

average out, several similar households need to be clustered. In previous lifestyle-studies, this is

81

implicitly solved by averaging over pre-defined household segments. In our approach (Figure 1),

82

we first form groups of similarly-behaving households and then in a second step derive

83

archetypes representing the average behavior of these groups. Finally, the environmental impacts

84

of these archetypes will be assessed by means of a hybrid life cycle assessment (LCA) that

85

sources environmental background data from both EEIOM and process-based life cycle

86

inventory databases. With a hybrid LCA framework, known issues with EEIOM (e.g. the


4

Page 5 of 38


87

“product quality problem”7,9,18,20,22,25) can be partly overcome by using physical functional units,

88

and EEIOM can complement missing data of process-based LCA7. Furthermore, the applied

89

LCA-modeling will also allow for the computation of different impact assessment methods and

90

not only carbon footprints or energy requirements as done in previous studies3,5,9,10,13,16–20,22,24,29.

91

Thus, the chosen LCA-approach might also help to reveal potential burden shifts induced by

92

planned environmental measures. Although suggested by several authors1,3,13,15, the application

93

of hybrid LCA is – to our knowledge – rare. That said, it would still be possible to directly

94

couple the archetypes with EEIOM or even more sophisticated macro-economic models as used

95

e.g. in21,30.

96

All computation steps will be described in more detail below (see also Figure 1).

97 98

Figure 1. Simplified flow scheme providing a synopsis of the whole modeling approach. For

99

each step, sections of the main article or the Supporting Information (SI) are indicated in which

100

more detailed descriptions can be found.


5


101

Page 6 of 38

2.1 Consumption Data

102

The main data source of this study is the Swiss HBS31 (2009-2011) which provides detailed

103

information on the characteristics and consumption behavior for 9734 households. Households

104

participating in the survey report on daily expenditures, income, and quantities of bought goods

105

(e.g. for food) during one month. In addition, they also report on periodic expenditures (e.g.

106

newspaper subscriptions), possession of durable goods and on extraordinary purchases or

107

revenues in the last few months (e.g. buying a car within the last year). Data on household

108

characteristics and on household members are also collected. For each household, the final

109

dataset comprises statistics on 20 different durable goods, 8 income categories (plus 4 on

110

aggregated levels), 19 household variables, 6 attributes for each household member and 356

111

(plus 175 on aggregated levels) consumption categories classified based on the United Nations’

112

“Classification of Individual Consumption according to Purpose” (COICOP)32. For consumption

113

areas, purchased amounts in liters or kilograms are available for 92 categories (plus 14 on

114

aggregated levels). Further information on the categorization and a list of survey attributes is

115

provided in the SI.

116

2.2 Pre-Processing of Consumption Data

117 118

The present section describes how variables needed for the computations in 2.3 (pattern recognition) and 2.4 (LCA) were created based on the original HBS-data.

119

Besides converting categorical variables to numerical data by a set of binary variables, several

120

count-statistics were created to better compare household member information (see SI).

121

Moreover, for some categories, more detailed information than specified in the HBS-data is

122

required to apply LCA (cf. section 2.4). This particularly concerns housing-related categories

123

and public transport demand. With regard to the first, especially tenants typically do not have full


6

Page 7 of 38


124

information on the exact breakdown of their utility bills into refuse and sewage collection, water

125

supply, electricity, and heating. To impute this missing information, a modeling approach using

126

K-Nearest-Neighbor-Regression33, Random-Forest-Regression34,35 and LASSO-Regression36

127

was employed (see SI). The predicted data was then converted to quantities by means of price

128

data37–40. These prices were retrieved as close as possible to the specific circumstances of each

129

household by taking into account household type, location, and survey year. The pre-processed

130

HBS-data was then validated against national statistics40–44 (cf. SI).

131

The demand for public transport is a second issue. While kilometers driven by car can be

132

estimated with liters of fuel purchased, HBS-information on public transport mainly relates to

133

season tickets and travel card expenditures, thereby lacking information on effectively driven

134

kilometers by public transport. Therefore, this demand was estimated for each household using

135

data from the Swiss Mobility Microcensus 201045, which provides detailed information on the

136

mobility behavior of the Swiss population (see SI).

137

2.3 Pattern Recognition and Clustering of Households

138

2.3.1 Preparation for Pattern Recognition

139

Due to seasonality and storage effects, the survey month might bias the results of individual

140

households in the HBS31. Consequently, the determination of household clusters with similar

141

characteristics and behavior requires the dataset to be pre-processed and filtered for household

142

features which make similar behavior identifiable independent of the month in which the survey

143

took place.

144

A flow scheme in SI-section 3.1 visualizes the following preparatory step. Each HBS-attribute

145

was judged separately whether it shall be included for pattern recognition or not. Note that

146

exclusion only concerns the pattern recognition steps in sections 2.3.2 and 2.3.3, while the


7


Page 8 of 38

147

deduction of archetypes in 2.3.4 will resort to all available attributes. Household characteristics

148

including the count-statistics of 2.2, durable goods statistics, periodic expenditures and revenues

149

were all considered in the pattern recognition. In contrast, daily purchase attributes were only

150

included if they are bought on a regular basis and not stored for more than a month. Otherwise, it

151

was appraised if the attribute’s inclusion within an aggregated level is reasonable (e.g. if an

152

aggregated attribute such as “fruits” shall rather be included than “grapes” which are rarely

153

bought from January to June in Switzerland).

154

In a next step, all candidate attributes were checked for seasonality. The application of

155

ANOVA46 and Kruskal-Wallis-tests47 revealed if the influence of the survey month is

156

statistically significant. In case of statistically significance, the respective attribute was corrected

157

for seasonality (see SI-section 3.1).

158

The final dataset for pattern recognition comprises 157 attributes in total, thereof 85

159

consumption and income categories.

160

2.3.2 Self-Organizing Map (SOM)

161

A two-tiered approach was applied to find patterns in the HBS-data based on which the

162

households shall be grouped. In a first stage, a self-organizing map (SOM)48,49 pre-conditioned

163

and reduced the dataset, which was then clustered in a second stage. Such two-step clusterings

164

demonstrated to perform well and are robust even in the case of noisy and high-dimensional

165

datasets50–53.

166

The SOM was proposed by Kohonen48,49 and belongs to the class of unsupervised artificial

167

neural networks. It generally combines vector quantization and non-linear projection to a lower-

168

dimensional space; usually to a discrete 2D-lattice of neurons. Thereby, the SOM learns the

169

patterns in the dataset in an ordered fashion and is thus able to preserve the topology of the data.


8

Page 9 of 38


170

This means that neighboring neurons in the map have similar characteristics. A prototype vector

171

with the same dimension as the vectors of the input dataset is associated with each neuron.

172

During the training phase of the SOM, this set of prototype vectors is optimized to become a

173

representative substitute for the original dataset. The resulting map is thus a reduced, but still

174

representative dataset that is not only smoothed with regard to noise but also facilitates the

175

recognition of patterns, be it for subsequent clustering algorithms or visually for the human eye

176

(see component maps in the SI).

177

For the present study, the SOM was tuned by generating several SOMs with different

178

parameters (number of neurons, arrangement of neurons, initial and final neighborhood radius,

179

neighborhood function, and number of epochs) based on literature recommendations48,49,54–56 and

180

then by choosing the model with a topographic error close to zero and the lowest quantization

181

error56. The topographic error evaluates the order of the map (e.g. if it is twisted) and represents

182

the share of samples in the input dataset for which the first and second closest neurons are not

183

adjacent in the map56, while the quantization error judges the representativeness and thus the

184

accuracy of the map56. The final map consists of 987 neurons arranged in a 21:47-planar-lattice

185

(see SI for more tuning information and a short introduction to SOMs).

186

2.3.3 Clustering

187

In the second stage of pattern recognition, we apply clustering algorithms on top of the SOM

188

to form groups of neurons. Since each clustering technique has its strengths and weaknesses, two

189

well-known approaches, which differ in their basic principles, were tested and evaluated: K-

190

Means57–59 and agglomerative clustering60,61. While k, the number of clusters to be built, was the

191

only tuning parameter for K-Means, agglomerative clustering was run in different combinations

192

of affinity metrics (e.g. Euclidean distance, L1-norm) and linkage criteria (Ward62 and average).


9


Page 10 of 38

193

The evaluation of clusters – and thus the “best” choice of clustering techniques and associated

194

parameters – is not a trivial task if the ground truth is unknown63. For the present study, we

195

mainly focused on two performance methods: Silhouette coefficients (S)64 and U-Matrix. S

196

relates the distances between one point in a cluster and all other points in the same cluster to the

197

distances from that point to all points in the second closest cluster. The so-called U-Matrix

198

(unified distance matrix) is an important visualization of a SOM and supports clustering on top

199

of the map. At the map position of each neuron, an U-Matrix depicts the sum of distances in the

200

high-dimensional space between the prototype vector of the respective neuron and the prototype

201

vectors of all adjacent neurons52,65,66. Large U-heights indicate that a neuron’s prototype vector is

202

distant from others, while small U-heights are associated with prototypes that are surrounded by

203

other vectors in the data space66. The U-Matrix thus suggests visually which neurons should be

204

grouped together to form clusters.

205

The quest for the best clustering was subdivided into two parts (see SI for details): First, a pre-

206

selection of parameters within each approach was conducted mainly based on S. Afterwards, a

207

detailed comparison of the two alternatives and fixing the number of clusters for both was based

208

on the U-Matrix. For this, the cluster borders were projected on the U-Matrix and the clustering

209

methods were judged by their ability to redraw the visible groupings of neurons in the U-Matrix.

210

Additionally, two other criteria broadened the information basis for the final decision: ANOVA-

211

tests were run for each attribute to obtain an impression if reasonable results are produced.

212

Second, the number of households per cluster was determined to get some indication about the

213

representativeness of the clusters. Considering all these criteria, an agglomerative clustering

214

technique with Ward-linkage, which produces 34 clusters, was finally selected. Furthermore, the

215

applied agglomerative clustering implementation67 allows for including connectivity constraints,


10

Page 11 of 38


216

meaning that only clusters which are adjacent on the map can be merged by the algorithm. This

217

ability was seen as another advantage over K-Means.

218

2.3.4 Deriving Consumption Archetypes

219

Considering the statistical basis of the analyses of the Federal Statistical Office31, we assumed

220

a cluster to be representative for a population’s group if at least 130 HBS-households are

221

member of this cluster. Since not all clusters built in 2.3.3 comply with this criterion, some post-

222

processing in the sense of manually merging clusters was needed. Indeed, this merging is also

223

justified by the dendrogram in the SI which shows a blur between 34 and 24 clusters, indicating

224

that some clusters are close to each other. Cutting the dendrogram at different positions might be

225

reasonable in the presence of sub-clusters51 (see SI for more reasons). Therefore, starting with 34

226

clusters, all clusters with less than 130 households were merged with adjacent clusters if these

227

merges happen in the dendrogram between 24 and 34 clusters. This resulted in 26 fair clusters

228

and 2 clusters with less than 130 households. These two clusters will continue to be part of the

229

subsequent analysis, but they will be marked accordingly.

230

The behavior-archetypes are now constituted by the centroids of the clusters. Note that the

231

computation of the clusters’ centroids follows the averaging-procedure of the Federal Statistical

232

Office and includes a representativeness-weight31. Furthermore, the attributes which were

233

filtered out in 2.3.1 are now re-used and also entered the vector of the archetypes.

234

2.4 LCA-Modeling

235

The final modeling step comprised the coupling of the archetypes’ demands with detailed life

236

cycle background data in order to quantify the environmental impacts of the archetypes’

237

consumption behavior. The overall functional unit for the LCA was chosen to be one year of

238

household consumption. The life cycle inventory data was extracted from three well-known and


11


Page 12 of 38

239

transparent databases in the following priority order: ecoinvent v3.368, Agribalyse v1.269, and

240

EXIOBASE v2.270,71.

241

The functional units of ecoinvent- and Agribalyse-activities require quantities instead of

242

expenditures. While section 2.2 prepared housing- and public-transport-related categories for this

243

purpose, conversions based on price data (e.g.37–39,72) or further information73 was necessary for

244

other consumption areas. Fortunately, for almost all food categories, quantities are directly

245

available in the HBS-dataset. The process models for processed food closely followed the

246

modeling of Walker et al.74, but adjusted to Swiss conditions. Generally, we always attempted to

247

adapt the process models as close as possible to the domestic conditions of Switzerland. For

248

instance, Swiss market activities for food were constructed based on the import statistics

249

provided by Scherer and Pfister75 and car fleets and energy mixes for heating technologies were

250

based on national statistics40,76.

251

The creation of process models with EXIOBASE-sectors including the conversion of

252

purchaser-prices (HBS) to basic-prices (EXIOBASE) generally followed the suggestions of

253

Steen-Olsen et al.3

254

Finally, it needs mentioning that only environmental impacts directly associated with a certain

255

household’s behavior were considered. For instance, only direct spending on education or health

256

care and thus retraceable to a specific household were taken into account. This leads to an

257

underestimation of impacts from “health” and “education” since the Swiss education system is

258

largely financed by the government, while health expenses are usually covered by insurances

259

whose premiums are not necessarily indicative of the health care utilization of households and

260

thus of the effectively caused impacts. Furthermore, governmental consumption was not re-


12

Page 13 of 38


261

distributed to households as done in other studies77,78. Details of the LCA-modeling are disclosed

262

in the SI.

263

3 Results and Discussion

264

3.1 Drivers of Clustering

265

The U-Matrix and the final clustering are presented in Figure 2. The two clusters not reaching

266

full representativeness are named “OA” and “OB”, while all other clusters are labelled with

267

random letters. The illustration of clusters on top of the SOM together with the distance

268

information given in the U-Matrix also reveals relationships between clusters. For instance, “N”

269

will be more similar to the adjacent “O” than to “J”.


13


Page 14 of 38

270 271

Figure 2. Top: U-Matrix of the SOM. Contours were inserted to improve visibility of distances.

272

Note that a pixel in the map corresponds to the map position of a neuron; Bottom: Final

273

clustering on top of the SOM. The clusters “OA” and “OB” are the clusters with less than 130

274

households.

275

Building upon the clustering selection procedure in 2.3.3, the quality of the final clustering was

276

further analyzed by a visual appraisal of the 95%-confidence intervals of the clusters’ means of

277

each attribute and by the ANOVA-tests (see section 2.3.3 and SI). These tests indicated that for

278

all attributes there is at least one cluster significantly differing from the others (largest p-value:

279

2.45E-14). ANOVA and the qualitative visual judgment support the reasonability of the final


14

Page 15 of 38


280

clustering. In addition, ANOVA provides a possibility to analyze the underlying drivers of the

281

clustering process. An attribute’s test statistics (F-values) can be considered as a measure of

282

“distinction power”. Among the top 20 distinctive attributes, there are, in particular, geographic

283

variables and employment status, but also variables related to household size, consumption areas,

284

and durable goods statistics. Attributes which are rated to be least important for forming the

285

clusters are either variables which concern only few households (e.g. university fees) or

286

consumption areas for which obviously less distinct patterns could be found (e.g. consumption of

287

“coffee, tea and cocoa”). All results can be found in the SI.

288

3.2 Individual Archetypes and their Interrelations

289

A simplified description of all archetypes is available in Table 1. In this table, the clusters’

290

centroids are categorized according to income and average number of persons per household.

291

However, note that the clusters emerged from many different – including also non-socio-

292

economic – attributes. The attributes presented in Table 1 and those used in the following

293

analysis are thus just of an indicative nature meant to help better grasp the archetypes (see also

294

additional details and results in the SI).

295

Table 1. Simplified categorization of archetypes (clusters’ centroids) along the axes of income

296

(vertical) and number of persons per household (horizontal). The numbers in parentheses show

297

income in Swiss Francs per month and average numbers of persons per household respectively.

298


15


Single-person households (1.1 - 1.3)

Small households (1.5 - 1.7)

Very high income (>19000)

Two-persons households

OB (around retirement)

High income (13000 - 19000)

Q (divorced middle-aged males)

Families (3.4 - 3.6)

Large families (>4)

S (selfemployed; some with primary school kids/ teenagers)

OA (homeowners with primary school kids/ teenagers)

L (with primary school kids/ teenagers)

C (rather young adults with babies)

J (homeowners with teenagers; "technophile")

K (SwissItalian; with primary school kids/ teenagers; homeowners)

F (SwissFrench; with teenagers; tenants)

A (homeowners with primary school kids)

P (SwissGerman; around retirement) X (SwissGerman; Zurich; around retirement, older than P)

Medium income (7800 - 9500)

O (SwissGerman; Zurich; well-mixed-notretired adults) R (Eastern Switzerland; well-mixed-notretired adults)

N (SwissGerman; Bern; well-mixed-notretired adults) Low income (< 5000)

Small families (2.4 - 3.2)

M (SwissFrench; around retirement)

Medium-high income (9500 - 13000)

Medium-low income (5000 - 7800)

D (young, unmarried; w/o children)

Page 16 of 38

G (North Western Switzerland; middle-aged adults w/o children)

Z (very young adults with babies)

U (Central Switzerland, well-mixed-notretired adults with children)

B (tenants with primary school kids/ teenagers; depend on other households; by trend: 1 adult and 2 children)

I (Swiss-French; well-mixed-notretired adults w/o children) W (SwissFrench; slightly older middleaged adults w/o children) E (Swiss-Italian; middle-aged adults w/o children)

Y (very old people)

T (Swiss-Italian; retired)

V (SwissGerman; retired)

H (old, widowed females)


16

Page 17 of 38


299

The environmental impacts of the individual archetypes are presented in Figure 3. Even

300

though the applied LCA-modeling enables for the computation of all environmental indicators

301

that are supported by the background life cycle inventory databases, only greenhouse gas

302

emissions (GHG) according to IPCC 2013 (100a)79 and total endpoints of ReCiPe 2008 (H,A)80

303

are discussed in the following (see SI for results of the Ecological Scarcity method81). Besides

304

allowing for comparison to other studies, GHG have the highest priority in political discussions

305

due to climate change, while the ReCiPe-endpoints shall help to calculate the overall

306

environmental relevance of household consumption.

307

In the scope of this article, only a few comparisons between some of the archetypes are

308

presented. Thereby, selected examples in the aspects of income, household size, age structure,

309

and footprint composition will be discussed. To support these comparisons, Figure 3 resumes

310

information from Table 1 and shows some general characteristics of the archetypes such as

311

income, number of persons and average age of adults and children.


17


Page 18 of 38

312 313

Figure 3. Top: Total annual impacts per archetype; colors illustrate income, while size of the

314

markers represents prevalence of the archetypes (number of households per cluster). The red


18

Page 19 of 38


315

dashed line depicts the prevalence-weighted average (AVG). Indicative archetype characteristics

316

are given in parentheses: P stands for average number of persons per archetype, while A provides

317

the average age of adults/children; Bottom: Bar plots showing the composition of the

318

environmental impacts on a per-capita basis. GHG (IPCC 2013, 100a) are shown in the left part

319

of the figure and results of ReCiPe 2008 (total endpoints, H,A) are on the right-hand side. Ranks

320

of total footprints are given in parentheses to facilitate the comparison of top and bottom figures.

321

In the upper part of Figure 3, a correlation between household income and total impacts can be

322

observed. However, a few high-income households (e.g. C (“family with babies”), D (“young,

323

unmarried couples”) and Q (“divorced, middle-aged males”)) are an exception to this. In addition

324

to income, household size is crucial for total emissions. This becomes apparent in the per-capita-

325

bar plots in the lower part of Figure 3 when we observe how the order of archetypes changes.

326

The archetypes with the largest incomes (OA (“very high-income family”), OB (“very high-

327

income retired couple”), S (“small families with self-employed persons”)) still spearhead the per-

328

capita figure. But OB outstrips OA, and except for OA and S, most other family-archetypes (e.g.

329

A (“family with primary school children”), B (“single-parent-families”), and Z (“young adults

330

with babies”)) – including some with high-income (e.g. J (“family with teenagers”) or C) – can

331

be found on the bottom of the per-capita-impact-ranking. In contrast, single-person households

332

(e.g. H, N, O, and R) show low total annual impacts, while in a per-capita-perspective these

333

archetypes move to the middle field. H (“old, widowed females”), together with V (“low-

334

income, retired couple”) and Y (“low-income, very old couple”), also reveal a certain generation

335

gap with regard to footprint composition: all three of these clusters – which are also neighbors in

336

the SOM – are relatively older in age, have low communications and mobility impacts, and have


19


Page 20 of 38

337

higher housing energy and health impacts compared to other archetypes. This is for instance in

338

contrast with the “young” D archetype, which exhibits high transport, but low housing, impacts.

339

Interestingly, there are also archetypes with similar socio-economic conditions but different

340

total emissions and footprint compositions. O and R are adjacent in the SOM and both are single-

341

person households with similar age and income range. However, archetype O is related to urban

342

households (Zurich), while R has its origin in the more rural Eastern Switzerland. R’s mobility

343

impacts are higher on account of larger emissions from car-driven distances. This is, however,

344

partly compensated by O’s large demand for air travel and taxi rides. Furthermore, O has larger

345

impacts induced by eating out.

346

Finally, the comparison of J and F reveals another interesting aspect. Both archetypes are

347

similar with regard to size, income and age (“families with teenagers”), per-capita footprints and

348

compositions. However, for devising targeted measures to abate housing-induced emissions,

349

there is one especially important difference between the two clusters: J-households are

350

homeowners, whereas F are tenants.

351

3.3 Archetypes in the Collective and Environmental Hotspots

352

The analysis of the archetypes can help policymakers in identifying target groups. For

353

instance, OA and OB have large footprints and have money available to invest in

354

environmentally beneficial technologies to cover e.g. their transport and housing demand.

355

However, even though the signaling effect to target these groups may be large, they are not very

356

prevalent. Consequently, it is also important to look at the overall picture and investigate how the

357

different archetypes contribute to cumulative impacts; which is in the case of OA and OB

358

together less than 2.5%. Since HBS-data can be assumed to be representative for Switzerland, we

359

used the number of households per cluster as an indication of prevalence. This information is


20

Page 21 of 38


360

also illustrated by the size of the markers in the scatterplots of Figure 3. The upper part of

361

Figure 4 depicts the prevalence-weighted contributions of both the archetypes and the

362

consumption areas. The lower part of Figure 4 then relates the archetypes’ prevalence-weighted

363

total impact shares to their per-capita-emissions in a simplified attempt to group the archetypes

364

based on the prospects to target them. It becomes apparent that the three family-archetypes A, J

365

and L, which are ranked 1st, 5th and 6th by frequency, are together responsible for about 27% of

366

total impacts. While the per-capita-emissions are rather low for A and J, their high contribution

367

to total impacts is due to their size (>4 persons per household on average) and their sheer

368

abundance. In contrast, L is not only frequent, but also causes large per-capita and total

369

footprints. These small, high-income and home-owning families with teenagers could be an

370

interesting target group for reducing consumption impacts. However, this is just one example

371

since L’s contribution is about 8% of total emissions and further archetypes need to be targeted

372

as well. Analyzing the main consumption areas in more detail, V (“low-income, retired couple”)

373

and D (“young couple”) appear to be important. Being among the three archetypes with largest

374

per-capita emissions for air travel and car trips, D has particularly large transport impacts and is

375

also responsible for high emissions in the category “restaurants and hotels”. In contrast, V shows

376

high health and housing impacts in addition to large per-capita food emissions. However, while

377

the lower part of Figure 4 classifies D as a “promising target”, V seems to be a borderline case

378

which requires more careful consideration.


21


Page 22 of 38

379


22

Page 23 of 38


380

Figure 4. Top: Prevalence-weighted contributions of the archetypes and main consumption areas

381

to total environmental impacts as well as prevalence-weighted contributions of archetypes and

382

consumption-subcategories within the main consumption areas (ordered by magnitude of

383

contribution). Except for the Swiss total, only contributions with a cumulative effect of about

384

50% are displayed. Bottom: per-capita emissions in relation to prevalence-weighted

385

contributions to total impacts. The divisive horizontal and vertical lines correspond to the

386

medians of the x- and y-values respectively. Archetypes mentioned in the text are highlighted in

387

red. [Non-dur gds.: Non-durable household goods; Pkg. holiday: Package holiday; SpS: Sporting

388

services; CulS: Cultural services; PC: Computer, office appliances and other peripherals]

389

The second perspective provided in Figure 4 (upper) identifies – just as many previous

390

studies1,3,7,14–16 – food, housing and transport as the most important consumption categories.

391

While transport and housing each account for about 25% of total GHG emissions, transport takes

392

the lead in ReCiPe-endpoints with 24%, closely followed by food and housing (each 21%).

393

Thereby, transport is clearly dominated by car trips (about 90%) and housing by energy use

394

(about 85%). For food, dairy products (especially semi-hard/hard cheese) and meat (mainly beef)

395

are of similar importance; each with shares around 30% in the GHG-perspective and about 25%

396

with regard to ReCiPe-endpoints.

397

The computed prevalence-weighted average for Switzerland totals at 9.0 t CO2-eq/person/year.

398

This is between the top-down study of Jungbluth et al.77 with 11.0 t CO2-eq/person/year and

399

8.6 t CO2-eq/person/year found by Girod and De Haan18 in another HBS-based study. Note that

400

both studies refer to a prior time period and that the original average of77 (12.8 t CO2-

401

eq/person/year) was adjusted to account for our study’s underestimation of health care and

402

educational services, and for not redistributing governmental consumption.


23


403

Page 24 of 38

3.4 Limitations of the Study

404

The presented archetypes were formed based on behavior and characteristics of households.

405

Thereby, the goal of our approach was to provide the most generalized basis for further

406

investigations. Yet, within these post-analyses – such as the application of LCA with a specific

407

environmental indicator – some clusters become seemingly very similar. Depending on the

408

policy goals, it may thus be reasonable to further group the archetypes according to their

409

environmental impacts.

410

Another limitation of this study pertains to the underreporting in the HBS. Although

411

participating households are closely supervised and receive advice from specialists82, previous

412

studies that have made use of consumer expenditure surveys revealed that underreporting is a

413

common issue3,22,29. This could also explain the slightly lower total carbon footprint assessed

414

here in comparison to the economy-wide study of Jungbluth et al.77

415

Finally, the applied hybrid LCA-modeling was done with great care and adjusted as much as

416

possible to Swiss conditions. Still, LCA-modeling always requires assumptions, average mixes

417

and simplifications that might affect the final results. Hereby, the uncertainties induced by the

418

conversion of expenditures to functional units via price data, the uncertainties arising from

419

converting HBS-purchaser-prices to EXIOBASE-basic-prices3,9,10,29, and the limited number of

420

biosphere flows in EXIOBASE need special mentioning. The latter is discussed in the SI and

421

may lead to a slight under-accounting of ReCiPe-endpoints. Follow-up research aims at

422

developing a framework to capture uncertainties involved in the archetype-approach.


24

Page 25 of 38

423


4 Outlook

424

The proposed archetype-approach is meant to deliver important insights into households’

425

consumption behavior for policymakers to derive and prioritize targeted measures. In this

426

context, our approach can also be used for identifying drivers of environmental impacts as done

427

in previous studies10,16,20,22,25,26. To demonstrate this application, a univariate correlation analysis

428

between environmental impacts and main household characteristics is presented in the SI. The

429

computation of such correlations, as well as studying the drivers for clustering and the

430

prevalence-weighted analyses, help to capture the big picture and support thus the identification

431

of general tendencies, hotspots and potential target groups of households. However, since the

432

overall trends do not necessarily apply to particular archetypes, the deduction of targeted actions

433

requires an in-depth understanding of the target groups and should thus be done on the basis of

434

individual archetypes. For this, the archetype-approach offers a predestined basis by allowing for

435

backtracking environmental impacts to the living conditions of real households and observable

436

consumption behavior.

437

While this study aimed to provide the basis for identifying strategies, the specific analyses

438

need to be done by local environmental policymakers according to their political agendas.

439

Thereby, they could follow the comprehensive framework proposed by Schanes et al.4 and also

440

consider suggestions from related research in behavioral economics and psychology28. The input

441

from these disciplines, which aim to understand motivational factors and cognitive biases, could

442

provide support to go even beyond “conventional” measures, such as taxes, regulations or

443

subsidies, and allow for profiling households. Based on this, personalized messages and

444

measures could be formulated which directly address different consumer groups in order to

445

effectively raise awareness and to encourage them to change towards more sustainable


25


Page 26 of 38

446

consumption patterns. But before investigating how people could be motivated to lower their

447

environmental impacts, it is of high importance to know which archetypes prevail in the local

448

policymakers’ sphere of influence. The prevalence-weighted analyses illustrates a first attempt in

449

this direction at a national level, but the archetypes can further be used as a basis for a

450

regionalized model. In follow-up research, the archetypes will be assigned to real households

451

based on the national census within a probabilistic-classification approach. In addition, this use

452

of archetypes will simultaneously be combined with other nationwide bottom-up models such as

453

the building energy model of Buffat et al.83 and agent-based mobility models as used by Saner et

454

al.84. The final model will estimate a realistic environmental profile for each real household in

455

Switzerland and thus provide information for analyses on any desired regional scale.

456

ASSOCIATED CONTENT

457

The following files are available free of charge: PDF and XLS file with detailed methods and

458

results.

459

AUTHOR INFORMATION

460

Corresponding Author

461

* Andreas Froemelt

462

ETH Zurich

463

Institute of Environmental Engineering

464

Ecological Systems Design

465

HPZ E32.2

466

John-von-Neumann-Weg 9


26

Page 27 of 38


467

CH-8093 Zurich

468

[email protected]

469

Phone: +41 44 633 68 40

470

Present Addresses

471

† Process Technology, BU Environmental Technology, Rittmeyer Ltd., Inwilerriedstrasse 57,

472

6340 Baar, Switzerland

473

Author Contributions

474

The manuscript was written through contributions of all authors. All authors have given approval

475

to the final version of the manuscript.

476

Funding Sources

477

SCCER Mobility, funded by Innosuisse

478

Notes

479

The authors declare no competing financial interest.

480

ACKNOWLEDGMENT

481

We would like to thank Christie Walker, Stephan Pfister and Adrian Haas for their important

482

inputs and Jonas Müller and Annina Brupbacher for their assistance. Our thanks also go to the

483

Federal Statistical Office, Kjartan Steen-Olsen and Richard Wood for their support. This

484

research project is part of the Swiss Competence Center for Energy Research SCCER Mobility

485

of the Swiss Innovation Agency Innosuisse.

486

REFERENCES


27


487

(1)

Page 28 of 38

Ivanova, D.; Stadler, K.; Steen-Olsen, K.; Wood, R.; Vita, G.; Tukker, A.; Hertwich, E. G.

488

Environmental Impact Assessment of Household Consumption. Journal of Industrial

489

Ecology 2016, 20 (3), 526–536.

490

(2)

491 492

Goals; E/2017/66; United Nations: New York, 2017. (3)

493 494

Secretary-General (United Nations). Progress towards the Sustainable Development

Steen-Olsen, K.; Wood, R.; Hertwich, E. G. The Carbon Footprint of Norwegian Household Consumption 1999-2012. Journal of Industrial Ecology 2016, 20 (3), 582–592.

(4)

Schanes, K.; Giljum, S.; Hertwich, E. Low carbon lifestyles: A framework to structure

495

consumption strategies and options to reduce carbon footprints. Journal of Cleaner

496

Production 2016, 139, 1033–1043.

497

(5)

498 499

Jones, C. M.; Kammen, D. M. Quantifying carbon footprint reduction opportunities for U.S. households and communities. Environ. Sci. Technol. 2011, 45 (9), 4088–4095.

(6)

Spangenberg, J. H.; Lorek, S. Environmentally sustainable household consumption: From

500

aggregate environmental pressures to priority fields of action. Ecological Economics

501

2002, 43 (2–3), 127–140.

502

(7)

503 504

Environ. Sci. Technol. 2005, 39 (13), 4673–4684. (8)

505 506 507

Hertwich, E. G. Life Cycle Approaches to Sustainable Consumption: A Critical Review.

Tukker, A.; Cohen, M. J.; Hubacek, K.; Mont, O. The Impacts of Household Consumption and Options for Change. Journal of Industrial Ecology 2010, 14 (1), 13–30.

(9)

Druckman, A.; Jackson, T. The bare necessities: How much household carbon do we really need? Ecological Economics 2010, 69 (9), 1794–1804.


28

Page 29 of 38

508


(10)

Baiocchi, G.; Minx, J.; Hubacek, K. The Impact of Social Factors and Consumer Behavior

509

on Carbon Dioxide Emissions in the United Kingdom. A Regression Based on Input-

510

Output and Geodemographic Consumer Segmentation Data. Journal of Industrial Ecology

511

2010, 14 (1), 50–72.

512

(11)

Minx, J. C.; Wiedmann, T.; Wood, R.; Peters, G. P.; Lenzen, M.; Owen, A.; Scott, K.;

513

Barrett, J.; Hubacek, K.; Baiocchi, G.; et al. Input-output analysis and carbon footprinting:

514

An overview of applications. Economic Systems Research 2009, 21 (3), 187–216.

515

(12)

O’Rourke, D.; Lollo, N. Transforming Consumption: From Decoupling, to Behavior

516

Change, to System Changes for Sustainable Consumption. Annual Review of Environment

517

and Resources 2015, 40, 233–259.

518

(13)

519 520

Systems Research 2011, 23 (1), 27–47. (14)

521 522

Hertwich, E. G. The Life Cycle Environmental Impacts of Consumption. Economic

Tukker, A.; Jansen, B. Environmental Impacts of Products: A Detailed Review of Studies. Journal of Industrial Ecology 2006, 10 (3), 159–182.

(15)

Di Donato, M.; Lomas, P. L.; Carpintero, Ó. Metabolism and Environmental Impacts of

523

Household Consumption: A Review on the Assessment, Methodology, and Drivers.

524

Journal of Industrial Ecology 2015, 19 (5), 904–916.

525

(16)

526 527 528

Hertwich, E. G.; Peters, G. P. Carbon Footprint of Nations: A Global, Trade-Linked Analysis. Environ. Sci. Technol. 2009, 43 (16), 6414–6420.

(17)

Bin, S.; Dowlatabadi, H. Consumer lifestyle approach to US energy use and the related CO2 emissions. Energy Policy 2005, 33 (2), 197–208.


29


529

(18)

530 531

Page 30 of 38

Girod, B.; De Haan, P. More or Better? A Model for Changes in Household Greenhouse Gas Emissions due to Higher Income. Journal of Industrial Ecology 2010, 14 (1), 31–49.

(19)

Girod, B.; De Haan, P. GHG reduction potential of changes in consumption patterns and

532

higher quality levels: Evidence from Swiss household consumption survey. Energy Policy

533

2009, 37 (12), 5650–5661.

534

(20)

535 536

Wier, M.; Lenzen, M.; Munksgaard, J.; Smed, S. Effects of household consumption patterns on CO2 requirements. Economic Systems Research 2001, 13 (3), 259–274.

(21)

Duarte, R.; Feng, K.; Hubacek, K.; Sánchez-Chóliz, J.; Sarasa, C.; Sun, L. Modeling the

537

carbon consequences of pro-environmental consumer behavior. Applied Energy 2016,

538

184, 1207–1216.

539

(22)

540 541

Weber, C. L.; Matthews, H. S. Quantifying the global and distributional aspects of American household carbon footprint. Ecological Economics 2008, 66 (2), 379–391.

(23)

Druckman, A.; Jackson, T. Household energy consumption in the UK: A highly

542

geographically and socio-economically disaggregated model. Energy Policy 2008, 36 (8),

543

3177–3192.

544

(24)

Jones, C.; Kammen, D. M. Spatial Distribution of U.S. Household Carbon Footprints

545

Reveals Suburbanization Undermines Greenhouse Gas Benefits of Urban Population

546

Density. Environ. Sci. Technol. 2014, 48 (2), 895–902.

547

(25)

Büchs, M.; Schnepf, S. V. Who emits most? Associations between socio-economic factors

548

and UK households’ home energy, transport, indirect and total CO2 emissions. Ecological

549

Economics 2013, 90, 114–123.


30

Page 31 of 38

550


(26)

Lenzen, M.; Wier, M.; Cohen, C.; Hayami, H.; Pachauri, S.; Schaeffer, R. A comparative

551

multivariate analysis of household energy requirements in Australia, Brazil, Denmark,

552

India and Japan. Energy 2006, 31 (2), 181–207.

553

(27)

554 555

Saner, D.; Beretta, C.; Jäggi, B.; Juraske, R.; Stoessel, F.; Hellweg, S. FoodPrints of households. The International Journal of Life Cycle Assessment 2016, 21 (5), 654–663.

(28)

Frederiks, E. R.; Stenner, K.; Hobman, E. V. Household energy use: Applying behavioural

556

economics to understand consumer decision-making and behaviour. Renewable and

557

Sustainable Energy Reviews 2015, 41, 1385–1394.

558

(29)

Ivanova, D.; Vita, G.; Steen-Olsen, K.; Stadler, K.; Melo, P. C.; Wood, R.; Hertwich, E.

559

G. Mapping the carbon footprint of EU regions. Environmental Research Letters 2017, 12

560

(5).

561

(30)

562 563

Distribution. Ecological Economics 2017, 136, 62–72. (31)

564 565

Sommer, M.; Kratena, K. The Carbon Footprint of European Households and Income

Bundesamt für Statistik (BFS). Haushaltsbudgeterhebung (HABE) 2009 bis 2011; Bundesamt für Statistik (BFS): Neuchâtel, Switzerland, Fed Regist., 2013.

(32)

United Nations Statistics Division. COICOP (Classification of Individual Consumption

566

According to Purpose) https://unstats.un.org/unsd/cr/registry/regcst.asp?Cl=5 (accessed

567

Nov 22, 2017).

568

(33)

569 570

Altman, N. S. An Introduction to Kernel and Nearest-Neighbor Nonparametric Regression. The American Statistician 1992, 46 (3), 175–185.

(34)

Breiman, L. Random Forests. Machine Learning 2001, 45 (1), 5–32.


31


571

(35)

Page 32 of 38

Genuer, R.; Poggi, J.-M.; Tuleau, C. Random Forests: some methodological insights;

572

Institut National de Recherche en Informatique et en Automatique: Le Chesnay Cedex,

573

France, 2008.

574

(36)

575 576

Statistical Society. Series B: Statistical (Methodological) 1996, 58 (1), 267–288. (37)

577 578

Federal Department of Economic Affairs Education and Research. Preisüberwachung http://www.preisvergleiche.preisueberwacher.admin.ch (accessed Apr 12, 2017).

(38)

579 580

Tibshirani, R. Regression Shrinkage and Selection via the Lasso. Journal of the Royal

Federal

Electricity

Commission

(ElCom).

Strompreis-Webseite

https://www.strompreis.elcom.admin.ch (accessed Jan 20, 2017). (39)

Federal

Statistical

Office

(BFS).

Landesindex

der

Konsumentenpreise.

581

Durchschnittspreise Energie; Bundesamt für Statistik (BFS): Neuchâtel, Switzerland, Fed

582

Regist., 2017.

583

(40)

584 585

für Energie (BFE): Bern-Ittigen, Switzerland, 2016. (41)

586 587

590

Schweizerischer

Verein

des

Gas-

und

Wasserfaches

(SVGW).

Trinkwasser

http://wasserqualitaet.svgw.ch/index.php?id=874 (accessed Nov 28, 2017). (42)

588 589

Bundesamt für Energie (BFE). Schweizerische Gesamtenergiestatistik 2015; Bundesamt

Bundesamt für Umwelt (BAFU). Abfallmengen und Recycling 2009 im Überblick; Bundesamt für Umwelt (BAFU): Bern, Switzerland, 2010.

(43)



32

Page 33 of 38

591


(44)

592 593


(45)

Bundesamt für Statistik (BFS); Bundesamt für Raumentwicklung (ARE). Mobilität in der

594

Schweiz. Ergebnisse des Mikrozensus Mobilität und Verkehr 2010; Bundesamt für

595

Statistik (BFS): Neuchâtel, Switzerland, 2012.

596

(46)

597 598

USA, 1973. (47)

599 600

Fisher, R. A. Statistical methods for research workers, 14th ed.; Hafner: New York, NY,

Kruskal, W. H.; Wallis, W. A. Use of ranks in one-criterion variance analysis. J. Am. Stat. Assoc. 1952, 47 (260), 583–621.

(48)

601

Kohonen, T. The Self-Organizing Map. Proceedings of the IEEE 1990, 78 (9), 1464– 1480.

602

(49)

Kohonen, T. K. Self-Organizing Maps; Springer Berlin Heidelberg, 2001; Vol. 30.

603

(50)

Dürrenmatt, D. J.; Gujer, W. Identification of industrial wastewater by clustering

604

wastewater treatment plant influent ultraviolet visible spectra. Water Science &

605

Technology 2011, 63 (6), 1153–1159.

606

(51)

607 608

Neural Networks 2000, 11 (3), 586–600. (52)

609 610

Vesanto, J.; Alhoniemi, E. Clustering of the Self-Organizing Map. IEEE Transactions on

Vesanto, J.; Sulkava, M. Distance matrix based clustering of the Self-Organizing Map. In Artificial Neural Networks - ICANN 2002; 2002.

(53)

Canetta, L.; Cheikhrouhou, N.; Glardon, R. Applying two-stage SOM-based clustering


33


Page 34 of 38

611

approaches to industrial data analysis. Production Planning and Control 2005, 16 (8),

612

774–784.

613

(54)

614 615

Vesanto, J.; Himberg, J.; Alhoniemi, E.; Parhankangas, J. SOM Toolbox for Matlab 5; Technical Report A57; Helsinki University of Technology: Helsinki, Finland, 2000.

(55)

Vesanto, J. Neural network tool for data mining: SOM toolbox. In Proceedings of

616

symposium on tool environments and development methods for intelligent systems

617

(TOOLMET 2000); 2000; pp 184–196.

618

(56)

Tan, H. Sen; George, S. E. Investigating learning parameters in a standard 2-D SOM

619

model to select good maps and avoid poor ones. In Australasian Joint Conference on

620

Artificial Intelligence; Springer Berlin Heidelberg, 2004; pp 425–437.

621

(57)

MacQueen, J. Some methods for classification and analysis of multivariate observations.

622

Proceedings of the fifth Berkeley symposium on mathematical statistics and probability

623

1967, 1 (14), 281–297.

624

(58)

625 626

Lloyd, S. Least squares quantization in PCM. IEEE Transactions on Information Theory 1982, 28 (2), 129–137.

(59)

Arthur, D.; Vassilvitskii, S. K-Means++: the Advantages of Careful Seeding. In

627

Proceedings of the eighteenth annual ACM-SIAM symposium on Discrete algorithms;

628

Society for Industrial and Applied Mathematics, Ed.; 2007; pp 1027–1035.

629

(60)

630 631

Anderberg, M. R. Cluster analysis for applications; Academic Press: New York, NY, USA, 1973.

(61)

Hartigan, J. A. Clustering algorithms; Wiley: New York, NY, USA, 1975.


34

Page 35 of 38

632


(62)

633 634

1963, 58 (301), 236–244. (63)

635 636

(64)

(65)

Lötsch, J.; Ultsch, A. Exploiting the Structures of the U-Matrix. In Advances in SelfOrganizing Maps and Learning Vector Quantization; Springer: Cham, 2014; pp 249–257.

(66)

641 642

Rousseeuw, P. J. Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics 1987, 20, 53–65.

639 640

Liu, Y.; Li, Z.; Xiong, H.; Gao, X.; Wu, J. Understanding and enhancement of internal clustering validation measures. IEEE Transactions on Cybernetics 2013, 43 (3), 982–994.

637 638

Ward, J. H. Hierarchical Grouping to Optimize an Objective Function. J. Am. Stat. Assoc.

Ultsch, A. U*-Matrix: a Tool to visualize Clusters in high dimensional Data; University of Marburg: Marburg, Germany, 2003.

(67)

Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel,

643

M.; Prettenhofer, P.; Weiss, R.; Dubourg, V.; Vanderplas, J.; Passos, A.; Cournapeau, D.;

644

Brucher, M.; Perrot, M.; Duchesnay, É. Scikit-learn: Machine Learning in Python. Journal

645

of Machine Learning Research 2012, 12, 2825–2830.

646

(68)

Wernet, G.; Bauer, C.; Steubing, B.; Reinhard, J.; Moreno-Ruiz, E.; Weidema, B. The

647

ecoinvent database version 3 (part I): overview and methodology. The International

648

Journal of Life Cycle Assessment 2016, 21 (9), 1218–1230.

649

(69)

650 651 652

Koch, P.; Salou, T. AGRIBALYSE: Methodology Report - Version 1.2; ADEME: Angers, France, 2015.

(70)

Tukker, A.; de Koning, A.; Wood, R.; Hawkins, T.; Lutter, S.; Acosta, J.; Rueda Cantuche, J. M.; Bouwmeester, M.; Oosterhaven, J.; Drosdowski, T.; Kuenen, J. Exiopol -


35


Page 36 of 38

653

Development and Illustrative Analyses of a Detailed Global MR EE SUT/IOT. Economic

654

Systems Research 2013, 25 (1), 50–70.

655

(71)

Wood, R.; Stadler, K.; Bulavskaya, T.; Lutter, S.; Giljum, S.; de Koning, A.; Kuenen, J.;

656

Schütz, H.; Acosta-Fernández, J.; Usubiaga, A.; Simas, M.; Ivanova, O.; Weinzettel, J.;

657

Schmidt, J. H.; Merciai, S.; Tukker, A. Global Sustainability Accounting-Developing

658

EXIOBASE for Multi-Regional Footprint Analysis. Sustainability (Switzerland) 2015, 7

659

(1), 138–163.

660

(72)

Federal

Statistical

Office

(BFS).

Landesindex

der

Konsumentenpreise.

661

Durchschnittspreise ausgewählter Produkte; Bundesamt für Statistik (BFS): Neuchâtel,

662

Switzerland, Fed Regist., 2017.

663

(73)

Weidema, B. P.; Bauer, C.; Hischier, R.; Mutel, C.; Nemecek, T.; Reinhard, J.; Vadenbo,

664

C. O.; Wernet, G. Overview and methodology - Data quality guideline for the ecoinvent

665

database version 3; ecoinvent report No. 1(v3); ecoinvent: St. Gallen, Switzerland, 2013;

666

Vol. 3.

667

(74)

Walker, C.; Gibney, E. R.; Hellweg, S. Comparison of Environmental Impact and

668

Nutritional Quality among a European Sample Population – findings from the Food4me

669

study. Scientific Reports 2018, 8 (1), 2330.

670

(75)

Scherer, L.; Pfister, S. Global Biodiversity Loss by Freshwater Consumption and

671

Eutrophication from Swiss Food Consumption. Environ. Sci. Technol. 2016, 50 (13),

672

7019–7028.

673

(76)

Federal Statistical Office (BFS). STAT-TAB: Die interaktive Statistikdatenbank;


36

Page 37 of 38


674 675

Bundesamt für Statistik (BFS): Neuchâtel, Switzerland, 2016. (77)

Jungbluth, N.; Nathani, C.; Stucki, M.; Leuenberger, M. Environmental Impacts of Swiss

676

Consumption and Production. A Combination of Input-Output Analysis with Life Cycle

677

Assessment; Federal Office for the Environment (FOEN): Bern, Switzerland, 2011.

678

(78)

Froemelt, A.; Mauchle, M.; Steubing, B.; Hellweg, S. Greenhouse Gas Emissions

679

Quantification and Reduction Efforts in a Rural Municipality. Journal of Industrial

680

Ecology 2018, 22 (1), 92–105.

681

(79)

Intergovernmental Panel on Climate Change (IPCC). Climate Change 2013: The Physical

682

Science Basis. Contribution of Working Group I to the Fourth Assessment Report of the

683

Intergovernmental Panel on Climate Change; Cambridge University Press: Cambridge,

684

United Kingdom and New York, NY, USA, 2013.

685

(80)

686 687

Goedkoop, M.; Heijungs, R.; Huijbregts, M.; De Schryver, A.; Struijs, J.; van Zelm, R. ReCiPe 2008; Ministrie van VROM: Den Haag, Netherlands, 2012.

(81)

Frischknecht, R.; Büsser Knöpfel, S. Swiss Eco-Factors 2013 according to the Ecological

688

Scarcity Method. Methodological fundamentals and their application in Switzerland;

689

Environmental studies no. 1330; Bundesamt für Umwelt (BAFU): Bern, Switzerland,

690

2013.

691

(82)

692 693 694

Bundesamt

für

Statistik

(BFS).

Haushaltsbudgeterhebung

2011:

Kommentierte

Ergebnisse und Tabellen; Bundesamt für Statistik (BFS): Neuchâtel, Switzerland, 2013. (83)

Buffat, R.; Froemelt, A.; Heeren, N.; Raubal, M.; Hellweg, S. Big data GIS analysis for novel approaches in building stock modelling. Applied Energy 2017, 208, 277–290.


37


695

(84)

Page 38 of 38

Saner, D.; Heeren, N.; Jäggi, B.; Waraich, R. A.; Hellweg, S. Housing and Mobility

696

Demands of Individual Households and their Life Cycle Assessment. Environ. Sci.

697

Technol. 2013, 47 (11), 5988–5997.

698 699

For Table of Contents Only

700


38

Using Data Mining to Assess Environmental ... - ACS Publications

Recommend Documents