Friend Recommendations in Social Networks using Genetic Algorithms and Network Topology Jeff Naruchitparames, Mehmet Gunes, Sushil J. Louis University of Nevada, Reno Evolutionary Computing Systems Lab (ECSL (excel)) http://ecsl.cse.unr.edu ([email protected])
15
Embed
Friend Recommendations in Social Networks using Genetic Algorithms and Network Topology
Friend Recommendations in Social Networks using Genetic Algorithms and Network Topology. Jeff Naruchitparames , Mehmet Gunes , Sushil J. Louis University of Nevada, Reno Evolutionary Computing Systems Lab (ECSL (excel)) http://ecsl.cse.unr.edu ([email protected]). Outline. - PowerPoint PPT Presentation
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Friend Recommendations in Social Networks using Genetic Algorithms
and Network Topology
Jeff Naruchitparames, Mehmet Gunes, Sushil J. Louis
University of Nevada, RenoEvolutionary Computing Systems Lab (ECSL (excel))
• Recommend friends on facebook• Customized to each user• Use– Friends of friends– Degree centrality– Pareto Optimal GA
• GA identifies useful “social” features– Feature selection
• How do we figure out if we are making progress?
Prior Work• Facebook seems to use a friend-of-friends
approach. • Analyze friend graphs to find cliques or
communities (Kuan)• Filter: GA used to optimize 3 parameters derived
from structure of social network. Then filter based on these parameters (Last CEC, Silva)
• …more• We also use a filtering approach based on features
identified by a pareto-GA
Jeff’s Friends
Approach – Successive filtering
• Consider friends of friends (fof)• Add users who have high degree centrality– Degree centrality = deg(v)/n-1– N is number of vertices
• Personalize recommendations based on N social features
• Which M features from these N?– N == 10 in this paper– GA chooses M
Ten Features (1/2)
1. Number of Shared Friends2. Number of friends in town3. Age Range4. General Interests
1. Number of shared likes, music
5. Common photos1. Number of shared photo tags
Ten Features (2/2)
• Number of shared events• Number of shared groups• Number of liked movies• Education– Same school with two year overlap
• Number of same: Religion and Politics
Caveats
• Preliminary work• 10 features 10 bits 1024 points in search
space. That’s easy for exhaustive search!• But we want to– Test approach on a small problem first– Then expand to N >> 10 features
Methodology• Representation
• Genetic Algorithm– Selects features to use for filtering– Pareto optimality principles to compare feature sets.
Pareto front tells you which feature sets work well• Best combination of features for each
central person through Pareto optimality
Feature
1 Present, 0 Absent
Pareto Genetic Algorithm
• Chromosome fitness is inverse pareto rank times number of friends
• Elitist GA, tournament selection• Single point crossover (0.92)• High mutation probability (0.89)• Populations size = 20• Number of generations = 30• Results averaged over 3 runs on 100 users
Performance comparison method
• 100 users• Remove 10 friends• See if system recommends those 10• Track number of friends correctly
recommended
Results
Conclusions and Future Work
• Pareto GA seems to help• Pareto friendships seem
promising as a representation
• Performance metric
• Lots of work left to do– Experiment with GA– Do we really need Pareto GA?– More features– Combinations with other
approaches
While you ask Questions?
http://ecsl.cse.unr.eduCI in RTS games: Research Assistantships