Leveraging Artificial Intelligence to Analyze Citizens' Opinions on Urban Green Space
Mohammadhossein Ghahramani, Nadina J. Galle, Fabio Duarte, Carlo Ratti, Francesco Pilla
LLeveraging Artificial Intelligence to Analyze Citizens’ Opinions onUrban Green Space
Mohammadhossein Ghahramani a , Nadina J. Galle a,b , Fábio Duarte b , Carlo Ratti b andFrancesco Pilla a a Spatial Dynamics Lab, University College Dublin, Ireland b Senseable City Laboratory, Massachusetts Institute of Technology, USA
A R T I C L E I N F O
Keywords :Sentiment AnalysisSupervised LearningNatural Language ProcessingImbalanced Classification
A B S T R A C T
Continued population growth and urbanization is shifting research to consider the quality of urbangreen space over the quantity of these parks, woods, and wetlands. The quality of urban green spacehas been hitherto measured by expert assessments, including in-situ observations, surveys, and remotesensing analyses. Location data platforms, such as TripAdvisor, can provide people’s opinion on manydestinations and experiences, including UGS. This paper leverages Artificial Intelligence techniquesfor opinion mining and text classification using such platform’s reviews as a novel approach to urbangreen space quality assessments. Natural Language Processing is used to analyze contextual informa-tion given supervised scores of words by implementing computational analysis. Such an applicationcan support local authorities and stakeholders in their understanding of—and justification for—futureinvestments in urban green space.
1. Introduction
Urban Green Space (UGS) such as parks, woods, andwetlands represent a fundamental component of any urbanecosystem. In addition to the many ecological, economic,and psychological benefits, since the 1800s, UGS have beenrecognized for their ability to offer refuge from pervasive airpollution, and congestion [2, 10, 43]. Today, the ecologicalbenefits of green in the city are well-documented, but thereis also a growing body of evidence of its positive impact onhuman health and well-being [43]. Green space offers cit-izens more opportunities for social contact and stress relief– whether impromptu or planned [48]. Studies show UGSshould be of critical importance to public mental health, es-pecially from an urban planning perspective [21, 40].For many citizens, UGS has become an extension of, orin many cases the replacement of, the traditional backyard,meaning more people are sharing less green space. Despiteappeals for green space’s place in the city’s master plansand worldwide urban population growth, UGS has decreasedin several cities [13, 19, 23, 33]. Lucrative urban develop-ment and construction are often to blame for its demise. Tomeet demand, studies have suggested the quality of greenspace significantly contributes to neighborhood satisfactionand well-being, independent of the quantity of green space[48]. How to measure the quality of UGS has been hotly ⋆ This paper is the result of a research project funded by ConnectingNature (Grant Agreement No. 730222) under the European Community’sFramework Program Horizon 2020 and the Netherlands Fulbright Associ-ation through the Fulbright U.S. Doctoral Student Program. This researchwas made possible by the support of the Senseable City Laboratory at theMassachusetts Institute of Technology, which hosted Nadina Galle duringher fieldwork from September 2019 to March 2020. [email protected] (M. Ghahramani); [email protected] (N.J. Galle); [email protected] (F. Duarte); [email protected] (C. Ratti); [email protected] (F. Pilla)
ORCID (s): debated in urban forestry and planning fields, with severalattempts made to streamline and standardize quality assess-ments of UGS [6, 8, 18, 20]. However, with current meth-ods relying on expert assessments, some warn it discreditsthe experience of local users; who are likely more qualifiedto assess their own UGS than outside experts [25].The definition of quality UGS is still contested, and theirrole remains undervalued. Measuring UGS quality is alsoa tedious process; observational techniques are often criti-cized as they require extensive repeat measurements at thesame location, incurring large time and cost expenditures[39]. Even when data is collected, it is quickly outdated,leaving progress out of reach [36].How can a city ensure it provides safe, inclusive, high-quality UGS for all? Emerging technologies are gaining trac-tion as a way to gain up-to-date information on—and en-gage local users in—the planning and improvement of UGS[14, 32, 17]. There are several quantitative approaches to an-alyzing UGS such as drones [31], satellite imagery [11], andGoogle Street View images [42], but there is a need for re-producible qualitative analysis. One such technology, Natu-ral Language Processing (NLP), combines computer scienceand linguistics to understand the language in a piece of text.One of its applications, Sentiment Analysis (SA), can ex-tract and categorize positive, negative, or neutral sentimentfrom a chunk of text. While NLP can work on any writtentext, performing SA on georeferenced crowd-sourced datasources such as TripAdvisor, Twitter, Yelp, Booking.com,and Airbnb have shown particular promise [5, 36, 26]. Ap-plications range from understanding consumers’ attitudes to-ward their products to the socioeconomic status of communi-ties to hospitality organizations’ performance [29, 30, 37]. Ithas been suggested that in most of these applications, senti-ment analysis should become a complementary tool for qual-ity assessment and evaluation [15].TripAdvisor is a particularly popular platform with a rich
M. Ghahramani et al.:
Preprint submitted to Elsevier
Page 1 of 10 a r X i v : . [ c s . S I] F e b everaging Artificial Intelligence to Analyze Citizens’ Opinions on Urban Green Space and publicly accessible database on attractions, destinations,and landmarks, including UGS [29]. While relevant to thisresearch, studies of the demographic makeup of TripAdvi-sor are limited. Some groups are likely over- and/or under-represented on TripAdvisor, but it is still advantageous overpopulation-based surveys, a costly and tedious method to ac-quire representative population samples. TripAdvisor offersa viable, complementary method to harvest local opinionand feedback on UGS.This paper presents a novel NLP application using Tri-pAdvisor to assess the quality of UGS. The corpus collectedwere TripAdvisor reviews of St. Stephen’s Green, the mostpopular public park in Dublin (Ireland). St. Stephen’s Green,in the middle of Dublin’s city center, is a 10-hectare parkwith over eight million visitors on an annual basis. The parkhas well-maintained facilities on the grounds, including over3 km of walking paths and public restrooms inside the park.Experimental, computational analyses were implementedvia two scenarios, and different phases have been included toaddress identifying the sentiment expressed in reviews. Theproposed method allows the extraction and interpretation ofsentiment with minimal human effort by applying ArtificialIntelligence (AI) and Machine Learning (ML) algorithms.The contributions of this work are as follows:• We present a novel application of NLP and text miningusing TripAdvisor to assess the quality of UGS.• We present how a self-contained sentiment analysismodel can be implemented to evaluate people’s atti-tudes toward various entities given a class imbalancedissue.The paper is organised as follows: some related work inthe field of sentiment analysis and opinion mining is pre-sented in Section 2; the proposed approach with its associ-ated discussions is presented in Section 3; Section 4 showsthe experimental results; Section 5 details the discussion;and Section 6 concludes the paper.
2. Related Work
The value of UGS remains underestimated due to a lackof information about what quality green space entails andhow existing spaces within the city score on important social-quality parameters. Measuring the quality of UGS is a te-dious process. Observational approaches are often criticizedas they require extensive repeat measurements at the samelocation, incurring large time and cost expenditures [6, 39].Even when data is collected, it is quickly outdated, leavingprogress out of reach.
A recent improvement is web-based civic participationplatforms. In an effort to gain insights into how people per-ceive a park’s quality, several cities have released apps. Am-sterdam recently launched “MyPark”, an app that asks theuser questions about specific areas of a specific park. Once the results are analyzed, the feedback is incorporated into apark redesign to better meet local user needs.FixMyStreet is another example. The map-based appacts as a liaison between residents and their local author-ity on problems such as potholes and broken street lightsneeding their attention. The app was launched in the UnitedKingdom (UK) and has proliferated across the country. FixMyS-treet also has an open-source platform that helps people runsimilar websites all over the world. Although mainly usedfor reporting common street problems, the app could also beused to highlight issues facing urban parks and woodlands.
Microblogging platforms challenge users to summarisetheir thoughts in a limited amount of words. Twitter, ar-guably the world’s microblogging pioneer, allows 280 char-acters per post (or “tweet”), a recent upgrade from the iconic140 characters they used to enforce. In her inspiring pa-per [39], Roberts proposes the use of crowdsourced, geo-tagged social media data, such as tweets, to inform how,when, and why people use UGS. This method overcomessome issues with previous approaches, such as report basedmethods, which are difficult to validate, and observationalmethods, which require multiple observations over differentdays and seasons to ensure reliability [39]. It can even beused to derive seasonal variation in physical activity in UGS[39]. Crowdsourcing data from Twitter offers an alternativeas it is publicly available and instantly accessible, incurringno additional time or costs.Yet, both web-based civic participation platforms andsocial media data face limitations. FixMyStreet, with over12,000 reports sent to UK councils every month, has muchless usability than Twitter, which recorded 17 million monthlyactive British users in the first quarter of 2018. It is unlikelyany significant amount of these tweets are actually about is-sues on the street, but it is a much broader data source.There are also socio-demographic concerns regarding theuser base of both FixMyStreet and Twitter. In 2017, [34] an-alyzed over 30,000 FixMyStreet reports, compared them to arange of socio-demographic indicators, and revealed crowd-sourced civic participation platforms tend to marginalize low-income and ethnically diverse communities.In the same way, the elderly population, who show lowerlevels of engagement with these forms of technology, are dis-regarded explicitly in such research [4]. This is especiallyconcerning as urban parks are supposed to be a shared pub-lic space for all ages. Roberts also reports Twitter data lacksdemographic information about Twitter users such as theirage, occupation, or ethnicity [39]. Although not crucial todetermine opinions, these parameters are useful for furtherexamination of where particular attitudes may originate. Ev-idently, there is a need for inclusive, unrestricted, unbiased,and freely-solicited opinions about UGS.
NLP is used to understand the language in a piece of textand reveal the sentiment behind it. The method combines
M. Ghahramani et al.:
Preprint submitted to Elsevier
Page 2 of 10everaging Artificial Intelligence to Analyze Citizens’ Opinions on Urban Green Space computer science and linguistics. In recent years, the popu-larity of virtual assistants like Siri, Alexa, and Google Homehas accelerated the demand for voice user interfaces. And,as such, increased research on how computers understandspeech and speak themselves.NLP can also work on written text, like user-generatedreviews on the world’s largest travel website, TripAdvisor.The open, online community reaches 390 million unique vis-itors each month and lists 465 million reviews and opinionsabout more than seven million accommodations, restaurants,and attractions in 49 markets worldwide. TripAdvisor is atreasure trove of sentiment. When writing a review, a re-viewer is prompted to describe their first-hand experiencecausing both tourists and locals to flock to TripAdvisor to ex-press their opinions. Whereas Twitter offers a platform forsharing an occasional opinion, TripAdvisor explicitly asksfor the sentiment.The overall user base of both TripAdvisor and Twitter isstill poorly understood. In 2007, Gretzel [22] found frequenttravel review readers tend to be younger, have slightly higherincomes, and are more likely to contribute to online content.They are also more likely to post reviews themselves. Thus,one could assume the reviewers share a similar demographicto the reader, at least in 2007.Gender differences can also play a role. Blumenthal [3]found little to no gender differences amongst reviewers onTripAdvisor in 2014. Twitter, on the other hand, did exhibitgender differences. According to Statista, an online statis-tics, market research, and business intelligence portal, dur-ing a 2018 study period, 42.8 percent of global Twitter userswere female, and 57.2 percent were male (Global TwitterUser Distribution by Gender 2018 | Statistic, n.d.).The use of Twitter tends to drop as age increases. Inthe United States, those under 50, especially those 18-29,are most likely to use Twitter. And only 6 percent of Twit-ter users constitute the 65+ age group. TripAdvisor’s Trip-Barometer report showed a slightly more even distributionof the travel site’s user base.Research to validate these demographic claims is lim-ited, and studies comparing TripAdvisor with Twitter’s userbase are non-existent. Although some groups remain over-and/or underrepresented on TripAdvisor, it is advantageousover Twitter as it generally covers a broader demographicspectrum. In fact, the only known method to encompass ageneral population is population-based surveys, where anexperiment is administered to a representative populationsample. However, this process is costly and lengthy, and assuch, TripAdvisor offers a viable, complementary method toharvest local opinion and feedback on UGS.So far, sentiment analysis using TripAdvisor as a datasource has only been applied in the hospitality and tourismsectors. Here, shallow NLP techniques are applied to extractsentiment [15] automatically. These simple expressions, whichare derived from the reviews, can be used to evaluate thequality of hotels or restaurants. García-Barriocanal’s pre-liminary study [15] was able to identify emotion types withreasonable effectiveness and suggested sentiment analysis using TripAdvisor reviews should become a complementarytool for hospitality evaluation.
Sentiment classification is an example of a supervisedmachine learning task, a process of assigning text documentsinto two or more predefined classes. In this process, an al-gorithm takes any observation (text document) as input andassigns a label from the class labels [16, 24]. Different data-driven supervised approaches have been used to deal withsuch a classification problem. Sentiment classification hasraised much attention in recent years and has undergone manychanges. Generally speaking, three techniques can be usedto construct a sentiment lexicon, i.e., dictionary-based, corpus-based, and hybrid methods. Dictionary-based methods useword matching based on the lexicon. However, since sen-timent words in the lexicon might be difficult to recognize,many texts cannot be analyzed by utilizing such classifiers.Corpus-based methods use labeled data, and lexicons are noteffectively taken into account in such approaches. To alle-viate the discussed shortcomings, a hybrid approach (i.e.,a combination of machine learning methods and lexicons)can help improve the sentiment classification performance.Since the text classification problem is a supervised learningtask in which the class observations is predicted based onsome feature values, a wide range of ML algorithms (e.g.,Support Vector Machine (SVM) [12, 1], Naive Bayes (NB)[12, 38], decision tree [12], random forest [12, 1], logisticregression, and neural networks [27, 47, 9, 7]) can be incor-porated.As explained, in this work, people reviews as to UGS aretaken into account. These texts are unstructured; thus, man-ually analyzing them can be tedious and time-consuming. Inthis type of data mining, people’s opinions, sentiments, andattitudes are analyzed. The main objective is to computerizethe process of reading reviews and evaluate them. It shouldbe mentioned that the most crucial task in sentiment analysisis the pre-processing phase, including different operations.Due to differences in data characteristics, these tasks mightdiffer from one sentiment analysis approach to another. Be-cause of the complexity of feature dependency, ML methodsmay achieve different results. Given this work’s characteris-tics, we aim to propose an appropriate approach to deal withvarious issues to be explained next. It is worth mentioningthat a self-contained model consisting of multiple phases isimplemented in this paper. In the last stage of the model (theclassification phase), different ML algorithms are tested, andtheir corresponding results are compared.
3. Method
There have been different approaches to perform senti-ment analysis. However, choosing a proper method is highlyrelated to the nature of a given work. This paper analyzespeople’s opinions and sentiments to identify different pos-itive and negative polarities on urban green space. Differ-ent from book articles and news reports, review texts are
M. Ghahramani et al.:
Preprint submitted to Elsevier
Page 3 of 10everaging Artificial Intelligence to Analyze Citizens’ Opinions on Urban Green Space
Figure 1:
Four phases of the sentiment analysis model used inthis work. often short and ambiguous. Various models, i.e., fully su-pervised and semi-supervised methods, have been consid-ered to analyze review comments. The methods in the for-mer category use manually labeled data. Their approachis very time-consuming to create lexicons manually. Somespecific supervised methods have been introduced to trainsentiment classifiers on emoticons and hashtags. Because ofsuch shortcomings, a semi-supervised model has been con-sidered in this paper. It should be noted that the classificationphase of the model is based on a supervised technique, whilean unsupervised method is used in the pre-processing phaseof the model. Most of the concerns related to opinion miningand sentiment analysis of reviews can be addressed by im-plementing effective pre-processing techniques. However,there are no effective pre-processing methods for all datasetsand algorithms. For instance, in this work, we deal with animbalanced classification issue since most comments in thedataset used are positive. A multi-layer approach consistingof different phases (i.e., web scrapping, data cleanings, im-balanced classification, and supervised ML) is implementedto address all concerns. Fig. 1 illustrates different phasesof the proposed model. The following sections present alldetails regarding each stage of the model.
TripAdvisor reviews for St. Stephen’s Green (Dublin,Ireland) were scarped using Selenium and Python. The pseu- docode is presented in Algorithm 1. The reviews were sub-sequently processed to focus on English texts, because of themass availability of English language text analysis tools anddictionaries. The reviews were collected from the period ofMay, 2006 to November, 2020, for a total of 16,613 reviews;in contrast, Dublin’s second most popular park had 4,753reviews for this time period. Of the St. Stephen’s Green re-viewers, 5,622 were from the United States, 3491 were fromIreland, 2,835 from the United Kingdom (UK), 796 fromCanada, 483 from Australia, 105 from Germany, 92 from theNetherlands, 91 from South Africa, 48 from New Zealand,43 from Denmark, 39 from Greece, 32 from United ArabEmirates, and 30 from China. The remaining 2,906 review-ers were from other countries, had misspelled their country,or left their location blank.The following review fields were extracted: review-title(written title of review); review-body (written review aboutthe destination); rate-value (1 is the lowest evaluation, 5 isthe best); review-location (where a reviewer is from); andreview-date (date review was written). Only review-bodyand rate-value data fields were used in this experiment. Thisdataset can be considered as a sequences of text, i.e., 𝐷 ={ 𝑋 , 𝑋 , ..., 𝑋𝑛 } where 𝑋 𝑖 refers to the 𝑖 𝑡ℎ review. Eachreview is also labeled as positive or negative, depending onits corresponding rate value. As the quality of data affects the analysis, it is essentialto employ a data pre-processing procedure. To that end, fea-ture extraction was performed, and a structured set from thereviews is created for the model-training purposes. A dimen-sionality reduction operation is also considered by applyingthe Term Frequency-Inverse Document Frequency (TF-IDF)technique [45]. These pre-processing steps help us convertunstructured text sequences into a structured feature space.Data cleansing operations were performed, and punctuationsand stop words were omitted. To make transformations (re-moving punctuations, stop words, and other cleansing oper-ations) implemented in this work, libraries from the NaturalLanguage Toolkit (NLTK) were used. This Python libraryhas been written for modeling text and provides various toolsfor loading and cleaning texts. This library’s different func-tions were used for filtering punctuation, stemming, normal-izing, extracting text from HTML, decoding Unicode char-acters, locating typos, and handling numbers.Data normalization techniques (e.g., Stemming and Lem-matisation) were applied, each review was converted to anumeric representation (corpus), and the 𝑛 -grams approach(with two different measures like Word Counts and TF-IDF)was implemented. The former is based on mapping morethan just one word (unigrams) onto the corpus. We havealso included word counts into our model. To that end, thenumber of times a given word or a sequence of words appearis counted. The latter, term TF-IDF, is a weighting measureto be used instead of word count representations. This mea-sure is considered to lessen the effect of implicitly commonwords in the corpus. The weight of a term in a review can M. Ghahramani et al.:
Preprint submitted to Elsevier
Page 4 of 10everaging Artificial Intelligence to Analyze Citizens’ Opinions on Urban Green Space
Algorithm 1:
Pseudo-code for extracting Tripadvisor reviews
Input :
𝐶𝑆𝑉 𝑓 𝑖𝑙𝑒 ← [Score, Date, Title, Review]; 𝑑𝑟𝑖𝑣𝑒𝑟 ← webdriver.Chrome(); 𝐹 𝑛 ← Selenium.find_element_by_xpath;
𝑈 𝑅𝐿 ← Output:
Review Comments Exception Function try: driver.find_element_by_xpath(xpath) except NoSuchElementException: return False return True n = number of web pages; HTML elements 𝑒 = ’taLnk ulBlueLinks’; 𝑒 = ’ui_bubble_rating bubble_’; 𝑒 = ’_34Xs-BQm’; 𝑒 = ’glasR4aX’; 𝑒 = ’IRsGHoPm’; 𝑒 = ’Dq9MAugU T870kzTX LnVzGwUB’; for 𝑖 ← , , ...., 𝑛 do if 𝐸𝑥𝑐𝑒𝑝𝑡𝑖𝑜𝑛𝐹 𝑢𝑛𝑐𝑡𝑖𝑜𝑛 ( ′′ ∕∕ 𝑠𝑝𝑎𝑛 [@ 𝑐𝑙𝑎𝑠𝑠 = 𝑒 ] ′′ ) then 𝑑𝑟𝑖𝑣𝑒𝑟.𝐹 𝑛 ( ′′ ∕∕ 𝑠𝑝𝑎𝑛 [@ 𝑐𝑙𝑎𝑠𝑠 = 𝑒 ] ′′ ) .𝑐𝑙𝑖𝑐𝑘 (); end 𝑑𝑓 ← 𝑑𝑟𝑖𝑣𝑒𝑟.𝐹 𝑛 ( ′′ ∕∕ 𝑑𝑖𝑣 [@ 𝑐𝑙𝑎𝑠𝑠 = 𝑒 ] ′′ ) ; 𝑛𝑢𝑚 ← 𝑙𝑒𝑛 ( 𝑑𝑓 ) ; for 𝑗 ← , , ..., 𝑛𝑢𝑚 do 𝑆𝑐𝑜𝑟𝑒 = 𝑑𝑓 [ 𝑗 ] .𝐹 𝑛 ( ′′ . ∕∕ 𝑠𝑝𝑎𝑛 [ 𝑐𝑜𝑛𝑡𝑎𝑖𝑛𝑠 (@ 𝑐𝑙𝑎𝑠𝑠, 𝑒 )] ′′ ) .𝑔𝑒𝑡 _ 𝑎𝑡𝑡𝑟𝑖𝑏𝑢𝑡𝑒 ( ′′ 𝑐𝑙𝑎𝑠𝑠 ′′ ) .𝑠𝑝𝑙𝑖𝑡 ( ′′ _ ′′ )[3]); 𝐷𝑎𝑡𝑒 = 𝑑𝑓 [ 𝑗 ] .𝐹 𝑛 ( ′′ . ∕∕ 𝑠𝑝𝑎𝑛 [@ 𝑐𝑙𝑎𝑠𝑠 = 𝑒 ] ′′ ) .𝑔𝑒𝑡 _ 𝑎𝑡𝑡𝑟𝑖𝑏𝑢𝑡𝑒 ( ′′ 𝑡𝑖𝑡𝑙𝑒 ′′ ); 𝑇 𝑖𝑡𝑙𝑒 = 𝑑𝑓 [ 𝑗 ] .𝐹 𝑛 ( ′′ . ∕∕ 𝑑𝑖𝑣 [@ 𝑐𝑙𝑎𝑠𝑠 = 𝑒 ] ′′ ) .𝑡𝑒𝑥𝑡 ; 𝑅𝑒𝑣𝑖𝑒𝑤 = 𝑑𝑓 [ 𝑗 ] .𝐹 𝑛 ( ′′ . ∕∕ 𝑞 [@ 𝑐𝑙𝑎𝑠𝑠 = 𝑒 ] ′′ ) .𝑡𝑒𝑥𝑡.𝑟𝑒𝑝𝑙𝑎𝑐𝑒 ( ′′ ∖ 𝑛 ′′ , ′′ ′′ ); end end Return: Score, Date, Title, Review.be defined as: 𝑤 ( 𝑟, 𝑡 ) = 𝑇 𝐹 ( 𝑟, 𝑡 ) ∗ 𝑙𝑜𝑔 ( 𝑛𝑑𝑓 ( 𝑡 ) ) (1)where 𝑛 is the number of reviews and 𝑑𝑓 ( 𝑡 ) is the numberof reviews consisting of the term 𝑡 in the corpus.Negation handling could be another challenging task forsentiment classification. However, since we deal with a two-class classification task, such concern can be easily addressed.The model negates the predicted class of observations, asthere are only two classes to choose from. Such negationrecognition can be a complicated process in cases where thereare more than two possible classes. In our case, the negationhandling procedure is considered as an Exclusive-OR prob-lem. As far as the negation scope detection is concerned, dif-ferent negation keywords are defined, and the regular expression-based NegEx method is used. Moreover, the negation im-plicitly is captured via n-grams.Although the dataset has been cleaned after performingthe explained operations, there is still a considerable con-cern. The most challenging data pre-processing task in thiswork is an imbalanced class issue. This procedure is usuallyregarded as a pre-processing task; however, we consider itas a separate task to be explained next due to its importancein our work. The observations that were labeled as negative are rela-tively rare as compared to the positive class (less than 10%).Positive and negative labels are determined according to thereviewers’ ratings. Should a rating be higher than four, thecorresponding review is considered as a positive one; other-wise, it is treated as a negative review (Table 1).Hence, we face an imbalanced classification issue. Inother words, positive class (the majority) outnumbered neg-ative class (the minority), and both classes do not make upan equal portion of our dataset. The conventional classifierssuch as Decision Tree ([41]) and Logistic Regression ([35])do not accurately measure model performance when facedwith imbalanced datasets. They usually have a bias towardsthe majority class, and the minority class observations aretreated as noise. To handle this issue, an SVM that performswell against highly imbalanced datasets are used to train ourmodel. This classifier is also equipped with a class weightmeasure to alleviate the situation. Moreover, a separate im-balanced classification phase is embedded into the model. Indoing so, different types of algorithms, i.e., multi-task learn-ing [28], adaptive sampling [44], and synthetic oversamplingmethod [46], were integrated and tested.Generally speaking, there are two distinctive approachesfor handling the mentioned issue: 1) skew-insensitive tech-niques and 2) re-sampling approaches. The former deals
M. Ghahramani et al.:
Preprint submitted to Elsevier
Page 5 of 10everaging Artificial Intelligence to Analyze Citizens’ Opinions on Urban Green Space
Table 1
Examples of positive and negative TripAdvisor reviews about St. Stephen’s Green (Dublin,Ireland).
Bubble_rating Review_body Label5 Stephens Green is a great place to visit nice walk around the park there also in the summertime musicplaying you can have a picnic there watch the ducks and the swans in the pond there is also boards givinga bit of history... positive5 Great place to relax in the city. Beautiful gardens and paths to walk around. If you need to just sit a bitthis is a great place to do so. positive4 St. Stephen’s Green park is perfect to step away from the hustle and bustle of Dublin. The scenery isbeautiful and calming. Then when you are refreshed you just step back into the action of the city. positive3 Only downside is anti-social behaviour. Always somebody hassling for money or asking for a smoke.Wouldn’t mind it on my own but with kids it’s terrible. Should be better patrolled. Europe’s... negative1 While on a visit to Dublin we brought our children to the park. The place is really nice but we were reallyshocked when we went to the playground. Near the playground entrance there were about 200 teensdrinking and causing trouble. negative2 While we were walking across the park, a young man tried to take my husband’s laptop. It was zippedinside a shoulder bag. I yelled and this person went away. negative with a class imbalanced problem by assigning a cost measureto the training data. The latter adjusts the original datasetsuch that a more balanced class distribution is achieved. Re-sampling methods ([46]) have become standard approachesand have been dominantly utilized recently. They can beclassified into different categories, e.g., sampling strategies,wrapper approaches, and ensemble-based methods. Imple-menting a proper method is crucial; otherwise, it can beproblematic, e.g., data loss and overfitting, and can resultin a poor outcome. This phase aims to balance class distri-bution relatively. As stated, three different techniques havebeen tested. We have found that a synthetic oversamplingalgorithm ([46]) performs better than the other two methods(i.e., adaptive sampling and multi-task learning). It is worthmentioning that the two other methods used are also com-putationally expensive. The synthetic oversampling algo-rithm creates synthetic samples based on the nearest neigh-bor approach. By implementing the method,
Failure class instances are synthetically created, and the distribution ismore balanced. The procedures are as follow:• Let 𝐴 be the set of all elements of the minority class.The algorithm detects k-nearest neighbors of all ob-servations ( 𝑆 ∈ 𝐴 ) of this class. In doing so, the Eu-clidean distance between each observation and otherelements is measured.• A sampling rate (e.g., 60%) is defined based on theimbalanced proportion. Given such a pre-defined rate,60% of k-nearest neighbors of each observation in theminority class are randomly selected. Let 𝐴 ′ be theset of k-nearest neighbors.• For each element in the obtained set ( 𝑆 ′ ∈ 𝐴 ′ ) thefollowing formula is used to create new samples. 𝑆 𝑛𝑒𝑤 = 𝑆 + 𝛼 ∗ | 𝑆 − 𝑆 ′ | (2)where 𝛼 is a random number between 0 and 1.The pseudo-code of the procedure integrated into the modelto handle class imbalanced issue is presented in Algorithm2. Algorithm 2:
Pseudo-code for handling class im-balanced issue
Input : 𝑚 ← number of minority classobservations; 𝑟 ← amount of over-sampling (%); 𝑘 ← number of neighbors; Output: r*m synthetic observations Exception Function 𝑓 = len(features) number of features ; 𝑆 ← [ ] observations in the minority class ; 𝑆 ′ ← [ ] synthetic observations ; Γ ← [ ]; for 𝑖 ← , , ..., 𝑚 do 𝐷 ← Compute Euclidean distance ; 𝐷.𝑠𝑜𝑟𝑡𝑒𝑑 () sort in an ascending order 𝐼𝑛𝑑𝑒𝑥 ← Find k-nearest neighbors and indices Γ .𝑎𝑝𝑝𝑒𝑛𝑑 ( 𝐼𝑛𝑑𝑒𝑥 ) 𝑃 𝑜𝑝𝑢𝑙𝑎𝑡𝑒 ( 𝑖, 𝑟, 𝐼𝑛𝑑𝑒𝑥 ) end 𝑝 ← a number between 1 and k 𝛼 ← a number between 0 and 1 for 𝑗 ← , , ..., 𝑓 do Λ = 𝑆 [ 𝑗 ] − 𝑆 ′ [ 𝑗 ]; 𝑆 𝑛𝑒𝑤 [ 𝑗 ] = 𝑆 [ 𝑗 ] + 𝛼 ∗ Λ; end Return 𝑆 𝑛𝑒𝑤 After all operations explained above are done, three moresteps are required before the pre-processed data is given to asupervised algorithm.1. Tokenisation: each review is broken into words (calledtokens).2. Vectorisation: each review is converted into a numericrepresentation (called corpus).3. Transformation: each review is transformed into onerow (including 0 or 1) where 1 is the word in the cor-pus corresponding to that column appearing in that re-view.
M. Ghahramani et al.:
Preprint submitted to Elsevier
Page 6 of 10everaging Artificial Intelligence to Analyze Citizens’ Opinions on Urban Green Space
Figure 2:
The support vectors given positive and negativeclasses.
Text normalization was used to convert text into moreconvenient, standard forms. Tokenisation was used to sepa-rate words from running text. Each review has a rate_valuebetween 1-5. Any rate_value of 1, 2, and 3 is considered anegative review; 4 and 5 are considered a positive review.Thus, there are two classes in this work, as the methods im-plemented are binary classification models. Then tokenizedwords are converted into a numeric representation, a processknown as vectorization. After the data is processed, two ap-proaches were applied: unigrams and n-grams. An 𝑛 -gramis a contiguous sequence of 𝑛 words collected from our re-views. When 𝑛 is equal to , it refers to as a unigram. Theircorresponding models are probabilistic language models forpredicting every word’s ratio (in a unigram approach) or se-quence of words (in an 𝑛 -gram approach). After all the de-scribed operations are done, the pre-processed data is trainedon a supervised ML method, i.e., SVM.SVM is incorporated as a discriminative classifier fordocument categorization in this work. As explained in theprior section, it is less sensitive to the class imbalanced prob-lems. This technique is based on the Structural Risk Min-imisation principle. SVM’s task is to learn and generalizean input-output mapping by finding separation between hy-perplanes defined by classes of data. In our case, the setof reviews is the algorithm input, and their respective la-bels are the output. SVM searches for a separating hyper-plane, which separates positive and negative reviews fromeach other with maximal margin; in other words, the dis-tance of the decision surface and the closest review is maxi-mal (Fig. 2).Let ( 𝑥 , 𝑦 ) , ( 𝑥 , 𝑦 ) , ..., ( 𝑥 𝑛 , 𝑦 𝑛 ) , 𝑦 𝑖 ∈ { 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒, 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒 } be our training observations. The SVM classifier is imple- mented by solving the following optimisation problem: 𝑚𝑎𝑥𝑖𝑚𝑖𝑠𝑒 𝑛 ∑ 𝑖 =1 𝜇 𝑖 − 12 𝑛 ∑ 𝑖,𝑗 =1 𝜇 𝑖 𝜇 𝑗 𝑦 𝑖 𝑦 𝑗 𝜙 ( 𝑥 𝑖 , 𝑥 𝑗 ) (3) 𝑓 ( 𝑥 ) = 𝑛 ∑ 𝑖 =1 𝑦 𝑖 𝜇 𝑖 𝜙 ( 𝑥 𝑖 , 𝑥 𝑗 ) + 𝜉 ∀ 𝑖 ∶ 0 ≤ 𝜇 𝑖 ≤ 𝐶 𝑎𝑛𝑑 𝑛 ∑ 𝑖,𝑗 =1 𝜇 𝑖 𝑦 𝑖 = 0 (4)where 𝜙 is a pisa kernel function, 𝜇 is a weight value, 𝜉 isa threshold and 𝐶 is a misclassification cost. The algorithmoffers an optimal hyperplane, which is a decision boundarybetween the two classes.
4. Results
Supervised machine learning approaches are about con-ducting algorithms that precisely project a given input fea-tures to an output space. Each of these methods operates intwo stages. First, an algorithm is trained based on a trainingdataset. Then, the algorithm is evaluated over various met-rics based on a test dataset. Splitting the dataset is essen-tial for an unbiased evaluation of prediction performance.Hence, the dataset used in this work was divided into twosubsets. The testing dataset includes reviews, consist-ing of positive and negative comments. As ex-plained above, this dataset is used for the evaluation of allmodels implemented in this work. It should be mentionedthat the imbalanced handling phase is implemented when thetraining data is fitted.Given the above discussion, the training set was appliedto train models. Computational analyses were implementedbased on two scenarios, i.e., a traditional approach and themodel proposed in this work. Both scenarios include all thedata handling steps explained earlier, i.e., data pre-processingand supervised learning. However, our proposed model in-cludes an additional imbalanced handling phase described inthe previous section. As far as the first scenario is concerned,various supervised algorithms, including Deep Neural Net-work (DNN), Recurrent Neural Network (RNN), QuadraticDiscriminant Analysis (QDA), and Random Forest (RF), aretested and their results are compared with the proposed model.Given the extracted features, all the mentioned algorithmswere fitted. After training all models, we have evaluatedthem to verify their applicability. Understanding how a modelperforms is essential to the use and development of text clas-sification methods. To do so, the area under the ReceiverOperating Characteristics (ROC) curves are used for com-paring the accuracy of algorithms. These curves reveal atrade-off between the true positive rate and the false positiverate. The evaluation metric is based on a confusion matrixthat comprises true positives (TP), false positives (FP), falsenegatives (FN), and true negatives (TN). The significance of
M. Ghahramani et al.:
Preprint submitted to Elsevier
Page 7 of 10everaging Artificial Intelligence to Analyze Citizens’ Opinions on Urban Green Space
Figure 3:
Proposed model ROC curves with and without im-balanced handling phase.
Figure 4:
Confusion matrix for the positive and negativeclasses. these four elements may vary based on the classification ap-plication. In this work, the fraction of correct predictionsoverall predictions is considered. 𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇 𝑃 + 𝑇 𝑁𝑇 𝑃 + 𝐹 𝑃 + 𝐹 𝑁 + 𝑇 𝑁 (5)Fig. 3 illustrates the ROC curve given the proposed modelin this paper with and without the imbalanced handling phase.As shown, the model’s predictive accuracy, assessed usingthe area under the curve (AUC), is over 97%. The confusionmatrix is also presented in Fig. 4.As stated, the proposed model has been experimentallyvalidated and compared with four different approaches. Theircorresponding performances have been evaluated accordingto their classification accuracies. The results are depictedin Fig 5. The ability of each method to accurately predictthe correct class is measured and expressed as a percent-age. ROC curves have been used to determine the predic-tive performance of the examined classification algorithms.
Figure 5:
ROC curves for different approaches.
Table 2
Comparisons of different classification metrics given 5 testedapproaches.Models Precision Recall F1 scoreProposed approach 0.971 0.997 0.983DNN-based model 0.946 0.993 0.968RNN-based model 0.92 0.994 0.955QDA-based model 0.901 0.991 0.942RF-based model 0.927 0.992 0.957
The area under a ROC has been considered as an evaluationcriterion to select the best classification algorithm. Whenthe area under the curve is approaching 1, it indicates thatthe classification was carried out correctly. We have alsotested three more metrics, i.e., Precision, Recall, and 𝐹 -Score (Table 2). The Recall metric is the measure of thecorrectly predicted positive reviews from all the actual pos-itive ones (Recall = 𝑇 𝑃𝑇 𝑃 + 𝐹 𝑁 ). Hence, it is a good indica-tor for evaluating models (given the cost of False Negatives)dealing with the imbalanced class issue.All experimental results show that our proposed modelis superior to those tested. The additional imbalanced han-dling phase incorporated improves the fit of the model.
5. Discussion
TripAdvisor reviews reveal a treasure trove of global,comparative data. To date, no other crowdsourced data waswidespread enough to allow for such comparison betweengreen spaces, both within the city and beyond. The mostcommon criticism of observational approaches was time andcost expenditures spent on repeat measurements at the samelocations ([6]). Like TripAdvisor, Twitter-based methodsovercame this hurdle; tweets can be captured easily and fre-quently, offering greater measurement and (even longitudi-nal) analysis opportunities, saving time and costs ([39]). How-ever, TripAdvisor, for a city’s most popular parks, offersmore data. The TripAdvisor reviews collected for this study
M. Ghahramani et al.:
Preprint submitted to Elsevier
Page 8 of 10everaging Artificial Intelligence to Analyze Citizens’ Opinions on Urban Green Space averaged out to be 80 reviews/month/park for St. Stephen’sGreen. Roberts ([39]) study collected about 11 tweets/month/parkfor her study area. Furthermore, tweets and reviews are notthe same; while a tweet can be about UGS, a TripAdvisorreview explicitly asks for a reviewer’s experience (i.e., a per-son’s sentiment).The elderly population, who show lower levels of en-gagement with technology in general, are specifically over-looked in crowdsourced data-based research ([5]). This isespecially concerning as UGS are intended to be a sharedpublic space for all ages. Roberts also reports Twitter datalack demographic information about Twitter users, such astheir age, occupation, or ethnicity ([39]). These parameters,although not crucial to determine opinions, are useful for fur-ther examination of where particular attitudes may originate.Research to validate these demographic claims is limited,and studies comparing TripAdvisor with Twitter’s user baseare non-existent. Although some groups remain over- and/orunderrepresented on TripAdvisor, there is an option to col-lect some demographic information such as gender, nation-ality, and age. We acknowledge the bias most crowdsourceddata has and understand it is both a contested and fertile re-search area.TripAdvisor enabled us to utilize the abundance of open-source reviews. The accessibility of the reviews makes theproposed method highly scalable—especially for popular parks.In Dublin, 33 of its 50 parks are listed on TripAdvisor. How-ever, besides the most popular St. Stephen’s Green (16,613reviews), Phoenix Park (4,753 reviews), and St. Anne’s Park(244 reviews), the remaining 30 parks have 1-66 reviews.Worldwide, thousands of UGS, from large to small, are listedon TripAdvisor. However, Dublin follows a similar patternas other cities, where the most popular parks have signifi-cant reviews, and the lesser common parks have significantlyfewer reviews. Therefore, we suggest the proposed methodto be used only on a city’s most popular parks, as a proxy forUGS in cities, and then compare UGS between cities world-wide. By leveraging machine learning techniques for opin-ion mining and text classification, hundreds of thousands ofopinions previously overlooked can now be heard in an effortto improve these vulnerable public spaces.
6. Conclusions
Research to inform both policy and design of UGS is crit-ical to protect these vulnerable areas while simultaneouslyensuring access to the potential health and well-being ben-efits these spaces provide. Green spaces play a pivotal roleacross all aspects of city life, and as cities densify, the im-portance of accurately and effectively measuring the qualityof UGS has never been greater.This paper presents an experiment’s results to use NLPto extract citizen opinion on the quality of UGS, a highlynovel application of automatic text classification on TripAd-visor reviews. The results indicate that the proposed methodperforms better, at accuracy, which is better than otherapproaches tested in this work. Citizens, collectively, can enact meaningful change byacting as "ground agents" and providing valuable insightsdirectly from the front lines. In this regard, citizens’ insightsare a goldmine of data that organizations can use to maketheir cities smarter. The results presented in this paper holdthe potential to harness those opinions and give urban plan-ners and local authorities greater choice to identify, analyze,and improve the sentiment behind specific UGS, and allowUGS comparisons between cities worldwide. References [1] Al Amrani, Y., Lazaar, M., El Kadiri, K.E., 2018. Random forest andsupport vector machine based hybrid approach to sentiment analysis.Procedia Computer Science 127, 511 – 520. doi: https://doi.org/10.1016/j.procs.2018.01.150 . pROCEEDINGS OF THE FIRST INTER-NATIONAL CONFERENCE ON INTELLIGENT COMPUTING INDATA SCIENCES, ICDS2017.[2] Barton, J., Rogerson, M., 2017. The importance of greenspacefor mental health. BJPsych Int. 14, 79–81. doi: .[3] Blumenthal, M., 2014. Reviewer demographics-facebook has more women, yelp has more men.http://blumenthals.com/blog/2014/07/31/reviewer-demographics-facebook-has-more-women-yelp-has-more-men/.[4] Chen, Y., Lee, B., Kirk, R.M., . Internet use among older adults.Engaging Older Adults with Modern Technology , 124–141.[5] Chen, Y., Liu, X., Li, X., Liu, X., Yao, Y., Hu, G., Xu, X., Pei, F.,2017. Delineating urban functional areas with building-level socialmedia data: A dynamic time warping (dtw) distance based k -medoidsmethod. Landscape and Urban Planning 160, 48–60.[6] Cohen, D.A., Sehgal, A., Williamson, S., Marsh, T., Golinelli, D.,McKenzie, T.L., 2009. New recreational facilities for the young andthe old in los angeles: Policy and programming implications. Journalof Public Health Policy 30, S248–S263.[7] Colón-Ruiz, C., Segura-Bedmar, I., 2020. Comparing deep learn-ing architectures for sentiment analysis on drug reviews. Journal ofBiomedical Informatics 110, 103539. doi: https://doi.org/10.1016/j.jbi.2020.103539 .[8] Daniels, B., Zaunbrecher, B.S., Paas, B., Ottermanns, R., Ziefle, M.,Rob-Nickoll, M., 2018. Assessment of urban green space structuresand their quality from a multidimensional perspective. The Scienceof the Total Environment 615, 1364–1378.[9] Dong, J., 2020. Financial investor sentiment analysis based on fpgaand convolutional neural network. Microprocessors and Microsys-tems , 103418doi: https://doi.org/10.1016/j.micpro.2020.103418 .[10] Douglas, O., Lennon, M., Scott, M., 2017. Green space benefits forhealth and well-being: A life-course approach for urban planning, de-sign and management. Cities 66, 53–62.[11] Fang, F., McNeil, B., Warner, T., Dahle, G., Eutsler, E., 2020. Streettree health from space? an evaluation using worldview-3 data and thewashington dc street tree spatial database. Urban Forestry & UrbanGreening 49.[12] Fitri, V.A., Andreswari, R., Hasibuan, M.A., 2019. Sentiment analy-sis of social media twitter with case of anti-lgbt campaign in indone-sia using naïve bayes, decision tree, and random forest algorithm.Procedia Computer Science 161, 765 – 772. doi: https://doi.org/10.1016/j.procs.2019.11.181 . the Fifth Information Systems Interna-tional Conference, 23-24 July 2019, Surabaya, Indonesia.[13] Fuller, R.A., Gaston, K.J., 2009. The scaling of green space coveragein european cities. Biology Letters 5, 352–355.[14] Galle, N.J., Nitoslawski, S.A., Pilla, F., 2003. The internet of na-ture: How taking nature online can shape urban ecosystems. TheAnthropocene Review 6, 279–287. doi: https://doi.org/10.1177/2053019619877103 .[15] Garcia-Barriocanal, E., Sicilia, M., Korfiatis, N., 2010. Exploringhotel service quality experience indicators in user-generated content:
M. Ghahramani et al.:
Preprint submitted to Elsevier
Page 9 of 10everaging Artificial Intelligence to Analyze Citizens’ Opinions on Urban Green Space a case using tripadvisor data., in: Mediterranean Conference on In-formation Systems, p. 200–205.[16] Ghahramani, M., Qiao, Y., Zhou, M.C., O’Hagan, A., Sweeney, J.,2020. Ai-based modeling and data-driven evaluation for smart man-ufacturing processes. IEEE/CAA Journal of Automatica Sinica 7,1026–1037. doi: .[17] Ghahramani, M., Zhou, M., Wang, G., 2020. Urban sensing basedon mobile phone data: approaches, applications, and challenges.IEEE/CAA Journal of Automatica Sinica 7, 627–637.[18] Gidlow, C.J., Ellis, N.J., Bostock, S., 2012. Development of the neigh-bourhood green space tool (ngst). Landscape and Urban Planning 106,347–358.[19] Giezen, M., Balikci, S., Arundel, R., 2018. Using remote sensingto analyse net land-use change from conflicting sustainability poli-cies: The case of amsterdam. ISPRS International Journal of Geo-Information 7. doi: https://doi.org/10.3390/ijgi7090381 .[20] Girling, C., Kellett, R., 2005. Skinny streets and green neighbor-hoods: Design for environment and community. Island Press.[21] Grahn, P., Stigsdotter, U.K., 2010. The relation between perceivedsensory dimensions of urban green space and stress restoration. Land-scape and Urban Planning 94, 264–275.[22] Gretzel, U., 2007. Online travel review study: Role and impact ofonline travel reviews. Laboratory for Intelligent Systems in Tourism.[23] Haaland, C., van den Bosch, C.K., 2015. Challenges and strategiesfor urban green-space planning in cities undergoing densification: Areview. Urban Forestry & Urban Greening 14, 760–771. doi: https://doi.org/10.1016/j.ufug.2015.07.009 .[24] Hu, S., O’Hagan, A., Sweeney, J., Ghahramani, M., 2020. A spatialmachine learning model for analysing customers’ lapse behaviour inlife insurance. Annals of Actuarial Science doi: https://doi.org/10.1017/S1748499520000329 .[25] Hur, M., Nasar, J.L., Chun, B., 2010. Neighborhood satisfaction,physical and perceived naturalness and openness. Journal of Environ-mental Psychology 30, 52–59. doi: https://doi.org/10.1016/j.jenvp.2009.05.005 .[26] J., V.Z.B.T.S.C., Verburg P. H., T.K.F., 2018. Aesthetic appreciationof the cultural landscape through social media: An analysis of re-vealed preference in the dutch river landscape. Landscape and UrbanPlanning 177, 128–137.[27] Jin, N., Wu, J., Ma, X., Yan, K., Mo, Y., 2020a. Multi-task learningmodel based on multi-scale cnn and lstm for sentiment classification.IEEE Access 8, 77060–77072. doi: .[28] Jin, N., Wu, J., Ma, X., Yan, K., Mo, Y., 2020b. Multi-task learningmodel based on multi-scale cnn and lstm for sentiment classification.IEEE Access 8, 77060–77072. doi: .[29] Kourtit, Nijkamp, Romao., 2019. Cultural heritage appraisal by vis-itors to global cities: The use of social media and urban analytics inurban buzz research. Sustainability 11, 3470. doi: https://doi.org/10.3390/su11123470 .[30] Ma, E., Cheng, M., Hsiao, A., 2018. Sentiment analysis – a reviewand agenda for future research in hospitality contexts. InternationalJournal of Contemporary Hospitality Management 30, 3287–3308.doi: https://doi.org/10.1108/ijchm-10-2017-0704 .[31] Nasi, R., Honkavaara, E., Blomqvist, M., Lyytikäinen-Saarenmaa, P.,Hakala, T., Viljanen, N., Holopainen, M., 2018. Remote sensing ofbark beetle damage in urban forests at individual tree level using anovel hyperspectral camera from uav and aircraft. Urban Forestry &Urban Greening 30, 72–83.[32] Nitoslawski, S.A., Galle, N.J., Van Den Bosch, C.K., Steenberg,J.W.N., 2019. Smarter ecosystems for smarter cities? a review oftrends, technologies, and turning points for smart urban forestry. Sus-tainable Cities and Society 51, 101770. doi: https://doi.org/10.1016/j.scs.2019.101770 .[33] Nowak, D.J., Greenfield, E.J., 2018. Declining urban and communitytree cover in the united states. Urban Forestry & Urban Greening 32,32–55. doi: https://doi.org/10.1016/j.ufug.2018.03.006 .[34] Pak, B., Chua, A., Moere, A.V., 2017. Fixmystreet brussels: Socio-demographic inequality in crowdsourced civic participation. Journal of Urban Technology 24, 65–87.[35] Pampel, F.C., 2000. Logistic regression: A primer. SAGE.[36] Plunz, R.A., Zhou, Y., Vintimilla, M.I.C., Mckeown, K., Yu, T.,Uguccioni, L., Sutto, M.P., 2019. Twitter sentiment in new york cityparks as measure of well-being. Landscape and Urban Planning 189,235–246. doi: https://doi.org/10.1016/j.landurbplan.2019.04.024 .[37] Rahimi, S., Mottahedi, S., Liu, X., 2018. The geography of taste:Using yelp to study urban culture. ISPRS International Journal ofGeo-Information 7, 376. doi: https://doi.org/10.3390/ijgi7090376 .[38] Rakhmanov, O., 2020. A comparative study on vectorization and clas-sification techniques in sentiment analysis to classify student-lecturercomments. Procedia Computer Science 178, 194 – 204. doi: https://doi.org/10.1016/j.procs.2020.11.021 . 9th International Young Scien-tists Conference in Computational Science, YSC2020, 05-12 Septem-ber 2020.[39] Roberts, H., Sadler, J., Chapman, L., 2017. Using twitter to investi-gate seasonal variation in physical activity in urban green space. Geo:Geography and Environment 4.[40] Roe, J.J., Thompson, C.W., Aspinall, P.A., Brewer, M.J., Duff, E.I.,Miller, D., Mitchell, R., Clow, A., 2018. Green space and stress:evidence from cortisol measures in deprived urban communities. In-ternational Journal of Environmental Research and Public Health 10,4086–4103.[41] Rokach, L., Maimon, O., 2005. Top-down induction of decision treesclassifiers—a survey. IEEE Transactions on Systems, Man and Cy-bernetics, Part C (Applications and Reviews) 35, 476–487.[42] Seiferling, I., Naik, N., Ratti, C., Proulx, R., 2017. Green streets-quantifying and mapping urban trees with street-level imagery andcomputer vision. Landscape and Urban Planning 165, 93–101.[43] Swanwick, C., Dunnett, N., Woolley, H., 2003. Nature, role and valueof green space in towns and cities: An overview. Built Environment29, 94–106. doi: https://doi.org/10.1016/j.cities.2017.03.011 .[44] Tra, V., Duong, B., Kim, J., 2019. Improving diagnostic performanceof a power transformer using an adaptive over-sampling method forimbalanced data. IEEE Transactions on Dielectrics and ElectricalInsulation 26, 1325–1333. doi: .[45] Yahav, I., Shehory, O., Schwartz, D., 2019. Comments mining with tf-idf: The inherent bias and its removal. IEEE Transactions on Knowl-edge and Data Engineering 31, 437–450. doi: https://doi.org/10.1109/tkde.2018.2840127 .[46] Yang, X., Kuang, Q., Zhang, W., Zhang, G., 2018. Amdo: Anover-sampling technique for multi-class imbalanced problems. IEEETransactions on Knowledge and Data Engineering 30, 1672–1685.doi: .[47] Yao, F., Wang, Y., 2020. Domain-specific sentiment analysis fortweets during hurricanes (dssa-h): A domain-adversarial neural-network-based approach. Computers, Environment and Urban Sys-tems 83, 101522. doi: https://doi.org/10.1016/j.compenvurbsys.2020.101522 .[48] Zhang, Y., Van den Berg, A.E., Van Dijk, T., Weitkamp, G., 2017.Quality over quantity: Contribution of urban green space to neighbor-hood satisfaction. International Journal of Environmental Researchand Public Health 14. doi: https://doi.org/10.3390/ijerph14050535 . M. Ghahramani et al.: