PathRec: Visual Analysis of Travel Route Recommendations
Dawei Chen, Dongwoo Kim, Lexing Xie, Minjeong Shin, Aditya Krishna Menon, Cheng Soon Ong, Iman Avazpour, John Grundy
PPathRec: Visual Analysis of Travel Route Recommendations
Dawei Chen * † , Dongwoo Kim * , Lexing Xie * † , Minjeong Shin * , Aditya Krishna Menon † * ,Cheng Soon Ong † * , Iman Avazpour ‡ , John Grundy ‡ *The Australian National University, † Data61/CSIRO, ‡ Deakin University{u5708856,dongwoo.kim,lexing.xie,minjeong.shin,aditya.menon,chengsoon.ong}@anu.edu.au{iman.avazpour,j.grundy}@deakin.edu.au
ABSTRACT
We present an interactive visualisation tool for recommendingtravel trajectories. This system is based on new machine learningformulations and algorithms for the sequence recommendationproblem. The system starts from a map-based overview, taking aninteractive query as starting point. It then breaks down contribu-tions from different geographical and user behavior features, andthose from individual points-of-interest versus pairs of consecutivepoints on a route. The system also supports detailed quantitativeinterrogation by comparing a large number of features for multiplepoints. Effective trajectory visualisations can potentially benefit alarge cohort of online map users and assist their decision-making.More broadly, the design of this system can inform visualisationsof other structured prediction tasks, such as for sequences or trees.
KEYWORDS
Route Visualisation, Travel Recommendation, Learning to rank
Sequence recommendation has emerged as an important frame-work for modelling diverse problems such as travel route and musicplaylist recommendation [3]. Unlike classical ranking algorithmswhere items are considered independently [7], a sequence recom-mendation algorithm requires modelling a structure between itemsand suggests a set of items as a whole. For example, consider recom-mending a trajectory of points-of-interest (POIs) in a city to a visitor.While a classical ranking algorithm can learn a user’s preferencefor each individual location, it may ignore the distances betweenthem and could suggest a longer trajectory than is optimal. Severalsequence recommendation algorithms have been proposed to solvethis problem and demonstrated superior performance compared toclassical ranking algorithms [3, 9]. Nonetheless, recommendationalgorithms for sequences and trajectories [1, 3] have many compo-nents and can be difficult for a user to understand. This is part of thegeneral challenge of introducing transparency and accountabilityfor machine learning algorithms [4].In this paper, we tackle the problem of sequence visualisation,specifically focussing on travel routes recommendation. A travelroute is a sequence of POIs, and the sequence recommendationproblem can be formulated as a structured prediction problem [3].Based on a diverse set of features for individual and pairs of POIs, wetrain the prediction model with trajectory data extracted from geo-tagged photos taken in Melbourne [1]. To visualise the suggestedroutes, we develop a novel tool that efficiently displays multiplesuggested routes, which helps users understand the process behind
Figure 1: Travel route visualisation system . Given a start-ing POI and the number of POIs to be visited, the systemrecommends multiple routes from travel history of tourists.Shown above: recommendation in central Melbourne. the recommendations. Specifically, our system decomposes a totalscore of each route into a set of features and their correspondingscores, and shows the total score as a stacked bar plot of the features.The system also visualises the differences between POIs in a singleroute to show how POIs in that route can exhibit vast diversity.This visualisation helps tourists who want diverse experiencesby choosing the best route among multiple recommendations. Gen-eralising to a broader class of routes, such a visualisation could alsohelp users of online mapping apps to make decisions on suggestedtravel routes, such as by trading off distance, traffic, and scenery. The travel route recommendation problem involves a set of POIs ina city. Given a trajectory query x = ( s , l ) , comprising a start POI s and trip length l , the goal is to suggest one or more sequences ofPOIs that maximise some notion of utility.Following [3], we first cast travel recommendation as a structuredprediction problem, which allows us to leverage the well-studiedliterature of structured SVMs (SSVM) [6]. From a visualisationperspective, an advantage of the SSVM is the explicit representationof feature scores in its final decision process. Specifically, we candisassemble the final score of a route into feature scores of each POIand the transition between two adjacent POIs. We use hand-craftedPOI features such as the category, popularity, and average visitduration of previous tourists. We also crafted transition featuressuch as the distance and neighbourhood of two POIs to maximisethe interpretability of the outcome. a r X i v : . [ c s . H C ] J u l igure 2: Visualisation of POI and transition scores for top10 recommended routes. Each bar from left to right repre-sents a relative score of each POI or transition along theroute. The length of stacked bars represents the total scoreof the suggested route. Our goal is to design an interactive visualisation system on top ofthe structured prediction framework. Figure 1 shows the overviewof a live demo system, which consists of five major components: amap to display the suggested routes, an input box for user query(upper left), a stacked score of routes (upper right), a POI list box(lower left), and a radar chart to compare features of multiple POIs(lower right). The role and the construction of the four major com-ponents, besides the main map, are as follows:
Query input : A query consists of a starting POI and a trip length.Users can choose the starting POI by clicking icons on the mapand can adjust the slide to set the trip length. In addition, threedifferent travelling modes (e.g. bicycling, driving and walking) aresupported, and we optimise the suggested routes for each mode.
Route score visualisation : The SSVM evaluates relevance scoresof POIs and transitions in a candidate route to the given query anduses the sum of the relevance scores to determine the ranks ofthe routes. To visualise the POI and transition scores, we adopt astacked bar representation [5], designed to support the visualisationof multi-attribute ranking. In Figure 2, the system decompose thescores of top 10 recommended routes into POI and transition scoresvia the stacked bar representation, where the size of each bar isproportional to the relevance score of the corresponding POI andtransition in the route. Note that the POI and transition scores arescaled differently to support better visual discrimination . For aseamless match between a route on the map and the correspondingPOI scores in the bar plot, we use the same colour for both POI scoreand POI icon on the map. We also allow users to select multiplerows to visualise the corresponding routes on the map. POI list : The POI list box provides the list of POI names and cat-egories along the recommended route. The list is sorted accordingto the suggested visiting order, and again, the same POI colour isused to match the corresponding POI on the map. On top of the list,the system also provides an estimated travel time and total distanceof the route. The POI list box is updated whenever a user selects adifferent route or the system makes a new recommendation. If morethan one route is selected, the system displays the information ofthe most recent chosen route.
POI feature visualisation : We further provide a radar chart toanalyse the variation between POIs in a single route. For example,in Figure 3, we compare two POIs (
Melbourne Aquarium and
QueenVictoria Market ) in terms of POI features and their importance in See Appendix for details at http://arxiv.org/abs/1707.01627
Figure 3: POI feature comparison between
Melbourne Aquar-ium and
Queen Victoria Market : the former scores higher on
Popularity and
Visits difference features whereas the latterscores higher on
Visits and
Popularity difference features. the suggested route. The radar chart shows the corresponding POIfeature scores when a user selects a route. In particular, the usercan check/uncheck any POI in the selected route, and the featurescores of all checked POIs will be shown in the chart.
In this demonstration, we detail an interactive route analyser whichhelps the interaction between users and a route recommendationsystem. The system benefits from the explicit feature constructionof the structured prediction model, and visualises recommendedroutes in terms of information on both the routes and the POIs.
Acknowledgements.
We thank the National Computational In-frastructure (NCI), supported by the Australian Government, forcomputational resources. This work is supported in part by the Aus-tralian Research Council via Discovery Project program DP140102185.
REFERENCES [1] Dawei Chen, Cheng Soon Ong, and Lexing Xie. 2016. Learning Points and Routesto Recommend Trajectories. In
Proceedings of the 25th ACM International onConference on Information and Knowledge Management . ACM, 2227–2232.[2] Dawei Chen, Cheng Soon Ong, and Lexing Xie. 2016. Learning Points and Routesto Recommend Trajectories (CIKM ’16) .[3] Dawei Chen, Lexing Xie, Aditya Krishna Menon, and Cheng Soon Ong. 2017.Structured Recommendation.
CoRR
IEEE transactionson visualization and computer graphics
19, 12 (2013), 2277–2286.[6] Thorsten Joachims, Thomas Hofmann, Yisong Yue, and Chun-Nam Yu. 2009.Predicting structured objects with support vector machines.
Commun. ACM
Computer
42, 8 (2009).[8] Ching-Pei Lee and Chuan-bi Lin. 2014. Large-scale linear rankSVM.
Neuralcomputation
26, 4 (2014), 781–817.[9] Kwan Hui Lim, Jeffrey Chan, Christopher Leckie, and Shanika Karunasekera.2015. Personalized tour recommendation based on user interests and points ofinterest visit durations (IJCAI ’15) .[10] N.B. Robbins. 2005.
Creating more effective graphs . Wiley-Interscience.[11] Ben Shneiderman. 1996. The eyes have it: A task by data type taxonomy for infor-mation visualizations. In
Visual Languages, 1996. Proceedings., IEEE Symposiumon . IEEE, 336–343.[12] Wikipedia. 2017. Parallel coordinates. https://en.wikipedia.org/wiki/Parallel_coordinates. (2017). Accessed: July 2017. athRec: Visual Analysis of Travel Route Recommendations
A ALTERNATIVE APPROACHES TO TRAJECTORY RECOMMENDATION
A number of approaches have been proposed to solve the trajectory recommendation problem. Lim et al. [9] formulated an optimisationproblem inspired by the travelling salesman problem, and Chen et al. [2] proposed to learn a RankSVM [8] model to rank POIs with respectto a query, in particular, the training objective ismin w w ⊤ w + C · n (cid:213) i = (cid:213) ( p , p ′ )∈R( x ( i ) ) max (cid:16) , − w ⊤ (cid:16) Φ ( x ( i ) , p ) − Φ ( x ( i ) , p ′ ) (cid:17)(cid:17) , (1)where w denotes the model parameters, Φ is a query-POI feature mapping, C > R( x ) is the set of POI pairs ( p , p ′ ) such that p is ranked above p ′ , e.g. POI p is observed more often than POI p ′ , with respect to query x . Lastly, the top-ranked POIs withrespect to the given query were taken to form a trajectory.Instead of ranking POIs, a Markov chain was learned from routes in the travel history [2], and recommending a trajectory is achievedwith either the classic Viterbi algorithm or an integer linear programming when repeated POI visits are discouraged. Further, Chen et al. [2]proposed to combine the POI ranks learned by RankSVM (1) and the transition preferences learned by a Markov chain using a heuristicwhich traded off between point affinity and transition preference. B SCORE NORMALISATION FOR VISUALISATION
We visualise multiple suggested trip by ranking them according to their scores from SSVM, in addition, we leverage the linear form of SSVMthat score of a trajectory is the sum of point and transition scores in the trip. As shown in Figure 2, scores of both POIs and transitions in atrip are visualised in a stacked bar plot, which helps users of the system to better understand what contributes to a suggested trip.To support better visual discrimination, we perform linear scaling on scores of suggested trajectories, scores of POIs and transitions.
B.1 Scaling overall trajectory scores
Specifically, we scale the trajectory scores for the top 10 suggested routes such that: • The first trajectory scores 100, • and the last (i.e. the 10-th) trajectory scores 10. B.2 Displaying unary and pairwise scores
We further scale the POI scores using the same scaling parameters as that of trajectory scores, however, to visualise the transition scores inthe stacked bar plot, we perform another linear scaling as the scores of POIs and transitions are in quite different range, in particular, thetransition scores are scaled linearly into the range [0.1, 1].
B.3 Displaying POI features
Besides scores of POIs and transitions of a suggested trip, we also visualise the scores for individual POI features (e.g. popularity, visitduration, category) in a radar chart, as shown in Figure 3. We perform another linear scaling such that scores of POI features are in the range[1, 10].
C CHOICES ON VISUALISATION PARADIGMS
We want to visualise a ranked list of suggested trips instead of a single choice, further, we would like to break down the score of a trip intothe contributions of its unary and pairwise components. Following the Visual Information-Seeking Mantra [11], we made the followingdesign choices for this visualisation system:
Overview first.
For a given query, as shown in Figure 1, the visualisation system shows the top-ranked trip on a map. The other componentsof the system, as described in Section 3, can display more details of these recommendations after a user zooms in, as described below.
Zoom and filter.
The system shows the scores of top-10 suggested trajectories and the scores of the corresponding POIs and transitions ina stacked bar plot, as shown in Figure 2. Further, a POI list box at the lower left side of the system interface displays the information of POIsincluding names and categories of the selected trip, as described in Section 3. Lastly, at the lower right side, contributions of individualfeatures of all POIs in the selected trip are compared, and Figure 3 shows an example which compares individual feature scores between twoPOIs.