User Personality and the New User Problem in a Context-Aware Point of Interest Recommender System
0 likes261 views
The document presents research on improving context-aware recommender systems (CARS) to address the cold start problem by utilizing user personality information. It demonstrates that incorporating personality traits enhances the accuracy of recommendations for new users compared to using demographic data alone. The proposed system, South Tyrol Suggests (STS), was tested in a low data environment and showed promising results in providing personalized recommendations.
User Personality and the New User Problem in a Context-Aware Point of Interest Recommender System
- 2. ENTER 2015 Research Track Slide Number 2 Agenda • Context Aware Recommender Systems (CARS) • New user problem in CARS – cold start • SoluIon: demographic and personality informaIon usage • Experimental comparison • Conclusions and future work
- 3. ENTER 2015 Research Track Slide Number 3 IntroducGon • Recommender Systems (RSs) are tools that support users decision making by suggesting products that can be interesting to them. • Examples of Recommender Systems:
- 4. ENTER 2015 Research Track Slide Number 4 State-of-the-art Technique • CollaboraGve Filtering: • Predicts unknown raIngs exploiIng raIngs given by users to items • The items with the highest predicted raIngs are recommended • If a new user has no raIng, it is impossible to compute any raIng predicIon for her (cold start problem). 3 4 2 5 3 ?
- 5. ENTER 2015 Research Track Slide Number 5 • Extends collaboraIve filtering by considering the contexts in which the items are experienced by the users • The Cold Start problem becomes even worse. Context-Aware RS (CARS) 3 4 2 5 3 ? 1 1 2 5 2 ? 3 ? 3 3 2 5 ? 3 ? 5 5 4 5 3 ? RS CARS
- 6. ENTER 2015 Research Track Slide Number 6 Related Works • Common soluIons of the cold-start problem in CARSs: – (Branhaufer et al. 2014) proposes a hybrid soluIon that exploits a selecIon of two CARS algorithms, each one suited for a parIcular cold- start situaIon, and switches between these algorithms depending on the detected situaIon (new user, new item or new context). – (Codina et al. 2013) proposes SemanIc Pre-Filtering (SPF) which exploits, in the recommendaIon process, not only contextual raIngs that exactly match the target context but also those there were acquired in semanIcally similar contexts. – (Zheng, et al. 2013) proposes weighIng contextual raIngs based on their similarity to a given target context to tackle cold start in CARS.
- 7. ENTER 2015 Research Track Slide Number 7 Our Approach • Cold start problem in CARS: – a new user, without raIng history, requests a recommendaIon to the system (new user problem) • Approach: – The system exploits the user's personality informaIon in the recommendaIon process to detect hidden factors modelling the preferences of the user • Outcome: – using personality informaIon, our CARS called STS could generate personalized recommendaIons for the new users, with higher accuracy compared to a baseline that uses demographic informaIon.
- 8. ENTER 2015 Research Track Slide Number 8 South Tyrol Suggests (STS) • A mobile Android context-aware RS that recommends places of interests (POIs) in the South Tyrol region, Italy. • The system was in an extreme cold-start situation (only 700 ratings for total of 27,000 POIs).
- 13. ENTER 2015 Research Track Slide Number 13 STS: RaGng ElicitaGon • By knowing the user's personality the system idenIfies and presents POIs that she has experienced, and can rate. • For each POI, STS presents 3 contextual factors, that user can specify freely, if she remembers.
- 14. ENTER 2015 Research Track Slide Number 14 STS: RecommendaGons • STS computes raGng predicGons for all POIs, using the users' personality and the raIngs they have given to POIs • The top POIs with the highest predicted raIngs are recommended to the user.
- 15. ENTER 2015 Research Track Slide Number 15 RecommendaGon Algorithm • Context-aware Matrix FactorizaGon (CAMF) (Baltrunas et al., 2011) extends matrix factorizaIon by incorporaIng baseline parameters: they model overall deviaIons of items‘ raIngs produced by the nature of the users, items, and contexts ˆruic1,...,ck = i + bu + bicj j=1 k ∑ + puqi T ī: average raIng of item i bu: baseline for user u bicj : baseline for contextual condiGon cj and item i pu: latent factor vector of user u qi: latent factor vector of item i
- 16. ENTER 2015 Research Track Slide Number 16 EvaluaGon: Goals • We are interested in: – comparing the accuracy of recommendaIons (based on CAMF model) for new users, by using either: • demographics informaIon • Or the Big-5 personality traits – IdenIfying the demographic a$ributes or Big-5 personality traits that generate more accurate recommendaIons for new users.
- 17. ENTER 2015 Research Track Slide Number 17 EvaluaGon: Dataset Total number of ratings 1,379 Number of users 239 Number of items 184 Number of contextual factors 14 Number of contextual conditions 56 Number of contextual situations 799 Number of demographic attributes 2 Number of personality attributes 5 New version of dataset is available for download (ResearchGate login required): h$ps://www.researchgate.net/publicaIon/305682479_Context-Aware_Dataset_STS_-_South_Tyrol_Suggests_Mobile_App_Data
- 19. ENTER 2015 Research Track Slide Number 19 EvaluaGon Methodology • Methodology: ten-fold cross-validaIon scheme (Braunhofer et al., 2014, Shani et al., 2008): – First, we randomly split the users in the data set into 10 subsets – Then, for 10 runs, we use the raIngs of one subset of users as tesIng set and the remaining ones as training set – In this way we create a test set of raIngs coming from users that have no raIngs in the training set, i.e., really cold (new) users – The predicIon model is trained on the training set and tested on the test set (raIng predicIons are compared to stored raIngs).
- 20. ENTER 2015 Research Track Slide Number 20 EvaluaGon Metrics • Mean Absolute Error – The lower the be$er – Measures the average absolute deviaIon of the predicted raIng from the user's true raIng: • Normalized Discounted CumulaGve Gain: – The higher the be$er – The recommendaIons for u are sorted according to the predicted raIng values, then DCGu is computed:
- 22. ENTER 2015 Research Track Slide Number 22 Conclusions • We have shown that: – Personality informaIon gives a be$er raIng predicIon model than demographic informaIon - which is a more common approach to tackle the cold-start problem – Using even a single personality trait (out of five) can sIll produce a significant improvement of the recommendaIon quality.
- 23. ENTER 2015 Research Track Slide Number 23 Future Work: Datasets • Such as MyPersonality: • Facebook app for taking psychometric tests • more than 6,000,000 personality profiles • more than 4,000,000 of Facebook profiles
- 24. ENTER 2015 Research Track Slide Number 24 Future Work: Social Mining • The personality of the users can be learnt from their interacIons in social networks (Bachrach et al., 2012) • Social profiles can even be used directly in collaboraIve filtering based RSs (Fernández-Tobías et al., 2014).
- 25. ENTER 2015 Research Track Slide Number 25 Open Issues: GamificaIon • Can we make the personality acquisiIon more fun? Shoot the ball to the personality trait that is strongest in you NeuroGcism ConscienGo us-ness Openness Extraversio n Agreeablen ess Big Five Personality Traits