An Empirical Case of a Context-aware Mobile Recommender System in a Banking Environment

Presented at the 3rd FTRA International Conference on Mobile, Ubiquitous, and Intelligent Computing (MUSIC 2012), Vancouver, Canada

Full paper available at: http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=6305818

Transcript

1. An Empirical Case of a Context-aware Mobile Recommender System in a Banking Environment Daniel Gallego Gabriel Huecas June 26, 2012
2. Background • Traditional recommender systems are based on subjective data: – Personal scores, biased tastes, etc. – This causes a trust problem in the results • New E-Commerce applications: – Use recommendation features based on real purchases (e.g. Amazon) – Increase confidence in the results • From E-Commerce to U-Commerce: – Due to the growth of the mobile environment 2
3. Research Motivation • Banks have rich real information about: – Customer purchases and their profiles – Economic trends • Bank customers usually have smartphones: – Provide context-awareness information (e.g. location) • We have developed a system to generate personalized recommendations about banking places: – A place is any entity where bank clients have paid with their credit cards 3
4. A Model for generating Context-aware Recommendation using Banking Data 4
5. Phase 1: Social Context Generation 5 Target user Banking Client Profiles Transactions and Places Target User Profile User Profile Clustering Transactions Assignment User’s Cluster Discovery Social Clusters Clusters Trends Map Phase I: Social Context Generation User’s Cluster Trends Map
6. 6 Phase 2: Location Context Filtering Mobile Device Context Information User’s Location Acquisition Location based Filtering Phase II: Location Context Filtering Geo-Located User’s Cluster Trends Map User’s location User’s Cluster Trends Map
7. 7 Lunch time Restaurants category Phase 3: User Context Filtering Phase III: User Context Filtering User context Ranking Generation Geo-Located User’s Cluster Trends Map Display Recommendation Personalized Recommendation User Feedback
8. Architecture and Implementation 8
9. Social Clusters: Evaluation • Tested in a collaboration with an important Spanish bank • Banking data provided with information on: – 2.5 million credit card transactions – 222,000 places information – 34,000 anonymous customer’s profiles between 48 and 55 years old 10
10. Social Clusters: Results 11 -10000 0 10000 20000 30000 40000 50000 60000 70000 46 48 50 52 54 56 Average credit card expense in one year (€) Average age Circle diameter equivalent to Social Cluster size
11. User acceptance: Evaluation • System deployed in the bank Labs: – Secure environment • Mobile client: – Allow customers to test the application • Online survey: – 100 bank customers – 2 scenarios: restaurants and supermarkets – Several properties evaluated in a Likert scale – Comments provided by test users 12
12. User Acceptance: Results • Overall positive attitude towards the system – High confidence in the recommendations • Users remarked privacy issues related to a real commercial exploitation 13 0% 20% 40% 60% 80% 100% Convenient Desirable Effective Reliable Useful User friendly 1 (strongly disagree) 2 (disagree) 3 (neutral) 4 (agree) 5 (strongly agree)
13. Conclusions • The system generates: – Context-aware recommendations – Using banking data – In mobile systems • Wide domain range of recommendation categories • Mobile prototype tested successfully in the Bank Labs environment • High confidence in the personalized recommendations generated: – Because of the banking data used 14
14. Future Work • Proactivity: – The system pushes recommendations to the user – When current situation seems appropriate – Without user explicit request • Multiple personalities in the system – “What kind of customer do you want to be today?” – Several profiles with different social clusters associated 15
15. Questions? Daniel Gallego @thanos_malkav dgallego@dit.upm.es http://danielgallegovico.es Thank you!
16. APPENDIX 17
17. Architecture and Implementation 18
18. Banking Data Anonymization • Client profiles: – <profileID, gender, age, avgExpensePerYear> • Bank transactions: – <profileID, placeID, payAmount, time, date> • Places: – <placeID, category, name, address, lat, lng> 19
19. Recommender technologies • Developed in Java • MySQL database to store the social clusters generated • Apache Mahout as machine learning engine 20
20. Recommender algorithms • Canopy + K-means to reduce the clustering process complexity (provided by Mahout) 1. Canopy clustering – Manhattan distance 2. K-means clustering – Euclidean distance 21 1 , = − 1 = − =1 , = − = ( − )2 =1
21. Web Service • Wrapper for the recommender • REST API with JSON format – Using Restlet framework • Independence between client and server technologies • Allow communication between the recommender and third-party applications 22
22. Mobile Client • Developed using Android: – Open platform – Allow free distribution of applications outside Play Store – JSON libraries available 23