Feelings’ Rating and Detection of Similar Locations, Based on Volunteered Crowdsensing and Crowdsourcing

In this paper, an innovative geographical locations’ rating scheme is presented, which is based on crowdsensing and crowdsourcing. People sense their surrounding space and submit evaluations through: (a) a smartphone application, and (b) a prototype website. Both have been implemented using the state-of-the-art technologies. Evaluations are pairs of feeling/state and strength, where six different feelings/states and five strength levels are considered. In addition, the detection of similar locations is proposed by maximizing a cross-correlation criterion through a genetic algorithm approach. Technical details of the overall system are provided so that the interested readers can replicate its components. The experimental results on real-world data, which also include comparisons with Google Maps Rating and Tripadvisor, illustrate the merits and limitations of each technology. Finally, the paper is concluded by uncovering and discussing interesting issues for future research.


I. INTRODUCTION
Rating websites and rating smartphone applications enable people to vote on or rate people, content, or other things.Typically they show content either in a random fashion or chosen by an algorithm and then ask users for a rating or assessment using a defined scale or feedback mechanism.Several examples exist in the recent literature.Hot or Not [1] allows users to rate photos posted by other users.WeRate-Dogs [2] is a Twitter and Facebook account for rating dog photos and videos.RateMyProfessors [3] allows college and university students to anonymously assign ratings to professors of different institutions around the world.RateMDs [4] allows people to submit and read reviews of doctors, dentists, psychologists, hospitals etc.Among their objectives, all these and other similar applications aim at helping people in their everyday choices, i.e., what to avoid and what to prefer.According to [5], ''. .

. we live in an era of heightened accountability and evaluation plays a major role in ensuring
The associate editor coordinating the review of this manuscript and approving it for publication was Dian Tjondronegoro.
that accountability''.Current trends show that in the near future we will be enabled to rate just about everything.
On the other hand, everyday several different things, events, activities and phenomena happen in several different geographical locations all over the world.Some of them can make us feel happy (e.g. a concert in a park, a street performance, a nice statue, camping, fishing, hiking, sightseeing, shopping etc.).Some others can make us feel sad, frustrated or in danger (e.g.street fighting, terrorism, war, dirt, street closures, road accidents, cruelty to animals, tornados etc.).Most people would avoid/prefer locations that generate bad/nice feelings.On the other hand, things in this world change very rapidly.In particular a specific geographical location could produce very nice feelings ''today'', but ''tomorrow'' something bad could happen at this same location.Of course ''today'' and ''tomorrow'' can be as close as ''now'' and ''after a split of a second'' (e.g.imagine a nice concert in a park which is disrupted by gun shooting or imagine unearthing survivors from debris after an earthquake.Feelings can change rapidly at the same location).
Inspired by the aforementioned real situations, we have developed and tested an innovative scheme, which enables VOLUME 7, 2019 This work is licensed under a Creative Commons Attribution 4.0 License.For more information, see http://creativecommons.org/licenses/by/4.0/people to rate in real time all geographical locations around the world.In this way, anyone and at any time can know the feelings that any location generates and then make her/his plans accordingly.In particular users can rate the locations they visit, by submitting an evaluation through their smartphone.Each evaluation consists of latitude, longitude, feeling/state and strength of feeling/state, for example (37.978117, 23.712161, Like, 4/5).Influenced by social media, our prototype allows for six different feelings/states (Like, Love, Haha, Wow, Sad, Angry) and five strength levels (1 to 5).The six feelings/states (reactions) were launched by Facebook on February 2016, after a long experimentation phase and collaboration with professors specializing in nonverbal communication.In this way, users can easily and quickly express how they feel.On the other hand, our users who interacted with the implemented prototype were mainly students between 18 and 26 years old and they were very familiar with Facebook; so, no training was required.All evaluations are aggregated to estimate an average feelings/states vector for each location.Additionally a novel scheme is also proposed, which estimates the similarity among different locations, by maximizing a cross-correlation criterion through a genetic algorithm approach.
Overall: (a) the rationale of connecting feelings and geographic locations is to enrich existing schemes, which consider only POIs, excluding at the same time the majority of earth's geographic locations from being evaluated.Furthermore feelings' evaluation is a very easy and quick way to evaluate locations.Our scheme can also be used for evaluating locations in the sea (about 71% of the Earth's surface is water-covered, and the oceans hold about 96.5% of all Earth's water).(b) the prototype has been used for a 90-days experimentation phase, but we hope that it is further developed by major organizations and offered as a new free service, since it extends Google Maps Rating 6] and Tripadvisor [7] mainly by: (i) evaluating all locations and not only POIs and (ii) evaluating locations using feelings/states.This is a much richer way to evaluate locations than the existing ones and it is also very easy and fast to use.To sum up, this work introduces three main innovations: • It enables users to evaluate locations using feelings/ states.Similar services such as Google Maps Rating [6] and Tripadvisor [7] do not use feelings, but just a score.
• To the best of the authors' knowledge there is not any other scheme that detects similar locations, in terms of feelings.Here it should also be stressed that the proposed scheme also faces the many-to-many location similarity problem (e.g. a group of users who have visited different locations, would like to find locations that match their preferences as a group and not individually.Examples are provided in Sections IV and VI).
• Since the overall vision is to provide a means to rate every square meter on our planet, the proposed approach enables users to evaluate any location and not only points of interest (POIs) e.g.parks, schools, government buildings, shops, restaurants, hotels etc..For example a user can evaluate a location on a road, during heavy traffic.

II. RELATED WORK
Even though locations' evaluation (mainly evaluation of POIs) is a more recent topic, the problems of locations' similarity and travel assistance have also been studied in the past.In particular, in [8] a computational method to formally assess the similarity of spatial scenes is proposed based on the ordering of spatial relations.One scene is transformed into another through a sequence of gradual changes of spatial relations.The number of changes required yields a measure that is compared against others, or against a pre-existing scale.In [9] the focus is put on estimating the similarity of a pair of discrete spatial objects, where only their location is considered.Location refers to the interrelation between position, shape and size of the objects.Excluded from comparison are thematic attributes, relations in scenes and matching.The work in [10] focuses on the functionality and application of GIS in tourism development projects.The paper argues that GIS can bring significant added value to decision making through data analysis, modeling and forecasting.In [11] a prototype augmented reality-based tour guide has been implemented.The system uses a micro-computer and an invisible spatial locating device while it also provides tour planning.In [12] a general framework for building candidate/ critique agents is presented and an Automated Travel Assistant is implemented for assisting users to plan their trips.The user initially provides some preferences regarding her/his trip and the system provides several choices that satisfy those preferences.Afterwards, the user notes favorable or unfavorable characteristics and the system responds to the user's actions.
On the other hand the recent literature focuses on volunteered geographic information (VGI) works for environmental monitoring [13], events reporting, human movement analysis, disaster management, travel planning etc..In particular Celino [14] proposes the adoption of a provenancebased Human Computation approach to aggregate and consolidate VGI, through employing an ontological formulation.Wang et al. [15] present a flood density estimation scheme that is based on fusing remote sensing data and data provided by users over social media.Fusion is performed by considering a target optimality criterion of choice.
In [11] humanitarian mapping is proposed that labels buildings, roads etc.. Labeling is accomplished by fusing multiple freely accessible crowdsourced geographic data.In [17] an extended random walker approach is developed for extracting target objects, by combining very-high-resolution images and VGI data.Foreground and background seeds are automatically selected with the assistance of VGI data.Verstockt et al. [18] describe a multimodal crowdsourcing bike-sensing setup for automatic geo-annotation of terrain types.The proposed system is mainly based on the analysis of VGI gathered by cyclists, using the accelerometer and GPS sensors of their smartphones.
In [19] an automatic approach for building up POIs resources is presented.The scheme is based on the types of POIs extracted from Google Maps and the street names are obtained from OpenStreetMap.Potential addresses and place names are retrieved through submitting queries to the Google search engine.In [20] automatic urban land-use classification is introduced, using crowdsourced geotagged images.
Thejaswini et al. [21] present a human mobility-based smartphone sensors sampling algorithm.The algorithm considers velocity of human mobility as an important parameter for improving sensing area coverage and reducing energy consumption.In [22] state-of-the-art algorithms are compared, designed for votes' aggregation in crowdsourcinglike settings.The crowdsourcing scenario focuses on the Cropland Capture game, which aims to map cultivated lands using around 170,000 satellite images.In [23] various quality measures and indicators for selected types of VGI and existing quality assessment methods are reviewed.In [24] latent characteristics of road segments are extracted from multi-domain user generated content.Extracted information is integrated in the form of a multi-criteria path-finding application.
Zhou et al. [25] propose a travel-planning tool by crowdsourcing multiple sources of user generated content, to provide customized information for tourists.The algorithm harvests hotel reviews from TripAdvisor, photo information from Flickr, and costs between destinations from Uber.In [26] user internal factors are explored to measure influence.The proposed user internal factors include user sentimental deviations and the review's reliability.In addition an attention mechanism is utilized that could auto-learn the weights of different factors.Through a case study on Yelp it was found that the proposed user sentimental deviations and the review's reliability are effective in improving the accuracy of rating predictions.
In [27] geo-tagged photos are segmented into different categories to find categorized POIs, using density-based clustering.Geo-tagged photos from Flickr are analyzed and various POIs patterns are reported.In [28] full use of the mobile users' location sensitive characteristics is made in order to carry out rating prediction.The work mines the relevance between user's ratings and user-item geographical location distances and the relevance between users' rating differences and user-user geographical location distances.Moreover, user-item geographical connection, user-user geographical connection, and interpersonal interest similarity, are fused into a unified rating prediction model.
In [29] the site selection problem is considered, by recommending locations satisfying special requirements.The work focuses on specific site selection of meteorological observation stations.The proposed algorithm not only recommends locations that provide more accurate prediction and cover more areas, but also minimizes the cost of building new stations.In [30] a framework to extract associative points-of-interest patterns from geo-tagged photos in Queensland, Australia is proposed.The framework combines clustering for points-of-interest detection, and association rules mining for associative points-of-interest patterns.
In [31] an LDA + Word2vec model is proposed to mine user interest.Then, a social user sentimental measurement approach is introduced.Finally user topic, user sentiment and interpersonal influence are fused into a recommender system based on probabilistic matrix factorization.
A very interesting and the most relevant -to our workscheme is proposed in [32], where sentimental attributes of POIs are detected through sentiment analysis of microblogs.Then POIs with certain sentiment attributes are recommended to different users.In particular, the sentiment of microblogs is analyzed by incorporating a lexicon-based method.HowNet Sentiment Dictionary is used to compute sentiment values.Each microblog is divided into clauses.Next, for each clause, the positive and negative words are detected, a process which leads to the calculation of the review sentiment value.Finally a sentiment-based POI recommendation model is proposed, where users' preferences are approximated by solving an optimization problem.The optimization problem considers user and POI latent features, sentimental similarity between pairs of POIs and geographical distance between pairs of users and POIs.In order to highlight the differences of our scheme from [32], the two schemes are compared on a conceptual level (and not quantitatively), since they focus on different things.In particular: 1) the proposed scheme collects locations' evaluations provided by the users in an explicit manner, while [32] (a) collects the locations based on check-ins (b) collects and analyzes the respective associated text of each check-in and (c) evaluates the location by performing sentiment analysis on the associated text.Here it should be mentioned that check-ins may be fake, since [32] does not have access to the GPS coordinates of the users' devices.Our scheme has access to the GPS coordinates of each device, since a prototype is implemented and installed on each device.Furthermore sentiment analysis algorithms face several challenging problems to evaluate the real feelings expressed by text.Our scheme avoids these problems, since each user selects her/his feeling/state.2) the proposed scheme uses 6 feelings/states for each location, while [32] has three different evaluations (positive, neutral, negative).Thus our scheme is more fine-grained and more accurate in feelings' expression than [32].3) the proposed scheme detects similar places without considering user profiles, while [32] considers user profiles.Thus [32] provides a more personalized service.4) both schemes consider the geographical distance of locations.The proposed scheme performs a prefiltering by excluding locations that are out of the adopted radius, while [32] incorporates the factor of geographical distance between user and POI in the optimization problem.5) the proposed scheme focuses on the many-to-many location similarity problem, while [32] focuses on the one-to-one problem (user-POI pairs, e.g.user 1 is matched to POI 1 , user 1 is matched to POI 2 , user 1 is matched to POI 3 etc.sequentially).In the case of [32], the many-to-many problem could somehow be described like: there is a group of friends that would like to select POIs to visit.They have different profiles and different preferences, but they would like to go out altogether.This is what our scheme can solve, which is very common in the real world (many people do not go out alone, but together with other people).Other relevant schemes include [33]- [38].However also these works do not consider feelings' rating or detection of similar locations.
On the other hand, the two predominant well-known services Google Maps Rating and Tripadvisor are highly related to the proposed scheme.In particular both services enable users to write reviews and rate places.In both services also, reviews may be taken down if others label them inaccurate or if they do not comply with the respective review policies of the two companies.Even though very interesting and extremely valuable, these two services could be further improved.In this work we implement an innovative real-time geographical locations' rating scheme, which proposes some fundamental enhancements to the aforementioned review and rating services.The overall approach is corollary and a capstone of our previous works [39]- [42], where crowdsensing, crowdsourcing and user generated content analysis methods have been incorporated in order to detect similar destinations, to annotate images from different geographical locations and to provide multimedia content to users, based on their geo-locations.

III. DEFINITIONS & METRICS
Let us denote as U the set of all registered users u i , i = 1, 2,. . ., N, of the proposed locations feelings' rating service: At each time instance, every user can express a feeling for the location she/he currently is, from the set of feelings/states: where fl 1 is the first feeling/state, e.g.happy, fl 2 the second feeling/state e.g.sad etc.Note that the current implementation of the proposed system allows for only one feeling/state per time instance (e.g. the user cannot send the feelings/states ''happy'' and ''angry'' at the same time instance).For each feeling/state a strength is also allocated, from the set of feelings'/states' strengths: For example ''happy, 5/5'' is stronger than ''happy, 3/5'', where the feeling/state is followed by a number corresponding to the feeling's/state's strength.
Let us now assume that user u i ∈U, rates a specific geographical location l at a specific time t j , where, in this paper, t j is equal to the Coordinated Universal Time plus two hours (UTC+2).The rating provided by u i is denoted as EV u i,l,t j (fl n , st o ), where fl n denotes the n-th feeling/state and st o denotes the o-th strength of the feeling/state.Axiom 1: Each u i can be only at one l for each specific t j .Axiom 2: Each u i can send only one rating for a specific l at each specific t j .
Axiom 3: Each l can be rated by several users u i ∈U for each specific t j .
Additionally each u i may visit one l many times and can submit different ratings each time.Furthermore our proposed rating service also allows for rapid mood changes and thus rapid rating changes.For example imagine that u i is in a football stadium and at time t j her/his team scores.At this time instance u i could evaluate l by (Fig. 1): Now imagine that after a while violent incidents breaks out in the football stadium.In this case u i could rate l by (Fig. 2): All ratings from all users, for a specific location l and for all time instances are gathered in the ratings set E l of this specific location: Set E l is by default sorted according to time i.e. from most recent to oldest rating, but, if needed e.g. for faster access to its data, it can also be sorted according to: (a) user id (b) rating i.e. from best to worst score (5/5, 4/5 . . ., 1/5), irrespectively of feeling/state, (c) feeling/state (e.g.happy, sad, angry etc.) and (d) score and expressed feeling/state i.e. happy 5/5, happy 4/5 etc, sad 5/5, sad 4/5 etc, angry etc.Let us now estimate the set of all appearances of a specific feeling/state at a specific strength in set E l : Time instances t p and t q can cover either all ratings or just ratings of a specific time interval.Then the average strength AvS of a specific feeling/state fl r at a specific location l is defined as: And the average normalized strength is given by: where card(.)denotes the cardinality of the corresponding set.In the general case 0 ≤AvN l fl r ≤ 1, where value 0 is possible only when the specific feeling/state does not appear in the ratings set E l .
Then for a location l, the average feelings/states vector is defined as: where AF l fl 1 , AF l fl 2 , . . ., AF l fl w represent the average values of the feelings/states included in set FL respectively, regarding location l and from the first time that the location was evaluated (t 0 ) until time instance t j .More specifically: where nc is a normalization coefficient so that sum AVF l 1 and it is calculated by: Here two things should be noted: (a) the average feeling vector for a location l can be estimated for any time interval [t i , t k ], by filtering out all evaluations from E l that do not belong to the specified time interval, (b) aggregation of all evaluations to provide the average feelings/states vector AVF l may lead to data loss.Data loss can be avoided only by avoiding data aggregation.However in this case the location descriptor will continuously increase every time a new evaluation will arrive and then we will have to fight against the curse of dimensionality.
Finally the set O AVF of all average feelings' vectors of all currently visited locations l i , i∈W= {1, . . ., n}, is formed as:

IV. DETECTION OF SIMILAR GEOGRAPHICAL LOCATIONS
In this paper similarity of geographical locations is estimated in terms of the feelings/states that the location has generated to its visitors and not in terms of geographical features (e.g.elevation, terrain types, physical factors of the environment etc.) or in terms of comments that users may send together with their feelings' rating.For example, even though Grand Canyon (Arizona, USA) and Whitehaven Beach (Whitsunday Islands, Australia) present completely different geographical features, they may be very similar in terms of feelings/states expressed by their visitors (e.g.Wow 5/5, Love 5/5).Additionally, the idea of estimating similar locations comes from the need of people to visit/avoid places that generate similar good/bad feelings/states.It can work like an automatic crowd-based place recommender/avoider, based on our previous (several) visits to different locations.
In order to find similar geographical locations, in this paper a correlation criterion is maximized, so that the selected locations match in terms of feelings/states expressed.In particular, the similar geographical locations are those with the maximum correlation among all considered locations (which may be a subset of all currently evaluated locations that are represented in O AVF ).Now let us suppose that a user has visited a location l i which is currently represented by vector AVF l i [t 0 ,t j ] and she/he wants to find out those T R locations (including l i ) that are most similar to l i and to the rest T R -1 locations.Parameter T R can either be given by the user in an interactive way or be a priori set.Assuming that the user does not limit the search to a specific distance from l i and without loss of generality, locations' similarity expands to all locations included in O AVF .Then the correlation coefficient of the average feelings' vectors is defined as (for simplicity time interval is omitted): where is the covariance of the two vectors, is the average feelings/states vector of O AVF and is the variance of AVF l i .In order to define a measure of correlation between T R average feelings' vectors, we first define the following index vector: where is the subset of W T R which contains all sorted index vectors b.Thus, each index vector b = (b 1 , . . ., b T R ) corresponds to a set of geographical locations' numbers.The correlation measure of the average feelings' vectors AVF l i , i = b 1 , . . ., b T R can then be defined as: Based on the aforementioned metrics and definitions, it is obvious that the procedure of searching for a set of T R maximally correlated average feelings' vectors is equivalent to searching for an index vector b that maximizes the correlation measure R F (b).This searching procedure is limited within subset Y, since index vectors are used to form sets of average feelings' vectors.In this case any permutations of b's elements will lead to the same vector sets.As a result, the set of the T R average feelings' vectors that present maximum correlation, can be represented by: Here it should be stressed and clarified that the proposed scheme does not only consider the one-to-one similarity problem but also the many-to-many similarity problem, which has not been considered, especially for geographic locations.
To make it more clear, two examples of the many-to-many similarity problem are provided: (a) a user may have evaluated locations A, B, C etc. in the past (many locations).Now the user may want to find similar locations to A and to B and to C etc. and not just to A, or just to B, or just to C etc.. To give an example: the user likes Mykonos, Santorini and Corfu.She/he would like to find other islands that are similar to Mykonos AND to Santorini AND to Corfu.She/He does not want to find islands that are only similar to Mykonos, or only similar to Santorini or only similar to Corfu.The proposed scheme returns locations D, E, F, G etc (many locations) in an optimal way.In this case the group of locations A, B, C, D, E, F, G etc. provides maximum correlation (maximum similarity).(b) there is a group of friends that would like to select POIs to visit.They have different profiles, different preferences and have provided different evaluations for the same POIs.But now they would like to go out altogether.Thus they would like to find places that will satisfy (as much as possible) all members of the group.This is what our scheme can solve, which is very common in the real world (many people do not go out alone, but together with other people).

A. MAXIMIZATION OF CORRELATION MEASURE: A GENETIC ALGORITHM APPROACH
In case of a limited number of evaluated locations, the maximum value of R F (b) can be estimated by exhaustive search.However, in reality this is not possible, since millions of different locations may be included in O AVF , while the multidimensional space Y includes all possible combinations of different locations.Complexity can be drastically reduced if a logarithmic search algorithm [43] is incorporated, which, however, will generally converge to sub-optimal solutions.For these reasons, a genetic algorithm (GA) [44] approach is adopted.In particular, possible solutions of the optimization problem, i.e., sets of evaluated geographical locations, are represented by chromosomes.The genetic material of In traditional GA schemes the initial populations are randomly generated.In this paper and in order to increase the possibility of locating sets of average feelings' vectors with significant correlation, we exploit the structure of AVF l and sort O AVF according to the prevailing feeling/state.Then the vectors that are in the neighborhood (according to the sorting procedure) of the vector under consideration are used for generating the initial population.
In our maximization problem and for a given population, the correlation measure R F (b) plays the role of the objective function that estimates the performance of all chromosomes b i , i = 1, . . ., L. However, objective values are mapped to fitness values by a fitness function that follows a linear normalization scheme.More specifically, the b i chromosomes are sorted in descending order of R F (b i ), since our aim is to maximize the objective function.Towards this direction let t(b i ) ∈ {1, . . ., L} represent the rank of chromosome b i , i =1, . . ., L. If we define an arbitrary fitness value fv c for the chromosome that provides the best performance, then the fitness of the i-th chromosome can be given by the following linear function: where fv d is a decrement rate.By this way the population's average objective value is mapped to the average fitness [45].
After calculating the fitness values, fv(b i ), i = 1, . . ., L, for all members of the current population, the parent selection procedure is then activated so that fitter chromosomes provide higher numbers of offspring and thus have better chances to survive during the next GA cycle.In this paper parent selection is accomplished by a proportionate scheme, which is implemented by the roulette wheel selection procedure [46].A set of new chromosomes is then produced by mating the selected parents and applying a crossover operator.The parents' genetic material is randomly combined to produce the genetic material of the offspring.A generalized uniform crossover approach is adopted by our scheme, where each parent gene is considered to be a potential crossover point.
Then the new chromosomes undergo mutation.In particular, each offspring b i is replaced by a randomly generated gene b i ∈ W= {1, . . ., T R }, if a probability test is passed.For a given population B(m), m≥0 and once new chromosomes are generated, the next population B(m+1) is formed by: (a) inserting these new chromosomes into B(m) and (b) deleting older chromosomes so that the next population also consists of L members.The exact number of replacements of the GA [44].All aforementioned details refer to one GA cycle.Many GA cycles should be carried out until convergence (usually when the best fitness remains constant).

V. SYSTEM IMPLEMENTATION
In order to evaluate the proposed scheme, a prototype website and a smartphone application were implemented.The main implementation steps are described below.
Step 1: One 4-CPU Intel Xeon E5-4610 v2 @ 2.3 GHZ ×64, 128 GB RAM, Windows server 2012 R2, server was used Step 2: A symmetric internet line of 50 Mbps (50 Mbps for uploading and downloading) was used Step 3: The free WAMPSERVER software [35] (apache web server, MySql Server) was installed on the server Step 4: The Content Management System Drupal version 7 was installed on the server.The Drupal architecture is based on a sequence of interconnected files, developed in Javascript, which offers huge automation capabilities, simplicity of design and great compatibility with all browsers.Additionally, the selection of Drupal instead of Joomla, Wordpress or other CMSs was mainly based on two factors: (a) feasibility, i.e. uniform configuration of the application and the blog for their greatest possible compatibility and (b) production of executable application files, suitable for android and iOS devices.The Drupalgap module, which satisfies the two above-mentioned factors, provides the following capabilities: 1) It creates an independent -fully compatible and interconnected -application with the underlying website.Joomla and Wordpress provide a similar feature, but this results in an application -packing the blog-web site in a responsive format (friendly version for mobile devices) 2) It creates distinct sub-modules of the application 3) It supports Interface Compatibility -Interconnection to Adobe's Free Phonegap Software, which converts the application into executable .apkfiles for Android operating systems and .apafor IOS operating systems (iphone, ipad) 4) It allows for further customization of the application itself through its open source code 5) It enables app simulation for Android and IOS devices in Google Chrome due to its interface compatibility with the browser's Ripple Emulator plug-in Step 5: The webpage was implemented providing six different feelings/states (Like, Love, Haha, Wow, Sad, Angry).Each feeling/state was accompanied by the respective emoticon and five strength levels (see Fig. 1 & 2).Towards this direction, Drupal's Fivestar module was properly modified.The webpage also offered several other features (e.g.New User Registration, Login, Forum, Post on Social Media, Download smartphone app, Detection of geographical location, Top four articles, Content Approval etc.).Also Drupal's Get Locations module for creating the latitude-longitude polygons was used, together with Google Map API v3 and a Geojson file (which contained pairs of coordinates in polygons.Polygons were coded in Javascript).
Step 6: The app for smartphones was designed so that it enables users to take several actions such as share taken photographs, evaluate locations using one of the six available feelings/states (evaluations could be accompanied by comments in text form), post on social media etc. PhoneGap was initially used to convert the application files into an executable .apkfile (for Android).In order to convert the application files into a .apafile (for iOS) a virtual Macintosh with Sierra OS was created on the server and Xcode was installed which, together with Phonegap, supported the conversion of files to a .apafile.
Due to space limitations, a small part of the modified Fivestar module and the Map Pageshow callback function are provided in Appendix A. Full code is available on request.

A. DATA ACQUISITION AND TOP LOCATIONS
In this section, the performance of the proposed scheme is investigated, while a comparison to Google Maps Rating and Tripadvisor is also provided.In particular the application has been publicized on the Social Media groups of the Department of Business Administration of the University of West Attica [47], [48].As of 14/12/2018, the two groups had more than 1,200 members in total, mainly students between 18 and 26 years old, which are technology-literate.
As a result, the .apkapp was downloaded and installed by 198 users while the .apaapp by 129 users.In total 327 users have participated in our experiments.Data were collected for 90 days, between 25/06/2018 and 03/08/2018 and between 26/10/2018 and 14/12/2018.Overall 3,129 unique locations have been evaluated, receiving 67,689 evaluations in total (∼752 evaluations/day, 93.5% of which between 11:00 a.m. and 23:00 p.m., UTC+2), i.e. 9.57 unique locations per user in total or 34.77 unique locations per day on average.
Here it should be mentioned that a manual process has been followed to provide the 3,129 unique locations, since in some cases different evaluations were received from very close locations (e.g.(latitude, longitude) pairs (37.978787, 23.712390) and (37.978788, 23.712416).Both locations are outside a concert hall in Athens).This process can be automated if experts manually segment areas into clusters.Alternatively, automatic clustering methods could be tested (Connectivity-based, Centroid-based, Distributionbased, Density-based etc.).Future research could examine such approaches and evaluate the error induced by each clustering method.Also case-specific studies can be carried out (e.g.clustering in a football stadium, clustering on a road, clustering on the sea surface etc.).For example in a football stadium, the one bleacher may accommodate the fans of one team and the other bleacher the fans of the other team.
In this case, even though the stadium may cover several different (latitude, longitude) pairs, in reality it could be split into two parts (two ''locations'' of different feelings/ states).Now regarding the 3,129 unique places, about 90.9% (2,847) of these were located in a radius of 30 km from the center of Athens.For clarity reasons, Figure 3 presents the top 50 locations in terms of total number of evaluations for the whole period.Each evaluation included a feeling/state among the six available, a strength among the five available and a timestamp (UTC+2).Figures 4(a , where [t 0 , t j ] is the 90 days interval.Here it should be mentioned that a top-evaluated location may not be included in a top-feeling location.This is due to the fact that locations which receive several different evaluations under different circumstances, may provide an AVF l [t 0 ,t j ] that is not among the top for a specific feeling/state.Additionally the habits of students are typically different from the habits of tourists, senior citizens, workers etc.This is why several of the evaluated locations include coffee shops, theaters, cinemas, entertainment places etc.

B. DETECTION OF SIMILAR LOCATIONS
As already stated, similarity of geographical locations is estimated in terms of the feelings/states that the location generates to its visitors at a particular time period and not in terms of geographical features (e.g.elevation, terrain types etc.).There are several different scenarios and challenges in detecting similar locations.In the simplest scenario: (a) a user is currently at a specific location, (b) an average feelings/states vector AVF l i [t 0 ,t j ] has been estimated for the specific location, where t 0 /t j is the time instance of the first/last evaluation that the specific location has received, (c) the user looks for the location l k , with an AVF l k [t 0 ,t j ] that maximizes r i,k (see Eq. ( 12)).
In this scenario the solution is provided by calculating the correlation between AVF l i [t 0 ,t j ] and all other vectors of O AVF and keeping the maximum correlation vector.In our experiments however more generic scenarios are presented.
First the long-term similarity scenario is examined in two different cases and then the short-term similarity is also considered.

1) LONG-TERM SIMILARITY: CASE I
In Case I of long-term similarity the experimental parameters were: (i) a user was currently at a specific location l i , (ii) the user was looking for T R -1 other locations that were not only similar to l i (in the long term) but also similar to each other, so that R F (b) (see Eq. ( 18)) is maximized.Towards this direction, initially an index has been assigned to each location according to the time of first evaluation (e.g. the first location that has been evaluated was assigned index ''1'' (l 1 ), the second location was assigned index ''2'' (l 2 ) etc.).In this way, in each experiment the vector b (see Eq. ( 16)) was easily formed.Next the number of locations has been selected to be T R = 6, mainly for two reasons: (a) suggestion of 5 similar places is not overwhelming to users, (b) the more places we suggest the less their correlation, as it can be observed in Figure 5.
In particular Figure 5 presents convergence of the GA, for the same initial location l 78 , which was randomly selected among the 3,129 unique locations, and for T R = 6, 7 and 8 respectively.Additionally, (a) 100 GA cycles have been carried out, (b) AVF l 78 [t 0 ,t j ] was estimated for t 0 = [15/11/2018, 00:01 a.m -UTC+2] and t j = [14/12/2018, 23:59 p.m. -UTC+2], (c) average vectors for all other locations have also been estimated for the same time interval, (d) a radius of 3 km from l 78 has been used (an area of ∼10.91 square miles or 28.26 square kilometers).This is due to the fact that usually, users look for similar locations near their current location.Thus even though locations presenting more similarity to the current location may exist outside the radius of 3 km, they are not considered.As a result the computational cost is also reduced, (e) the correlation measure is normalized in the interval [0, 1], (f) every time the GA population includes eighty (80) chromosomes (different locations), while at each GA iteration, thirty parents are selected, (g) for possibly accelerating the convergence time of the GA, the initial population included the top eighty (80) locations in terms of correlation (r 78,k -see Eq. 12).

2) LONG-TERM SIMILARITY: CASE II
For the experiment of Case II a significant number of repetitions have been carried out (results are provided for 560 repetitions).In this case the experimental parameters were: (i) a group of friends has planned to go out together, (ii) these friends had different preferences but they were looking for locations that would possibly satisfy all of them, (iii) for each repetition, each one of the friends provided to the similarity algorithm two locations that she/he preferred most, (iv) T R = 6, 7 and 8 was examined based on the size of each group, (v) each group of friends selected a meeting location as the center of search, (vi) all the rest of the parameters were similar to Case I (i.e. 100 GA cycles were performed, a radius of 3 km around the center of search was used, the GA population was eighty chromosomes etc.).
Convergence patterns of the GA algorithm were similar to Figures 5 and 6.The average correlation measure R F (b) for all 560 repetitions was equal to 0.33, 0.25 and 0.19 for T R = 6, 7 and 8 respectively.Furthermore in each repetition of the experiment, each member of the group of friends evaluated the proposed location(s) that the group has visited (to check if  the proposed locations were really satisfactory or not).By this way the performance of the location similarity algorithm was also assessed by the end users.Aggregated average results of AVF l are presented in Table 1 for the three similar locations estimated by the proposed similarity algorithm.
As it can be observed in Table 1, user satisfaction reaches 88% for the first estimated similar location (L 1 ), 82% for the second (L 2 ) and 80% for the third (L 3 ).

3) SHORT-TERM SIMILARITY
Finally another experiment has also been performed regarding short-term similarity, where in this case: (a) a user is currently at a specific location l i , (b) the user is looking for T R -1 other locations that are not only similar to l i (in the short-term) but also similar to each other, so that R F (b) is maximized.This scenario examines short-term decisions.In our case a user has gotten stuck in the traffic, she/he expresses angriness and would like to avoid locations that currently make other users angry.Differences with the longterm scenario include: (a) an average vector for the current location is not estimated, but only the current evaluation of the specific user is considered, (b) average vectors for all other locations are estimated for a very short term interval.
In our experiments and due to the limited number of users, the short term interval is set to the last one hour, ending at the current evaluation of the specific user (in the presented results, data for the time interval Of course within one hour several things may change (especially in traffic), and a shorter interval may be more appropriate.However in this case many more users and evaluations per minute are needed to provide accurate results, (c) in the case of traffic jams, only locations on roads should be examined.However the current version of the proposed scheme does not distinguish among location types (e.g.theater, road, parking, supermarket, restaurant etc.).The rest of the parameters were the same as in the long-term scenario, except of the GA population (which was limited to 37, since only 37 evaluations were made in the aforementioned time interval, in a radius of 3 km from the evaluation of the user under examination) and the number of parents (12 in this case).Results of the GA convergence are shown in Figure 6, where convergence is faster compared to the long-term scenario (the examined set has fewer samples).Moreover the five similar locations are mapped in Figure 7 and their (latitude, longitude) pairs are shown in Table 2.As it can be observed, only one of the five similar locations is related to a road.Thus, addition of location type to the maximization process could solve this problem.

C. COMPARISON WITH GOOGLE MAPS RATING AND TRIPADVISOR
To the best of the authors' knowledge, there is not any other real-time feelings'/states' locations rating and similarity estimation algorithm/platform.Google Maps Rating and Tripadvisor are the two predominant well-known services.Both enable users to evaluate every possible location, by using a five-level rating scale.In particular Google Maps Rating provides a five-stars scale, while Tripadvisor provides a fiveconcentric circles scale.There is not any provision for expression of feelings.Both approaches estimate and show only the average evaluation score.Also both approaches show the total number of reviews.Google Maps Rating provides a ''Sort by'' field, where reviews can be sorted by ''Most relevant'', ''Newest'', ''Highest rating'' and ''Lowest rating''.Tripadvisor offers four review-search fields ''Traveler rating'' (with sub-fields: Excellent, Very good, Average, Poor, Terrible), ''Traveler type'' (with sub-fields: Families, Couples, Solo, Business, Friends), ''Time of year'' (with sub-fields: Mar-May, Jun-Aug, Sep-Nov, Dec-Feb) and ''Language''.
In order to compare the proposed approach to Google Maps Rating and Tripadvisor the following assumptions are made: (a) Sad and Angry feelings in the proposed approach correspond to 1-star/concentric circle ratings (b) Like, Love, Haha and Wow with strength one/two/three/four/five, correspond to 1/2/3/4/5-star/concentric circle ratings respectively.Next two specific cases are presented.In the first case the National Archaeological Museum of Athens is examined for a short interval [16/ Figure 8 shows the plots of the average evaluation scores of the three compared schemes and for the specific time period.Here it should be mentioned that: (a) the average score is per day and it is estimated based on unequal numbers of evaluations among the three schemes, (b) in the proposed scheme no evaluation was made on 17/07/2018 and linear interpolation was used for this specific date.Additionally in Figure 9 the boxplot of the total number of evaluations per scheme is provided.
As it can be observed from Figures 8 and 9, even though the average scores of the three approaches are similar, timebehaviors are different.In particular and in order to explain these scoring pattern variations, every year the Athens Open Air Film Festival is organized in Athens.Last year the 8th  Athens Open Air Film Festival had planned to show the film ''The Remains of the Day'' (James Ivory, 1993) at the National Archaeological Museum of Athens, on the 18 th of July 2018.However, on the 18 th of July 2018 riots had erupted in the neighborhood of the National Archaeological Museum of Athens, leading to the cancellation of the film projection.As a result, the users have provided bad evaluations for this location (1.3/5.0 for the proposed approach).On the other hand the average score for this day (18/7/2018) were 4.5/5.0 for Google Maps Rating and 5.0/5.0 for Tripadvisor.Trends at this specific location were not captured by Google Maps Rating or Tripadvisor, since both are mainly used by tourists who have different habits, interests and perception of the city compared to locals.This is why the evaluations (and thus the respective patterns are different).On the 19 th of July 2018 the score of the proposed approach returns back to the normal levels (4.6/5.0).The two aforementioned analyzed cases (i.e.National Archaeological Museum of Athens and Syntagma Square of Athens) provide a comparison among the schemes in qualitative terms.In order to also quantitatively compare the three schemes, initially 1,091 locations among the 3,129 unique locations have been distinguished.The 1,091 locations have been evaluated by all three schemes.From them, 327 locations have been excluded from our statistical results, due to very limited numbers of evaluations.The remaining 764 locations have been examined for unexpected events (either good or bad), during the total evaluation period.In 196 locations no unexpected events happened, thus 568 locations were considered for statistical evaluation.In these 568 locations, 1,818 events happened during the evaluation period, 691 of which were positive and 1,127 negative.Figure 12 shows the results, where linear prediction is incorporated as the baseline method.Here it should be mentioned that evaluations of all three schemes (Google, Tripadvisor, Proposed) are provided to the linear prediction method.In this experiment we assume that an event is correctly detected only when the evaluation score changes for at least 40% (either increases or decreases) compared to the average of the previous and next evaluation, for a one-day time window.In particular Figure 12(a): (a) exhibits the success of the compared schemes in detecting positive events (TP: True Positive), (b) exhibits the success of the compared schemes in detecting negative events (TN: True Negative) and (c) exhibits the overall success of the compared schemes in detecting both positive and negative events (TA: True All) On the other hand Figure 12(b): (a) presents false positives (FP: False Positive), in case when a scheme detects a positive event that did not happen (b) presents false negatives (FN: False Negative), in case when a scheme detects a negative event that did not happen and (c) exhibits the overall failure of the compared schemes (false positive and false negative events -FA: False All).As it can be observed the proposed scheme provides the best performance both in correctly detecting events (66.83%) and in making the fewest mistakes (15.4%).Linear prediction follows with (25.46% and 18.53% respectively).The percentages for Google Maps Rating are 19.58% and 21.01% while Tripadvisor is very close with 17.99% and 18.92% respectively.Here it should be mentioned that the performance of linear prediction heavily depends on the provided evaluations.Overall both Google Maps Rating and Tripadvisor do not capture short-term trends of places, a functionality that the Proposed scheme conceptualizes and offers.Additionally both schemes mainly attract tourists, who have different habits, interests and perception of the city compared to locals.Of course Google Maps Rating and Tripadvisor are very useful and important services, since they provide average scores and many different comments that can help people to schedule their activities.Finally the rapid mood change for a specific location, does not upgrade/downgrade the location, which may not/may be a very worthy place to visit.It just provides a hint to visit/avoid the specific location at the specific time.

VII. CONCLUSION
In this paper an innovative locations' rating scheme is presented.Feelings/states evaluations, volunteered crowdsensing and crowdsourcing and similarity among locations are the main topics of this research.Experimental results on real world data, as well as comparisons with Google Maps Rating and Tripadvisor, illustrate the merits and limitations of each technology.In the future many more related aspects could be examined.For example: (a) how about considering personalization and user modeling?Algorithms like the ones presented in [26] and [29] could be examined.(b) what happens when a user has several different portable devices ?Can she/he send several votes from the same location (skewing the ranking) ?Location-based authentication methods could be adopted.(c) what happens in case the system receives 10,000 ratings/second ?Stress tests should be carried out.(d) What about privacy of the users ?Privacy algorithms could be examined (e) Finally, more feelings/states could be such as danger, fear, anxiety, boredom, guilt, shyness, surprise, trust, self-confidence etc. so that we get a much more accurate picture of what happens in each location.
chromosomes consists of location numbers (indices).Thus index vectors of the form b = (b 1 , . . ., b T R ) ∈ W T R represent chromosomes and each index vector follows an integer number encoding scheme.More specifically integer numbers are used for representing the genes b i ∈W, i = 1, . . ., T R .At cycle zero, an initial population of L chromosomes, B(0)=(b 1 , . . ., b L ) is generated and used for the creation of new populations B(m), m>0.In order to create B(m) at the m-th cycle, different operations are performed on B(m-1).These operations (described next) are carried out again and again, until B(m) converges to an optimal solution.

FIGURE 3 .
FIGURE 3. Top 50 locations in terms of total number of evaluations (marked with a red circle ''T50'').

FIGURE 4 .
FIGURE 4. (a) Top 5 locations of the feeling/state ''Like'' (b) top 5 locations of the feeling/state ''Love'' (c) top 5 locations of the feeling/state ''Haha'' (d) top 5 locations of the feeling/state ''Wow'' (e) top 5 locations of the feeling/state ''Sad'' (f) top 5 locations of the feeling/state ''Angry''.In all cases the respective emoticon is used to pinpoint the location.

FIGURE 5 .
FIGURE 5. Convergence of the GA, for T R = 6, 7 and 8 -long term experiment.

FIGURE 6 .TABLE 1 .
FIGURE 6. Convergence of the GA, for T R = 6 -short term experiment.TABLE 1.Average vector AVF l for the three similar locations (L 1 , L 2 , L 3 ) estimated by the proposed similarity algorithm.

FIGURE 7 .
FIGURE 7. The five similar locations (marked with S) in a radius of 3km.Current location is marked with the location icon in black background.

FIGURE 9 .
FIGURE 9. Boxplot of the evaluation scores of the national archaeological museum of athens, from Google Maps rating, tripadvisor and the proposed scheme (time period: 16-20/07/2018).

FIGURE 11 .
FIGURE 11.Boxplot of the evaluation scores of the national archaeological museum of athens, from Google Maps rating, tripadvisor and the proposed scheme (time period: 16-20/07/2018).

FIGURE 12 .
FIGURE 12. (a) Detection of True Positive, Negative and All Events for all schemes (b) Detection of False Positive, Negative and All Events for all schemes.

TABLE 2 .
The (latitude, longitude) pairs of the five similar locations of Figure7.