In this paper we propose a new approach for improving the personalization of POIs recommender system. Existing context-aware POIs recommender systems usually take into account only peripheral contextual variables. We present Ricochet, an ontology-based system that refines the recommendation results by implementing an inter-POI parameter that we call the “complementarity”. We show how this new parameter can generate more effective recommendations. Our experiments are grounded using data from the location-based social network (LBSN) Yelp.com.
Recommender systems have changed the way people find products, information and even other people. They provide personalized recommendations and predictions over a large amount of information.
Places of interest, also called points of interest (POIs), are geographical marks that represent a certain importance for people because they play a specific role in the city. For example, places where we eat (restaurant), where we sleep (hotel), where we spend a good moment (bar) or where we participate in cultural activities (museum, theater).
With the rapid growth of location-based social networks (LBSNs), POIs recommender
systems are becoming increasingly popular. Various types of approaches can be found
both in academic literature (such as context-aware approach [
The research questions driving our work are: why do people go from one POI to another? Is there a link between these two POIs?
The contributions of this paper are two-fold: A set of recommendation criteria for constructing a relevant POIs recommender system, derived from a qualitative user study, based on interviews of 12 users of LBSN applications such as Foursquare and Yelp. An ontology of POIs compatible with Yelp taxonomy of place categories. A Semantic Web-based approach that is capable of re-ranking Yelp’s recommendations by taking into account the complementarity parameter.
The rest of the paper is organized as follows. In section 2, we present the state of the art. In section 3, we detail the “Ricochet” system. In section 4, we present the evaluation methodology and we report the results. In section 5, we summarize the outcome of this work and we mention some future work. 2
The POIs recommender systems are a recent but important research domain that attracts contributions from academic research institutions and companies constructing novel user applications.
In [
In [
In [
Except for academic papers, there are several commercialized applications like Foursquare, Yelp, Facebook places etc. They usually allow users to do check-ins. Recommendations in these LBSNs usually combine the collaborative filtering and the context. On Foursquare, some recommendations are based on the user’s and the user’s friends’ checkin history. For example, “You haven’t been here yet”, “Your friends have checked-in here” etc. Foursquare has released a new recommender system recently. Our work was conducted before. These proprietary applications do not use Semantic Web technologies and do not consider the complementarity parameter between POIs. In this paper we demonstrate the importance and the advantage of doing so. 3
In this section, we describe the Ricochet system. The presentation consists of the following parts: criteria of recommendation, construction of OntoPOI, recommendation engine. 3.1
We conducted user interviews in order to understand the important elements of POI choice. According to the responses of our interviewees, we found 3 types of criteria that people take into account when choosing a POI: contextual criteria, intrinsic criteria of POIs and criteria of complementarity between POIs.
Contextual criteria indicate the weather, the moment of the day etc. Intrinsic criteria of POIs indicate characteristics of POIs, for example, the opening hours, the popular hours, the price, the atmosphere, the comfort, the decoration, the location, the type of food, the quality of food, the quality of the service, the smells etc. Criteria of complementarity indicate relations between POIs and reasons why we go from one POI to another. In order to represent the knowledge about POIs and to do intelligent inferences, we decided to construct an ontology. 3.2
Like all ontologies, OntoPOI contains two basic components: classes and properties. The class "Thing" has four sub-classes: "Place", "Context", "Characteristic" and "Complementarity". "Context", "Characteristic" and "Complementarity" correspond with the three types of criteria of recommendation. The sub-classes of "Place" are the taxonomy of entertainment POIs. As we wanted to realize some experiments with Yelp’s data, we decided to use the same taxonomy as Yelp: "Active Life", " Arts and Entertainment ", " Beauty and Spas ", "Food", "Nightlife" and "Restaurant". OntoPOI is thus compatible with Yelp’s data and can reorganize them in an intelligent way thanks to the good inferring capacity of the logic-based ontology. OntoPOI can be downloaded at: http://sepage.com/ontology/OntoPOI_Ricochet.owl
In [
Arts & Entertainment= "feeling artsy" College & Education="feeling nerdy" Nightlife="feeling like a party animal" Great Outdoors="feeling outdoorsy" Shops="feeling shopaholic" Food="feeling hungry" Home / Work / Other="feeling workaholic"
Our work is aligned to this idea and takes it further. We go to a POI because we have a certain feeling at a given moment. This feeling can be caused by the POI that we just visited and the POI where we decide to go can satisfy this feeling. This feeling can be also interpreted as a need or a sensation. These feelings/needs/sensations can be physiological, as the hunger, the thirst and the elimination. They can also be physical or intellectual. The complementarity can be seen as the link between the POI that causes a need and the POI that satisfies this need. For example, the hunger can be caused by a POI where we make physical efforts and can be satisfied by a POI where we eat. The relation of complementarity between two POIs can be interpreted in the following way: POI 1 causes
Feeling satisfies
POI 2
In [
To represent this in the ontology, we used OWL Full4. We created two Object Properties “is caused by” and “is satisfied by” which have the class “Complementarity” as domain and rdfs:Class as range. The class "Complementarity" has four instances: "high intensity", "moderately high intensity", "moderately low intensity", "low intensity". We defined, for each instance, their values of the two properties by applying the process represented on Figure 2. We created some other Object Properties: "is the exposition of" (inside or outside) and "is the time for". We created some Datatype Properties: "has address", "has city", "has phone", "has postal code" etc which are basic information about POIs. 3.3
Recommendation engine We used several tools: Yelp API, Jena API and Google Maps API. Firstly, the Java program gets data by accessing to Yelp API that returns 50 POIs near the current location of the user. Secondly, the information of these POIs are translated into RDF triples and stored in Jena RDF repository with OntoPOI. Thirdly, according to the context and the check-in information of the user, we generate SPARQL queries, in order to determine the adequacy of each POI with regards to the user’s situation, according to 3 criteria. The total adequacy is calculated according to the following formula: Total point = Point of the weather criterion + Point of the moment of day criterion +3 * Point of the complementarity criterion.
Ricochet recommends the 10 highest-rated POIs in descending order. The recommended POIs are marked on a local map by using Google Maps API. This process is represented on Figure 3.
Location Context/
Check-in Rating POIs Recommendation GoogleMAP
Translation of Yelp’s data into
RDF Triple Query Processing Inference
Storage of data in RDF repository with OntoPOI In order to evaluate the effectiveness of our approach, we compared the recommendations produced by two variants of Ricochet. The first (R1) is that described above, the second (R2) is not complementarity-aware. Using these two variants, we could measure if the proposed complementarity-aware system can improve the user perceived relevance of the recommendations with the same data. The evaluations took place in Paris in France with
the participation of 10 persons. We varied the day time (morning, afternoon, night) and the place (not only downtown). We asked the evaluators to choose a POI nearby in our database and to imagine that they just visited it. We obtained the recommendations made by the two variants. Then, we invited our evaluators to rate each recommended POI by referring to the following scale: 0: I won’t go there (not relevant); 1: I hesitate (partially relevant); 2: I’ll go there without hesitation (definitely relevant). To gauge the relevance, we used the following metrics:
Precision was used to evaluate the quality of the recommendations. It is the number of relevant recommended POIs divided by the total number of recommended POIs. The POI having the score 2 counts for 1 relevant POI, 1 counts for 0.5, 0 doesn’t count.
Relevant recommended POIs Precision =
Total recommended POIs
Recall was used to evaluate the quantity of POIs extracted. We modified the traditional recall. We showed not all the database but only 10 recommended POIs. The evaluators judged the relevance of these shown POIs. We cannot know their appreciation about the non-shown POIs. The modified recall is the number of relevant recommended POIs divided by the total number of relevant recommended POIs of the two variants.
Relevant recommended POIs Recall (modified) =
Total relevant recommended POIsof the 2 variants
normalized Discounted Cumulative Gain (nDCG) was used to evaluate the quality
of the ranked list. The nDCG value of a ranking list at position n is calculated in the
following way:
log(j)
where ( ) is the rating of the j-th document in the ranking list, and the normalization
constant Zn is chosen so that the perfect list gets a nDCG score of 1 ([
N(n) ≡ Zn ∑
{2r(j) − 1 n j=1 2r(j) − 1, j = 1 , j > 1 R1 is the first variant, R2 the second. 1-10 are the numbers of the ten evaluations. We can see clearly that the performance of R1 is generally better than that of R2 in terms of the quality, the quantity and the ranking. For 1 (a park), the two performances are the same. It was the dinner time. During the mealtime, the rank of the food POIs is largely elevated.
The rest of the evaluations were done outside the mealtime where Ricochet privileged the complementarity parameter. The evaluators were pleased to be recommended complementary POIs. For 2 (a coffee), 8 (a massage center), we are well rested and often want a more intense activity, R2 recommended a beauty spa, a day spa, a massage, a tea room which have a similar intensity, R1 recommended only more intense POIs like a park, a dance studio, a museum etc. For 4 (a cinema), R2 recommended another cinema. For 5 (a gym club), 7 (a swimming pool), 10 (a tennis court), we normally feel tired, but R2 recommended still wearying POIs like a swimming pool, a gym club, a bike rental, a museum, R1 recommended more relaxing POIs, like a juice bar, a cinema, a cosmetic beauty supply, a coffee-tea. For 3 (a museum) and 6 (a hair salon), 9 (a musical venue), the performances are close. As before, Ricochet recommended complementary POIs that have a different intensity. But it turned out that the proposed intensity was not always suitable. People often need to change rhythm between two activities. But every individual has his own rhythm. The change of the rhythm seems to be submitted to a modulation according to the psyche of the individual. For example, some people like only the restful activities, for them, we would do better to eliminate the tiring activities. Others like all the activities, light as intense. For the latter, the range of the recommendations would deserve to be more refined. In our current system, there is only one rule on the change of the rhythm. However the appropriateness of a POI with regards to the rhythm is important and we intend to further personalize the taking into account of the rhythm, potentially by combining our approach with a machine learning approach.
To conclude this evaluation, even though the number of evaluators was limited, the results showed evidence that considering the complementarity can improve the relevance of the recommended POIs. In our future work we intend to conduct a more complete evaluation with more users and more specific situations. 5
In this paper, we presented Ricochet an ontology-based POIs recommender system and illustrated the advantage of this approach and the importance of implementing the complementarity parameter when recommending POIs.
In spite of good critics of our evaluators, there are still several weak points in our system to be improved in a future work.
Firstly, being dependent to Yelp API, we cannot have access to all recommendable
POIs but to a pre-selection made by Yelp. This is also the reason why we didn’t compare
our results against the baseline ordering from Yelp. This influences on the application of
the complementarity parameter and thus on the quality of the results. Secondly, we
identified but did not include yet the intrinsic criteria in our system. And nevertheless, these
criteria can potentially be useful for refining the results. Thirdly, it could be possible to
cover a more complete set of contextual information like in [
The POIs recommender systems are adopted by the modern society. An ongoing effort should be made to help people discover more personalized and relevant POIs.