Understanding the Impact of Weather for POI Recommendations ⇤ Christoph Trattner Alexander Oberegger Know-Center, Austria TUG, Austria Lukas Eberhard Denis Parra Leandro Marinho TUG, Austria PUC, Chile UFCG, Brasil ABSTRACT with millions of subscribers doing millions of check-ins ev- POI recommender systems for location-based social network eryday all over the world2 . This vast amount of check-in services, such as Foursquare or Yelp, have gained tremen- data, publicly available through Foursquare’s data access dous popularity in the past few years. Much work has been APIs, has inspired many researchers to investigate human dedicated into improving recommendation services in such mobility patterns and behaviors with the aim of assisting systems by integrating di↵erent features that are assumed users by means of personalized POI (point of interest) rec- to have an impact on people’s preferences for POIs, such as ommendation services [15, 16]. time and geolocation. Yet, little attention has been paid to Problem Statement. The problem we address in this the impact of weather on the users’ final decision to visit a paper is the POI recommendation problem. Hence, given recommended POI. In this paper we contribute to this area a user u and their check-in history Lu , i.e., the POIs that of research by presenting the first results of a study that aims they have visited in the past, and current weather condi- to recommend POIs based on weather data. To this end, we tions C = {c1 , . . . , c|C| }, where ci are weather features such extend the state-of-the-art Rank-GeoFM POI recommender as temperature, wind speed, pressure, etc., we want to rec- algorithm with additional weather-related features, such as ommend the POIs L̂u = {l1 , . . . , l|L| } that they will likely temperature, cloud cover, humidity and precipitation inten- visit in the future that are not in Lu . sity. We show that using weather data not only significantly Objective. Most of the existing approaches on POI rec- increases the recommendation accuracy in comparison to the ommendation exploit three main factors (aka contexts) of original algorithm, but also outperforms its time-based vari- the data, namely, social, time and geolocation [5, 10, 15]. ant. Furthermore, we present the magnitude of impact of While these approaches work reasonably well, little atten- each feature on the recommendation quality, showing the tion has been paid to weather, a factor that may potentially need to study the weather context in more detail in the light have a major impact on users’ decisions about visiting a of POI recommendation systems. POI or not. For example, if it is raining in a certain place in a certain period of time, the user may prefer to check-in indoor POIs. Keywords In this paper we contribute to this area of research by POI Recommender Systems; Location-based services; Weather- presenting the first results of a recently started project that Context exploits weather data to recommend, for a given user within a given city, the POIs that they will likely visit in the fu- ture. To this end, we extract several weather features based 1. INTRODUCTION on data collected from forecast.io such as temperature, cloud Location-based social networks (LBSN) enable users to cover, humidity or precipitation intensity, and feed it into a check-in and share places and relevant content, such as pho- state-of-the-art POI recommender algorithm called Rank- tos, tips and comments that help other users in exploring GeoFM [10]. The reason why we decided to build our ap- novel and interesting places in which they might not have proach on top of this algorithm is twofold: (i) Rank-GeoFM been before. Foursquare1 , for example, is a popular LBSN has shown to outperform other strong baselines from the lit- ⇤Corresponding author: ctrattner@know-center.at erature and (ii) it is very easy to extend it with additional 1 contextual data. https://foursquare.com/ Research Questions. To drive our research the follow- ing three research questions were defined: • RQ1. Do weather conditions have a relation with the check-in behavior of Foursquare users? • RQ2. Is it possible to improve current POI recom- mendation quality using these weather features? Copyright held by the author(s). • RQ3. Which weather features provide the highest im- RecTour 2016 - Workshop on Recommenders in Tourism held in conjunc- pact on the recommendations? tion with the 10th ACM Conference on Recommender Systems (RecSys), 2 September 15, 2016, Boston, MA, USA. https://foursquare.com/infographics/10million City #Check-Ins #Venues #Users Sparsity Sym. Description Minneapolis 37,737 797 436 89.1% U set of users u1 , u2 , ..., u|U | Boston 42,956 1141 637 94.3% L set of POIs l1 , l2 , ..., l|L| Miami 29,222 796 410 91.0% F Cf set of classes for feature f Honolulu 16,042 410 173 77.4% F set of weather feature classes f1 , f2 , ..., f|F Cf | ⇥ latent model parameters containing the learned weights Table 1: Basic statistics of the dataset. {L(1) , L(2) , L(3) , U (1) , U (2) , F (1) } for locations, users and weather features. Xul |U | ⇥ |L| matrix containing the check-ins of users at POIs. Contributions. To the best of our knowledge, this is the Xulc |U | ⇥ |L| ⇥ |F Cf | matrix containing the check-ins of users at first paper that investigates in detail the extent to which POIs at a specific feature class c. D1 user-POI pairs: (u, l)|xul > 0. weather features such as temperature, cloud cover, humidity D2 user-POI-feature class triples: (u, l, c)|xulc > 0. or precipitation intensity impact on users’ check-in behav- W geographical probability matrix of size |L|x|L| where wll0 iors and how these features perform in the context of POI contains the probability of l0 being visited after l has been visited according to their geographical distance. wll0 = (0.5+ recommender systems. Although there is literature showing d(l, l0 )) 1) where d(l, l0 ) is the geographical distance between that POI recommender systems can be improved by using the latitude and longitude of l and l0 . some kind of weather context such as e.g. temperature, it is WI probability that a weather feature class c is influenced by not clear yet, how much they add or what type of weather feature class c0 . wicc0 = cos sim(c, c0 ). Nk (l) set of k nearest neighbors of POI l. feature is the most/least useful one. Another contribution yul the recommendation score of user u and POI l. of this paper is the introduction of a weather-aware rec- yulc the recommendation score of user u, POI l and weather fea- ommender method that builds upon a very strong state-of- ture class c. I(·) indicator function returning I(a) = 1 when a is true and 0 the-art POI recommender system called Rank-GeoFM. The otherwise. method is implemented and embedded into the very pop- ✏ margin to soften ranking incompatibility. ular recommender framework MyMediaLite [7] and can be w learning rate for updates on weather latent parameters. g learning rate for updates on latent parameters from base ap- downloaded for free from our GitHub repository (details in proach. Section 8). E(·) a function that turns the rating incompatibility Outline. The structure of this paper is as follows: In Sec- Incomp(yulc , ✏), that counts the number of locations l0 2 L that should be ranked lower than l at the current tion 2 we highlight relevant related work in this field. Sec- weather context c and user Pu but are ranked higher by the tion 3 describes how we enriched Rank-GeoFM with weather model, into a loss E(r) = r 1 i=1 i . data. Section 4 describes the experimental setup and presents ucll0 function to approximate the indicator function with a contin- 1 uous sigmoid function s(a) = 1+exp( . ucll0 = s(yul0 c + results from our empirical analysis. Section 5 presents in- a) ✏ yulc )(1 s(yul0 c + ✏ yulc )) sights on the results obtained with our weather-aware rec- |L| b n c if the nth location l0 was ranked incorrect by the model the ommender approach. Finally, Sections 6 and 7 conclude the |L| expactation is that overall b n c locations are ranked incor- paper with a summary of our main findings and future di- rect. rections of the work. g, µ auxiliary variable that save partial results of the calculation of the stochastic gradient. 2. RELATED WORK With the advent of LBSNs, POI recommendation rapidly Table 2: The notations used to describe Rank-GeoFM and became an active area of research within the recommender the incorporation of the weather context. systems, machine learning and Geographic Information Sys- tems research communities [2]. Most of the existing research works in this area exploit some sort of combination between in a more recent and state-of-the-art algorithm, and we some (or all) of the following data sources: check-in history, also provide details of which weather features contribute social relations (e.g. friendship relations), time and geolo- the most to the recommender performance. In an exten- cations [1, 5, 6, 8, 10, 13, 15]. While these di↵erent sources sion of their initial work, Braunhofer et al. [4] implemented of data (aka contexts) a↵ect the user’s decision on visiting and evaluated a context-aware recommender system which a POI in di↵erent ways, weather data, which according to uses weather data. They find that the model which lever- common sense may have a great influence on this decision, ages the weather context outperformed the version without are still rarely used. it. Although more similar to our current work, they did not Martin et al. [11] proposed a mobile application which provide a detailed feature analysis as the present article. architecture considered the use of weather data to person- In summary, compared to previous works which have used alize a geocoding mobile service, but no implementation or weather as a contextual factor for recommendation systems, evaluation was presented. A similar contribution was done we provide detailed information about the recommendation by Meehan et al. [12], who proposed a hybrid recommender algorithm and we contribute an implementation extending system based on time, weather and media sentiment when a state-of-the-art matrix factorization model exploiting rich introducing the VISIT mobile tourism recommender, but weather data. Moreover, we also provide details on how the they neither implemented nor evaluated it. weather features were exploited by it, as well as a detailed Among the few works that have actually used weather analysis about the impact of the features on the recommen- in the recommendation pipeline, Braunhofer et al. [3] intro- dation performance. duced a recommender system designed to run in mobile ap- plications for recommending touristic POIs in Italy. The au- thors conducted an online study with 54 users and found out 3. RECOMMENDATION APPROACH that recommendations that take into consideration weather Our recommendation approach is built upon a state-of- information were indeed able to increase the user satisfac- the-art POI recommender algorithm named Rank-GeoFM tion. Compared to this work, our implementation is based [10], a personalized ranking based matrix factorization method. Algorithm 1: Rank-GeoFM with weather context transforming continuous values of weather features (e.g., temperature) into intervals might alleviate this problem. Input: check-in data D1 , D2 , geographical influence matrix Hence, a mapping function is introduced (see Equation 1) W , weather influence matrix W I, hyperparameters ✏, C, ↵, and learning rate g and w that converts the weather features into interval bins. |F Cf | Output: parameters of the model defines the number of bins for the current weather feature. ⇥ = {L(1) , L(2) , L(3) , U (1) , U (2) , F } We will refer to these bins as feature classes. The best re- 1 init: Initialize ⇥ with N (0, 0.01); Shu✏e D1 and D2 sults were obtained with |F Cf | = 20 (validated on hold-out randomly data). 2 repeat ⌫ 3 for (u, l) 2 D1 do (value min(f )) · (|F Cf | 1) 4 approach from Li et al. [10] cf (value) = (1) (max(f ) min(f )) 5 end 6 for (u, l, c) 2 D2 do To extend the original Rank-GeoFM approach with weather 7 Compute yulc as Equation 3 and set n = 0 context, three additional latent factors are introduced that 8 repeat 9 Sample l0 and c0 , Compute yul0 c0 as are represented by matrices in a K-dimensional space. The Equation 3 first one is for incorporating the weather-popularity-score 10 n++ that models whether or not a location is popular with re- 11 until I(xulc > xul0 c0 )I(yulc < yul0 c0 + ✏) = 1 spect to a specific weather feature class and is named L(2) 2 or n > |L| R|L|⇥K , where K denotes the size of the latent parameter 12 if I(xulc > xul0 c0 )I(yulc < yul0 c0 + ✏) = 1 space. Furthermore, a matrix L(3) 2 R|L|⇥K is introduced then ⇣j k⌘ to model the influence between two feature classes. In other |L| 13 ⌘=E n ucll0 words, L(3) softens the borders between the particular fea- 14 g= ⇣P ⌘ ture classes. The third latent parameter F (1) 2 R|F Cf |⇥K (1) P (1) is then used to parametrize the feature classes of the spe- c⇤ 2F Cf wic0 c⇤ fc⇤ c+ 2F Cf wicc+ fc+ cific weather feature. In addition to the latent parameters, (1) (1) (2) (2) a Matrix W I 2 R|F Cf |⇥|F Cf | is introduced for storing the 15 fc fc w ⌘(ll0 ll ) (3) (3) probability that a weather feature class c is influenced by 16 ll ll w ⌘g feature class c0 . Denoting xulc as the frequency that a user (2) (2) 17 ll 0 ll 0 w ⌘fc u checked-in POI l with the current weather context c, this 18 (2) ll (2) ll + w ⌘fc probability is calculated as follows: 19 end P P l2L xulc xulc 0 u2U 20 Project updated factors to accomplish wicc0 = qP P qP P (2) constraints 2 2 21 end u2U l2L xulc u2U l2L xulc0 22 until convergence (1) , L(2) , L(3) , U (1) , U (2) , F (1) } To calculate the recommendation score for a given user u, 23 return ⇥ = {L POI l and weather feature class c, Equation 3 is introduced, where yul denotes the recommendation score as computed in Li et al. [10]. We have selected Rank-GeoFM over other alternatives, be- (1) X (1) yul = u(1) u · ll + u(2) u · wll⇤ ll⇤ cause it has been shown to be a very strong POI recom- l⇤ 2Nk (l) mender method compared to other approaches often cited X (3) (2) (3) (1) in the literature. In Li et al. [10] the authors compared yulc = yul + fc(1) · ll + ll · wicc⇤ fc⇤ Rank-GeoFM against twelve other recommender methods, c⇤ 2F C showing that Rank-GeoFM significantly outperforms strong Algorithm 1 describes how we incorporated the weather generic baselines, such as user-KNN, item-KNN CF, WRMF, context features into the base Rank-GeoFM approach. Tak- BPR-MF [7] as well as specialized POI recommender meth- ing the initialization and the hyperparameters from the orig- ods, such as BPP [17]. Another reason for choosing Rank- inal approach, we first iterate over all pairs of users and POIs GeoFM is related to its ability to easily accommodate addi- (u, l) 2 D1 , where D1 is the set of all check-ins and do the tional features, such as the ones that we plan to use in this adjustments of the latent parameters as described in Li et work. The aim of Rank-GeoFM is to learn latent parameters al. [10]. that model the relationship between the context of interest We then introduce an iteration over all triples (u, l, c) 2 D2 in order to adjust the Table 2 describes the symbols used in the recommender latent parameters on the incorrect ranked venues according algorithm. For each type of contextual data considered, la- to the specific weather context. This adjustment is necessary tent model parameters are introduced. The prediction score because the algorithm might rank a triple (u, l, c) correctly of a triple is then made based on this where on the other hand (u, l, c0 ) might be ranked incor- learned latent parameters. The parameters are trained us- rectly. The adjustments are then done accordingly to the ing a fast learning scheme introduced by the authors that is base algorithm in lines 6-20. based on Stochastic Gradient Descent (SGD). During our studies we found that with a learning rate To add the weather context into Rank-GeoFM, the weather of g = .0001, as used in Li et al. [10], the algorithm did features’ values needed to be discretized. This was done to not converge. The reason for that is that the adjustments reduce data sparsity. For example, if we considered tem- are done on a higher granularity for each (u, l, c) triple and perature as a real number, most of the check-ins concerning not just on the (u, l) level. Henceforth, we introduce a new specific temperature values would probably be zero. Thus, learning rate parameter w = .00001 for the weather con- (a) Cloud cover (b) Visibility (c) Moonphase (d) Precipitation intensity (e) Pressure (f) Temperature (g) Humidity (h) Windspeed Figure 1: Check-in distributions over the eight weather features. 4.1 Datasets The dataset we used in this study was obtained from the work of Yang et al. [14]. It is a Foursquare crawl comprising user check-in data from April 2012 to September 2013. The original dataset contains more than 33 million check-ins from 415 cities in 77 countries. However, before dealing with our problem on such a large scale, we decided to first concentrate our investigation on a small set of US cities. We selected four cities that could represent some weather variety in order (a) “Austrian Restaurant” (b) “Farm” to investigate whether our model is robust to such variety of weather conditions (see Figure 3). Table 1 provides an overview of the check-in statistics of the four target cities chosen for our experiments: Minneapolis, Boston, Miami and Honolulu. Concerning the weather information, we have used the API of forecast.io3 to collect, for each tuple present in our dataset, their corresponding weather informa- tion. For that, we need to pass the following request to the API: (c) “Ski Area” (d) “Ice Cream Shop” https://api.forecast.io/forecast/APIKEY/LAT,LON,TIME Figure 2: Examples of check-in distributions over di↵erent For the purposes of our analysis, we obtained eight weather types of places in Foursquare. On the left hand side, places features, namely, cloud cover, visibility, moon phase, precip- where people check-in at lower temperatures are shown and itation intensity, pressure, temperature, humidity and wind on the right higher temperature places are featured. speed, for all places and time-stamps in our dataset that are provided by forecast.io. text, for which stable results could be observed (validation 4.2 Data Analysis on hold-out data). Similarly to Li et al. [10], we found in Figure 1 shows the probability distributions of check-ins our experiments that the best values of the hyperparameters for each of the eight weather features used. Notice that the are as follows (validated on hold-out data): ✏ = .3, C = 1.0, distributions of pressure, temperature, humidity and wind ↵ = = .2, and K = 100 as used for the dimensions of the speed resemble a normal distribution (see the colored ap- matrices L(1) , L(2) and L(3) . proximation curve). Moreover, while moon phase seems to follow a uniform distribution, which indicates that it will likely not help the recommendation model, the distribution 4. EXPERIMENTAL SETUP of precipitation is very skewed, showing that users have a In this section we describe in detail our experimental setup, strong preference to check-in places when there is low pre- i.e., the datasets we used, a brief characterization of these cipitation intensity (i.e., not raining), indicating that this datasets concerning the weather features used, and the eval- feature might have a good discriminative power. uation protocol we have chosen to conduct our study. 3 https://developer.forecast.io/docs/v2 (a) Cloud cover (b) Visibility (c) Moonphase (d) Precipitation intensity (e) Pressure (f) Temperature (g) Humidity (h) Windspeed Figure 3: Weather feature variability (sorted) measured via standard deviation over cities. Left: cities with lowest variability. Right: cities with highest variability. (a) Cloud cover (b) Visibility (c) Moonphase (d) Precipitation intensity (e) Pressure (f) Temperature (g) Humidity (h) Windspeed Figure 4: Mean weather feature values (sorted) for POI categories with standard errors. In addition to this, Figure 2 illustrates the check-in dis- the original Rank-GeoFM approach, that takes into account tribution as a function of temperature in four di↵erent POI both the check-in history of users and geographical influence. categories. As highlighted in this figure, di↵erent patterns We also compare to the time-based method of Rank-GeoFM, occur depending on the category chosen. While people pre- that was also introduced in Li et al. [10]. fer to check-in in e.g., “Austrian Restaurants” or “Ski Areas” Metric. As evaluation metric NDCG@k (Normalized when the temperature is low, “Ice Cream Shops” or “Farms” Discounted Cumulative Gain) with k = 204 was chosen, as are preferred when temperatures are higher. we want to predict the top-k POIs for a user. Figure 3 shows how the weather features vary in each city of the original Foursquare dataset. Notice that with the exception of moon phase, all the features present a depen- 5. RESULTS dency regarding the city where they are measured, indicat- Figure 5 shows the results of our o✏ine experiment. As ing that a di↵erent recommendation model should proba- shown, in all cases Rank-GeoFM enriched with our pro- bly be trained for each di↵erent city. Moreover, in general, posed weather features significantly outperforms the orig- weather shows a higher variability in the north of the US inal Rank-GeoFM algorithm, which answers RQ2. For all and a very low variability in the south that peaks in the pairwise-comparisons (recommenders with weather context island Honolulu which shows almost no variability in terms vs. without) a standard t-test showed that the p-values were of weather. Figure 4 shows the di↵erent mean values of always smaller than p < .001. What is even more interest- the eight weather features over the POI categories. With ing to note is the performance of Rank-GeoFM that utilizes the small overlapping of the standard error of the means the time feature as contextual factor. As highlighted, in all it’s revealed that indeed categories have a distinct popular- cases, Rank-GeoFM with weather features, such as visibil- ity across various weather feature values. Even moon phase ity and precipitation intensity outperforms the time-based shows a divergent category popularity at its tails. variant, showing that indeed weather conditions may help After this analysis we can confidently state that there is to improve the recommendation quality. indeed a relation between the weather conditions and the We also highlight the fact that certain weather features check-in behavior of Foursquare users, which answers our perform better than others and this ranking seems to be first research question (RQ1) stated at Section 1. city dependent. This can be clearly observed in Figure 5, where the results of Rank-GeoFM with each weather feature 4.3 Evaluation is shown. This answers RQ3, showing which features provide Protocol. To evaluate the performance of our algorithm, the highest gain in recommendation quality. For example, we have chosen the same evaluation protocol as described in Honolulu the best performing feature is precipitation in- in the original Rank-GeoFM paper [10]. Hence, we split the tensity, while in Minneapolis visibility seems to work best dataset (according to the time line) into training, validation among all investigated weather features. Similar patterns and test sets for each city by adding the first 70% of the can be observed for other features, such as temperature or check-ins of each user to the training set, the following 20% cloud cover, changing their relative importance across the to the test set and the rest to the validation set (=10%). four cities. These observations are in line with the results in The training set was then used to learn the latent model Figure 1, showing a strong tendency of check-ins into POIs parameters. During the training phase of the algorithm, under certain weather conditions. However, what is also in- the validation set was used to tune the algorithm conver- teresting to note is the good performance of the moon phase gence. When convergence was observed (typically around feature, which appeared to be uniformly distributed in gen- 3,000 – 5,000 iterations with fast learning scheme enabled), 4 Please note, that we have also run simulations with k = 5 the training was stopped and the learned parameters were or 10, with similar trends in the results as obtained with k = used to evaluate the model on the test set. 20. However, due to limited space, they were not included Baselines. As baselines for our experiments, we used into this paper. (a) Minneapolis (b) Boston (c) Miami (d) Honolulu Figure 5: Recommender accuracy for the eight di↵erent weather context features (sorted by importance) compared to Rank- GeoFM without weather context (denoted as “Baseline”). For further comparison the time-aware version of Rank-GeoFM is included, denoted as “Time”. The red dotted line denotes the baseline. eral (cf. Figure 1). Hence, it appears, that at the level of places (which answers RQ1). Furthermore, we use the pro- locations there is indeed a strong preference for check-ins in posed weather features within a state-of-the-art POI recom- di↵erent phases of the moon. In a recent research, Kohyama mender and we were able to increase the recommender accu- et a. [9] found a relation between moon phase, tidal varia- racy in comparison to the original method that does not use tion, humidity and rainfall. Notably, we found a positive weather data (thus answering RQ2). Furthermore, our ex- relation by analyzing these data based on check-ins, finding periments revealed that the weather context is more useful a small but positive correlation between moon phase and than the context of time and, that the weather features used precipitation intensity, humidity, cloud cover and pressure, in this work are city-dependent. Finally, our study showed as seen in the last row of the correlation matrix shown in (see RQ3) that among the considered weather features, pre- Figure 6. Although further analysis should be performed to cipitation intensity and visibility are the most significant establish a link between our study and theirs, this might be ones to improve the ranking in a weather-aware POI recom- a possible explanation regarding the e↵ect of moon phase in mender system. our POI recommendation model. Finally, the relative performance improvement over the original Rank-GeoFM also seems to be location dependent. 7. FUTURE WORK Hence, while our approach work to a great extent better Currently, our work only investigates one weather feature compared to the baseline for Miami and Honolulu, the dif- at a time. Investigating di↵erent hybridization or context- ferences are less pronounced for Minneapolis. One reason for aware recommender system (CARS) methods and other con- this observation could be that there are more POIs available text variables will be therefore a task to be conducted in our showing similar weather profiles. However, to further con- future work. Furthermore, it will help to investigate in more firm these hypotheses, additional analyses are needed. detail, how the algorithm performs on the whole Foursquare dataset, as more interesting patterns across cities may oc- 6. CONCLUSIONS cur. Finally, we would like to extend our investigations also at user levels, since the current ones concentrate only on the In this paper we presented our preliminary findings on weather profiles of the POIs. how weather data may a↵ect users’ check-in behavior and how this information can be used in the context of a POI recommender system. As our preliminary analyses on the 8. OPEN SCIENCE Foursquare check-in data showed, the weather factors have In order to make the results obtained in this work repro- indeed a significant impact on the people’s check-in behav- ducible, we share code and data of this study. The proposed ior, showing di↵erent check-in profiles for di↵erent kinds of method Rank-GeoFM with weather context is implemented [5] C. Cheng, H. Yang, I. King, and M. R. Lyu. Fused matrix factorization with geographical and social influence in location-based social networks. In Proc. of AAAI, pages 17–23, 2012. [6] G. Ference, M. Ye, and W.-C. Lee. Location recommendation for out-of-town users in location-based social networks. In Proceedings of the 22Nd ACM International Conference on Information & Knowledge Management, CIKM ’13, pages 721–726, New York, NY, USA, 2013. ACM. [7] Z. Gantner, S. Rendle, C. Freudenthaler, and L. Schmidt-Thieme. MyMediaLite: A free recommender system library. In In Proc. of RecSys’11, 2011. [8] H. Gao, J. Tang, X. Hu, and H. Liu. Exploring temporal e↵ects for location recommendation on location-based social networks. In Proceedings of the Figure 6: Correlation matrix for the eight weather features 7th ACM Conference on Recommender Systems, investigated (*p < 0.5, **p < 0.01, ***p < 0.001). RecSys ’13, pages 93–100, New York, NY, USA, 2013. ACM. [9] T. Kohyama and J. M. Wallace. Rainfall variations with the help of the MyMediaLite framework [7] and can induced by the lunar gravitational atmospheric tide be downloaded for free from our GitHub repository5 . Fur- and their implications for the relationship between thermore, the data samples used in the experiments can be tropical rainfall and humidity. Geophysical Research requested for free via email to the corresponding author. Letters, 43(2):918–923, 2016. 2015GL067342. [10] X. Li, G. Cong, X.-L. Li, T.-A. N. Pham, and Acknowledgements S. Krishnaswamy. Rank-geofm: A ranking based geographical factorization method for point of interest This work is supported by the Know-Center. The Know- recommendation. In Proc. of SIGIR’15, pages Center is funded within the Austrian COMET Program - 433–442, New York, NY, USA, 2015. ACM. managed by the Austrian Research Promotion Agency (FFG). [11] D. Martin, A. Alzua, and C. Lamsfus. A Contextual The authors Denis Parra and Leandro Marinho were sup- Geofencing Mobile Tourism Service, pages 191–202. ported by CONICYT, project FONDECYT 11150783 and Springer Vienna, Vienna, 2011. EU-BR BigSea project (MCTI/RNP 3rd Coordinated Call) [12] K. Meehan, T. Lunney, K. Curran, and respectively. A. McCaughey. Context-aware intelligent recommendation system for tourism. In Pervasive 9. REFERENCES Computing and Communications Workshops (PERCOM Workshops), 2013 IEEE International [1] J. Bao, Y. Zheng, and M. F. Mokbel. Location-based Conference on, pages 328–331. IEEE, 2013. and preference-aware recommendation using sparse [13] I. Nunes and L. Marinho. A personalized geo-social networking data. In Proceedings of the 20th geographic-based di↵usion model for location International Conference on Advances in Geographic recommendations in lbsn. In Proceedings of the 2014 Information Systems, SIGSPATIAL ’12, pages 9th Latin American Web Congress, LA-WEB ’14, 199–208, New York, NY, USA, 2012. ACM. pages 59–67, Washington, DC, USA, 2014. IEEE [2] J. Bao, Y. Zheng, D. Wilkie, and M. Mokbel. Computer Society. Recommendations in location-based social networks: [14] D. Yang, D. Zhang, and B. Qu. Participatory cultural A survey. Geoinformatica, 19(3):525–565, July 2015. mapping based on collective behavior in location [3] M. Braunhofer, M. Elahi, M. Ge, F. Ricci, and based social networks. ACM Transactions on T. Schievenin. STS: design of weather-aware mobile Intelligent Systems and Technology, 2015. in press. recommender systems in tourism. In Proceedings of [15] M. Ye, P. Yin, W.-C. Lee, and D. L. Lee. Exploiting the First International Workshop on Intelligent User geographical influence for collaborative Interfaces: Artificial Intelligence meets Human point-of-interest recommendation. In Proc. of Computer Interaction (AI*HCI 2013) A workshop of SIGIR’11, pages 325–334. ACM, 2011. the XIII International Conference of the Italian [16] H. Yin, Y. Sun, B. Cui, Z. Hu, and L. Chen. Lcars: a Association for Artificial Intelligence (AI*IA 2013), location-content-aware recommender system. In Proc. Turin, Italy, December 4, 2013., 2013. of KDD’13, pages 221–229. ACM, 2013. [4] M. Braunhofer, M. Elahi, F. Ricci, and T. Schievenin. [17] Q. Yuan, G. Cong, and A. Sun. Graph-based Context-aware points of interest suggestion with point-of-interest recommendation with geographical dynamic weather data management. In Information and temporal influences. In Proc. of CIKM’14, pages and communication technologies in tourism 2014, 659–668. ACM, 2014. pages 87–100. Springer, 2014. 5 https://github.com/aoberegg/WPOI