An essential task for tourists having a pleasant holiday is to have a well-planned itinerary with relevant recommendations, especially when visiting unfamiliar cities. Many tour recommendation tools only take into account a limited number of factors, such as popular Points of Interest (Pois) and routing constraints. Consequently, the solutions they provide may not always align with the individual users of the system. We propose an iterative algorithm in this paper, namely: BtRec (Bert-based Trajectory Recommendation), that extends from the PoiBert embedding algorithm to recommend personalized itineraries on Pois using the Bert framework. Our BtRec algorithm incorporates users' demographic information alongside past Poi visits into a modified Bert language model to recommend a personalized Poi itinerary prediction given a pair of source and destination Pois. Our recommendation system can create a travel itinerary that maximizes Pois visited, while also taking into account user preferences for categories of Pois and time availability. Our recommendation algorithm is largely inspired by the problem of sentence completion in ngalgoatural language processing (Nlp). Using a dataset of nine cities of diferent sizes, our experimental results demonstrate that our proposed algorithms are stable and outperforms many other sequence prediction algorithms, measured by recall, precision, and ℱ1-scores.
When planning a trip to foreign countries, the typical approach taken by most visitors is
to refer to guidebooks/websites for organizing their daily itineraries, or some may employ
tour recommendation systems that provide popular points of interest (Pois) based on their
popularity [
The subsequent sections of this paper are structured as follows: Section 2 presents background on Tour Recommendations and discusses the state-of-the-art in itinerary prediction. Section 3 provides a formal definition of the problem and introduces the notations used in our solution. Section 4 describes our experimental framework and outlines the baseline algorithms used for solution evaluation. In Section 5, we summarize our findings and discuss potential future extensions of this research.
We first introduce two problems in tourism-related recommendations: the itinerary planning
and next-location prediction. Itinerary planning involves scheduling activities to maximize
the trip experience within pre-set budgets [
Sequence prediction is a fundamental problem that involves the prediction of the next word
in a sequence based on previously observed words [
Bert models The Transformer model with its efective self-attention mechanism is popular
and has been widely adopted in Nlp and computer vision [
Machine Learning algorithms have been proposed to recommend popular Pois [
In this section, we introduce the tour recommendation problem and provide a list of the symbols
and terms used in Table 1. Given a city as the query with | | Pois, we denote a traveler, ∈ ,
with check-in records as a sequence of (, ) tuples, ℎ = [(1, 1), (, 2)... (, )],
2
for all ∈ Poi and denotes as the time marker of the photo posted to Location-Based Social
Networking. The problem addressed in this paper is to recommend a personalized sequence of
Pois for a traveler who is more likely to visit a given city, based on a set of historical trajectories.
The starting and ending Pois, denoted as 1 and ∈ Poi, respectively, are also provided in
the problem.
3.1. PPoiBert - A Refined Bert Model for Personalized Itinerary Prediction
Previous works in itinerary prediction have demonstrated that Bert can be utilized as an
itinerary prediction model by solving a series of Mlm problems [
Itinerary prediction in PPoiBert During the training of our PPoiBert model, the algorithm takes as input a list of User-IDs and their past trajectories for selecting subsequent Pois. To achieve this, we enhance the training algorithm by including more users’ demographic information into the corpus, as illustrated in Fig. 1. Subsequently the PPoiBert model is then trained with some preset epochs and hyper-parameters for predicting itinerary. To make a personalized itinerary recommendation, the PPoiBert algorithm first selects a similar user in the training dataset, ¯ as a reference for making itinerary recommendations. This is done by solving a series of Mlm queries for all users in the training data and select the user resulting in maximum Mlm score (Line 1 & 2 in Algorithm 1). The process of predicting a personalized itinerary is to solve a series of Mlm queries using the reference user in the Mlm queries, by utilizing the Unmask function in Algorithm 1. The remaining part of recommending personalized itineraries is to repeatedly query for the most relevant Poi between the source and destination Pois with ¯ as a preference model and insert predicted Poi into the predicted itinerary in Equation (2). 3.2. BtRec - Bert based Personalized Trajectory Recommendation Using
Demographic Information The proposed PPoiBert algorithm, introduced in Section 3.1, utilizes information from past trajectories based on a reference user-ID in the training dataset to make predictions by considering the preferences of similar users. BtRec extends the PPoiBert by fine-tuning the prediction algorithm by considering their demographic information, such as cities and countries, in the training of our embedding model, in addition to the past trajectories of users. This is achieved by modifying the corpus and context sequence of the PPoiBert model (as described in Section 3.1), incorporating additional information that may influence the decision-making process for selecting the next Pois to visit. Specifically, each sentence in the training data is supplemented with ‘word’s representing the user’s own city/country after the occurrence of the user-ID, as shown in Fig. 1. The aim is to provide relevant training examples to learn the relationship between Pois and users in diferent locations. We then develop personalized Bert models for making "Hyde Park", "Library" , "King's Park" ... ...
POI-BERT Masking Model
itinerary predictions based on users’ demographic locations and other relevant constraints were discussed. We evaluate the accuracy and efectiveness of our proposed algorithm in Section 4. Prediction of reference user The initial phase of BtRec involves finding a reference user for making recommendation based on the preference profiles from the training dataset, similar to the -Nearest Neighbors algorithm [19]. The algorithm iterates through all user-IDs to ifnd the most similar to the query specification, which is represented as a (, ) pair. This process involves solving a series of simple Mlm problems to identity a suitable reference user from the training set, as outlined in Equation (1); a reference-user is assigned by finding the user that maximized the Mlm prediction score, shown in Equation (1). A personalized itinerary is predicted using a pair of (source,destination) Pois in Equation (2).
let
= ArgMax∈¯ ︀( Unmask(“{[CLS], , , [MASK], , , [SEP]}′′)︀ (1) (, ) = ArgMax(︀ Unmask(︀ “[CLS], , , .., [MASK], .., , , [SEP]”))︀ ∀¯∈¯ (2)
The dataset used in our experimentation comprises photos uploaded to the Flickr platform, encompassing the trajectories of 5,654 users from nine popular cities [20]. The photos also include meta-data encompassing details such as the date, time, and GPS locations. By sorting the photos in the dataset based on time and mapping them to the relevant Pois using their GPS locations, we can reconstruct the travel trajectories of the users. This process generates sequences of time-sensitive Pois that represent the users’ trajectories in time. 1
To further evaluate the eficacy of our proposed algorithm on larger data, we incorporated datasets from Melbourne and Vienna [21]; they consist of about 52K photos from 260 Pois in these two cities. Our datasets have been divided into three distinct sets: Training, Validation, and Testing datasets. Initially, we sorted all photos according to their Trajectory-IDs based on their last check-in times in ascending order. To generate the Training Dataset, we set aside 1Source code is available at: https://nxh912.github.io/BTRec_RecTour23/
Description Registered city/country of Poi in Step- of ’s trajectory source Poi destination Poi theme label of Poi- e.g. ‘Museum’, ‘Park’,.. etc.
Poi sequence as a trajectory Predicted Poi sequence Total time allocated for the recommended trajectory ¯ User-’s preference profile ¯ set of ¯ ∈ ¯, in the training
dataset list of check-ins from user- sorted by time-stamps as a
trajectory, i.e. = {1..} the first 70% of trajectories based on their associated photographs. The subsequent 20% of trajectories were assigned to the validation set, while the remaining data was assigned to the testing dataset. This method of segregating the data helps to prevent the issue of having a trajectory being present in multiple datasets [17].
Data: 1, : starting/ending Poi Ids ,
: time budget of itinerary ;
Result: Predicted Poi IDs c beglri“lene∀e[tptCeL∈a′St←]←,,1,1,[M(AUSnKm];,aus,pk(,),)[;SEP]”, ℎ for l[e∀“M[tACS∈LKS{]]2,,..′|,′←,|′,−′,11′,},d1,o,......,,′′,,′′,,′−, 1,,[− S1E,P]”; end let ← ArgMax(Unmask()) ; until < ∑︀ ∈ (); return ; end
Algorithm 1: Itinerary Prediction Algorithm in BtRec
We compared the performance of our algorithm with the state-of-the-art algorithms for mining sequential patterns. Specifically, we identified the following algorithms for mining sequential patterns for performance comparison: • Spmf algorithms - the software package consists of several data mining algorithms implemented, such as Cpt, Cpt+, Tdag and Markov Chains [22, 23, 24, 25, 26, 27]. • SuBSeq: the algorithm employs compression data structures to eficiently store and manipulate the sub-sequences as a “Succinct Wavelet Tree” data structure [28]. • PoiBert: It relies on the general Bert model to generate predictions in choosing Pois [17].
Additionally, it employs bootstrapping to gauge the lengths of Poi visits by estimating duration of visits in the Pois. We also performed hyper-parameter tuning to obtain our prediction of Test dataset as described in Section 4.3.
Some baseline algorithms (such as Cpt and SuBSeq) solely predict the subsequent token (as Poi,) our sequence prediction task involves iteratively predicting additional tokens (as Poi) until the pre-set time limit specified by the user is reached. To compare the efectiveness of our proposed and baseline algorithms, we conducted all experiments under identical conditions outlined in Section 4.3, whereby the algorithms also shared the same datasets for training, validation and testing. 4.3. Experiment Methodology and Setup We performed experiments on nine cities from the Flickr dataset. We considered all trajectories in the dataset as the ground-truth dataset for our predictions, and we used the source/ destination Pois of each trajectory as inputs to our recommendation algorithm. Therefore, we filtered past trajectories with at least 3 Pois. To evaluate the performance of our models, we conducted a comparison of the accuracy with diferent sequence prediction models against various baseline algorithms. The accuracy of our algorithm is evaluated using the Validation and Test sets as follows: (i) For each trajectory in the dataset, which we refer to as the history-list, we considered the first and last Pois as the source and destination Pois for the query itinerary. (ii) the time limit for the query is regarded as the time interval between the first and last photos of each trajectory. (iii) Recommend the intermediate Pois of the trajectory within a specified time frame. To gauge the efectiveness of our models, we compared them with various sequence prediction models listed in Section 4.2. The accuracy of these models was assessed using Validation and Test sets. For each trajectory in the dataset, referred to as the history-list, we identified the first and last Pois as the source and destination Pois for the itinerary prediction query. The time allocated for the query was determined as the time diference between the first and last photos of each trajectory. We evaluated the performance of the proposed BtRec prediction algorithm by using the precision ( ), recall (ℛ), and ℱ1 scores [17]. Tuning of hyper-parameters To find the optimal hyperparameters for our experiments, we trained diferent prediction models in PoiBert, PPoiBert and BtRec using various epochs, ranging from 1 to 100, on the training dataset. Next, we employed these trained models to predict itineraries in our validation dataset. The model with the highest average ℱ1 score from the validation dataset was then selected. Finally, we reported the prediction accuracy using the chosen model to generate recommendations in the test dataset. These experiments ensure that we solely rely on the trained model to verify its validity in predicting new data.
We assessed the efectiveness of our proposed algorithm in various cities by comparing the actual Pois trajectories (constructed travel histories based on the chronological ordering of photos) as the ground truth dataset values of the itinerary predictions. The accuracy of the predicted itineraries was compared in terms of average ℱ1 scores in Table 2. The PoiBert algorithm achieved 62.32% on average, the proposed BtRec algorithm significantly outperforms the baseline algorithms with an average ℱ1 score of 63.55% on our datasets. Compared to the actual trajectories (ground-truth data), both PPoiBert and BtRec recommend itineraries with high ℱ1 scores, suggesting a good balance between the true positives and false positives in the predictions. Our proposed PPoiBert algorithm can recommend tour trajectories that are more personalized and relevant to users’ preferences. Additionally, our proposed BtRec algorithm further enhances the prediction of Poi itinerary with users’ demographic information into the embedding model. Our BtRec algorithm predicted a tour itinerary that further outperforms other baseline algorithms with an average ℱ1 score: 61.45%. Even without parameter tuning, the BtRec algorithm achieves an average ℱ1 score of 58.05% across all datasets and hyperparameters, while the PPoiBert algorithm achieves an ℱ1 score of 56.45% on average. Average Recall(ℛ)/ℱ1/Precision() scores of prediction algorithms in Test datasets (%)
Budapest
Delhi Edinburgh
Glasgow
Melbourne
Osaka
Perth
Toronto
Vienna
All cities 3 these algorithms cannot find new Poi, except from the starting and ending Pois. Hence, they have a precision scores of 100%.
and recommended trajectory.
In this paper, we propose BtRec designed to suggest a sequence of Pois that enables tourists to plan an optimal schedule while considering factors such as locality, time constraints, and individual preferences. Our approach involves constructing and training a Bert-based language model to fine-tune the recommendation system. This process utilizes training, validation, and test datasets to ensure accurate and personalized recommendations. By leveraging the power of Bert classification, we aim to provide tourists with a more refined and context-aware itinerary planning. Additionally, we designed an iterative method to generate Pois based on users’ interests and demographic information. By analyzing just a pair of source and destination Pois, our iterative algorithm, BtRec, accurately identifies users’ preferences for selecting subsequent Pois during their tours by analyzing the statistics of uploading (potentially) photos over the frame of their visits to Pois. Extensive experiments conducted on nine cities showed that our proposed algorithm, which considers frequencies of photos, and locality of Pois with other users’ demographic information, outperforms nine baseline algorithms in terms of average ℱ1 scores. A potential extension is to measure in average hit-rate@k for the observed Acknowledgments: This research is funded in part by SUTD under grant RS-MEFAI-00005-R0201. The computational work was partially performed on resources of the NSCC and the Social AI Studio. s11227-016-1925-2. [17] N. L. Ho, K. H. Lim, Poibert: A transformer-based model for the tour recommendation problem, in: 2022 IEEE International Conference on Big Data (Big Data), 2022, pp. 5925– 5933. doi:10.1109/BigData55660.2022.10020467. [18] A. Agarwal, L. W. Dietz, Recommending the duration of stay in personalized travel recommender systems, in: Proceedings of the Workshop on Recommenders in Tourism (RecTour 2022) Seattle, WA, USA 2022, volume 3219 of CEUR Workshop Proceedings, CEURWS.org, 2022, pp. 1–20. URL: https://ceur-ws.org/Vol-3219/paper1.pdf. [19] K. S. Beyer, J. Goldstein, R. Ramakrishnan, U. Shaft, When is ‘nearest neighbor’ meaningful?, in: Proceedings of the 7th International Conference on Database Theory, ICDT ’99, Springer-Verlag, Berlin, Heidelberg, 1999, p. 217–235. [20] K. H. Lim, J. Chan, C. Leckie, S. Karunasekera, Personalized trip recommendation for tourists based on user interests, points of interest visit durations and visit recency, Knowledge and Information Systems 54 (2018) 375–406. [21] P. Padia, K. H. Lim, J. Cha, A. Harwood, Sentiment-aware and personalized tour recommendation, in: 2019 IEEE International Conference on Big Data (Big Data), IEEE, 2019, pp. 900–909. [22] T. Gueniche, P. Fournier-Viger, V. S. Tseng, Compact prediction tree: A lossless model for accurate sequence prediction, in: H. Motoda, Z. Wu, L. Cao, M. Zaiane, Osmar nd Yao, W. Wang (Eds.), Advanced Data Mining and Applications, Springer Berlin Heidelberg, Berlin, Heidelberg, 2013, pp. 177–188. [23] G. et al, Cpt+: Decreasing the time/space complexity of the compact prediction tree, in: T. Cao, E.-P. Lim, Z.-H. Zhou, T.-B. Ho, D. Cheung, H. Motoda (Eds.), Advances in Knowledge Discovery and Data Mining, Springer International Publishing, Cham, 2015, pp. 625–636. [24] V. N. Padmanabhan, J. C. Mogul, Using predictive prefetching to improve world wide web latency, COMPUTER COMMUNICATION REVIEW 26 (1996) 22–36. [25] Laird, Saul, Discrete sequence prediction and its applications., Machine learning 15 (1994) 43–68. [26] J. Cleary, I. Witten, Data compression using adaptive coding and partial string matching, IEEE Transactions on Communications 32 (1984) 396–402. doi:10.1109/TCOM.1984. 1096090. [27] Pitkow, P. Pirolli, Mining longest repeated subsequences to predict world wide web surfing, in: 2nd USENIX Symposium on Internet Technologies & Systems, USENIX Association, Boulder, CO, 1999. [28] K. et al, Succinct bwt-based sequence prediction, in: S. Hartmann, J. Küng, S. Chakravarthy, G. Anderst-Kotsis, A. M. Tjoa, I. Khalil (Eds.), Database and Expert Systems Applications, Springer International Publishing, Cham, 2019, pp. 91–101.