Raasta - Intelligent User Recommendation System

Presentation on algorithm design and proposed model for the development of a Multi-criteria Recommendation System

Transcript

1. Raasta Intelligent User Recommendation System Under the guidance of: Dr. MPS Bhatia(COE) Mr. Kunwar Singh(ECE) Developed by: Mrinal Ahlawat (100/EC/12) Sahil (140/EC/12) Astha Arya (242/CO/12)
2. Introduction • Recommendation systems predict the preference of a user. • Multi-criteria recommendation system can be used to search for required device in an IOT network. • Designed an algorithm and implemented a user to user recommendation system.
3. Traditional Recommendation Systems Item 1 Item 2 …. Item M User 1 User 2 …. User N MxN User vs Item Matrix Through the ratings of neighbouring users for an item, the recommender system can predict the rating or preference of this user for an item.
4. The value of ratings user u gives to item i is calculated as an aggregation of some similar user’ s rating of the item. Similarity calculation plays a critical role in finding a set of users on which aggregation functions can be used.
5. User Check-in Behaviour Number of Check-in vs Time graph [1]
6. We will be making a 2-D matrix of data for each individual user USER CHECK-IN BEHAVIOUR MATRIX This matrix adds a temporal dimension to reflect the time correlation of user behaviour. [1] Proposed Model
7. If there are M users, N check in spots, T check-in time intervals. Each user has a time spot check-in matrix. The study object is thus a three-dimensional matrix R. [1]
8. • Traditional Method of similarity calculation cannot be applied to individual user’s 2D matrix as : ◦ The matrix is sparse ◦ The computation would be time expensive • Hence, we need a dimension reduction technique.[2]
9. • The regression line shows the best approximation of the original data with a 1-Dimensional object. [4] • A high dimensional, highly variable set of data points can be reduced to a lower dimensional space, which is easier to analyse. Singular Value Decomposition
10. • It is a matrix factorization technique which takes a rectangular matrix A of size MxN and returns a 1D vector which can uniquely identify the matrix A. [3] • We apply SVD to spot-time matrix of each user (Ri). • The diagonal values of Σ generated uniquely defines Ri
11. Cosine Similarity Function Now, we can find the Location score by finding the similarity between the SVD vector for two users :
12. Friendship Score(FS) is based on the friendship network of the user. The score, with the initial value of 1000 is divided by a factor of 10 for each depth we traverse in the graph.
13. Likeness score(LS) is based on the number of similar interests of two users. For two users, if A and B represent the likeness vectors: The Online Score(OS) is given by:
14. Along with the user behaviour extracted from the spot-time matrix we will also take the traditional online score as a parameter. The final recommendations will be based on the following formula:-
15. Adaptiveness of the Algorithm The value of alpha is automatically adjusted according to user feedback of the recommendation: • Send request to the recommended user • Reject the recommendation
16. Other Applications • Recommendation of users for applications where location and user behaviour can be fetched. e.g. -Location-Based Mobile Network • Recommendation of users for applications creating Special Interest Groups. • Matchmaking algorithm on matrimonial websites.
18. Get My Location Fetch Twitter-Trends Like Trends as per interest Generate Recommendations Helps in determining the likeness score for each user Fetched using Twitter Rest API which returns Top K Trending Topics The current location of the user is plotted on Google Map. Using the model proposed earlier, top K user recommendations are generated
19. DEMONSTRATION
20. [1] QIAO Xiuquan, SU Jianchong, ZHANG Jinsong, XU Wangli, WU Budan, XUE Sida,CHEN Junliang, Recommending Friends Instantly in Location- based Mobile Social Networks, China Communication, February 2014 [2] Lu Yang, Anilkumar Kothalil Gopalakrishnan, A Collaborative Filtering Recommendation Based on User Profile and User Behaviour in Online Social Networks, Computer Science and Engineering Conference (ICSEC), 2014 International [3] Badrul M. Sarwar, George Karypis, Joseph A. Konstan, John T. Riedl, Application of Dimensionality Reduction in Recommender System--A Case Study, GroupLens Research Group / Army HPC Research Centre. [4] Osman Nurġ Osmanli, A Singular Value Decomposition Approach For Recommendation Systems, IEEE Internet Of Things, July 2010. References
21. Thank You!