Identifying salient topics for personalized place similarity

Abstract

The ability to find similar places is an important component to geographic information retrieval applications as varied as travel recommendation services, marketing analysis tools, and socio-ecological research. Using generative topic modelling on a large collection of place descriptions, we can represent places as distributions over thematic topics, and quantitatively measure similarity for places modelled with these topic signatures. However, existing similarity measures are context independent; in cognitive science research there exists evidence that when people perform similarity judgments they will weigh properties differently depending on personal context. In this paper we present a novel method to re-weight the topics that are broadly associated with a place, based on users' interests inferred from sample place similarity rankings. We evaluate the method by training topics associated with texts about places, and perform a user study that compares user-provided similar places to those provided by automatically personalised place rankings. The results demonstrate improved correspondence between user rankings and automated rankings when personalised weights are applied.

Introduction

as mixtures of topics derived from probabilistic topic modelling on a corpus of place descriptions. These descriptions can take many forms, including travel entries, encyclopaedia articles, newspaper articles, and social media postings. These data sources are plentiful and all provide insight into the experiences of people in various places and, in particular, attributes that are commonly described for those places. Topics are trained using latent Dirichlet allocation (LDA) topic model inferencing on a set of these crowdsourced documents (Blei et al., 2003) . Topic model training applies Bayesian inference to arrive at a representation of each place as a probability vector over a finite set of topics. Once we have a topic distribution for each place, we can compare their similarities in terms of the relative entropy of these topic distributions. This measure captures the similarity of the places as derived from an aggregation of descriptions from many people. In other words, the similarity of two places is defined as similarity of the prototypical or “average” representations of the places. There is psychological evidence that for many kinds of categories a prototypical instance of the category can be represented as an average from a set of exemplars in this manner (Rosch, 1978) .

However, there is also evidence that when performing similarity (or dissimilarity) judgements on a set of stimuli, people will assign higher salience to some properties than others (Medin et al., 1993) . In conceptual space theory, it is proposed that these salience differences can be modelled as weights on the quality dimensions on which these similarity measurements are based (Ga¨rdenfors, 2000; Raubal, 2004) . When people judge the similarity of two places, they choose a set of salient properties on which to compare the two places. This set of properties will be smaller than the total set of possible properties on which one could possibly compare the two places (Tversky, 1977) . In addition, the set of properties that are salient will vary from person to person and depend on the places being compared. The challenge, therefore, is to identify which weighting to use in a particular context. This feature selection problem becomes one of further reducing the dimensionality of the topic space to a subset of topics that are relevant to a user in a specific context.

In this paper we present a methodology to 1) generate a prototypical mixture of topics associated with places based on crowdsourced natural language descriptions, 2) identify the topics of interest to an individual user, and 3) provide a personalised ranking of similar places to a source place based on the inferred interests of the user. We propose a method to automatically re-weight the topics using a sample similarity ranking for a small subset of the data.

In the following section, we provide background on computing place, personalised search and ranking, and topic models. Section 3 describes our method to identify salient topics. Following that, Section 4 presents the results of a user evaluation of the method. Finally, we conclude and point to future work. 2

Background

2.1

Place This section provides background material on computing place, personalised search and ranking, place recommendation, and topic modelling algorithms used in this paper.

Several researchers have noted that place is a subjective, experiential phenomenon that is socially constructed (Relph, 1976; Tuan, 1977) . Simply put several of the characteristics of places that are salient to people cannot be captured by a spatial footprint and structured data, such as population or median income. The difficulty in developing place representations that take into account common-sense understandings has meant that, to a great degree, place has been represented in a limited manner in geographic information systems (GISs). However, there has been a renewed interest in recent years in modelling place more holistically as it becomes clear that traditional models are insufficient for the needs of many technologies (Jordan et al., 1998; Winter et al., 2009) . An increasingly vast amount of data is available online that describe places and types of places in natural language. They come in a variety of forms: e.g., travelogues, encyclopaedia entries, microblogs, literary works, social media sites, etc. These documents provide us with an extraordinary opportunity to understand places not in terms of quantitative, tabular data, but rather in terms of how people write about them. 2.2

Personalised search and ranking Personalised search and ranking on the World Wide Web has been an ongoing research area for several years (Pitkow et al., 2002; Teevan et al., 2005; Micarelli et al., 2007) . Pitkow et al. (Pitkow et al., 2002) have identified two main strategies for personalising search: 1) query augmentation, which uses additional contextual information to refine the query; and 2) result processing, which filters the search results based on a user model. In addition to algorithmic approaches to personalisation, user interface design can play a large role in how well search results can be personalised (Teevan et al., 2005) .

Much of the personalisation research relies on using previous search behaviour to develop a model of the user’s interests and using this model to personalise search results on the web (Teevan et al., 2005) . Developing a user profile based on implicit information has several advantages over asking the user for explicit feedback, because not only does gathering explicit feedback impose an added burden on the user but also users can provide inaccurate information (Speretta and Gauch, 2005) . In mobile search, user-based personalisation can be augmented by location-based personalisation and these factors can interrelate to model how a user’s interests change based on location (Bouidghaghen et al., 2011) . Formal ontologies can be used to align user interests that are inferred from search history with well-defined concepts described in a semantic hierarchy (Daoud et al., 2007) .

Once there is a model that associates user categories with specific interests, it is possible to use different methods of characterising user similarity to provide personalised results for categories of users (McKenzie et al., 2013) . Guy et al. (Guy et al., 2010) have explored comparing user similarity in social media along 3 different axes of people, things, and places. 2.3

Semantic similarity in GIScience The importance of context in semantic similarity measurement is a well-studied research topic in geographic information science (Raubal, 2004; Rodr´ıguez and Egenhofer, 2004; Janowicz et al., 2011) . Commonly, context is modelled by applying weights to the attributes that used to describe geospatial features; however, finding the appropriate weights is a challenge ( Keßler, 2012 ). Semantic referencing is an algorithm schema that uses the idea of control similarities provided by a user to semi-automatically calibrate factor weights for similarity measurement of geographic entities (Janowicz et al., 2010) . The methods presented here are a form of semantic referencing.

The problem addressed in this paper (personalised similarity of places) is distinct from previous work on personalised web ranking, since we focus on the specific problem of finding similar places. Although individuals have different impressions of places, due to the nature of places being manifested in the physical world, there are broad regularities in the types of activities and features that people write about a given place. Because of these regularities, a crowdsourced representation of a place will have broad applicability. Thus, the search for places is not an information retrieval task on an individual document or artefact but rather a similarity search for places, represented as aggregations derived from multiple documents. Personalised place similarity remains a relatively unexplored research area outside of smaller-scale investigations in tourism and sense-of-place research (Moore and Graefe, 1994; Kaltenborn, 1998; Jorgensen and Stedman, 2006) and sociological investigations of place attachment (Stedman, 2002) .

As with prior user behaviour for web personalisation, examples of similar places provide a mechanism to infer the interests of the user without requiring him or her to explicitly enter interests. Furthermore, the sample ranking of similar places – and which is provided directly by the user in this evaluation – can also be gathered from social data. 2.4

Topic models Probabilistic topic models, such as latent Dirichlet allocation (LDA), provide methods to characterise the documents in a natural language corpus as multinomial mixtures of topics (Blei et al., 2003) . Each topic is further modelled as a discrete probability distribution over all the words in a corpus. The top most probable words for a topic are usually semantically interpretable by humans. E.g., a topic with the top terms “wine, red, drink, cheese” can be interpreted as being about the activity of wine tasting (Adams and McKenzie, 2013) .

LDA models the generation of each document in the corpus as an iterative process; first, a topic is chosen based on the topic distribution for the document; second, a word is randomly selected from that topic. This is repeated for each term in the document. Since each word is drawn randomly and is not based on the previous word, word order makes no difference in the LDA model. Assuming this generative process for how the documents were created, the words of an existing set of documents become the observed variables in the model and Bayesian inference is used to infer the most-likely topics to have generated the data. This inference can be performed programmatically in an approximate manner using a markov chain Monte Carlo Gibbs sampling algorithm (Griffiths and Steyvers, 2004) .

Apart from discovering the latent topics in a corpus in an unsupervised manner, this dimensionality reduction allows one to compare the thematic similarity of two documents without requiring that the documents share exact terms. A natural way to measure the similarity of documents is the calculate the Kullback-Leibler (KL) divergence between the two topic distributions, P and Q (see Equation 1) (Steyvers and Griffiths, 2007) . (1) (2) DKL(P k Q) =

X ln i where M = (P + Q)=2. These methods of measuring similarity ignore that the importance of individual topics will vary from user to user.

Several extensions to LDA have been developed to characterise the mixture of topics associated with a location or place (Wang et al., 2007; Eisenstein et al., 2010; Hao et al., 2010; Sizov, 2010; Yin et al., 2011; Hong et al., 2012) . In these models the generative process for creating a document is extended and some additional form of geographic evidence (either location or place name) is used to condition topics based on that evidence. Thus, a distribution of topics can be associated not only with a document but also a place or location. The technique employed here for evaluation is to represent the topic mixture for a place as a weighted average of the topic probability distributions of all the descriptions for the same place (Adams and Janowicz, 2012) . Since the methods described in this paper for weighting topics are post-hoc operations (i.e. assigned after the topic model inferencing), they are broadly applicable to these other flavours of topic models.

3 Identifying salient topics

In this section, we present a method for personalising the weights on the LDA topics, given a source place s and a set of N target places ft1; t2; : : : ; tN g, and a user specified ordering for those target places based on their similarity to the source place. For each place there is a discrete probability distribution for topics, and for any given individual topic the strengths of that topic for both source and target places can be compared. For example, let the following be a sample user ranking of three cities in terms of similarity to New York City:

1. Chicago

2. Los Angeles 3. Houston

The set of salient topics ZS is defined as a subset of all topics, such that z 2 ZS if and only if the Kendall’s rank correlation coefficient between the user-provided ordering and the topic strengths for topic z is positive (Kendall, 1938) . The Kendall’s measures rank correlation as a function of the number of concordant and discordant pairs in a set of joint observations. A pair of observations (i; j) of two variables (x; y) is concordant if xi < yi ^ xj < yj or xi > yi ^ xj > yj . They are discordant if xi < yi ^ xj > yj or xi > yi ^ xj < yj . Letting C be the number of concordant pairs and D be the number of discordant pairs in a sample and n be the number of observations, Kendall’s is defined in Equation 3.

Thus, the relative weight ( i) for each topic is zero for negative correlations. For positive correlations the weight is the correlation normalised such that all positive correlations sum to 1. See Table 2 for an example.

Kendall’s -B is an extension to Kendall’s to deal with situations where there are tied rankings. This will occur with topics when one or more documents have a value of 0:0 for the topic strength. When a model is trained using a large user ordering

Chicago Los Angeles

Houston

Kendall’s Relative weight

topic 1

Chicago (0.0) Los Angeles (0.22)

Houston (0.6) 1.0 0.75

topic 2 Los Angeles (0.22)

Chicago (0.4) Houston (0.5) 0.33 0.25

topic 3 Los Angeles (0.0)

Houston (0.1) Chicago (0.4) -0.67 0.0 number of topics this will occur often. Kendall’s -B (Equation 4) handles tied values by changing the denominator of the equation.

B = p(n0 nc

nd n1)(n0 n2) ; where n0 = n(n 1)=2; n1 = Pi ti(ti 1)=2; n2 = Pj uj (uj 1)=2; nc = number of concordant pairs; nd = number of discordant pairs; ti = number of tied values in the ith group of ties for the first quantity; and uj = number of tied values in the jth group of ties for the second quantity. 3.1

Using Kendall’s

as product For each salient topic, zi, an associated salience weight, wi, is assigned to the topic equal to the Kendall’s correlation. Given a set of salience weights w1; w2; : : : wn we can define relative weights for each topic 1; 2; : : : n, where i = Pw1niw . Let be this vector of relative weights, a weighted Kullback-Liebler divergence A is now defined in Equation 5. (4) (5) DKL( ; P k Q) = X iP (i) ln i

The divergence is weighted not only on the probabilities of the topic variables but also with the salience weight based on the ordering, and it is thus a weighted average of the logarithm difference between the probabilities. From this new divergence function, the weighted Jensen Shannon divergence A is defined in Equation 6.

1 1

J SD( ; P k Q) = 2 DKL( ; P k M ) + 2 DKL( ; Q k M ); (6) where M is defined as in the normal JS divergence.

This weighted measure can be interpreted as a form of weighted sum model in a multi-criteria decision analysis.The salience weights drawn from a sample ranking can be viewed as a personalised general measure of topic saliency for comparing all places. Table 3 shows a comparison of Jensen Shannon divergence and weighted Jensen Shannon divergence A calculations given the user ranking specified above. Note, in this example the weighted ordering matches the user specified ordering; however, it is not guaranteed to match in all cases, since the salience weights are based on a comparison of rankings, not actual similarity values of the probabilities. 3.2

Using Kendall’s to alter topic weights An alternative method re-weights the topic probabilities associated with each place and uses the traditional JS divergence measure to obtain a context-dependent similarity. As with the previous method only positive correlations are used. Letting z be the topic vector for a place, Algorithm 1 describes the steps for re-weighting topics. Each topic probability, zi, is multiplied by the weight for that topic, i, and then the new topic vector is normalized to sum to 1 (see Algorithm 2). Algorithm 1 Pseudocode for re-weighting topics from user provided ranking.

sum 0.0 for all zi do if wi > 0 then zi0 wi zi sum sum +zi0 else

zi0 0:0 end if end for for all zi0 do

zi0 zi0= sum end for

This new topic distribution for a place has a very intuitive interpretation. The re-weighted probability of topic i, zi0, is a function of the prototypical probability of the topic, zi, for the place, conditioned on the probability that this specific user is concordant with the average user for this topic. This re-weighted distribution has the advantage that it maintains the desired mathematical properties when comparing similarities using KL and JS divergences (i.e., a minimum divergence of 0 and the square root of the JS divergence is a metric). In Table 3 the Jensen Shannon result using this method is called the weighted Jensen Shannon B measure.

city

Chicago

Los Angeles

Houston

JS In this section, we present the results of a human participants test to evaluate the efficacy of using sample rankings to prime salience weights on individual topic dimensions. LDA topic models were trained on two corpora of georeferenced place descriptions: 1) a set of 200,000 georeferenced Wikipedia articles and 2) a set of 275,000 travel blog entries.1 Prior to performing topic modelling, standard stemming and cleaning of the documents was performed to remove html tags and other noise in the text. The travel blog entries were matched to places by the authors according to a fixed geographic hierarchy, e.g., Orlando, Florida, United States. Each georeferenced Wikipedia article was matched to a named place by intersecting the location associated with the article (based on coordinates template in the article) with the spatial footprint of the named places. Thus, each article is linked to one place. It is possible that the association between text and a place can be refined further using more sophisticated techniques; however, this still remains a open research problem and not the focus of the current work (Vasardani et al., 2013) .

The topic signature for a place was calculated by taking all the documents with a relation to the place and averaging over their topic vectors. Thus, topics that have a relatively high probability in a large number of documents associated with a place will be associated with the place (e.g., in travel blog entries a theme park topic will have high value for Orlando, FL, but a skiing topic will not). 4.1

Design The user study was designed using the Amazon Mechanical Turk system to gather several place similarity rankings.2 A qualification test was presented to the Mechanical Turk users to help ensure high-quality (non-spam) results (Kittur et al., 2008) . This test was a simple filter that gathers user information and asks a set of simple questions about geographic knowledge. The primary goal of the survey was to determine the degree to which a user-provided ranking on a set of cities can be used to inform the system on the “interests” of the user more generally, and thus tailor search results.

The study focused on comparing 30 U.S. cities rather than other types of places in order to maximise participant familiarity. Although initial variants of the test were done using cities from around the world, it was difficult to find users who had broad familiarity with multiple cities from around the world. Therefore, the tests were limited to residents of the U.S., and a good sample of participants with familiarity of target cities could be gathered. Table 4 shows all the cities that were compared.

Mechanical Turk participants were asked to rank order 7 cities in terms of similarity to another city. Users were also asked to enter how the places are similar as well as describe their familiarity with each of the 8 cities. Table 5 shows two sample responses for cities that are ranked in similarity to Los Angeles.

In total 96 participants provided 5 rankings each, leading to 480 rankings total. Out of the 480 rankings, in 135 cases (28.1%) the participant was ‘not’ familiar with the main city being compared, in 205 cases (42.7%) the participant was 1downloaded from http://www.travelblog.org 2http://www.mturk.com Algorithm 2 Pseudocode for re-weighting topics from user provided ranking when not all topics are informed by the ranking.

sum 0.0 sumOthers 0.0 for all zi do if zi 2 ZNI then zi0 zi sumOthers sumOthers +zi else if wi > 0 then zi0 wi zi sum sum +zi0 else

zi0 0:0 end if end for for all zi0 do if zi0 2= ZNI then

zi0 zi0 (1 sumOthers )= sum end if end for

Atlanta

Austin Baltimore Boston Charlotte Chicago Cleveland Dallas Denver Detroit

Houston

Kansas City Las Vegas Los Angeles Miami

Minn.-St. Paul

Nashville New Orleans New York City Philadelphia

Phoenix

Pittsburgh Portland, OR Salt Lake City San Diego

San Francisco San Jose Seattle St. Louis

Washington D.C. ‘somewhat’ familiar, and in 140 cases (29.2%) the participant was ‘very’ familiar. In 50 cases (10.4%) the participant had lived in (or near) the city and in 290 cases (60.4%) the participant had visited the city. 4.2

Spearman’s footrule distance Spearman’s footrule distance is a measure of correspondence (or disarray) between two rankings, similar to Spearman’s and Kendall’s . Let Sn be the set of all permutations of the set of n integers f1; : : : ; ng. From (Diaconis and Graham, 1977) , the Spearman’s footrule distance, Dn is defined as in Equation 7.

n Dn( n; n) = X j n(i) i=1 n(i)j ; (7) where n and n are elements of Sn, i.e., different permutations. Spearman’s footrule is used here rather than Spearman’s rho, because we want to be able to consider the average footrule distance between several automated rankings and user provided rankings, which is not valid for correlations. 4.3

Average Spearman’s footrule results In Figure 1 a chart is shown of the average Spearman’s footrule distance between the human participant rankings and automated similarity calculations. Looking over all responses, the clearest outcome was that for source places with which the participant is not familiar; generally, a participant’s assessment is primarily made based on distance rather than other properties. When the participant is somewhat or very familiar with the source place there is an increase in the average Spearman’s footrule distance between distance-based rankings.

Looking at Figure 1 it is abundantly clear that similarity ranking based on all travel blog topics or all Wikipedia articles does not correspond in any definitive way with user-provided rankings. In part this is because there is no “average person” – there is very little correspondence between different participants about the appropriate ranking of similar cities. In Miami Minneapolis-St. Paul Dallas Austin Salt Lake City Cleveland Portland, OR

Miami

Dallas Austin Portland, OR Minneapolis-St. Paul Salt Lake City Cleveland addition, individuals have a limited set of properties that they use to compare places as opposed to the large number of different topics found in large text corpora. This is not a problem with the methodology, however, because the task that the system is to perform is not necessarily to match individual human reasoning but rather to allow exposure of patterns and similarities in place description topics that are otherwise opaque to an individual user (or human participant). 4.4

Automatic topic weighting evaluation Despite this lack of correspondence, we can use the participant studies to evaluate the degree to which it is valid to assume a user has a set of ‘interests’ (i.e., a salient subset of topics) that are common across different place comparisons and whether we can identify those interests in terms of topics automatically from one or more sample rankings. For example, looking back at Table 5, these two participants consider very different properties when comparing the cities for similarity. Note, however that these participant-provided properties do not necessarily reflect all the properties used by the participant to discriminate between places – e.g., it does not explain the particular ordering of Austin, Portland, and Minneapolis-St. Paul by participant 2, because the same property (“music and performing arts”) is given for all of them.

In order to do this evaluation we test whether on average automatic topic re-weighting using the techniques in the previous section result in smaller Spearman’s footrule distances for the other rankings by the same participant. In other words, we evaluate the degree to which a single ranking can be extrapolated more generally to other rankings of cities. One sample ranking from the participant is used to calculate weights on the topics and new city rankings are calculated using the weighted JS divergence for each of the other city comparisons done by the same participant. For each ranking the Spearman’s footrule distance is calculated between the new weighted ranking and the participant-provided ranking. This process is then repeated for each sample ranking by exchanging which one sets the weights. From this cross-validation an average footrule distance can be calculated for the weighted ranking for a participant, and this can be further averaged over all participants. Figure 1 shows this average footrule distance over all participants when using the weighted JS divergence B technique. The weighted JS divergence A has a similar result, though the overall average footrule distance tends to be slightly higher.

The average Spearman’s footrule distance shows a marked decrease over the unweighted results for Wikipedia and travel blogs that were shown in Figure 1, which indicates that this method does help to identify general topics of interest for a person. One very interesting result is that the matching to the participant rankings is distinctly better when the participant is not very familiar with the source city being ranked against. One explanation for this result is that people who are very familiar with a place will have very specific impressions of the place and thus properties that are salient for that place do not transfer over from other places. Whereas when people are not familiar with a place there is a stock set of “background interests” that they use to help compare places. This indicates that automatic topic weighting will be particularly useful when done as part of a system designed to enable users to learn and explore knowledge about geographic places with which they are not already very familiar. Weighted  Wiki  topics   Weighted  T.B.  topics  

Distance  

Popula;on   Wikipedia  topics   Travel  blog  topics   Very   Somewhat   Not   0   From these results, the automatic weighting of topics based on a sample ranking can legitimately be generalised to other rankings. An interesting alternative would be to extract interests of a user implicitly from social network or other data and apply those to topic weights, but that remains beyond the scope of the current research (Kelly and Teevan, 2003; Ricci et al., 2011) . On a case-by-case basis we also examined whether the reasons participants give for similarity correspond to high shared values on specific LDA topics, but apart from anecdotal evidence, it was difficult to do a quantitative evaluation on these data, because the mapping of participant-provided reasons and topics is highly subjective.

Table 6 shows the topics in Wikipedia (from 1200 topics) and travel blogs (from 1500 topics) for Los Angeles that increase the most (by ratio) based on re-weightings from the two sample participant rankings shown in Table 5. What is remarkable about the topics shown in Table 6 is that despite the apparent difference between the automatically-determined most-salient topics and the participant-provided reasons, the automatic weighting does significantly increase the concordance between the system rankings and the participant rankings. One explanation is that the topics that are weighted with high salience represent background properties that factor in the participants’ conceptualisations of city categories but which are not at the forefront of their conscious comparisons of individual cities. 5

Conclusion

People judge the similarity or difference of places based on different contextual factors, such as their personal interests. Geographic information systems that take advantage of these factors and can provide places that are similar to places known to a user have many potential applications, including travel recommendation services, marketing analysis tools, and socio-ecological research tools. In this paper we presented a new method to automatically identify the topics that are salient to a user when performing similarity judgment. Topics can be any set of features associated with a place, such that a place is represented as a probability vector of topic values. These topic values form a topic signature that is generally associated with a place, and we demonstrated how probabilistic topic modelling can be used to generate such probability vectors for places. Topic 827 Topic 1242 Topic 904

Topic 761 Topic 834 Topic 890

Topic 909 Topic 793 Topic 1228

Topic 417 Topic 164 Topic 847

Similar to Los Angeles ranking 1 – Travel blog

game,basebal,team,play,watch,sport,hockey,fan,stadium,player restaur,order,food,tabl,meal,menu,eat,waiter,serv,dinner money,pay,cost,expens,price,onli,cheap,pound,buy,free

Similar to Los Angeles ranking 1 – Wikipedia

book,work,publish,life,wrote,histori,year,mani,author,writer cathol,bishop,dioces,roman cathol,roman,john,father,priest golf,cour,club,golf club,hole,countri club,countri,locat,golf cours

Similar to Los Angeles ranking 2 – Travel blog

mile,road,stop,highway,gas,motel,state,sign,rout,trip extrem,complet,entir,exact,time,veri,actual,howev,made,ani histor,site,build,visit,histori,tour,histor site,area,museum,mani

Similar to Los Angeles ranking 2 – Wikipedia side,east,east side,west,north,south,locat,north side,west side polic,offic,polic offic,polic station,polic depart,depart,enforc,forc turkish,ottoman,turkey,greek,byzantin,ottoman empir,sultan

Similarity calculations based on probability distributions are commonly context-neutral and only measure the relative entropy or JS divergence of the distributions. We presented a novel approach to use a small, user-provided sample ranking of similar places to automatically re-weight topic weights. This new re-weighting can be used to serve up personalised similarity values between places based on the topics that are of most interest to a user. Since this re-weighting of topic values is done after the initial training used to discover topics associated with places, it is fully compatible with a variety of different methods to create semantic signatures associated with a place, not solely topic modelling as used here.

We evaluated our method with a Mechanical Turk user study and showed that using a small sample ranking of similar places (i.e., a set of control similarities) results in a larger correspondence between automated place similarity rankings and personal users’ rankings. Personalised place recommendation and similarity search remains a relatively unexplored research area. A follow-up study to better understand user motivation in performing place similarity will be a valuable next step. Future work will also involve exploring social media data and other sources such as search history to automatically provide sample similar places for a given user.

Proceedings of the National Academy of Sci

References

Adams , B. and K. Janowicz ( 2012 ). On the geo-indicativeness of non-georeferenced text . In J. G. Breslin, N. B. Ellison , J. G. Shanahan , and Z. Tufekci (Eds.), ICWSM, pp. 375 - 378 . The AAAI Press.

Adams , B. and G. McKenzie ( 2013 ). Inferring thematic places from spatially referenced natural language descriptions . In D. Sui, S. Elwood , and M. Goodchild (Eds.), Crowdsourcing Geographic Knowledge , pp. 201 - 221 . Springer Netherlands.

Banerjee , S. S. and R. R. Dholakia ( 2008 ). Mobile advertising: does location-based advertising work ? International Journal of Mobile Marketing 3 ( 2 ), 68 - 75 .

Blei , D. M. , A. Y. Ng , and M. I. Jordan ( 2003 ). Latent Dirichlet allocation . Journal of Machine Learning Research 3 , 993 - 1022 .

Bodenhamer , D. J. ( 2010 ). The potential of spatial humanities . In D. J. Bodenhamer , J. Corrigan , and T. M. Harris (Eds.), The Spatial Humanities: GIS and the Future of Humanities Scholarship , pp. 14 - 30 . Indiana University Press.

Bouidghaghen , O. , L. Tamine , and M. Boughanem ( 2011 ). Personalizing mobile web search for location sensitive queries . In A. B. Zaslavsky , P. K. Chrysanthis , D. L. Lee , D. Chakraborty , V. Kalogeraki , M. F. Mokbel , and C.-Y. Chow (Eds.), Mobile Data Management (1) , pp. 110 - 118 . IEEE.

Daoud , M. , L. Tamine , M. Boughanem , and B. Chebaro ( 2007 ). Learning implicit user interests using ontology and search history for personalization . In Proceedings of the 2007 international conference on Web information systems engineering , WISE' 07 , Berlin, Heidelberg, pp. 325 - 336 . Springer-Verlag.

Diaconis , P. and R. L. Graham ( 1977 ). Spearman's footrule as a measure of disarray . Journal of the Royal Statistical Society. Series B (Methodological) 39 ( 2 ), 262 - 268 .

Eisenstein , J. , B. O 'Connor , N. A. Smith , and E. P. Xing ( 2010 ). A latent variable model for geographic lexical variation . In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing , EMNLP '10, Stroudsburg , PA, USA, pp. 1277 - 1287 . Association for Computational Linguistics.

Ga ¨rdenfors, P. ( 2000 ). Conceptual Spaces: The Geometry of Thought. A Bradford Book . MIT Press.

Griffiths , T. L. and M. Steyvers ( 2004 ). Finding scientific topics . ences 101(Suppl. 1) , 5228 - 5235 .

Grossner , K. E. , M. F. Goodchild , and K. C. Clarke ( 2008 ). Defining a digital earth system . Transactions in GIS 12(1) , 145 - 160 .

Guy , I. , M. Jacovi , A. Perer , I. Ronen , and E. Uziel ( 2010 ). Same places, same things, same people?: mining user similarity on social media . In Proceedings of the 2010 ACM conference on Computer supported cooperative work , CSCW '10 , New York, NY, USA, pp. 41 - 50 . ACM.

Hao , Q. , R. Cai , C. Wang , R. Xiao , J.-M. Yang , Y. Pang , and L. Z. 0001 ( 2010 ). Equip tourists with knowledge mined from travelogues . In M. Rappa, P. Jones , J. Freire , and S. Chakrabarti (Eds.), WWW, pp. 401 - 410 . ACM.

Hong , L. , A. Ahmed , S. Gurumurthy , A. J. Smola , and K. Tsioutsiouliklis ( 2012 ). Discovering geographical topics in the twitter stream . In A. Mille , F. L. Gandon , J. Misselis , M. Rabinovich , and S. Staab (Eds.), WWW, pp. 769 - 778 . ACM.

Janowicz , K. , B. Adams , and M. Raubal ( 2010 ). Semantic referencing - determining context weights for similarity measurement . In S. I. Fabrikant, T. Reichenbacher , M. J. van Kreveld , and C. Schlieder (Eds.), GIScience , Volume 6292 of Lecture Notes in Computer Science, pp. 70 - 84 . Springer.

Janowicz , K. , M. Raubal , and W. Kuhn ( 2011 ). The semantics of similarity in geographic information retrieval . Journal of Spatial Information Science (2) , 29 - 57 .

Jones , C. B. and R. S. Purves ( 2008 ). Geographical information retrieval . International Journal of Geographical Information Science 22 ( 3 ), 219 - 228 .

Jordan , T. , M. Raubal , B. Gartrell , and M. J. Egenhofer ( 1998 ). An affordance-based model of place in GIS . In T. Poiker and N. Chrisman (Eds.), Proceedings of the 8th International Symposium on Spatial Data Handling (SDH'98) , Vancouver, Canada, pp. 98 - 109 .

Jorgensen , B. S. and R. C. Stedman ( 2006 ). A comparative analysis of predictors of sense of place dimensions: Attachment to, dependence on, and identification with lakeshore properties . Journal of Environmental Management 79 , 316 - 327 .

Kaltenborn , B. P. ( 1998 ). Effects of sense of place on responses to environmental impacts: A study among residents in svalbard in the Norwegian high Arctic . Applied Geography 2 ( 18 ), 169 - 189 .

Kelly , D. and J . Teevan ( 2003 ). Implicit feedback for inferring user preference: a bibliography . SIGIR Forum 37 , 18 - 28 .

Kendall , M. G. ( 1938 ). A new measure of rank correlation . Biometrika 30 ( 1 /2), pp. 81 - 93 .

Keßler , C. ( 2012 ). What is the difference? a cognitive dissimilarity measure for information retrieval result sets . Knowl. Inf. Syst . 30 ( 2 ), 319 - 340 .

Kittur , A. , E. H. Chi , and B. Suh ( 2008 ). Crowdsourcing user studies with mechanical turk . In CHI '08: Proceeding of the twenty-sixth annual SIGCHI conference on Human factors in computing systems , New York, NY, USA, pp. 453 - 456 . ACM.

Kumar , R. and S. Vassilvitskii ( 2010 ). Generalized distances between rankings . In M. Rappa, P. Jones , J. Freire , and S. Chakrabarti (Eds.), WWW, pp. 571 - 580 . ACM.

Larson , R. ( 1996 ). Geographic information retrieval and spatial browsing . In L. C. Smith and M. Gluck (Eds.), Geographic Information Systems and Libraries: Patrons, Maps, and Spatial Information , pp. 81 - 124 .

McKenzie , G. , B. Adams , and K. Janowicz ( 2013 ). A thematic approach to user similarity built on geosocial check-ins . In D. Vandenbroucke, B. Bucher , and J. Crompvoets (Eds.), Geographic Information Science at the Heart of Europe, Lecture Notes in Geoinformation and Cartography , pp. 39 - 53 . Springer International Publishing.

Medin , D. L. , R. L. Goldstone , and D. Gentner ( 1993 ). Respects for similarity . Psychological Review 100 ( 2 ), 254 - 278 .

Micarelli , A. , F. Gasparetti , F. Sciarrone , and S. Gauch ( 2007 ). The adaptive web . Chapter Personalized search on the world wide web , pp. 195 - 230 . Berlin, Heidelberg: Springer-Verlag.

Moore , R. L. and A. R. Graefe ( 1994 ). Attachments to recreation settings: The case of rail-trail users . Leisure Sciences 16 , 17 - 31 .

Pitkow , J. , H. Schu¨tze, T. Cass, R. Cooley , D. Turnbull , A. Edmonds , E. Adar, and T. Breuel ( 2002 ). Personalized search . Commun. ACM 45 ( 9 ), 50 - 55 .

Ranganathan , A. and R. H. Campbell ( 2002 ). Advertising in a pervasive computing environment . In M. S. Viveros, H. Lei , and O. Wolfson (Eds.), Workshop Mobile Commerce, pp. 10 - 14 . ACM.

Raubal , M. ( 2004 ). Formalizing conceptual spaces . In A. C. Varzi and L. Vieu (Eds.), Formal Ontology in Information Systems, Proceedings of the Third International Conference (FOIS 2004 ), pp. 153 - 164 . IOS Press.

Relph , E. ( 1976 ). Place and Placelessness . Pion.

Ricci , F. , L. Rokach , and B. Shapira ( 2011 ). Introduction to recommender systems handbook . In F. Ricci, L. Rokach , B. Shapira , and P. B. Kantor (Eds.), Recommender Systems Handbook , pp. 1 - 35 . Springer.

Rodr´ıguez, M. A . and M. J. Egenhofer ( 2004 ). Comparing geospatial entity classes: an asymmetric and context-dependent similarity measure . International Journal of Geographical Information Science 18 ( 3 ), 229 - 256 .

Rosch , E. ( 1978 ). Principles of categorization . In E. Rosch and B. B. Lloyd (Eds.), Cognition and Categorization , pp. 27 - 48 . Hillsdale (NJ), USA: Lawrence Erlbaum Associates.

Sizov , S. ( 2010 ). Geofolk: latent spatial semantics in web 2.0 social media . In B. D. Davison , T. Suel , N. Craswell , and B. Liu (Eds.), WSDM, pp. 281 - 290 . ACM.

Speretta , M. and S. Gauch ( 2005 ). Personalized search based on user search histories . In Web Intelligence , 2005 . Proceedings. The 2005 IEEE/WIC/ACM International Conference on, pp. 622 - 628 .

Stedman , R. C. ( 2002 ). Toward a social psychology of place : Predicting behavior from place-based cognitions, attitude, and identity . Environment and Behavior 34 ( 5 ), 561 - 581 .

Steyvers , M. and T. Griffiths ( 2007 ). Probabilistic topic models . In T. Landauer, D. Mcnamara , S. Dennis , and W. Kintsch (Eds.), Handbook of Latent Semantic Analysis . Lawrence Erlbaum Associates.

Teevan , J., S. T. Dumais , and E. Horvitz ( 2005 ). Personalizing search via automated analysis of interests and activities . In SIGIR '05: Proceedings of the 28th annual international ACM SIGIR conference on Research and development in information retrieval , New York, NY, USA, pp. 449 - 456 . ACM.

Tuan , Y.-F. ( 1977 ). Space and Place: the Perspective of Experience . The Regents of the University of Minnesota.

Tversky , A. ( 1977 ). Features of similarity . Psychological Review 84 ( 4 ), 327 - 352 .

Vasardani , M., S. Winter , and K.-F. Richter ( 2013 ). Locating place names from place descriptions . International Journal of Geographical Information Science 27 ( 12 ), 2509 - 2532 .

Wang , C. , J. Wang , X. Xie , and W.-Y. Ma ( 2007 ). Mining geographic knowledge using location aware topic model . In R. Purves and C. Jones (Eds.), GIR, pp. 65 - 70 . ACM.

Winter , S. , W. Kuhn , and A. Kru ¨ger ( 2009 ). Guest editorial: Does place have a place in geographic information science? Spatial Cognition and Computation 9 , 171 - 173 .

Yin , Z. , L. Cao , J. Han, C. Zhai , and T. S. Huang ( 2011 ). Geographical topic discovery and comparison . In S. Srinivasan, K. Ramamritham , A. Kumar , M. P. Ravindra , E. Bertino, and R. Kumar (Eds.), WWW, pp. 247 - 256 . ACM.