Towards the Recommendation of Personalised Activity Sequences in the Tourism Domain Gunjan Kumar, Houssem Jerbi, and Michael P. O’Mahony Insight Centre for Data Analytics, School of Computer Science, University College Dublin, Ireland {firstname.lastname}@insight-centre.org ABSTRACT lifelogging and urban computing domains, where activities included In this paper we consider the problem of recommending sequences socialising, eating, etc. and modes of transport, respectively. In of activities to a user. The proposed approach leverages the order as this paper, we extend the activity recommendation framework to well as the context associated with the user’s past activity patterns address the task of recommending a sequence of activities to the to make recommendations. This work extends the general activity user. Moreover, we apply our framework to the tourism domain, recommendation framework proposed in [16] to iteratively recom- where a recommended sequence of activities might be, for example, mend the next sequence of activities to perform. We demonstrate visiting a zoo, eating Italian food, and then listening to live music. the efficacy of our recommendation framework by applying it to the Our work is motivated by the assumption that people tend to tourism domain and evaluations are performed using a real-world repeat similar patterns of activities under similar circumstances (checkin) dataset. [29]. Hence, in order to infer the next activities for a user, it is important to consider the activity patterns performed in the past. At CCS CONCEPTS the same time, the context surrounding these activities significantly affects the next activities the user performs. The importance of • Information systems → Recommender systems; Decision modelling context has been recognised in both tourism [6, 17, 18] support systems; Spatial-temporal systems; as well as recommender systems research [1]. Context is particularly important in tourism as the user is predominantly mobile [10]. For KEYWORDS example, features such as the time of day, location and weather can Sequence Recommendation, Recommender Systems, Activity Rec- determine whether a user visits a particular amusement park in the ommendation, Activity Timeline Matching city or not. In recommender systems research, the task of recommending 1 INTRODUCTION sequences is comparatively under-explored [13, 27]. However, there exists works, particularly for points of interest/itinerary (LBSN) Internet and digital technologies have significantly influenced the [20, 30, 34] and music playlists [2, 4, 8, 22, 23, 27] recommendation, tourism sector in the last decade resulting in a steady growth in which address this task. A popular approach for modeling sequences e-tourism [7]. Users now have easy access to vast amounts of infor- has been Markov-based models [4] and all-kt h -order Markov mod- mation on the web which assists them to plan trips, make reserva- els [3, 9, 25, 28]. However, in general, these approaches are not tions, and purchase products etc. However, the number of available suitable for modelling sequences of activities with multiple features choices have increased so rapidly that it has become difficult to find or context and are limited to the Markov assumption which does the right information at the right time. Thus, recommender systems, not apply in all cases [4]. An alternative hierarchical graph-based which have found immense success in e-commerce, have the po- approach to capture sequences and geographical hierarchies in lo- tential to play a crucial role in e-tourism by providing personalised cation trajectories is presented in [19]. This is further enhanced in and relevant content to users [5, 14, 24, 27]. [35] by modeling location popularity and user experiences to mine To provide useful recommendations, it is essential to capture the popular travel sequences across users in a non-personalised man- behaviour and needs of users, which has been particularly challeng- ner. Similarly, graph-based models have been used for collaborative ing in e-tourism [26]. However, as digital technologies have now itinerary recommendation [33]. However, these approaches do not permeated our daily lives to a great extent, many aspects of our lives capture the context information associated with user activities. can now be easily recorded in digital format. For example, physi- The key distinguishing characteristic of our work is that the cal activities performed, locations visited and media consumed by model captures both the past activities of users, together with the users can be recorded using mobile devices [12]. Moreover, mobile context associated with these activities, in order to recommend personal assistants, such as Google Now and Microsoft Cortana, are the next sequence of activities for users to perform. The main capable of passively recording the digital activities of users. These contributions of this work can be summarised as follows: recordings, which contain the activity patterns and preferences of users, can facilitate the development of personalised recommender systems capable of generating recommendations at the right time • The extension of the generic activity recommendation frame- and in the right way for a given user and context [11, 31, 32]. work in [15, 16] to recommend the next sequence of activi- In our previous work [15, 16], we proposed a generic activity ties that should be performed by users. For this, an iterative, recommendation framework to recommend the next activity to content-based recommendation approach is proposed, which perform to a user. Our approach was applied successfully in the takes the sequence as well as the features associated with RecTour 2017, August 27th, 2017, Como, Italy. 26 Copyright held by the author(s). previous activity occurrences into consideration to build the Algorithm 1: SeqNCSeqRec recommendation model (Section 2); • The application of our proposed algorithm to the tourism Input: User, u; user’s past timeline, T ; recommendation time, RT ; domain. Experiments using a location checkin dataset [21] current activity object, aoc ; N -count value, N demonstrate the efficacy of our approach in recommend- Output: a recommended timeline (sequence) Tr ec of L activity ing sequences given a diverse variety of activities and user objects, Tr ec =< aor ec 1 , aor ec 2 , ...aor ec i ..., aor ec L > activity patterns (Section 3). 1. Extract the current timeline Tc from T ; the final element of Tc is aoc 2. Tr ec ← < > 2 RECOMMENDATION APPROACH 3. i ← 1 In this section, we formulate the problem of recommending the 4. while i ≤ L do next sequence of activities to a user. These activities can be, for ex- 5. Extract candidate timelines T from T (each j ample, eating Italian food, shopping at a bookstore, listening to live Tj ∈ T ends with an activity object ao f such j music, etc. The proposed content-based sequence recommendation that ao f .name = aoc .name) algorithm leverages sequential patterns in a user’s past activities 6. R←{} as well as the contextual information (for example, time of day, 7. for each Tj ∈ T do location, weather, etc.) associated with each activity occurrence. j R ← R ∪ ao f +1 8. for each ao ∈ R do 2.1 Problem Formulation Compute Score(ao) We introduced the concept of an activity object and an activity time- 9. aor ec i .name ← top-1(ao.name : ao ∈ R) line in [15]. An activity object, aoi , refers to a single 10. Compute and assign features to aor ec i  occurrence of Tr ec ← append(Tr ec , aor ec i ) an activity and consists of a set of features, aoi = vi1 , vi2 , ..., vim , 11. which describe the context surrounding that particular occurrence 12. Tc ← append(Tc , aor ec i ) of the activity. For example, an activity object can refer to an in- 13. RT ← aor ec i .time stance of ‘a visit to a zoo’ (i.e. the activity name) with associated 14. i ←i +1 contextual features, such as time of day, geo-location, weather, popu- 15. return Tr ec larity of the location, etc. An activity timeline (or timeline for short) for a user is then a chronological sequence of all activity objects From this set of scored activity objects, the top-1 activity name performed by that user, T =< ao 1 , ao 2 , ..., aon >. with the highest score is returned as the name for aor ec i in Tr ec (Step 9). The values for the other features of aor ec i are then com- puted (Step 10) based on the average values for each feature from 2.2 Recommendation Algorithm the user’s past timeline. For example, if the recommended activity The proposed recommender is based on previous work [16], in name is eating ‘Italian Food’, the time at which this activity should which the past activities performed by a user were modelled as a occur (aor ec i .time) is calculated as follows. The median difference timeline, T , and the objective was to recommend the next activity between all occurrences of ‘Italian Food’ and the immediately pre- to a user to perform. Here, we extend this approach to recom- ceeding activity in the user’s past timeline is calculated; aor ec i .time mend the next sequence of activities for users to perform, Tr ec =< is then given by the current recommendation time (RT ) plus this aor ec 1 , aor ec 2 , ..., aor ec L >. difference. Referring to Algorithm 1, a sequence of activities at a given rec- Before the next iteration of the algorithm, aor ec i is appended ommendation time (RT ) are generated as follows. The most recent to the current timeline Tc (and becomes the current activity object activity object performed by the user, referred to as the current in the next iteration) (Step 12) and the recommendation time (RT ) activity object, aoc , is initialised as the activity object occurring is set to aor ec i .time (Step 13). Thus, the L activity objects in the at time RT in the user’s timeline. The current timeline, Tc , is then recommended timeline Tr ec are generated in L iterations. extracted from the user’s timeline; it consists of the subsequence d(Tj , Tc ) − min d(Tp , Tc ) of the N activity objects occurring prior to aoc and ends with aoc Tp ∈T (Step 1). Score(ao) = 1 − . (1) max d(Tp , Tc ) − min d(Tp , Tc ) The recommendation of each activity object aor ec i in Tr ec is Tp ∈T Tp ∈T performed iteratively (Step 4) as follows (see [16] for details). For each previous occurrence in the user’s timeline of an activity with 2.2.1 Distance between Timelines. For the purpose of determin- the same name as aoc (e.g. ‘Italian Food’), a candidate timeline (Tj ) ing the similarity between two timelines T1 and T2 , the two-level is extracted (Step 5). Let T be the set of all candidate similarity algorithm proposed in our earlier work [15] is used. This  timelines in algorithm first computes the minimum cost of rearranging the ac- a given iteration. A two-level edit distance d(. , .) between each candidate and the current timeline is computed [15]; based on these tivities to achieve the same activity sequence and then aligns the distances, a score (Eqn. 1) is assigned to the activity that occurs values of the features of the corresponding activity objects. See [15] immediately after each candidate timeline Tj in T (Steps 7–8). for further details on this approach. RecTour 2017, August 27th, 2017, Como, Italy. 27 Copyright held by the author(s). 2.2.2 N-count matching. The matching unit determines the length of the subsequences to be considered when calculating the distances 9 30 between timelines. The SeqNCSeqRec algorithm uses the N -count Median of % agreement 8 Algorithm Algorithm 25 matching approach as proposed in [16]. Thus, the N activity ob- 7 SeqNCSeqRec SeqNCSeqRec 20 over users jects in the timeline preceding the current activity object form the 6 BiGramSeqRec BiGramSeqRec current timeline (and likewise for candidate timelines). Note that 5 PopSeqRec 15 PopSeqRec the optimal value of N for each user will differ, depending on the 4 degree of repetition and regularity of activities performed by each. 3 10 2 5 1 3 EVALUATION 0 0 We first describe the dataset used to construct activity timelines 1 2 3 1 2 3 for users and the experimental methodology employed. This is Sequence length (k ) Sequence length (k ) followed by an evaluation of the proposed N -count based sequence (a) (b) recommender. Figure 1: Median percentage agreements for recommended 3.1 Dataset sequences for SeqNCSeqRec and baseline algorithms using For our experiments, we used a subset of the Gowalla checkins timelines constructed from categories at (a) level 2 and (b) dataset [21]. The complete dataset obtained contains around 36 level 1 in the hierarchy. million checkins, 2.8 million locations and 0.3 million users. Every checkin is bound to a specific location and timestamp. A subset of sequence of categories of length 3 was generated at different recom- these locations have categories assigned to them, such as, ‘Italian mendation times (RT s), which corresponded to the end time of each Food’, ‘Bookstore’, ‘City Park’, etc. These locations also have contex- activity object in the test timeline. Recommendation performance is tual features such as latitude, longitude, number of users checking evaluated using agreement @ k (k = 1, 2, 3) which is the percentage in to it, number of photos taken at the location, etc. In relation to of RT s for a user where the first k categories in the recommended our recommendation framework, each of the location categories is sequence and the actual sequence are an exact match. considered as an ‘activity name’ and the recommendations made For the computation of two-level edit distances between time- will be sequences of these categories. Hence, for evaluation, we lines, the following operation costs and feature weights were used: select only those checkins locations which have assigned categories. c ins = cdel = 1, and c sub = 2 ; wcat eдory = 2, w st ar t −t ime = 1, Further, categories are organised in a three-level hierarchy, con- wpopul ar ity = 1, wlocat ion = 1. These weights were set according sisting of 7, 134 and 151 level 1, 2, and 3 categories, respectively. For to their hypothesised importance from the perspective of compar- example, the level 1 category ‘Food’ has child categories ‘African’, ing timelines; for example, the weight associated with updating ‘American’, ‘Asian’, ‘Coffee Shop’, etc. at level 2, while ‘Coffee Shop’ the category was set to the highest value since this is clearly a key has child categories ‘Starbucks’ and ‘Dunkin Donuts’ at level 3. consideration when computing distances between timelines. See Given our objective is to recommend activities (categories) to users, [15] for details on the two-level edit distance approach. we consider level 2 categories as the most suitable level of gran- ularity, and hence any checkin locations with level 3 categories are assigned the parent category at level 2. As such, the names of 3.3 Recommendation Performance activity objects in user timelines are given by the level 2 categories The performance of our proposed sequence-based N -count se- of the locations checked in to by users. quence recommendation algorithm (SeqNCSeqRec) is compared Since the characteristics of the timelines on weekdays and week- to the following baselines: ends are different, here we considered data corresponding to week- • The bi-gram-based sequence recommender (BiGramSeqRec) days only. To address multiple consecutive checkins by users at the is based on the Markov assumption that the next activity same location, we merged such checkins for a given user if they had depends only on the current activity. For each user, the fre- the same category, were less than 600 meters apart and occurred quency of occurrence of all activity name (category) bi-grams within an interval of 10 minutes. Further, we selected only those in the user’s timeline are computed. For a given RT , a se- users which have checkin data for at least 50 days with a minimum quence of activity objects is recommended iteratively as of 10 checkins per day. The sampled dataset had 916 users with 2.7 per SeqNCSeqRec except that, at each iteration, the most million checkins in total. The median number of checkins per day frequently occurring bi-gram beginning with the current for users varied from 11–134, while the median number of distinct activity name is identified, and the recommended activity categories of checkins per day for users varied from 4–58. is simply that of the second element of this bi-gram. Such Markov-based approaches have proved to be quite successful 3.2 Methodology in modelling sequences in previous studies [9]. An offline evaluation was conducted for the proposed recommenda- • For a given user and RT , at each iteration of the algorithm, tion approach. Each user’s complete timeline was split into training the popularity-based sequence approach (PopSeqRec) recom- and test timelines, where the test timeline contained data for the mends the activity that the user performed most frequently most recent 20% of available days. For each user, a recommended at that time in the past. RecTour 2017, August 27th, 2017, Como, Italy. 28 Copyright held by the author(s). 3.3.1 Algorithm Performance. Figure 1(a) shows the median per- centage agreements (k = 1, 2, 3) over all users for the proposed Se- Mean % agreement 2.5 qNCSeqRec recommender and the two baselines. For SeqNCSeqRec, the results shown correspond to the optimal value of N -count for Group1[0] each user. It is clear from these results that the proposed approach 2.0 Group2[1,4] significantly outperforms the baseline approaches. For example, Group3[5+) SeqNCSeqRec improves upon BiGramSeqRec by 16.98%, 45.38%, and 1.5 129.3% for recommended sequences of length 1, 2, and 3, respec- tively, and improves upon PopSeqRec by more than 100% in all cases. Differences in results between the proposed and baselines 0 1 2 3 4 6 8 10 12 algorithms are statistically significant (Wilcoxon-Mann-Whitney N−count rank sum test) at the p<.05 level. The results also indicate that performance declines when larger sequences are recommended, Figure 2: Mean percentage agreement for recommended se- which is to be expected, given the increased challenges involved in quences over users in each group. making such recommendations. While the above findings are promising, it can be seen that the percentage agreements achieved by all algorithms are relatively low; for example, the percentage agreement is 9.5% for sequence current activity only (N -count = 0); Group 2: next activities are lengths of 1 using SeqNCountSeqRec. We make the following obser- based on the current activity and a small number of past activities vations in this regard. Firstly, as described in the previous section, (N -count lies in the interval [1,4]); and Group 3: next activities are in order to generate a sequence of recommendations, only the top-1 based on the current activity and a larger number of past activities recommended activity is considered at each iteration of the SeqNC- (N -count = 5+). SeqRec algorithm. In addition, the evaluation is based on only a In this experiment, users were assigned to one of the above single recommended sequence being made to users, which clearly groups based on the range in which their optimal N -count value represents a strict approach. appears (optimal in the sense that best percentage agreement was Secondly, while many (although not all) level 2 categories are seen for sequences of length 3). Overall, 421, 374 and 121 users semantically similar, they are not considered a match according to were assigned to Groups 1, 2 and 3, respectively. Results are shown the evaluation metric. For example, consider the level 2 categories in Figure 2. It can be seen that the mean recommendation perfor- ‘Mexican’ and ‘South American/Latin’ which relate to dining and mance for Group 1 users (46% of all users) was significantly lower are children of the level 1 category ‘Food’. From a recommenda- than that seen for users in the other groups. This finding is to be tion perspective, these different types of dining experiences are expected, since it indicates that it is easier to recommend sequences clearly related and (arguably) should represent a match. Thus, we of activities to users which are more consistent in their activity also evaluate our recommender when all checkin locations are patterns. Thus, it can be concluded that adopting a personalised mapped to level 1 categories in the hierarchy – i.e. user timelines approach for users, by selecting the optimal N -count value for each are constructed from activity objects with names given by the level user, is important. While it is not feasible to determine this value by 1 categories of locations checked in to by users. The results are experiment for large user bases, an approach to automatically learn shown in Figure 1(b). While similar trends as before are seen, the a suitable value for individual users such as proposed in previous percentage agreements achieved are much greater; for example, work [16] can be applied. over 30% for SeqNCSeqRec compared to the previous 9.5% for se- quences of length 1. The ‘true’ performance of the recommender lies somewhere in between these values (since not all level 2 cate- 4 CONCLUSIONS AND FUTURE WORK gories are semantically related); a further analysis of this matter is In this paper, we have expanded on our previous work to suggest left to future work. sequences of activities for users based on past activity patterns. Notwithstanding the strict evaluation metric used in this work, the 3.3.2 Performance across Users. A key intuition behind our ap- proposed approach shows promising performance and outperforms proach is that the next activities performed by individual users the baseline algorithms considered. In future work, we will inves- depends, to a lesser or greater extent, on their past activity pat- tigate collaborative approaches in which candidate timelines will terns. In the proposed SeqNCSeqRec recommendation algorithm, be drawn from the activities of other users in the system. Further, the number of past activities to be considered when generating rec- we will consider new approaches to suggest sequences of activities ommendations is determined by the N -count value (see Section 2.2). (for example, using RNNs) and investigate the recommendation In previous work [16], where the task was to recommend a single of context (for example, where, when, with whom etc.) associated activity to users, it was seen that the optimal N -count value varied with each of the suggested sequence of activities. across users. In this section, we investigate whether a similar affect is seen when recommending sequences of activities to users. As per [16], we hypothesise three distinct groups of users to 5 ACKNOWLEDGMENTS capture the degree to which past activity patterns reflect future The Insight Centre for Data Analytics is supported by Science activity performance – Group 1: next activities are based on the Foundation Ireland under Grant Number SFI/12/RC/2289. RecTour 2017, August 27th, 2017, Como, Italy. 29 Copyright held by the author(s). REFERENCES [23] Brian McFee and Gert RG Lanckriet. 2012. Hypergraph Models of Playlist Dialects. [1] Gediminas Adomavicius, Bamshad Mobasher, Francesco Ricci, and Alex Tuzhilin. In Proceedings of the 13th International Conference on Music Information Retrieval 2011. Context-Aware Recommender Systems. AI Magazine 32, 3 (2011). (ISMIR). [2] Jie Bao, Yu Zheng, David Wilkie, and Mohamed Mokbel. 2015. Recommendations [24] Antonio Moreno, Aida Valls, David Isern, Lucas Marin, and Joan Borràs. 2013. in Location-based Social Networks: A Survey. Geoinformatica 19, 3 (July 2015), SigTur/E-Destination: Ontology-based Personalized Recommendation of Tourism 525–565. https://doi.org/10.1007/s10707-014-0220-8 and Leisure Activities. Engineering Applications of Artificial Intelligence 26, 1 (Jan. [3] Thorsten Bohnenberger and Anthony Jameson. 2001. When Policies Are Better 2013), 633–651. https://doi.org/10.1016/j.engappai.2012.02.014 Than Plans: Decision-theoretic Planning of Recommendation Sequences. In Pro- [25] James Pitkow and Peter Pirolli. 1999. Mining Longest Repeating Subsequences ceedings of the 6th International Conference on Intelligent User Interfaces (IUI ’01). to Predict World Wide Web Surfing. In Proceedings of the 2nd Conference on ACM, New York, NY, USA, 21–24. https://doi.org/10.1145/359784.359829 USENIX Symposium on Internet Technologies and Systems - Volume 2 (USITS’99). [4] Geoffray Bonnin and Dietmar Jannach. 2014. Automated Generation of Music USENIX Association, Berkeley, CA, USA, 13–13. http://dl.acm.org/citation.cfm? Playlists: Survey and Experiments. Comput. Surveys 47, 2, Article 26 (Nov. 2014), id=1251480.1251493 35 pages. https://doi.org/10.1145/2652481 [26] Francesco Ricci. 2002. Travel Recommender Systems. IEEE Intelligent Systems [5] Joan Borràs, Antonio Moreno, and Aida Valls. 2014. Intelligent Tourism Recom- (2002). mender Systems: A survey. Expert Systems with Applications 41, 16 (2014), 7370 – [27] Francesco Ricci, Lior Rokach, Bracha Shapira, and Paul B. Kantor. 2010. Recom- 7389. https://doi.org/10.1016/j.eswa.2014.06.007 mender Systems Handbook (1st ed.). Springer-Verlag New York, Inc., New York, [6] Matthias Braunhofer, Mehdi Elahi, Francesco Ricci, and Thomas Schievenin. NY, USA. 2013. Context-Aware Points of Interest Suggestion with Dynamic Weather [28] Guy Shani, David Heckerman, and Ronen I. Brafman. 2005. An MDP-Based Data Management. Springer-Verlag, Cham, 87–100. https://doi.org/10.1007/ Recommender System. Journal of Machine Learning Research 6 (Dec. 2005), 978-3-319-03973-2_7 1265–1295. http://dl.acm.org/citation.cfm?id=1046920.1088715 [7] Dimitrios Buhalis. 2003. eTourism: Information technology for strategic tourism [29] Chaoming Song, Zehui Qu, Nicholas Blumm, and Albert-László Barabási. 2010. management. Pearson Education. Limits of Predictability in Human Mobility. Science 327, 5968 (2010), 1018–1021. [8] Shuo Chen, Josh L. Moore, Douglas Turnbull, and Thorsten Joachims. 2012. https://doi.org/10.1126/science.1177170 Playlist Prediction via Metric Embedding. In Proceedings of the 18th ACM SIGKDD [30] Chih-Hua Tai, De-Nian Yang, Lung-Tsai Lin, and Ming-Syan Chen. 2008. Recom- International Conference on Knowledge Discovery and Data Mining (KDD ’12). mending Personalized Scenic Itinerary with Geo-tagged Photos. In IEEE Interna- ACM, New York, NY, USA, 714–722. https://doi.org/10.1145/2339530.2339643 tional Conference on Multimedia and Expo. 1209–1212. https://doi.org/10.1109/ [9] Mukund Deshpande and George Karypis. 2004. Selective Markov Models for ICME.2008.4607658 Predicting Web Page Accesses. ACM Transactions on Internet Technology 4, 2 [31] Iis P. Tussyadiah and Dan Wang. 2016. Tourists’ Attitudes toward Proactive (May 2004), 163–184. https://doi.org/10.1145/990301.990304 Smartphone Systems. Journal of Travel Research 55, 4 (2016), 493–508. https: [10] Damianos Gavalas, Charalampos Konstantopoulos, Konstantinos Mastakas, and //doi.org/10.1177/0047287514563168 Grammati Pantziou. 2014. Mobile recommender systems in tourism. Journal of [32] Artem Umanets, Artur Ferreira, and Nuno Leite. 2014. GuideMe – A Tourist Network and Computer Applications 39 (2014), 319 – 333. https://doi.org/10.1016/ Guide with a Recommender System and Social Interaction. Procedia Technology j.jnca.2013.04.006 17 (2014), 407 – 414. https://doi.org/10.1016/j.protcy.2014.10.248 Conference on [11] Ramanathan Guha, Vineet Gupta, Vivek Raghunathan, and Ramakrishnan Srikant. Electronics, Telecommunications and Computers – CETC 2013. 2015. User Modeling for a Personal Assistant. In Proceedings of the Eighth ACM [33] Hyoseok Yoon, Yu Zheng, Xing Xie, and Woontack Woo. 2010. Smart Itinerary International Conference on Web Search and Data Mining (WSDM ’15). ACM, New Recommendation Based on User-generated GPS Trajectories. In Proceedings of the York, NY, USA, 275–284. https://doi.org/10.1145/2684822.2685309 7th International Conference on Ubiquitous Intelligence and Computing (UIC’10). [12] Cathal Gurrin, Alan F. Smeaton, and Aiden R. Doherty. 2014. LifeLogging: Springer-Verlag, Berlin, Heidelberg, 19–34. http://dl.acm.org/citation.cfm?id= Personal Big Data. Foundations and Trends in Information Retrieva 8, 1 (June 1929661.1929669 2014), 1–125. https://doi.org/10.1561/1500000033 [34] Hyoseok Yoon, Yu Zheng, Xing Xie, and Woontack Woo. 2012. Social Itinerary [13] Jonathan L. Herlocker, Joseph A. Konstan, Loren G. Terveen, and John T. Riedl. Recommendation from User-generated Digital Trails. Personal Ubiquitous Comput. 2004. Evaluating Collaborative Filtering Recommender Systems. ACM Trans- 16, 5 (June 2012), 469–484. https://doi.org/10.1007/s00779-011-0419-8 actions on Information Systems 22, 1 (Jan. 2004), 5–53. https://doi.org/10.1145/ [35] Yu Zheng, Lizhu Zhang, Xing Xie, and Wei-Ying Ma. 2009. Mining Interesting 963770.963772 Locations and Travel Sequences from GPS Trajectories. In Proceedings of the 18th [14] Katerina Kabassi. 2010. Personalizing Recommendations for Tourists. Telematics International Conference on World Wide Web (WWW ’09). ACM, New York, NY, and Informatics 27, 1 (2010), 51 – 66. https://doi.org/10.1016/j.tele.2009.05.003 USA, 791–800. https://doi.org/10.1145/1526709.1526816 [15] Gunjan Kumar, Houssem Jerbi, Cathal Gurrin, and Michael P. O’Mahony. 2014. Towards Activity Recommendation from Lifelogs. In Proceedings of the 16th International Conference on Information Integration and Web-based Applications & Services (iiWAS ’14). ACM, New York, NY, USA, 87–96. https://doi.org/10.1145/ 2684200.2684298 [16] Gunjan Kumar, Houssem Jerbi, and Michael P. O’Mahony. 2016. Personalised Recommendations for Modes of Transport: A Sequence-based Approach. The 5th ACM SIGKDD International Workshop on Urban Computing (UrbComp 2016) (2016). [17] Carlos Lamsfus, David Martin, Zigor Salvador, Alex Usandizaga, and Aurkene Alzua-Sorzabal. 2009. Human-Centric Ontology-Based Context Modelling In Tourism. Mediterranean Conference on Information Systems (2009). [18] Carlos Lamsfus, Dan Wang, Aurkene Alzua-Sorzabal, and Zheng Xiang. 2015. Going Mobile. Journal of Travel Research 54, 6 (2015), 691–701. https://doi.org/10. 1177/0047287514538839 [19] Quannan Li, Yu Zheng, Xing Xie, Yukun Chen, Wenyu Liu, and Wei-Ying Ma. 2008. Mining User Similarity Based on Location History. In Proceedings of the 16th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems (GIS ’08). ACM, New York, NY, USA, Article 34, 10 pages. https://doi. org/10.1145/1463434.1463477 [20] H. Liu, L. Y. Wei, Y. Zheng, M. Schneider, and W. C. Peng. 2011. Route Discovery from Mining Uncertain Trajectories. In 11th IEEE International Conference on Data Mining Workshops. 1239–1242. https://doi.org/10.1109/ICDMW.2011.149 [21] Xin Liu, Yong Liu, Karl Aberer, and Chunyan Miao. 2013. Personalized Point- of-interest Recommendation by Mining Users’ Preference Transition. In Pro- ceedings of the 22nd ACM International Conference on Information & Knowl- edge Management (CIKM ’13). ACM, New York, NY, USA, 733–738. https: //doi.org/10.1145/2505515.2505639 [22] Brian McFee and Gert RG Lanckriet. 2011. The Natural Language of Playlists. In Proceedings of the 12th International Conference on Music Information Retrieval (ISMIR). RecTour 2017, August 27th, 2017, Como, Italy. 30 Copyright held by the author(s).