WO2013041517A1

WO2013041517A1 - A method to generate a personalized tourist route

Info

Publication number: WO2013041517A1
Application number: PCT/EP2012/068325
Authority: WO
Inventors: Xavier Amatriain; Antonio CEBRIAN; Miquel RAMIREZ; Neal Lathia
Original assignee: Telefonica, S.A.
Priority date: 2011-09-22
Filing date: 2012-09-18
Publication date: 2013-03-28

Abstract

A method to generate a personalized tourist route. The method of the invention comprises automatically performing the next steps: - building a generic city model of a specific city comprising a list of Points of Interest, or POIs, from geolocated user-defined metadata information from a plurality of users; - filtering the information contained in said generic city model according to a user model generated with user-defined metadata information of a specific user, to obtain a personalized city model for said specific user; and - generating said personalized tourist route from said personalized city model.

Description

A method to generate a personalized tourist route

Field of the art

The present invention generally relates to a method to generate a personalized tourist route and more particularly to a method that comprises building a generic city model of a specific city comprising a list of Points of Interest, personalizing said generic model city according to the user's interests and generating said personalized tourist route according to user context variables and other constraints.

Prior State of the Art

The tourist domain is without a doubt one of the domains that has changed the most dramatically since the popularization of the Web. The traditional business model has changed in many ways since now most commercial transactions are performed online. Consumers value the possibility of accessing and comparing endless options. This allows them to choose to the smallest detail and even customize or personalize their travel plans. Furthermore, users have become much more informed and can combine this seemingly endless commercial choice with access to online resources about tourist destinations, activities, and sites to visit. When planning a city tour, a typical user will be confronted with three different tasks: first, there is the need to select the places to visit. This selection process can be simplified by visiting the most popular ones. But, ideally, one would like to personalize this selection to include places that relate to personal or group interests - i.e. sports, arts, museums, outdoors... Once these Points of Interest (POI) are selected, the tourist needs to factor in constraints such as time or budget availability or other contextual variables such as weather or transportation costs. Finally, the tourist needs to define a plan where all of these factors are taken into account while still allowing for some serendipity.

The availability of so much information and options might seem to be only beneficial to users. However, it is a known fact that the information overload caused by having so many choices might, in fact, result in the consumer making less and worse choices. The way to counter this effect is to offer tools that automatically filter information according to the user interests, context and needs. Recommender systems are a class of such tools in which a user utility function is optimized to an implicit query given by the user profile and possibly context. But in a tourist setting, it is not enough to filter and recommend places according to context and interest: Recommendations should also be set into an appropriate sequence that takes into account traveling constraints. Mobility Tracking and Activity Recognition

The main input to these systems is users' location, which tends to provide insight into the usage and characteristics of built environments; for example, Girardin et al. [1] were able to classify locations as tourist or work areas based on aggregated mobile phone usage. Mobile recommender systems, however, are not fully detached from their web-based equivalents. Semantic descriptions and geographic "hotspots" have been identified based on how web users are annotating geo-tagged Flickr photographs [2].

Peebles et al. [3] develop a framework for learning user-behaviour from crowd- sourced sensor data. Imperfect labels are regrouped via similarity measures. The framework was evaluated by deploying a set of users who logged their activities via audio tags.

Geo-Spatial Analysis and POI Discovery from Web Data

The development of Web 2.0 services has triggered the explosion of user- generated content that is uploaded to the web. More importantly, this data is often rich with annotations describing the user's location (e.g., latitude/longitude meta-data) and activities (such as user-input tags). The research in this domain spans two fields. First, the meta-data can be used to analyse the physical structure of urban environments. Crandall et al. [5] examined the collective photographing behaviour of the world's tourists to identify points of interest and rank the locations that the world is "paying attention to". These datasets can then be used to deduce trip-related information [6] and augment image search results [7]. The second field examines the implicit structure that emerges as user-generated content is uploaded to the web, such as the social network between photographers [8] and the collective behavior of web users who tag content [9]. The overall theme here is that of examining structure: user generated content tells us both about the world we live in and how we connect to one another.

Tourist Dynamics and Itinerary Generation

The exploration of how tourists navigate urban environments is a natural extension to the above: once POIs have been discovered, paths between them can be computed. Systems such as the one proposed by Choudhury et al. [10] aim to automatically generate travel itineraries that "reflect the wisdom of the touring crowds". To do so, the authors delve further into the data to not only identify POIs, but also estimate visit times and infer paths and transit times between them. The authors propose a pipeline of multiple heuristics that first extracts tourist trip data from photos to generate a POI graph and then uses the graph to generate intra-city tours. The GUIDE system [1 1], on the other hand, operates on pre-defined POIs and generates recommendations that are tailored to the locations that are being recommended (e.g., opening times). More generally, the intuition behind these systems is that they should be context-aware. In the case of the CyberGuide system [12], context-awareness meant incorporating a user's current and historical location.

Personalisation and Location/Mobile Recommendation

There is a growing body of literature about location and mobile recommendations. In practice, these are two kinds of system that often overlap. The first are systems that recommend locations, used to find nearby restaurants [14], shops [15], museum exhibits [16] or leisure activities [17]. The second are systems that recommend while on the move; these are mobile interfaces to, for example, movie recommender systems [13]. In practice, it is often the case that systems that recommend where to go are available when on the move, in order to both capture relevant location information and to be available in the precise moments when users would need them most.

In the context of the GeoLife (http://research.microsoft.com/en- us/projects/geolife/) project, Zheng et al. have built and tested activity and location mobile recommender systems for urban spaces [18] [19]. They build recommendations by reasoning on GPS traces of people's movements, which exploit correlations between locations or activities. The raw data is not readily available for this task: POIs are discovered by thresholding user's movements into stay points and then clustering the set of points into stay regions; extra information is also retrieved from POI databases. The evaluation of this work centered on users who had at least a 3-month location history recorded. To that extent, GeoLife targets locals rather than, as we do here, tourists. Tintarev, Amatriain and Flores [20] targeted tourists in a field study exploring the potential of personalisation techniques to improve their experiences as they navigate cities. While they found users followed more recommendations when directed toward non-personalized POIs, personalization lead to less repeat visits (of popular locations) and greater discovery and novelty of the locations explored. These results, along with similar studies, highlight the interplay between how knowledgeable users are of the places they visit and their propensity to follow "off the beaten track" recommendations. In fact, user's cumulative satisfaction was higher when recommendations were tailored to their preferences: these field studies lay a solid foundation to the system we propose here.

There are a variety of other works that have addressed the topic of personalization in tourist guides, including [22] [23] and [24].

In current works on POIs, the analyses are mainly used to gain insight into the world's physical and social structure, rather than used for a particular application scenario (such as recommending travel itineraries). Moreover, they are focused on collective behaviour rather than tailoring location discovery to individual user's needs. Aggregate values will not highlight underlying user-centric differences (e.g., preferences) and reflect popular, rather than personalized, location discovery. As it will be later explained, the process of finding "generally accepted" POIs that represent a city is different from the process of finding all POIs that can be valuable to some of the users of a particular application such as an itinerary recommendation engine.

Works on tourist dynamics and itinerary generation look at the problem of itinerary generation for city tourists. However, the existing proposals inherently do not incorporate user preferences or personalisation and they are tailored toward visiting popular locations. The underlying assumption is that all tourists will be interested in the same things.

Related to the Personalisation and location mobile recommendations, the literature highlights shortcomings on two main points: (a) there is no fully comprehensive system, that incorporates POI discovery with city and user modelling to compute personalized itineraries, (b) few systems are easily generalizable to any location, since they rely on pre-existing POI databases or manual user input.

Description of the Invention

It is necessary to offer an alternative to the state of the art which covers the gaps found therein, particularly related to the lack of proposals which really constitute a generic, flexible and automatic approach to personalized tourist route recommendation as well as non heuristic-based approaches that assume that each tourist is interested in different things. To that end, the present invention provides a method to a method to generate a personalized tourist route.

On contrary to the known proposals, the method of the invention, in a characteristic manner it comprises automatically performing the next steps:

- building a generic city model of a specific city comprising a list of Points of

Interest, or POIs, from geolocated user-defined metadata information from a plurality of users;

- filtering the information contained in said generic city model according to a user model generated with user-defined metadata information of a specific user, to obtain a personalized city model for said specific user; and

- generating said personalized tourist route from said personalized city model. Other embodiments of the method of the first aspect of the invention are described according to appended claims 2 to 15, and in a subsequent section related to the detailed description of several embodiments.

Brief Description of the Drawings

The previous and other advantages and features will be more fully understood from the following detailed description of embodiments, with reference to the attached drawings, which must be considered in an illustrative and non-limiting manner, in which:

Figure 1 shows a complete scheme and flow diagram of the steps to be performed in order to obtain a personalized tourist route, according to the method of the present invention.

Detailed Description of Several Embodiments

The present invention deals with the automatic generation of optimized and personalized tourist routes, more particularly, relates to a completely automatic recommender system for suggesting a sequential ordered list of selected Points-Of- Interest (POIs) taking into account contextual information, user interests and other constraints.

The present invention serves to solve the aforesaid problems by providing a tourist recommender system that is able to recommend personalized and contextualized sightseeing tours derived from publicly available web resources, taking into account contextual information, user interests and other constraints.

The method provided by the system, starting from the construction of a city model up to the final personalized recommendation is automatic and generic to any city, user, and context:

- First, the system automatically builds a city model or generic list of POIs by extracting geolocated user-defined tags (metadata information) from collaborative (content sharing) and other publicly available web resources (curated content selected by city experts).

- Second, the system uses the same resources to automatically derive the user interests regarding tourist activities. In order to do this, the system analyses the tags that a given user has introduced as metadata information in the above-mentioned web resources.

- Finally, the system use a formal tourist route planning approach that is able to take the previous two inputs and generate a meaningful (optimized) and personalized city tour recommendation. Using a given user model, the system can then filter out non-relevant POIs and establish the personalized benefits of visiting the rest. At the end, this process generates a personalized city model that is tailored to the given user and might be also influenced by contextual variables and other user-defined constraints (e.g.: time, budget, etc.).

The goal of the present invention is a fully automated and generic approach to generating personalized tourist routes. The main components of the approach and general architecture of the proposed solution is described next, according to the scheme shown in Figure 1.

The two data inputs are two datasets of tagged images: one for a given city, and the other one for a target user. These datasets will be used to extract the tourist model of the city as well as the user model containing the interests and preferences of the target user. It is important to note that the proposed solution is by no means tied to a particular data source. What the system needs ultimately is a set of tagged locations for the city and a set of weighted tags for the user. For the locations, it will be necessary to provide the location plus a list of (possibly) weighed tags that describe the location. The descriptive tags can include anything going from high-level concepts such as "architecture" to concrete labels such as "Gaudi". The important issue is that it is necessary that this set of tags intersect with the ones used to describe the users' interests. The matching between users and places will be done by measuring the distance between these tags.

The city model is obtained through a clustering process that will output a general tourist model of the city. This model, which is computed off-line and is generic to all users, is made of Points of Interest (POI's) that are considered relevant, together with their location and associated tags with frequencies.

The clustering process works by taking a (possibly large) set of locations and grouping them into the relevant POI's. This grouping is performed by defining a distance measure between locations. This distance measure takes into account both the geographical distance as well as (possibly) the semantic distance as measured by the cosine distance between tags. Once the distance measure is defined, several approaches to clustering are possible, taking into account that Hierarchical Clustering might work best.

Tags related to locations may be processed using standard text processing techniques such as stop word removal, stemming, and the use of thesaurus. Relative frequencies are computed by using the Term Frequency Inverse Document Frequency (TFIDF) approach.

Before clustering, it shall be determined what part of the dataset corresponds to tourists and filter out those that don't. In order to do this, a heuristic approach is used, where a tourist is defined as a person who has stayed at most D days in the city. Once this subset of the original dataset is obtained, the cluster algorithm takes in two control parameters that will define what is considered a relevant POI. The first one determines the minimum number of users that need to agree on a given location to consider it is relevant. The higher this parameter, the more it will tend to focus on popular sights. The second parameter defines the minimum radius for relevant POI's. One of the important issues when clustering locations together into POI's is to decide when two locations in fact define two different POIs or the same one. This parameter allows controlling the amount of geo-granularity wanted for the city model. These parameters need to be manually set and depend on the final characteristics wanted for the application (e.g. more variety of POIs vs. more focus on popular ones).

The user dataset is converted into a user's interest model in a quite straightforward manner: tags and frequencies are extracted and use those to describe the user interests. For example, if a user has tagged 50 out her 100 pictures with the word "architecture" the conclusion is that she is very interested in "architecture". Note that in this case, no more information— e.g. location of pictures is needed. Tags and frequencies from the user model are used in two different processes. On the one hand, the non-existence of given tags in the user model is used to filter-out non-relevant POI's from the city model. Thus, a generic city model made up of many POI's possibly relevant to any user is replaced by a personalized city model with only those POI's that might be relevant to the target user. On the other hand, tag frequencies are used in both the user and city models to infer the relative interest of each relevant POI. Vectors for each POI are constructed using the standard "Vector Model" where tags are weighed by their TF-IDF value. Tags for the user are weighed by the relative interest the user has in each "concept". This can be done automatically by accessing some content that the user has generated (e.g. Twitter profile) and again applying the TF-IDF approach. But it can also be done manually by asking the user to weigh her relative interest in a set of pre-defined tags by means of an initial user survey.

Once constructed a personalized model of the given city for the target user with cost weighs for POIs, personalized tourist routes that observe a number of constraints can be generated. Any constraint related to the user context, such as weather or time of the day can be used to generate different personalized versions of the city model by weighing the POI tags. However, three constraints are directly input to the planning process: start and end location, and available time. These constraints are input to the planner, together with the personalized city model in order to generate the final personalized city tour for the target user, given those constraints. Several kinds of automatic planners are possible. In the present invention an instance of the LAMA planner [26] is implemented.

The proposed solution is generic enough that it can observe these as well as any other constraint — e.g. we could easily extend those constraints to observe contextual variables such as the current weather, or user needs such as how hungry or thirsty she is.

In one particular embodiment of the invention, the system provides a method that consists in the analysis and clustering of user pictures uploaded to a photo sharing web service (e.g. Flickr or Picasa) in order to derive the so-called Points of Interest (POI). Apart from the location of these POIs, they are also characterized by analysing the collaborative tags attached to them. In another embodiment of the invention, the city model extracted from this wisdom of the crowds is complemented with curated city models automatically extracted from expert city guides obtained from online resources.

An embodiment of the invention that uses Flickr data to derive both the model of the city and the user has been implemented. The solution consists of a set of backend software artefacts and a frontend using Google Maps that renders the final information about POIs and routes.

The backend is a software framework written in a mix of Scala and Java, where Scala was used primarily for architecting the solution and Java for reusing some legacy code components intended mainly for crawling the sites. The framework fires some daemons that periodically query Flicker for new geo-positioned data. This data is downloaded into a file system that stores the raw data consisting of the coordinates and the associated tags and comments. Another process runs in background monitoring this file system and once some changes are detected, the before mentioned process for POI generation is run and the results are stored in an internal MySQL database.

A frontend is developed for easy querying and visualization of the generated information. It is a web application coded in Javascript that utilizes Google Maps Javascript API as the canvas for displaying geographical information. This web application communicates with our backend services through REST APIs serving JSON objects with the representation of the POIs and the tourist routes. Those JSON objects are transformed in the browser into suitable objects that the Google Maps API can understand. Advantages of the invention

In summary, the main contributions of the present invention are:

- A generic, flexible and automatic approach to personalized tourist route recommendation

- A novel approach to construct city models by clustering tagged user photos and discovering POI's

- A way to filter POI's according to a user profile to construct personalized city models

- A way to use formal planning methods for deriving a personalized constrained city tour

A person skilled in the art could introduce changes and modifications in the embodiments described without departing from the scope of the invention as it is defined in the attached claims. ACRONYMS

POI Point Of Interest

TF-IDF Term Frequency-Inverse Document Frequency

REFERENCES

[1] Fabien Girardin, Josep Blat, Francesco Calabrese, Filippo Dal Fiore, and Carlo Ratti. Digital footprinting: Uncovering tourists with user-generated content. IEEE Pervasive Computing, 7(4):36B43, 2008

[2] Tye Rattenbury, Nathaniel Good, Mor Naaman: Towards automatic extraction of event and place semantics from flickr tags. SIGIR 2007: 103-1 10 [3] D. Peebles and H. Lu and N. D. Lane and T. Choudhury and A. T. Campbell, "Community-Guided Learning: Exploiting Mobile Sensor Users to Model Human Behavior", in Proceedings of the Twenty-Fourth AAAI Conference on Artificial Intelligence (AAAI-10). 2010 [4] J. Froehlich, M. Chen, I. Smith, and F. Potter. Voting With Your Feet: An Investigative Study of the Relationship Between Place Visit Behavior and Preference. In ACM Ubicomp, 2006.

[5] D. Crandall, L. Backstrom, D. Huttenlocher, and J. Kleinberg. Mapping the World's Photos. In WWW, Madrid, Spain, April 2009.

[6] Adrian Popescu and Gregory Grefenstette. Deducing Trip Related Information from Flickr. In WWW, Madrid, Spain, April 2009. [7] L. Kennedy and M. Naaman. Generating Diverse and Representative Image Search Results for Landmarks. In WWW, Madrid, Spain, April 2008.

[8] S. Golder. Measuring Social Networks with Digital Photograph Collections. In ACM Conference on Hypertext and Hypermedia, 2008.

[9] C. Marlow, M. Naaman, D. Boyd, and M. Davis. Position Paper, Tagging, Taxonomy, Flickr, Article. In Collaborative Web Tagging Workshop (WWW), 2006. [10] M.D. Choudhury, M. Feldman, S. Amer-Yahia, N. Golbandi, R. Lempel, and C. Yu. Automatic Construction of Travel Itineraries using Social Breadcrumbs. In ACM Hypertext, Ontario, Canada, June 2010.

[1 1] K. Cheverst, N. Davies, K. Mitchell, A. Friday, and C. Efstratiou. Developing a Context-Aware Electronic Tourist Guide: Some Issues and Experiences. In ACM CHI, The Hague, The Netherlands, 2000.

[12] G.D. Abowd, C.G. Atkeson, J. Hong, S. Long, R. Kooper, and M. Pinkerton. Cyberguide: a Mobile Context-Aware Tour Guide. Wireless Network, 3,2007.

[13] B. Miller, J.A. Konstan, and J. Riedl. PocketLens: Toward a Personal Recommender System. In ACM TOIS, July 2005.

[14] F. Ricci and Q. N. Nguyen. Critique-Based Mobile Recommender Systems. OGAI Journal, 2005.

[15] Y. Takeuchi and M. Sugimoto. An Outdoor Recommendation System Based on User Location History. In ACM Ubicomp, 2006. [16] F. Bohnert, D.F. Schmidt, and I. Zukerman. Spatial Processes for Recommender Systems. In Proceedings of the 21 st Int Joint Conference on Artificial Intelligence, Pasadena, California, July 2009.

[17] V. Bellotti, B. Begole, E.H. Chi, and et al. Activity-Based Serendipitous Recommendations with the Magitti Mobile Leisure Guide. In ACM CHI, Florence, Italy, April 2008.

[18] Y. Zheng, L. Zhang, X. Xie, and W. Ma. Mining Interesting Locations and Travel Sequences From GPS Trajectories. In WWW, Madrid, Spain, April 2008.

[19] V.W. Zheng, Y. Zheng, X. Xie, and Q. Yang. Collaborative Location and Activity Recommendations with GPS History Data. In WWW, Raleigh, North Carolina, April 2010. [20] N. Tintarev, X. Amatriain, and A. Flores. Off the Beaten Track: A Mobile Field Study Exploring the Long Tail of Tourist Recommendations. In MobileHCI, Lisbon, Portugal, September 2010. [21] M. Modshing, R. Kramer, K. ten Hagen, and U. Gretzel. Effectiveness of Mobile Recommender Systems for Tourist Destinations: A User Evaluation. In AAAI Spring Symposium, Vancouver, Canada, 2007.

[22] M. van Setten, S. Pokraev, and J. Koolwaaij. Context-Aware Recommendations in the Mobile Tourist Application COMPASS. In Adaptive Hypermedia and Adaptive Web-Based Systems, Eindhoven, The Netherlands, 2004.

[23] L. Ardissono, A. Goy, G. Petrone, M. Segnan, and P. Torasso. Intrigue: Personalized Recommendation of Tourist Attractions for Desktop and Handset Devices. Applied Artificial Intelligence, 17, 2003.

[24] L.T. Lin C.H. Tai, D.N. Yang and M. S. Chen. Recommending Personalized Scenic Itinerary with Geo-Tagged Photos. In ICME, Hannover, Germany, June 2008. [25] S. C. Johnson. Hierarchical clustering schemes. Psychometrika, 32(3):241{254, 1967.

[26] S. Richter and M. Westphal. The LAMA planner: guiding cost-based anytime planning with landmarks. Journal of Artificial Intelligence Research. Volume 39 Issue 1 , September 2010

Claims

1. - A method to generate a personalized tourist route, characterised in that it comprises automatically performing the next steps:

- generating said personalized tourist route from said personalized city model.

2. - A method as per claim 1 , wherein said user model comprises descriptive tags of images from said specific user and tag frequencies of said descriptive tags.

3. - A method as per claim 1 or 2, wherein said geolocated user-defined metadata information is extracted from generic tags of images of said specific city from said plurality of users.

4. - A method as per claim 3 when depending on claim 2, comprising calculating tag frequencies of said descriptive tags and/or said generic tags by means of a Term Frequency Inverse Document Frequency approach.

5.- A method as per any of previous claims, comprising retrieving said geolocated user-defined metadata information and said user-defined metadata information from collaborative, content sharing and/or publicly available web resources.

6. - A method as per any of previous claims, comprising building said generic city model off-line and by means of a clustering process, said generic city model further comprising the location of POIs of said list of POIs.

7. - A method as per claim 6, wherein said clustering process makes use of a distance measure between locations, said distance measure comprising geographical distance and/or semantic distance between said generic tags.

8. - A method as per claim 6 or 7, comprising performing said clustering process by means of a Hierarchical Clustering.

9. - A method as per any of previous claims, comprising generating said list of POIs according to a minimum number of users that have agreed on considering a given location as relevant and/or to the minimum radius between two or more of said POIs.

10. - A method as per claim 9, wherein each of said users has stayed in said given location a minimum number of days, said given location being associated to one of said POIs.

1 1 . - A method as per any of previous claims 3 to 10 when depending on claim 2, comprising filtering-out said POIs that are not related to any of said descriptive tags of said user model in order to obtain said personalized city model.

12. - A method as per any of previous claims 3 to 1 1 when depending on claim 2, comprising constructing a plurality of vectors for each of said POIs containing cost weights according to Term Frequency-Inverse Document Frequency, or TF-IFD, values of said generic tags and to TF-IFD values of said descriptive tags or to a rating of said descriptive tags provided by said given user.

13. - A method as per claim 12, comprising using said plurality of vectors in order to obtain said personalized city model.

14. - A method as per any of previous claims comprising generating said personalized tourist route according to user context variables.

15. - A method as per claim 15, wherein said context variables are at least one of: start location, end location, available time, budget, weather, and time of the day.