A Case Study of Active, Continuous and Predictive Social Media Analytics for Smart City

Abstract

Imagine you are in Milano for the Design Week. You have just spent a couple of days attending few nice events in Brera district. Which of the other hundreds of events spread around in Milano shall you attend now? This paper presents a system able to recommend venues to the visitors of such a city-scale event based on the digital footprints they left on Social Media. By combining deductive and inductive stream reasoning techniques with visitor-modeling functionality, this system semantically analyses and links visitors' social network activities to produce high-quality recommendations even when information about visitors' preferences for venues and events is sparse.

Keywords

• Social Media Analytics
• User Modeling
• Stream Reasoning
• Link Prediction
• Recommendation
• User Engagement

-

Thanks to the widespread adoption of smart phones, location-based Social Media capture an increasingly accurate and up-to-date digital re ection of our cities. In particular, they allow to observe a constantly reshaping network of people that interacts with places around the city, while talking about a variety of topics.

Being able to predict in real time changes in such a network is of major interest for many applications supporting multiple city stakeholders. Short-term predictions may be of interest for tourists and citizens, while medium-term predictions may be useful for city-scale service managers. Tourists may receive precise and timely recommendations to explore places of their interest while sightseeing. Citizens may be informed to avoid means of public transport that are predicted to be crowded. Bike-sharing managers may plan for the availability and the rebalancing of bikes in the stalls based on the predictions of people moving from one neighbourhood to another.

To reach these goals, a system has to ful l the following requirements: R.1 build from one or more social media sources a network whose nodes represent people, places and topics while edges represent the presence of people in places, and the interest of people in topics R.2 track the evolution of such network in real time to detect changes, both in terms of nodes in the graph and in terms of edges R.3 predict the possible changes in the graph R.4 use the prediction to address speci c needs, such as resource planning

In a previous work [ 3 ], we proposed a Continuous Predictive Social Media Analytics (CP-SMA) solution that addressed the four requirements above. The key innovation of CP-SMA is its ability to predict the appearance of edges in the graph when little or no training information about existing edges is available.

Such conditions are typical for city-scale events (CSE) like the Milano Design Week (MDW). Geographically limited to the city of Milano, the MDW featured thousands of events attended by half a million visitors in half a thousand venues that serve as temporary exhibition centres. Large part of the visitors come from abroad. The venues hosting the events are temporary. Even associations between visitors and venues are not applicable across di erent editions of the event, because visitors, venues and events change every year.

CP-SMA was shown in the lab to accurately predict associations between visitors of MDW 2013 and its venues after only 3 days of observations [ 3 ]. In this paper, we report a real-time case study of CP-SMA deployed for MDW 2014, in collaboration with fuorisalone.it6, with the purpose of recommending venues to visitors who were detected to talk about MDW on Twitter. We evaluate the quality of the prediction by checking if the predicted associations are observed after the prediction is made. Our evaluation provides evidence that visitor-venue links can be discovered with a good balance between precision and recall, and sentiment annotations can be added to micro-posts with good accuracy. The evaluation also implicitly demonstrates the ability to build historical and eventrelated topical pro les of visitors by semantically analysing their social network activities.

The remainder of the paper is organised as follows. Section 2 presents the architecture of the solution deployed for MDW 2014. Section 3 illustrates the middleware, which we used to connect the components. Section 4 addresses R.1 and R.2, describing the solution used to identify visitors and to track the evolution of their associations with places (see also [ 4 ]). Section 5 describes the solution used to address R.1 and R.2 with regards to discovering links between visitors and topics. Section 6 details the robust machine learning based predictive model that addresses R.3 by predicting new visitor-venue links (see also [ 3 ]). Section 7 describes the strategy we adopted to engages visitors directly through social media (addressing R.4). Section 8 brie y compares our contribution with related work. Finally, Section 9 discusses the results obtained and the lessons learnt in this case study, concluding with future work. 6 Fuorisalone.it is the o cial Web portal of Milano Design Week.

Architecture

For this case study we designed a loosely-coupled architecture based on Streaming Linked Data principles [ 6 ] and provenance ontology [ 20 ], so that di erent components could communicate and scale independently from each other. Figure 1 illustrates the architecture. The gray looping arrow illustrates the overall ow of information from Twitter, throughout the components of our solution and back to Twitter. The ow starts from the Twitter Streaming API7 that continuously sends to our Social Listener (SL) all the tweets posted from a given area (in this case the Milano urban area), in order to detect visitor-venue links and the most active users. It loops back to the Twitter REST API8, from which the Visitor Modeller (VM) fetches more tweets of visitors in order to build their pro les. The data ow then goes through the Visitor-Venue Recommender (VVR), which predicts visitor-venues links, and it heads back to the REST API of Twitter that is used by the Visitor Engager (VE) to send recommendations to visitors.

More speci cally, the SL component registers geo-bound queries in the Twitter Streaming API. This starts pushing to the SL tweets posted within those geographical areas. The SL analyzes the tweets, links those with positive sentiment to the venues of the CSE, identi es the most active users and stores the results in the Blackboard, using the APIs exposed by the Blackboard Mediator (BM). The Blackboard is the central component of our solution, which decouples in space and time the intercommunication of all our components. The VM periodically looks up visitors recently recorded by the SL in the blackboard. It subsequently checks if such visitors have been previously observed and, if this is 7 https://dev.twitter.com/docs/api/streaming 8 https://dev.twitter.com/docs/api/1.1 not the case, it fetches all their tweets from the Twitter REST API. This component then extracts the semantic entities (from DBpedia) contained in each tweet, computes a pro le of the user as weighted links between visitors and entities, and adds this pro le to the Blackboard. The pro le is periodically updated with the most recent tweets and added to the Blackboard. The VVR periodically fetches the most recent visitor-venue links and visitor pro les from the Blackboard. Then for each visitor, the VVR predicts the top 10 most probable visitor-venue links not previously observed and adds them to the Blackboard. The VE periodically looks up the most recently predicted links and use them as recommendations for each visitor. If a recommendation concerns a visitor that has not received any message yet, the VE invites through Twitter such a visitor to follow its Twitter account. If the visitor is already a follower, the VE sends them a Direct Message with the computed recommendations. 3

Blackboard and Blackboard Mediator

The heterogeneity of the components and of the data they produce requires an integration middleware. Moreover, since the components are distributed, the data needs to be remotely managed, in a scalable and robust way. Last but not least, data elements are produced over time and, in this speci c scenario, the more recent they are, the more frequently they will be accessed. As a consequence, after a long enough period of time they can be archived preventing the data management system to require in nite memory.

These requirements can be addressed by a blackboard-based communication approach [ 17 ]: a exible communication model that does not require any prior restriction on who can place what information on the blackboard and when they do so. RDF represents a suitable data model to be used with blackboards since it poses no restriction on the type of information that can be exchanged. The Streaming Linked Data Format, proposed in a previous work [ 4 ], can be used to deal with data that is produced by distributed components, and changes frequently over time. This format has two main elements: the graphs that are continuously published by the components (named instantaneous graphs or, shortly, iGraphs), and a graph (called stream graph or, shortly, sGraph) that lists the iGraphs and annotates them with who produces each iGraph (using the PROV ontology [ 20 ] property prove:wasAttributedTo) and when (prove:generatedAtTime).

Our Blackboard is built on Jena TDB (as RDF triple store), Jena Fuseki (as SPARQL server for querying, updating and managing RDF graphs over HTTP protocol) and a Java API (namely Blackboard Mediator) that hides the complexity to invoke Fuseki and provides a single access point to the blackboard.

The Blackboard Mediator (BM) abstracts the communication between the components and the blackboard by o ering a Java API to the data producers and consumers. The three main methods are: add new graph, retrieve the list of graphs inserted in a given time period by a given agent (i.e. the component), and fetch a graph. In particular, the BM exposes a service to add a new iGraph g by encapsulating two steps: 1) the creation of a tracking sGraph for g in the default graph, and 2) the insertion of g. The unique identi er of g consists of the base URI, the agent and the arrival timestamp. The uniqueness of the agenttimestamp pair is enforced by the API.

We ran two di erent evaluation experiments accessing the blackboard using the BM component. The rst experiment measured the time needed to randomly access a single graph by URI, varying the number of iGraphs in the blackboard. The second one measured the time for retrieving the list of iGraphs in three different time windows (3 minutes, 30 minutes and 180 minutes wide, respectively) varying the number of iGraphs in the blackboard.

As illustrated in Figure 2, the time taken for randomly accessing a single iGraph is independent from the number of iGraph stored in the blackboard (Figure 2 A) and the time required to retrieve a list of iGraphs URI increases linearly with the number of iGraph in the blackboard (the number of iGraph in the system is presented using di erent symbols in the diagram) (Figure 2 B.1), but is not sensitive to the timespan of the window requested (Figure 2 B.2). The e cacy of the blackboard mediator as an integration middleware is empirically demonstrated by this evaluation. The linear growth of the access time with the amount of iGraph is not problematic because in our system iGraphs can be deleted from the blackboard when they are older than a prede ned threshold, thus preventing an in nite growth. 4

Social Listener

The SL is the container of a pipeline of components devoted to collecting, translating, decorating and analysing a generic data stream from social sources (see Figure 3). It addresses Requirements R.1 and R.2.

The Twitter Adapter represents the entry point of the pipeline, it establishes a connection with the Twitter Streaming API to collect the raw tweets and to transform them in a structured data format, using RDF as a common data model. The structured representation of a single tweet include the hashtags, the URLs, the creator, etc. in a structured way that eases the operations in the blocks downstream.

The next element in the pipeline is the CSE Entities Decorator [ 5 ]. This component links tweets to venues of the CSE. It uses an aggregation of four lexical similarity metrics that compares content and hashtags of each tweet to the names of the venues and the titles of the events hosted in each venue (thus also exploiting the semantic relationship between venues and events). As shown in the table below, the manual assessment of the correctness of the discovered links provides experimental evidence of the e ectiveness of our CSE Entities Decorator.

Type of Link Number of Links Correctness Visitor-Venue 761 72%

Visitor-CSE 3723 94%

The decoration phase is concluded by the Sentiment Decorator [ 22 ] that adds the sentiment information to the tweet. It relies on a dictionary-based classi cation method, and applies the same processing for tweets written in English and in Italian. More speci cally, our method uses English and Italian sentiment dictionaries derived from SentiWordNet [ 13 ] and MultiWordNet [ 19 ] semantic lexicons, where words are attributed with positive, negative and objective scores. These scores are combined into a single sentiment score for each tweet, by considering negation and modi er keywords, and also taking into account positive and negative emoticons. Within a speci c domain (as the one of CSEs) simple sentiment expression, even including irony and idioms can be captured. Still, complex sentiment expressions (rarely encountered in tweets mentioning venues) go beyond pattern-based analysis and require more robust methods that are not optimised for fast text processing because they need more than one-pass.

To evaluate the performance of our method, we randomly selected a third of the tweets liked to a venue, and asked two human experts to label their opinion polarities, obtaining 187 tweets with veri ed annotations. The annotated subset contained approximately equal amounts of positive, negative and objective (neutral) tweets, making the baseline classi er accuracy 0:37, when labeling all tweets as objective. We compared our method using various dictionaries, and also included the results obtained by SentiStrength [ 21 ] method for English, however equipped with Italian dictionaries (provided by the authors). The results of our experiments are reported in the table below, where we measured the overall classi cation accuracy, and the recall for individual classes, while utilizing polarity thresholds that yield the best accuracy.

Approach Baseline IT SentiStrength IT Only IT + EN IT + EN + Stem

Accordingly, the best performance is reached when our method utilizes both English and Italian sentiment dictionaries together with stemming. The accuracy in this case reaches 0:67, being 30% above the baseline and 15% above SentiStrength. What is more important, our method reached 0:73 for the recall of negative sentiments, reducing their misclassi cation as talksAboutPositively relation, which is of the main interest for the venue recommender.

In the Analyze phase two di erent continuous queries are performed over the decorated data: the rst one extracts the visitor-venue links for the VVR component and the second one extracts the visitor ids for the VM component. Sentiment plays an important role in this phase, only the tweets with a positive sentiment are passed to the VVR and VM components. In this way the recommendation and the visitor analysis is performed only on satis ed visitorsThe Blackboard Translator performs the last step of the process writing the results of the Social Listener on the blackboard using the Blackboard Mediator APIs. 5

Visitor Modeller

The VM performs monitoring and pro ling of visitors based on their social media activities. The VM monitors selected visitors (the visitors that have been detected by the SL) by crawling their social media posts, and subsequently analyses them with di erent techniques to obtain a pro le. It complements the SL in addressing Requirements R.1 and R.2.

The VM is composed by the following modules (see also Figure 4): a Front-end that reads visitor ids from the blackboard; a Scheduler that schedules the analysis of the ids; a Crawler that crawls visitors' tweets; an Analyzer that extracts entities from the tweets; and a Pro ler that generates a pro le for each visitor.

To guarantee high scalability, all components (except for the Pro ler) have been designed to be replicable, and connected by a load balancing mechanism implemented by a message queuing sub-system9. Once a visitor id has been read by the Front-end and scheduled for analysis by the Scheduler, the Crawler collects all visitor's tweets, using Twitter's public REST APIs. The Analyser then extracts the entities contained in the tweets. For example, from the tweet \Temperley 9 We use ActiveMQ http://activemq.apache.org/

London at London Fashion Week Spring 2014 http://t.co/mXFdvmCWXR", the VM detects that London Fashion Week refers to the entity http://dbpedia. org/resource/London_Fashion_Week.

The Analyser can use di erent Named Entity (NE) extractors. For this case study we used the probabilistic version of DBpedia-Spotlight10. This tool analyses the text by \spotting" possible candidates for entities, and disambiguating each candidate. NE tools usually provide also some values related to the likelihood that the result is correct. In Spotlight's case, this is not really a con dence value, but more an indication of how well the entity was disambiguated from alternative entities in DBpedia; the candidate could still be wrongly spotted (e. g. not be an entity). In particular, Spotlight provides two values for each found entities: assuming that there is an entity in the text, how likely this is the entity ranked rst by Spotlight, normalized with respect to all possible entities (the similarityScore), and how likely the second ranked entity is with respect to the rst ranked one (the percentageOfSecondRank ).

The Analyser lters extracted entities below a certain threshold. To obtain a single value from Spotlight, we adopted the following score for each entity: (1 0:5 percentageof secondrank) similarityscore, the rationale being that if an entity is very likely (similarityscore=1) but a second one is equally likely, both score 0.5. Since we observed that almost 80% of the entities had a score 0.9, we ltered out entities with lower score to have relatively high quality without losing too much in recall (1.090.237 entities were left).

To evaluate the quality of the extraction, we estimated the proportion of entities correctly extracted by manually verifying their presence in the tweets. In total we extracted 1.090.237 entities from 1.879.187 tweets. We sampled the entities extracted to estimate the proportion p^ of correct entities. Considering that such a proportion follows a binomial distribution that for large population can be approximated by a normal distribution, we used Wald's method to establish the sample size given the con dence interval conf and the error range err, such as that for the real estimate p we have that p 2 (p^ err; p^ + err) in 10 https://github.com/dbpedia-spotlight/dbpedia-spotlight/wiki conf% of the times:

n = qnorm(coenrfr2)2p^(1 p^) (1) where qnorm is the quantile function of the normal distribution. We chose err=0.025 (i. e. 2.5%) and conf=95%. From an initial exploratory evaluation with sample size of 300 entities we obtained an initial estimate for p^ to be around 0.74, which gives a sample size of 833. We therefore sampled (with replacement) the data set and carried out the evaluation presented in the table below.

Entities Samples Correct Precision Range Sup Range Inf Con dence 1090237 833 615 73.83% 71.33% 76.33% 95.00%

To be noted that this evaluation only measures the precision of the extraction, and not the recall. Evaluating recall on our data set is much harder since it requires the evaluator to de ne the set of the entities that are contained in a tweet, while verifying that an entity is contained in a tweet is much easier.

Finally, the Pro ler creates a visitor pro le. For each visitor the Pro ler builds both an historical and an event pro le. Historical pro les are based on tweets posted before the monitored event, while event pro les are based on tweets during the event. A visitor pro le is de ned as a set of pairs (e; w(u; e)), where e is an entity and w(u; e) is the weight of e in u's pro le. Initially our strategy has been to select the most common entities (we set the limit at max 10 entities) for each visitor and weighting them based on the number of occurrences. We are planning to evaluate alternative strategies, such as ltering out entities that are not discriminative (e. g. because they are too common, such as Milano), and use the class of entities to have more overlapping among visitor pro les. 6

Visitor-Venue Recommender

This section addresses the requirement R.3: predict the possible changes in the graph (see Section 1). We brie y describe the component Visitor-Venue Recommender (VVR) and evaluate it on the MDW data.

The VVR predicts visitor-venue links that are very likely to occur in the future. A statistical machine learning approach named SUNS [ 15 ] is applied in the VVR. SUNS can be viewed as a regularized matrix factorization approach11. This approach is robust with respect to sparsity of data and can incrementally improve predictions when more data become available. The regularization can reduce the model's sensitivity to its parameters. The model of the VVR is trained using visitors' pro les generated by the VM and visitor-venue links detected by the SL. It is then applied to recommending venues that will be potentially visited by visitors in the upcoming days. In the following updates, when new visitor-venue links or new or updated visitors' pro les are available, the model is retrained taking new information into account. This capability enables the VVR to cope with social media stream dynamics. 11 In recent years matrix factorization based approaches have been using in many recommendation systems, e.g. Net ix Prize.

The evaluation of the VVR in this paper is focused on the capability of capturing changes of social media stream over time. During the MDW the VVR processed 842 unique links between 512 visitors and 198 venues. This data is summarized in the table below. It is important to clarify what the training data is and what the test data is. Considering the update on April 8th at 4pm, the VVR fetched from the Blackboard 105 new unique links and 65 new visitors were observed. Until that time 100 visitors had 154 links associated with venues they talked about.12 This was the training data exploited at this update time. In subsequent updates until the end of MDW, 22 visitors had 42 more links to venues.

Update Time 2014-04-08T04:00:00 2014-04-08T16:00:00 2014-04-09T04:00:00 2014-04-09T16:00:00 2014-04-10T04:00:00 2014-04-10T16:00:00 2014-04-11T04:00:00 2014-04-11T16:00:00 2014-04-12T04:00:00 2014-04-12T16:00:00 2014-04-13T04:00:00 2014-04-13T16:00:00 sum

In order to evaluate the quality of the predictions in the VVR, we compared SUNS with random guessing (Random), Pearson correlation coe cient (Pearson) and most talked venues (MostTalked). Random assigns randomly a likelihood to every venue not talked yet by a visitor; Pearson makes recommendations based on the visitors' similarity; and MostTalked encodes the popularity of venues in the community observed by counting how frequent the venues were mentioned and recommending the venues ordered by the frequency to all visitors. In addition, we combined SUNS with MostTalked (SUNS+MostTalked), which intends to model both venue popularity and visitor preferences. The evaluation measures used were: precision@topN = T rueP ositive=N and recall@topN = T rueP ositive=(T rueP ositive + F alseN egative).

Figures 5 (a) and (b) show the recall@5 and @10 recommendations in 11 updates from April 8th at 4am to April 13th at 4am, 2014. The random ranking failed in most cases to recommend correct venues. Pearson performed better than Random and its recall value increased slowly over time. The performance of SUNS kept increasing too, especially in the last update. Surprisingly, MostTalked outperformed all other approaches and it was clearly overtaken by SUNS only in a few updates, when considering the top 5 recommendations. The combined approach SUNS+MostTalked performed in most updates better than both approaches. Figures 5 (c) and (d) show the precision@5 and @10 recommendations. In contrast to the recall, the precision decreased against the top number 12 Note that a visitor can talk about the same venue several times and that we report here the distinct count.

(a) (c) (b) (d) of recommendations and it was lower than the recall. That is because the average number of the ground truth links per visitor is not greater than 2, i.e. T rueP ositive + F alseN egative <= 2 (see the right two columns in the table above), whereas precision is calculated by dividing the top number 5 resp. 10.

Our evaluation uncovered several interesting ndings. First, the VVR reaches, on average for all visitors, recall over 20% at top 5 and more than 30% at top 10 (applying SUNS+MostTalked). If we consider the fact that many events were held for several days at the same venue so that the venues which had been talked by visitors in the past could be recommended again and again, the resulting VVR models performed fairly well. Second, in many updates, the more information is available, the higher quality of recommendations are provided. However, this increase is not observed at all updates. That is because the di culty degree of making recommendation does not only depend on the available data, but also on many other aspects in a particular update, such as the sparsity of the links regarding the number of visitors and the number of venues and the community structure (e.g. active users, user clusters). Last, but not the least, the popularity of venues dominates the performance of the VVR. The visitor similarity, explored by both SUNS and Pearson, did not contribute as much as it usually does in social networks. 7

Visitor Engager

The purpose of the VE is to automatically contact relevant Twitter users and provide them with the possibility to view and rate their personalized recommendations (addressing Requirement R.4). The VE is a Java enterprise application build on top of Errai framework and is publicly available at https: //github.com/WISDelft/AttendeeEngager. It consists of four main components: 1) The Front-end consisting of an informative landing page of the VE and a personal dashboard, 2) The TwitterManager that handles sending of Twitter messages and monitoring of followers of the designated account, 3) The RecommendationManager which uses the Blackboard Mediator to retrieve recommendations calculated by the VVR, and 4) The QueueManager which reliably handles all communication between the components using the same underlying queue technology as for the Visitor Modeller.

The engagement process (see Figure 6) is started by receiving the Twitter user IDs of new users, as detected by the Social Listener. Using the @fuorisalonetips Twitter account, which was o cially supported by the organization, the VE sent each user only a welcome and (if necessary) a reminder message via a Tweet mentioning the user, a so-called mention. The welcome mention looked like: @joosterman Welcome to #fuorisalone Milano Design Week <. . . >. If the user started following the account, they would be considered opted-in and received the link to their personal dashboard. To minimize the amount of unsolicited tweets, we chose to send recommendations only to users who had actively opted in. At any time a user could opt-out via the personal dashboard or stop following the designated account to become opted-out, after which they would not receive any further messages. Opted-in users received direct messages (a private message) containing recommendations and the information that their dashboard was updated.

Using their Twitter account or dashboard link users could visit their personal dashboard (see Figure 7). The dashboard showed recommended venues grouped by day, and each venue could be expanded to see events taking place there. Thumb-up and thumb-down buttons were shown next to each venue. To assess the quality of the recommendations, in terms of users perception of usefulness and interest in the item recommended, we de ned two ratings: a user expanding a venue (implicit rating) and clicking on a thumb button (explicit rating).

There are two Twitter policies that apply to the VE: rstly, the harassment policy, that prevents an account to send large numbers of unsolicited mentions, and secondly the spam policy, that prevents an account to send many messages of duplicate content. Trying to comply to these policies we implemented a timer for sending messages such that messages were sent with an interval of 5 minutes. We also created 5 di erent versions of each message.

However, despite our e orts, the account was temporary suspended for violating the harassment policy. After the suspension we were able to again send messages but at a much slower pace. The system managed to send welcome messages to 179 users of which only 2 opted in and only 5 opted out resulting in a response rate of 3.9%. These results are in line with previous empirical studies [ 16 ] reporting a 4% response rate. The system did not send reminder messages because, due to the limitations imposed by Twitter, we gave higher priority to sending messages to new users. Two users visited the dashboard of which one used it to opt-out. Neither user used the dashboard to rate the recommended venues. As a result, it was clear that visitors did not appreciate to be contacted on Twitter, even when using an o cial Twitter account supported by the organization of the event (Fuorisalone). Therefore, in the future we need to devise alternative ways for visitor engagement.

Related Work

Several e orts in literature are related to the solution we adopted for this case study, although there are clear di erences with our approach.

City venues and their categories have been used in order to characterise a city's districts, also with the goal of recommending particular areas to visitors. [ 18 ] uses Foursquare checkins and Foursquare venue categories to classify di erent areas of the city, and to pro le users based on the categories of the venues that they visit. A similar work is [ 11 ], which clusters together areas in a city based on similar users checking in the venues of that area, based on the assumption that both people and places de ne the character of an area (what they call Livehoods). Both the venue and the visitor pro le are based on Foursquare venue categories. A step further is represented by [ 12 ], where the authors de ne venues also based on the pro le of the visitors visiting those venues, using Facebook data to pro le visitors. They claim that using semantics extracted both from places and from the online pro les of people who frequent those places improves city areas and venues classi cation. Di erently from our solution, these approaches rely on the availability of data about the categories of venues. Moreover, venues are considered as having a static \character", i.e. a xed category. In our case, during a City Scale Event venues change their character and this data is not available from Location Based Social Networks such as Facebook or Foursquare.

Other e orts have tried to relate tweets to the geographical location from where they were posted. [ 14 ] predicts location of a tweet, using a generative model based on location-dependent language models. Similarly, also [ 9 ] propose a probabilistic framework for estimating a Twitter user's city-level location based on the content of the user's tweets. Such approaches do not achieve the precision needed in our approach, considering that they are able to place users within 100 miles of their actual location. Also using the geolocation information of social streams, but for a di erent purpose, [ 10 ] analyses check-ins as a way to study user mobilities with a large scale perspective (millions of check-ins).

Related to the problem of linking a tweet to a venue, [ 8 ] uses entity extraction from tweets in order to enrich the static description of Point of Interests with dynamic information contained in the tweets. Again, this approach operates at city scale, while we operate at venue scale. Moreover, our problem is not to characterise an event, but to know who exactly is attending it.

A similar approach to ours, [ 1 ] also uses social streams and other usergenerated sources such as GPS information to understand the urban context and generate recommendations, using a semantic approach based on Semantic Web technologies (Pellet, OWL and SWRL). Di erent from our work, their user modeling is not based on user interests but on user preferences for mobility (such as avoid crowded place to go to work), since their recommendations are related to route planning. Therefore, their user models have a di erent content. Similarly, in the eld of Smart Cities and mobility, the CityPulse project13 aims to create a distributed framework for semantic discovery, processing and interpretation of 13 http://www.ict-citypulse.eu/ large-scale real-time Internet of Things and relevant social data streams, with the goal of knowledge extraction in a city environment [ 7 ]. In particular,\Extracting City Tra c Events from Social Stream' deals with extracting tra c events from twitter streams [ 2 ]. 9

Conclusions and Future Work

The growing amount of digital footprints that visitors of city-scale events leave on social media enables to build a graph of visitors attending events hosted in venues. This graph is continuously reshaping and the ability to predict in real time its evolution can foster innovative social media analytics services. In our previous work [ 4,3 ], we showed that our deductive and inductive stream reasoning techniques, once combined with visitor-modeling functionalities, can produce high-quality predictions of links in such a graph even when information about visitors' preferences for venues and events is sparse. In this paper, we report on our rst attempt to experimentally o er a real-world service for the visitors of Milano Design Week 2014. The case study con rms the ability of our system to semantically analyze and link visitors' social network activities to venues of a city-scale event. The case study also con rms the ability to build representative visitor pro les. The task of predicting links between visitors and venues is as hard as expected in the lab. The predictions the system produces are in line with our expectation (i.e. they are better than any base line methods in the state of the art). Also the integration middleware based on a Streaming Linked Data implementation of the blackboard paradigm showed satisfactory performances. The most critical task was engaging the visitors. We monitored visitors in Twitter and we also tried to engage them by using Twitter. Unfortunately, the task was much more challenging than expected. Even if we were granted the permission to use a Twitter account related to MDW and to host the application under a sub-domain of the o cial website of MDW, only few users opted to follow us and, thus, received our recommendations. Moreover, Twitter anti-spam policies resulted more powerful than our countermeasures.

In our future work, we intend to improve the correctness of the method employed in semantically analysing and linking social media to venues and topics. We will investigate the e ect of additional meta data (e.g. organizers of events) on the quality of recommendations. An important challenge we intend to take on in the future is to predict the appearance or disappearance of nodes from the people-venues-events graph. Moreover, we intend to deliver our recommendations through an o cial mobile application of a city-scale event so that we are perceived as a trustable source of recommendations.

References

1. Alvarez-Napagao , S. , Tejeda-Gomez , A. , Oliva , L. , Garcia-Gasulla , D. , Codina , V. , Vazquez-Salceda , J. : Urban context detection and context-aware recommendation via networks of humans as sensors . In: CARE ( 2014 )

2. Anantharam , P. : Extracting city tra c events from social streams ( 2014 ), https: //osf.io/b4q2t/wiki/home

3. Balduini , M. , Bozzon , A. , Della Valle , E. , Huang , Y. , Houben , G.J.: Recommending venues using continuous predictive social media analytics . IEEE Internet Computing 18 ( 5 ) ( 2014 )

4. Balduini , M. , Della Valle , E. , Dell'Aglio , D. , Tsytsarau , M. , Palpanas , T. , Confalonieri , C. : Social listening of city scale events using the streaming linked data framework . In: ISWC (2) . pp. 1 { 16 ( 2013 )

5. Balduini , M. , Della Valle , E. , Nasi , G. , Calegari , G.R.: A quest for lightweight methods to dynamically linking micro-posts to entities of a city-scale event . In: Proc. of Workshop on Context, Interpretation and Meaning ( 2014 )

6. Barbieri , D.F. , Della Valle , E.: A proposal for publishing data streams as linked data - a position paper . In: LDOW ( 2010 )

7. Bischof , S. , Karapantelakis , A. , Sheth , A. , Mileo , A. , Barnaghi , P. : Semantic modelling of smart city data . In: W3C Workshop on the Web of Things Enablers and services for an open Web of Devices . pp. 1 { 5 . Berlin, Germany ( 2014 )

8. Cano , A.E. , Varga , A. , Ciravegna , F. : Volatile classi cation of point of interests based on social activity streams . In: Proc. Of SDOW ( 2011 )

9. Cheng, Z. , Caverlee , J. , Lee , K. : You are where you tweet : A content-based approach to geo-locating twitter users . In: ICKM . pp. 759 { 768 . ACM ( 2010 )

10. Cheng, Z. , Caverlee , J. , Lee , K. , Sui , D.Z. : Exploring millions of footprints in location sharing services . In: Proc of the 5th Int Conference on Weblogs and Social Media . pp. 81 { 88 . The AAAI Press ( 2011 )

11. Cranshaw , J. , Hong , J.I. , Sadeh , N. : The livehoods project : Utilizing social media to understand the dynamics of a city . In: The 6th Int. AAAI Conf. on Weblogs and Social Media . AAAI Press, Dublin, Ireland ( 2012 )

12. Del Bimbo , A. , Ferracani , A. , Pezzatini , D. , D'Amato , F. , Sereni , M. : Livecities : Revealing the pulse of cities by location- based social networks venues and users analysis . In: WWW . pp. 163 |- 166 . Seoul, Korea ( 2014 )

13. Esuli , A. , Sebastiani , F. : Sentiwordnet: A publicly available lexical resource for opinion mining . In: LREC '06 ( 2006 )

14. Hong , L. , Ahmed , A. , Gurumurthy , S. , Smola , A.J. , Tsioutsiouliklis , K. : Discovering geographical topics in the twitter stream . In: WWW . pp. 769 |- 778 ( 2012 )

15. Huang , Y. , Tresp , V. , Bundschus , M. , Rettinger , A. , Kriegel , H.P. : Multivariate prediction for learning on the semantic web . In: ILP . pp. 92 { 104 ( 2010 )

16. Mahmud , J. , Zhou , M.X. , Megiddo , N. , Nichols , J. , Drews , C. : Optimizing the selection of strangers to answer questions in social media . CoRR abs/1404 . 2013 ( 2014 )

17. Nii , H.P. : Blackboard systems, part one: The blackboard model of problem solving and the evolution of blackboard architectures . AI Magazine 7 ( 2 ), 38 { 53 ( 1986 )

18. Noulas , A. , Scellato , S. , Mascolo , C. , Pontil , M. : Exploiting semantic annotations for clustering geographic areas and users in location-based social networks . In: The Social Mobile Web ( 2011 )

19. Pianta , E. , Bentivogli , L. , Girardi , C. : Multiwordnet: developing an aligned multilingual database . In: Proc. of the 1st Int. Conf. on Global WordNet ( 2002 )

20. T. Lebo et al.: Prov-o: The prov ontology . W3C Recommendation ( 2013 )

21. Thelwall , M. , Buckley , K. , Paltoglou , G. , Cai , D. , Kappas , A. : Sentiment strength detection in short informal text . J. Am. Soc. Inf. Sci. Technol . 61 ( 12 ), 2544 {2558 (Dec 2010 )

22. Tsytsarau , M. , Palpanas , T. , Denecke , K. : Scalable detection of sentiment-based contradictions . In: DiversiWeb Workshop ( 2011 )