A Generic Semantic-based Framework for Cross-domain Recommendation

Transcription

1 A Generic Semantic-based Framework for Cross-domain Recommendation Ignacio Fernández-Tobías, Marius Kaminskas 2, Iván Cantador, Francesco Ricci 2 Escuela Politécnica Superior, Universidad Autónoma de Madrid, Spain ign.fernandez0@estudiante.uam.es, ivan.cantador@uam.es 2 Faculty of Computer Science, Free University of Bozen-Bolzano, Italy mkaminskas@unibz.it, fricci@unibz.it

2 Contents Cross-domain recommendation Case study: adapting music recommendation to points of interest A semantic-based framework for cross-domain recommendation Semantic-based knowledge representation Semantic graph-based recommendation algorithm Preliminary results Future work

3 Contents 2 Cross-domain recommendation Case study: adapting music recommendation to points of interest A semantic-based framework for cross-domain recommendation Semantic-based knowledge representation Semantic graph-based recommendation algorithm Preliminary results Future work

4 Cross-domain recommendation 3 Recommender systems can help users to make choices, by proactively finding relevant items or services, taking into account or predicting the users tastes, priorities and goals The vast majority of the currently available recommender systems predict the user s relevance of items in a specific and limited domain

5 Cross-domain recommendation 4 In some applications, it could be useful to offer the user joint personalized recommendations of items belonging to multiple domains In an e-commerce site, we may suggest movies or videogames based on a particular book bought by a costumer In a travel application, we may suggest cultural events may interest a person who has booked a hotel in a particular place In an e-learning system, we may suggest educational websites with topics related to a video documentary a student has seen Potential benefits Offering diversity and serendipity Addressing the user cold-start problem (on the target domain) Mitigating the sparsity problem

6 Cross-domain recommendation 5 Some real applications do already recommend items from different domains, but their recommendations rely on statistical analysis of popular items, without any personalization strategy, or most of them only exploit information about the user preferences available in the target domain

7 Cross-domain recommendation 6 Research questions [Winoto & Tang, 2008]. At community level, are there correlations between user preferences for items belonging to the different domains of interest? 2. At individual level, can we build a recommendation model where each user s preferences in source domains are used to predict/adapt her preferences in target domains? 3. How should we evaluate the effectiveness of cross-domain item recommendations? [Winoto & Tang, 2008] Winoto, P., Tang, T If You Like the Devil Wears Prada the Book, Will You also Enjoy the Devil Wears Prada the Movie? A Study of Cross-Domain Recommendations. New Generation Computing 26(3),

8 Contents 7 Cross-domain recommendation Case study: adapting music recommendation to points of interest A semantic-based framework for cross-domain recommendation Semantic-based knowledge representation Semantic graph-based recommendation algorithm Preliminary results Future work

9 Case study: adapting music recommendation to points of interest 8 Recommending music artists that suit places of interest (POIs) Mobile city guide soundtrack Adaptive music playlist in a car [Braunhofer et al., 20] Braunhofer, M., Kaminskas, M., Ricci, F. 20. Recommending Music for Places of Interest in a Mobile Travel Guide. 5th ACM Conference on Recommender Systems.

10 Case study: adapting music recommendation to points of interest 9 In a previous work [Kaminskas & Ricci, 20], emotional tags were used to manually annotate places and music Emotional tags can be used to find matching between music and places of interest e.g. a monument and a music track may be described as strong and triumphant [Kaminskas & Ricci, 20] Kaminskas, M., Ricci, F. 20. Location-Adapted Music Recommendation Using Tags. 9th International Conference on User Modeling, Adaptation and Personalization,

11 Case study: adapting music recommendation to points of interest 0 In this work, we aim at automatically finding semantic relations between POIs and music artists We propose to explore the Web of Data (Linked Data) to find such relations Specifically, we propose to exploit DBpedia, the Linked Data version of Wikipedia DBpedia can be considered as a core ontology in the Web of Data Connected to many other ontologies Describing and linking more than 3.5 million concepts from a large variety of knowledge domains

12 Case study: adapting music recommendation to points of interest In this work, we aim at automatically finding semantic relations between POIs and music artists We propose to explore the Web of Data (Linked Data) to find such relations Specifically, we propose to exploit DBpedia, the Linked Data version of Wikipedia DBpedia can be considered as a core ontology in the Web of Data Connected to many other ontologies Describing and linking more than 3.5 million concepts from a large variety of knowledge domains

13 Case study: adapting music recommendation to points of interest 2 Issues to investigate, identified in [Winoto & Tang, 2008]. Correlations between user preferences for items of the different domains Correlations between POIs and music were established through tags in [Kaminskas & Ricci, 20] 2. Recommendation model to predict/adapt user preferences across domains This paper addresses this particular issue, presenting a semantic-based framework to support cross-domain recommendation 3. Evaluation of cross-domain recommendation effectiveness Future work

14 Contents 3 Cross-domain recommendation Case study: adapting music recommendation to points of interest A semantic-based framework for cross-domain recommendation Semantic-based knowledge representation Semantic graph-based recommendation algorithm Preliminary results Future work

15 A Semantic-based framework for cross-domain recommendation 4 Goal: finding semantic relations between a given POI and music artists Example: music artists related to the Vienna State Vienna State Wolfgang Amadeus Mozart Identified relations: Geographical: artists who were born, died or lived in Vienna Time-based: artists who were born, died or lived in the year (decade, century) the State of Vienna was built Category-based: artists who belong to music categories that are related through keywords to architecture structures/styles identified with the building of the of Vienna Tags: artists annotated with tags also assigned to the of Vienna

16 A Semantic-based framework for cross-domain recommendation A directed Acyclic Graph (DAG) representing semantic relations between concepts in two domains 5 Vienna Austria Arnold Schoenberg POI instance class CITY Mozart MUSIC ARTIST State of Vienna 9th century Brahms Bizet TIME houses opera composers Ballet venues ballet Ballet composers ARCHITECTURE CATEGORY KEYWORD MUSIC CATEGORY

17 A Semantic-based framework for cross-domain recommendation 6 The previous graph can be considered as a particular instance of a semantic class/category network The selection of classes and relations is guided by experts on the domains of interest and knowledge repositories CITY located in was born, died, lived in POI was built TIME was born, died, lived in MUSIC ARTIST belongs to ARCHITECTURE CATEGORY has keyword KEYWORD keyword of MUSIC CATEGORY subcategory of subcategory of

18 A Semantic-based framework for cross-domain recommendation 7 As a proof of concept, we have built our approach by exploiting DBpedia ontology in two stages:. Manually identifying DBpedia classes and relations belonging to the domains of interest to define the semantic-based knowledge representation 2. Automatically obtaining related DBpedia instances according to the classes and relations identified in the first stage Semantic framework 2 Semantic network Wolfgang Amadeus Mozart POI MUSIC ARTIST Vienna State

19 Contents 8 Cross-domain recommendation Case study: adapting music recommendation to points of interest A semantic-based framework for cross-domain recommendation Semantic-based knowledge representation Semantic graph-based recommendation algorithm Preliminary results Future work

20 Semantic graph-based recommendation algorithm 9 In the semantic network, a final score for each concept can be computed by weight spreading strategies Initial weight values for concepts and relations must be established Vienna Austria Arnold Schoenberg Mozart State of Vienna 9th century Brahms Bizet houses opera composers Ballet venues ballet Ballet composers

21 Semantic graph-based recommendation algorithm 20 = Vienna Austria Arnold Schoenberg Mozart = Bizet State of Vienna 9th century Brahms = houses opera composers = Ballet venues ballet Ballet composers

22 Semantic graph-based recommendation algorithm 2 Vienna Austria Arnold Schoenberg Mozart Bizet State of Vienna 9th century Brahms =0.2 houses opera composers =0.2 Ballet venues ballet Ballet composers

23 Semantic graph-based recommendation algorithm 22 Vienna Austria Arnold Schoenberg Mozart Bizet State of Vienna 9th century Brahms houses opera composers =0.08 Ballet venues ballet Ballet composers 0.2 =0.08

24 Semantic graph-based recommendation algorithm =.86 Vienna Austria =.048 Arnold Schoenberg =0.048 Mozart =0.09 Bizet State of Vienna 9th century Brahms houses opera composers Ballet venues ballet Ballet composers 0.08

25 Semantic graph-based recommendation algorithm =.86 Vienna Austria =.048 Arnold Schoenberg =0.048 Mozart =0.09 Bizet State of Vienna 9th century Brahms houses opera composers Ballet venues ballet Ballet composers 0.08

26 Semantic graph-based recommendation algorithm 25 The initial weights of an edge in the graph can depend on the relevance of the linked instances and of the corresponding semantic classes These relevance values could be assigned in different ways V( I, I') f rel r ( I, I'),rel r( CI, CI ') Class relevance Domain expert e.g. a city is more informative to link a POI than a keyword Relation relevance Entity semantic similarity e.g. co-occurrences of concepts Mozart and Vienna within a document collection Instance relevance User profile e.g. an interest in Mozart s compositions the relevance for Mozart gets higher

27 Semantic graph-based recommendation algorithm In general, the weight of an instance not only depends on its relevance value and that of its class, but also inductively on the weights of the predecessors in the network I,, I W( I) g rel e( I),rel e( CI ); W( I),, W( Ik ); V( I, I),, V( I k k, I) 26 To preliminarily test our approach we have implemented a simple retrieval algorithm computing weights by linear combination V( I, I') rel r( I, I') ( ) rel r( C I, CI '), [0,] k W ( I) W ( I p ) V ( I p, I) ( ) rele ( CI ), [0,] p

28 Contents 27 Cross-domain recommendation Case study: adapting music recommendation to points of interest A semantic-based framework for cross-domain recommendation Semantic-based knowledge representation Semantic graph-based recommendation algorithm Preliminary results Future work

29 Preliminary results Example: Vienna State (Vienna, Austria) 28

30 Preliminary results Top 0 musicians for Vienna State 29 Arnold Schoenberg Wolfgang Amadeus Mozart Emil von Reznicek Alban Berg Ludwig van Beethoven Antonio Vivaldi Giovanni Felice Sances Fritz Kreisler Georg Christoph Wagenseil Antonio Salieri Music artist Top music genres Born/Death countries Date Classical Avant-garde Classical Instrumental Classical Classical Contemporary Classical Instrumental Classical Baroque Classical Baroque Classical Violin Classical Baroque Classical Italian Austria USA Austria Austria Austria Germany Hungary Austria Germany Austria Italy Austria Italy Austria Austria USA Austria Austria Italy Austria 20th century 8th century 20th century 20th century 9th century 8th century 7th century 20th century 8th century 9th century

31 Preliminary results 30 Example: found relations between Vienna State and Wolfgang Amadeus Mozart PLACE OF INTEREST: Vienna State CITY: Vienna, Austria MUSIC ARTIST: Wolfgang Amadeus Mozart ARCHITECTURE CATEGORY: houses KEYWORD: opera MUSIC CATEGORY: composers MUSIC ARTIST: Wolfgang Amadeus Mozart TAG: energetic MUSIC CATEGORY: composers MUSIC ARTIST: Wolfgang Amadeus Mozart TAG: sentimental MUSIC CATEGORY: composers MUSIC ARTIST: Wolfgang Amadeus Mozart MUSIC GENRE: classical MUSIC ARTIST: Wolfgang Amadeus Mozart ARCHITECTURE CATEGORY: Theatres TAG: animated MUSIC GENRE: classical MUSIC ARTIST: Wolfgang Amadeus Mozart

32 Preliminary results Example: Wembley Stadium (London, UK) 3

33 Preliminary results Top 0 musicians for Wembley Stadium 32 Beady Eye (Oasis band members) house The Woe Betides Skunk Anansie The Fallen Leaves Ivyrise Plastic Ono Band (John Lennon & Yoko Ono) We Are Balboa Goldhawks Teddy Thompson Music artist Top music genres Born/Death Countries Date Rock British Indie Rock British Rock Grunge Rock Female vocalist Garage Acoustic Rock Alternative Experimental Avant-garde Indie Rock Female vocalist Rock British Folk British UK (origin) UK (origin) UK (origin) UK (origin) UK (origin) UK (origin) UK (origin) Spain-UK (origin) UK (origin) UK USA

34 Preliminary results Example: found relations between Wembley Stadium and Beady Eye 33 PLACE OF INTEREST: Wembley Stadium CITY: London, United Kingdom MUSIC ARTIST: Beady Eye TIME: 2007 MUSIC ARTIST: Beady Eye ARCHITECTURE CATEGORY: Music venues ARCHITECTURE CATEGORY: Rock music venues KEYWORD: rock MUSIC CATEGORY: Indie rock MUSIC ARTIST: Beady Eye MUSIC CATEGORY: Rock music MUSIC ARTIST: Beady Eye TAG: strong MUSIC CATEGORY: Rock music MUSIC ARTIST: Beady Eye

35 Preliminary results 34 Automatic extraction of data from DBPedia for an input city Modular and extensible implementation of the framework Dataset 3098 POIs located in 2 European cities 47.5 POIs/city 697 architecture categories 229 are directly linked to POIs Avg..4 categories/poi 09 keywords describing 8 different architecture categories Avg.. keywords/category 568 music artists 6 music categories 309 directly linked to artists (avg..7 categories/artist) 5 related to keywords (avg..2 keywords/category) Time data for 64.72% of the POIs

36 Contents 35 Cross-domain recommendation Case study: adapting music recommendation to points of interest A semantic-based framework for cross-domain recommendation Semantic-based knowledge representation Semantic graph-based recommendation algorithm Preliminary results Future work

37 Future work 36 Evaluation user study Are semantically relations between POIs and music artists really appreciated by users in a recommendation scenario? Do users find cross-domain recommendations meaningful, and prefer them over nonadapted music suggestions? Providing personalized recommendations Cascade strategy Obtaining semantically related artists to the input POI Ranking (adding, removing) artists with a recommender based on the user s preferences

38 Future work 37 Initializing entity and relation weights Exploiting data statistics to estimate the popularity of the semantic entities and relations Exploring several weight spreading strategies Constrained Spreading Activation Node in/out degrees Weight propagation thresholds Path length thresholds Flow Networks Ford-Fulkerson s algorithm to find maximum network flow Semi-automatic defining the semantic framework Automatically exploring DBpedia to identify relevant entities and relations describing the domains of interest

39 A Generic Semantic-based Framework for Cross-domain Recommendation Ignacio Fernández-Tobías, Marius Kaminskas 2, Iván Cantador, Francesco Ricci 2 Escuela Politécnica Superior, Universidad Autónoma de Madrid, Spain ign.fernandez0@estudiante.uam.es, ivan.cantador@uam.es 2 Faculty of Computer Science, Free University of Bozen-Bolzano, Italy mkaminskas@unibz.it, fricci@unibz.it