Improved Recommendation of Photo-Taking Locations using Virtual Ratings Mesut Kaya Derek Bridge Insight Centre for Data Analytics Insight Centre for Data Analytics University College Cork, Ireland University College Cork, Ireland mesut.kaya@insight-centre.org derek.bridge@insight-centre.org ABSTRACT In this paper, we study the task of recommending geolo- We consider the task of collaborative recommendation of cations to users, based on the data found in photo-sharing photo-taking locations. We use datasets of geotagged pho- sites. In other words, the items in our recommender system tos. We map their locations to a location grid using a geo- are geolocations. The recommender may help a user dis- hashing algorithm, resulting in a user × location implicit cover new places where they can enjoy good views or nice feedback matrix. Our improvements relative to previous settings, suitable for photo-taking. work are twofold. First, we create virtual ratings by spread- Collaborative recommendation of geolocations for photo- ing users’ preferences to neighbouring grid locations. This taking has been explored in recent work by Phan et al.[13] makes the assumption that users have some preference for Briefly, they use a cartographic hashing function to map the locations close to the ones in which they take their pho- latitude and longitude coordinates associated with photos tos. These virtual ratings help overcome the discrete na- to rectangular bins: photos taken from within the same bin ture of the geohashing. Second, we normalize the implicit have the same hash key. The user × location implicit feed- frequency-based ratings to a 1-5 scale using a method that back matrix uses the hash keys for locations. The rating by has been found to be useful in music recommendation al- a user for a location is given by the proportion of her geo- gorithms. We demonstrate the advantages of our approach tagged photos whose coordinates map to that bin; ratings with new experiments that show large increases in hit rate are therefore in (0, 1]. Phan et al. compared an item-based and related metrics. nearest-neighbours recommender with three matrix factor- ization methods. They ran experiments measuring RMSE, with non-negative matrix factorization having lowest RMSE. CCS Concepts In this paper, we, like Phan et al., are concerned with rec- •Information systems → Recommender systems; ommending locations. We use geohashing rather than Phan et al.’s cartographic hashing to map the places where users Keywords took the photos to buckets. Then we make two main contri- butions. First, we create virtual ratings by spreading users’ photography; geohashing; implicit ratings preferences to neighbouring grid locations. This makes the assumption that users have some preference for locations 1. INTRODUCTION close to the ones in which they take their photos. These The advent of digital cameras and smart phones has rev- virtual ratings help overcome the discrete nature of the geo- olutionised photo-taking. We now create more multimedia hashing. Second, we normalize the implicit frequency-based content than ever before; the content is more heavily con- ratings to a 1-5 scale using a method that has been found to textualised, e.g. with increasingly accurate data from phone be useful in music recommendation algorithms. We evaluate sensors such GPS receivers; and sharing the content has the effect of these two innovations separately and together never been easier nor more commonplace. using experiments that measure hit rate and related metrics, Web sites, such as Flickr and Facebook, where multime- rather than RMSE. dia content is shared, implicitly capture contextualised per- In Section 2 we review the related work; in Section 3 we sonal preferences over the places that people like to create present our proposed method; and in Section 4 we give the content of this kind, i.e. the places where they take photos. experimental results. The preference data is a resource from which we can build personalized and contextualized recommender systems [10]. 2. RELATED WORK There is an amount of previous research in recommending locations to users, e.g. [11, 3, 16, 21]. For the most part, this work is concerned with point-of-interest (POI) recom- mendation. For photo-taking, by contrast, we are not di- rectly interested in recommending POI locations; instead, we want to be able to recommend locations that may give Copyright held by the author(s). views of POIs (and nice settings for other photos). The lo- RecTour 2016 - Workshop on Recommenders in Tourism held in conjunc- tion with the 10th ACM Conference on Recommender Systems (RecSys), cations that we recommend may even be far from any POIs. September 15, 2016, Boston, MA, USA. Hence, following [13], we recommend rectangular cells in the coordinate space. Phan et al. map latitude and longi- tude coordinates to rectangular bins using a method of their own invention, which they call Cartographic Sparse Hashing (CASH) [1]. Their method has a parameter, r, the resolu- tion. At the Equator, bins will be r metres wide and r metres high. Note, however, that bins will be taller than r metres the further they are away from the Equator due to the curvature of the Earth. The resulting hash key is a 64- bit integer whose high bits are the hash of the longitude and whose low bits are the hash of the latitude. In more recent work, they use CASH within an activity recommender [1]. There are other location recommenders that also work in a coordinate space. For example, Liu et al. try to predict the next location that a user will visit [12]. Interestingly, they, along with Yuan et al. [21], also consider the role of time in location recommendation, which may also be relevant to Figure 1: Geohashing photo-taking, but which we do not investigate further here. Shared photos have been used as a data source for pur- poses such as POI detection [20], tag recommendation [17], the Equator are 152.9 metres wide and 152.4 metres high. photo-taking location detection [5], and route recommenda- It is interesting to note that Phan et al. run experiments tion [14]. Some work specifically uses Flickr data, just as we where they vary the resolution of their hashing method. Al- and Phan et al. use in our work; for example, Zheng et al. tering the resolution, however, does not just affect sparsity, recommend Flickr interest groups to users [22]. But none of it also alters the items (bins) that get recommended. Ar- this work, other than Phan et al.’s, uses this kind of data to guably, this should not be a parameter that one alters to recommend photo-taking locations. minimise error. It is instead something one should fix at The literature also contains descriptions of systems that a granularity that users find useful for photo-taking. As we assist with photo composition, e.g. [2, 15]. Bourke et al. de- said, we fix precision at 7, which give buckets that are about scribe what they call the social camera, which recommends 150 × 150 metres, which we think is an appropriate size for a list of popular photos that were taken near to the user’s photo-taking location recommendation. Had we used a pre- location in similar lighting conditions. The user can choose cision of 6, buckets would be approximately 1km by 600 one of these recommended photos, which will then be used metres, which is clearly too large for useful recommenda- as the basis for assistance with camera settings and framing tions. A case can be made for precision of 8 (about 40 by [2]. Rawat proposes a system called ClickSmart that can 20 metres) but anything higher is probably too small (e.g. provide real-time advice about scene composition and cam- precision of 9 recommends 5 by 5 metre locations). era settings using rules learned from social media images It is important also to say that the size of the buckets, [15]. which we use to recommend where to take photos, is unre- Our use of virtual ratings is similar in spirit to the ap- lated to the size of what might be photographed. From a proach of fuzzy event modelling proposed by Hidasi and Tikk bucket on the south bank of the River Thames, for example, [7]. They use a similar idea to model continuous contexts in a user might capture a panoramic shot of the north London factorization algorithms. skyline or she might zoom in on a pigeon eating a discarded hamburger. This emphasises the point too that recommend- 3. PROPOSED METHOD ing photo-taking buckets is not the same as recommending In this section, we explain our proposed approach. The POIs. In the same example of a bucket on the south bank of approach consists of geohashing, followed by the spreading of the River Thames, there might be a POI in the same bucket users’ preferences by creating virtual ratings in neighbouring (e.g. the London Eye) but the user might be taking photos buckets, followed by the conversion of implicit feedback to of POIs in a different bucket in the distance (e.g. Big Ben) 1-5 ratings. Finally, we use a collaborative recommendation or may not be taking photos of specific POIs at all (e.g. algorithm on the resulting feedback matrix. skylines and pigeons). After hashing, we have an initial user × location ratings 3.1 Geohashing matrix, where locations are buckets and ratings are based on We do not use Phan et al.’s CASH method, preferring frequencies. Figure 1 shows an example. Suppose a user u to use geohashing, which is more common. Both have the has taken six photos in six different locations. Suppose loc1 , same effect: they divide the surface of the Earth into a grid loc2 and loc5 are geohashed to the same bucket g1 . The of rectangular cells (called bins in CASH and buckets in ratings matrix contains triples such as hu, g1 , 3i, meaning geohashing); a hash function takes in latitude and longitude that user u has taken three photos in bucket g1 . and maps them to one of the cells of the grid. Geohashing maps latitude and longitude into a geohash key of up to 12- 3.2 Creating virtual ratings characters. Coordinates that map to the same key are in the One problem with hashing to a rectangular grid is its dis- same cell of the grid (bucket). The scheme is hierarchical: cretization of coordinate space. In Figure 1, for example, prefixes of the hash key designate larger cells that include taking a photo at loc2 is taken as positive feedback for that those designated by extensions of the prefix. The size of the point in space and others near it. But the rating is recorded prefixes is known as the precision. In this paper we take only for bucket g1 . The geohashing results in us recording 7-character long prefixes. These designate buckets that at no positive feedback for the nearby points in g2 . Our solution to this problem is to create virtual ratings in the user ×location matrix by spreading the original frequen- Table 1: Recommender configurations (0, 1] ratings 1-5 ratings cies to neighbouring buckets. First, we decide which buckets to spread to. We may spread to zero, one or more of the eight No virtual ratings 1 5 neighbouring buckets. We only spread from a bucket to a Virtual ratings 1-VR 5-VR neighbour if the bucket contains a photo-taking event that is close enough to the neighbour. We calculate the geodesic distance between the coordinates of the photo-taking events of music listening: to convert how often a user listens to a in the bucket and the centre of the neighbour.1 Only if the track into a 5-point scale. Unlike the kinds of 5-point ex- minimum of these distances is smaller than a threshold value plicit rating scales used in book and movie recommenders on ∆ will we create a virtual rating. For example, in Figure 1, the web, for Celma’s normalized ratings, a rating of 1 does the rating for g1 will only be spread to g4 if the distance not necessarily mean that the user dislikes the item; rather, between loc1 and the centre of g4 (this being smaller than the fact that the item was listened to at all implies some the distances from loc2 and loc5 ) is smaller than ∆. level of positive feedback, but less enthusiastic positive feed- Next, we decide the value of the virtual rating. Its value is back than that associated with higher points on the scale. a discounted version of the one that is being spread. Follow- It seems appropriate to use Celma’s method for the implicit ing [21], we use a power law distribution to model the pref- frequency-based ratings that we have in our photo-taking erence of a user for a neighbouring bucket as a function of scenario. the minimum distance we calculated previously. This maps a distance of 0 to a weight of 1.0 and it maps the maximum 3.4 Recommendation algorithm distance (∆) to a weight of 0.0. The neighbour’s virtual At this point, we have a normalized ratings matrix. Our rating is the product of the weight and rating (frequency) goal, given a user and bucket for which the user has no rat- associated with the source bucket. ing, is to predict the user’s rating. For this, we use matrix There is, however, the issue of how to aggregate ratings factorization to transform users and buckets into the same that ‘arrive’ in a bucket from different sources. For such latent factor space. We choose to use matrix factorization cases, we use the simple heuristic that the virtual rating is since it is widely used for collaborative recommenders and the maximum of the ratings arriving from different sources. a form of matrix factorization was the best performing ap- For example, in Figure 1, bucket g4 receives two virtual rat- proach in [13]. Specifically, for the matrix factorization we ings. One comes from g1 : it is g1 ’s rating (3) discounted use Koren et al.’s SVD [9], solving the objective function by an amount based on the distance d14 . The other comes using stochastic gradient descent. We have not ‘swapped in’ from g5 : it is g5 ’s rating (1) discounted by an amount based different recommender algorithms since our focus is on mea- on distance d34 . The larger of these will be taken as the suring the contributions made by the virtual ratings and the virtual rating for g4 . Note that the same calculation is used different forms of normalisation. (We do, however, compare even if g4 already contained a rating of its own: its new rat- against three baseline recommender systems – see below.) ing is the maximum of its original and the two discounted virtual ratings. Since the virtual ratings are discounted by 4. EXPERIMENTS an amount based on distance, only in exceptional cases will they replace an existing rating. 4.1 Datasets Spreading virtual ratings to neighbouring buckets does We collected the data used in this work from the photo- not, of course, enlarge what is being recommended. Recom- sharing website flickr.com by using its API. We searched for mendations continue to be made at the level of individual geotagged photos taken in London and Dublin in 2015 to cre- buckets. ate two datasets. After geohashing, we discarded users who had ratings for fewer than five buckets. The final dataset for 3.3 Rating normalisation London contains 112,671 photos taken by 978 unique users. Phan et al. normalize the frequency-based ratings to the Users have an average of 115 photos. The final dataset for range (0, 1] [13]. They do this by dividing the frequency Dublin contains 54,082 photos taken by 1,567 users and users (the number of photos taken by a user in a bin) by the total have an average of 34 photos. number of photos taken by that user. However, we found that by their approach 98% of the normalised ratings lie 4.2 Recommenders between 0 and 0.1. This has several problems: it implies We compared four configurations of the recommender, de- low preference for 98% of all locations in which a user took pending on whether ratings were normalized to (0, 1] (as in a photo; it gives a very skewed distribution; and it means [13]) or to 1-5 (as we propose) and depending on whether that a recommender that is evaluated using RMSE can do virtual ratings are used or not. The names of these four con- well by always predicting a number between 0 and 0.1. figurations are given in Table 1. It follows that the system Instead, we follow Celma’s method [4] to convert implicit called ‘1’ is closest to the one described in [13]: ratings are feedback to a 1-5 rating scale. Following Celma, we compute normalised to (0, 1] and virtual ratings are not used. The the Complementary Cumulative Distribution of the frequen- main difference with [13], as mentioned in Section 3.1, is cies in a user’s profile. Then, items (buckets) that fall in the that we are using geohashing where they used their CASH top 80 − 100% of the distribution are given a rating of 5, method. items that fall into the 60 − 80% range are given a rating We also compare against three baseline recommenders, of 4, and so on. Celma proposed his method in the context POP H, POP ALL, and HOME, which we now describe. 1 We use the geopy Python library for this purpose: https: //pypi.python.org/pypi/geopy 4.2.1 Popularity-based recommenders We compare our methodology with a baseline recommender ones for which a recommender will be rewarded if it recom- that recommends the most popular items for which the user mends them in the experiments. In [6], where they assume a has no rating. It is well-known that, in recommender sys- conventional 1-5 scale, the test set contains only the highly- tems in general, recommending the most popular items can rated items from the probe set (i.e. ones rated 4 or 5). How- be a highly competitive baseline [18]. For the kinds of 5- ever, we return to the point that we have made before that, point explicit rating scales used in book and movie recom- for Celma-style normalisation of frequency data, a rating of menders on the web, a popularity-based recommender typ- 1 or 2 does not necessarily denote dislike. If we want to test ically recommends to a user those items that she has not recommender performance on all liked items, then we must rated and that have the greatest number of high ratings (4 include in the test set all probe items, irrespective of their or 5 stars) [18]. In other words, the recommender only rec- values of their ratings. ommends items that lots of people like. Accordingly, we In fact, we have chosen to run all experiments twice: in include in our experiments a baseline that we refer to as one set of experiments, from the probe set we create a test POP H (‘H’ for ‘high’): in computing popularity, it counts set by retaining only test items where the user’s normalised only those users who have given the item a rating of 4 or rating is 4 or 5; for the other set of experiments, we take all 5. But, as we have already mentioned in Section 3.3, on the of the probe set as test set. rating scales that we are using, low ratings (e.g. 1 and 2) are There are no virtual ratings in the test sets and, where not signals of dislike: they show a track was listened to (in different recommenders are compared (even ones that nor- the case of music) or a photo was taken (in our case), which malise to (0, 1]), they are compared using the same sets of is positive feedback, albeit not as positive as a rating of 4 test items. or 5. Hence, we also include another baseline recommender, For each test item, we randomly select 1000 other buckets POP ALL. In computing popularity, it recommends items for which the user has no rating. We predict the user’s rated by the greatest number of users, irrespective of the ratings for all 1001 buckets and then sort them by descending values of the ratings. predicted rating. We then recommend the top-k (k = 10), which may or may not include the test item. 4.2.2 Home location-based recommender This means that experiments are being done on an item Using a Flickr dataset, Van Laere et al. conclude that a by item basis for each item in the test set, rather than on a user is more likely to take pictures in locations that are closer user by user basis. We must define our metrics accordingly. to home [19]. It follows then that we can build a baseline For each test item, we measure whether the test item is recommender that recommends to a user locations for which in the top-k or not. If it is, we call this a hit and record the she has no rating and that are close to her home location. total number of hits, H. From this, we calculate the hit rate Some Flickr users provide a textual description of their (or recall): home location in their Flickr profile. For the users in our H datasets, we wrote a crawler to visit their profiles and obtain HR = (1) |Test| their home location descriptions, if given. (Of course, not all of these descriptions will be correct, and this may reduce where Test is the test set. the performance of this baseline recommender.) We convert We also calculate the average reciprocal hit-rank (ARHR), the textual description to a geolocation (latitude and lon- which in our setting we define as follows: gitude).2 For the London dataset, we were able to obtain 1 X 1 the home geolocations of 40.52% of the users; for the Dublin ARHR = (2) dataset, it was 39.5% of the users. |Test| r ∈Test, rank rui ui rank rui 6=0 The baseline recommender, which we refer to as HOME, works as follows. For each user whose home location coordi- where rank rui is the position of this test item in the top-k nates are known, we calculate the geodesic distance between (1 ≤ rank rui ≤ k) or zero if this test item was not recom- their home location and the centres of the buckets (items). mended in the top-k. Then we recommend the closest buckets for which the user Although ARHR considers the positions of the hits, nei- has no rating. For users whose home location coordinates ther ARHR nor HR gives information about the original are not known, we default to recommending the most pop- rating as well as the position. It is common to discount ular buckets in the dataset, as POP ALL would do. this based on the logarithm of the rank. Hence, inspired by discounted cumulative gain (e.g. [8]), we define average 4.3 Methodology and metrics discounted gain (since, with only one test item at a time, it Phan et al. calculate the RMSE between the predicted and is not cumulative), as follows: actual ratings for the members of a test set. We chose to base our experiment on a newer methodology, emphasising 1 X rui if rank rui = 1 ADG = rui if rank rui > 1 recommendation over prediction [6]. (In any case, because |Test| r ∈Test, log2 (rank rui ) ui two of our systems normalise to (0, 1] and two normalise to a rank rui 6=0 discrete 1-5 scale, we cannot directly compare their RMSEs.) (3) We used 5-fold-cross-validation using 80% for training and 20% for a probe set. From the probe set, we construct a 4.4 Parameter ∆ test set. The idea is that the test set will contain items from When using virtual ratings, there is the parameter ∆, the probe set that the user liked, and therefore these are which needs to be set for the experiments. With small val- 2 We do this by using the geopy python library, which uses ues of ∆, there are fewer virtual ratings than with larger the Google Maps V3 geocoder: https://developers.google. values of ∆. We tried values of 0 (which is the same as no com/maps/documentation/geocoding/intro virtual ratings), 75, 150, 225 and 300. In Figures 2 and 3, Figure 2: HR, ARHR and ADG for 5-VR with varying Figure 4: HR ∆, London dataset Figure 5: ARHR Figure 3: HR, ARHR and ADG for 5-VR with varying ∆, Dublin dataset we show the hit rate, average reciprocal hit-rank and av- erage discounted gain for one of the system configurations (5-VR) with varying values of ∆ for the London and Dublin datasets, respectively. As can be seen, HR, ARHR and ADG tend to increase as ∆ increases but then, as ∆ becomes too big and so ratings are being spread too far, HR, ARHR and ADG level off or even fall. The Figures show that the most competitive value for ∆ for this system configuration for both datasets is 225. The results for 1-VR (not shown) follow a similar pattern, but its most competitive value for ∆ is 300 for both datasets. Figure 6: ADG These are the values we use for ∆ in the results that we show in the next section. bines both our innovations. It out-performs all recommenders 4.5 Results on all metrics across both datasets. We now compare the seven recommenders, i.e. the three There are also some other patterns in the results. For both baselines (POP H, POP ALL and HOME) and the four con- datasets and all evaluation metrics, of the four non-baselines, figurations of our recommender, which depend on the nor- system 1 is the worst; adding virtual ratings (1-VR) makes malisation scheme and whether virtual ratings are used or an improvement; scaling to 1-5 (system 5) is better again. not (1, 1-VR, 5 and 5-VR). As per the previous section, the But the best-performing configuration is 5-VR, where we two configurations that use virtual ratings use their most scale to 1-5 and use virtual ratings. competitive values for ∆. Another pattern is that the baselines are more competi- Figures 4, 5 and 6 show the hit rate, average reciprocal tive on the Dublin dataset than the London dataset. They hit-rank and average discounted gain respectively. never out-perform system 5-VR, but on the Dublin dataset, The worst-performing system overall is the one designated POP H and POP ALL both out-perform systems 1, 1-VR 1, corresponding roughly to the system in [13]. Not only is and 5. The performance of the HOME baseline is more it out-performed by the other three configurations (1-VR, 5 mixed. and 5-VR), it is out-performed by the three baselines (except For reference, Table 2 shows the values on which the Fig- on the London dataset, where it has very slightly higher HR ures in this section are based. than HOME does, but even here it is out-performed by the In fact, as explained in Section 4.3, we chose to run two other two baselines). set of experiments, differing in the way in which the test set The best-performing system is 5-VR — the one that com- was constructed from the probe set. In one case, all ratings Table 2: Results for experiments where test set equals probe set Dataset Metric HOME POP H POP ALL 1 1-VR 5 5-VR London HR 0.02 0.0282 0.0294 0.0222 0.0536 0.0691 0.0816 London ARHR 0.0077 0.0098 0.0106 0.005 0.0197 0.0255 0.0301 London ADG 0.0287 0.0394 0.0411 0.0243 0.0732 0.0971 0.1131 Dublin HR 0.0595 0.0936 0.0963 0.0191 0.0603 0.0821 0.1203 Dublin ARHR 0.0215 0.0324 0.0354 0.005 0.0195 0.0272 0.0461 Dublin ADG 0.085 0.1308 0.138 0.0255 0.0842 0.1169 0.1788 from the probe set are placed into the test set. These are the in part by a research grant from Science Foundation Ireland experiments whose results were shown in the Figures and in (SFI) under Grant Number SFI/12/RC/2289. Table 2. In the other case, only highly-rated items from the probe set are included in the test set. The results for this 7. REFERENCES set of experiments are given in Table 3. Comparing the two [1] P. Bhargava, T. Phan, J. Zhou, and J. Lee. Who, tables, we see that numerically the results in the second set what, when, and where: Multi-dimensional of experiments are higher. But, for the most part, the story collaborative recommendations using tensor about which systems out-perform each other remains the factorization on sparse user-generated data. In Procs. same. In particular, 5-VR remains far and away the best of of the 24th Int. Conf. on the World Wide Web, pages the recommenders. 130–140, 2015. [2] S. Bourke, K. McCarthy, and B. Smyth. The social 5. DISCUSSION & CONCLUSIONS camera: Recommending photo composition using We presented an approach to recommending geolocations contextual features. In Procs. of the Second Workshop to users for photo-taking. We geohashed coordinates to on Context-Aware Recommender Systems, 2010. cells in a rectangular grid, and used these as the items in [3] X. Cao, G. Cong, and C. S. Jensen. Mining Significant an implicit feedback matrix. We investigated two innova- Semantic Locations from GPS Data. Proc. VLDB tions. One was to create virtual ratings in neighbouring Endow., 3(1-2):1009–1020, 2010. cells. The other was to normalise ratings using a method [4] Ò. Celma. Music Recommendation and Discovery: developed for music recommenders. Our experiments, mea- The Long Tail, Long Fail, and Long Play in the suring hit rate, average reciprocal hit-rank and average dis- Digital Music Space. Springer, 2010. counted gain, showed that the two innovations together out- [5] D. J. Crandall, L. Backstrom, D. Huttenlocher, and performed all other configurations and popularity-based and J. Kleinberg. Mapping the world’s photos. In Procs. of home location-based baseline recommenders. the 18th Int. Conf. on the World Wide Web, pages There are many avenues for future work. For a start, we 761–770, 2009. can test on other datasets and on recommender algorithms [6] P. Cremonesi, Y. Koren, and R. Turrin. Performance other than SVD. There remains an open question about the of recommender algorithms on top-n recommendation precision of the geohashing. Here we are using precision of 7, tasks. In Procs. of the 4th ACM Conf. on resulting in buckets that are about 150 metres by 150 metres. Recommender Systems, pages 39–46, 2010. The only way to determine whether this is the best choice or [7] B. Hidasi and D. Tikk. Approximate modeling of whether higher precision (smaller buckets) would be better continuous context in factorization algorithms. In is through a user trial. It may even be that precision should Procs. of the 4th Workshop on Context-Awareness in be personalised or adaptive in some other way, and this could Retrieval and Recommendation, pages 3–9, 2014. be investigated in future work. [8] K. Järvelin and J. Kekäläinen. Cumulated gain-based Finally, there is the opportunity to integrate other factors evaluation of IR techniques. ACM Transactions on into the work. Presently, we recommend only the photo- Information Systems, 20(4):422–446, 2002. taking location. Date and time may also be important: per- [9] Y. Koren, R. Bell, and C. Volinsky. Matrix haps a location may offer better candidate subjects for pho- factorization techniques for recommender systems. tos in certain seasons, on certain dates or at certain times of Computer, 42(8):30–37, 2009. day. Related to date and time are the photo-taking condi- tions: a bucket may be a better location, e.g., when the sun [10] J. J. Levandoski, M. Sarwat, A. Eldawy, and M. F. has risen, when the weather is not overcast or when the sun Mokbel. Lars: A location-aware recommender system. is not glaring (although, of course, for some photographers In Procs. of the IEEE 28th Int. Conf. on Data these difficult conditions provide opportunities for the exer- Engineering, pages 450–461, 2012. cise of their photographic talents). Some users might also [11] B. Liu, Y. Fu, Z. Yao, and H. Xiong. Learning want assistance with the choice of subject, camera settings Geographical Preferences for Point-of-interest or composition. For this, it may be possible to integrate our Recommendation. In Procs. of the 19th ACM work with the kind of work done on the social camera [2] SIGKDD Int. Conf. on Knowledge Discovery and and ClickSmart systems [15]. Data Mining, pages 1043–1051, 2013. [12] Q. Liu, S. Wu, L. Wang, and T. Tan. Predicting the Next Location: A Recurrent Model with Spatial and 6. ACKNOWLEDGEMENTS Temporal Contexts. In Procs. of the 30th AAAI Conf. This publication has emanated from research supported on Artificial Intelligence, 2016. Table 3: Results for experiments where test set contains highly-rated probe items Dataset Metric HOME POP H POP ALL 1 1-VR 5 5-VR London HR 0.0272 0.0413 0.0402 0.0315 0.0698 0.0932 0.1083 London ARHR 0.0104 0.0137 0.0146 0.0071 0.0261 0.0361 0.0416 London ADG 0.0657 0.0923 0.0938 0.0577 0.165 0.2258 0.26 Dublin HR 0.0759 0.1193 0.1241 0.0289 0.081 0.1139 0.1529 Dublin ARHR 0.0292 0.0442 0.0479 0.0084 0.028 0.0403 0.0629 Dublin ADG 0.1812 0.2809 0.2962 0.0618 0.185 0.2638 0.3828 [13] T. Phan, J. Zhou, S. Chang, J. Hu, and J. Lee. pages 327–336, 2008. Collaborative Recommendation of Photo-Taking [18] H. Steck. Item popularity and recommendation Geolocations. In Procs. of the 3rd ACM Multimedia accuracy. In Proceedings of the fifth ACM conference Workshop on Geotagging and its Applications in on Recommender systems, pages 125–132. ACM, 2011. Multimedia, pages 11–16, 2014. [19] O. Van Laere, S. Schockaert, and B. Dhoedt. [14] D. Quercia, R. Schifanella, and L. M. Aiello. The Georeferencing Flickr resources based on textual shortest path to happiness: Recommending beautiful, meta-data. Information Sciences, 238:52–74, 2013. quiet, and happy routes in the city. In Procs. of the [20] Y. Yang, Z. Gong, and L. H. U. Identifying Points of 25th ACM Conf. on Hypertext and Social Media, Interest by Self-tuning Clustering. In Procs. of the pages 116–125, 2014. 34th Int. ACM SIGIR Conf. on Research and [15] Y. S. Rawat. Real-time assistance in multimedia Development in Information Retrieval, pages 883–892, capture using social media. In Procs. of the 23rd ACM 2011. International Conference on Multimedia, pages [21] Q. Yuan, G. Cong, Z. Ma, A. Sun, and N. M. 641–644, 2015. Thalmann. Time-aware point-of-interest [16] Y. Shi, P. Serdyukov, A. Hanjalic, and M. Larson. recommendation. In Procs. of the 36th Int. ACM Nontrivial Landmark Recommendation Using SIGIR Conf. on Research and Development in Geotagged Photos. ACM Trans. Intell. Syst. Technol., Information Retrieval, pages 363–372, 2013. 4(3):47:1–47:27, 2013. [22] N. Zheng, Q. Li, S. Liao, and L. Zhang. Which photo [17] B. Sigurbjörnsson and R. van Zwol. Flickr Tag groups should I choose? A comparative study of Recommendation Based on Collective Knowledge. In recommendation algorithms in Flickr. Journal of Procs. of the 17th Int. Conf. on the World Wide Web, Information Science, 36(6):733–750, 2010.