Predicting YouTube Popularity Using Latent Social Approach

A Latent Social Approach to YouTube Popularity Prediction Amandianeze Nwana Prof. Salman Avestimehr Prof. Tsuhan Chen

Statistics Up to 60% of all videos are watched through YouTube 1 9.85North American Traffic2 Other YouTube 90.15 YouTube Traffic2 27 Mobile 73 PCs and Laptops 1 YouTube Traffic Characterization: A View From the Edge; P Gill, Arlitt, Li, Mahanti 2 http://www.sandvine.com

Statistics: Campus View The most popular videos globally typically only account for about 1% of the videos viewed on campus daily 1 Correlation coefficient between global popularity and local popularity is too low 2 Number of Requests 1 >1 1 YouTube Traffic Characterization: A View From the Edge; P Gill, Arlitt, Li, Mahanti 2 http://www.sandvine.com

Typical Request Patterns Romnesia speech goes viral Sports Highlights Music Videos Source: UMass Amherst YouTube trace dataset (1 week)

Main Idea lol .did you see that video?

Main Idea Can a record of the transactions reveal information about the network structure and graph? Can the network structure and a record of the transactions predict future trends?

Traditional Caching Requests Requests Gateway Router YouTube Server Local Network Response Cache Response

Predictive (Social) Caching Requests Transactions Gateway Router Cache Transactions Latent Network

Goal

Challenge Transactions Gateway Router Latent Network

Estimating the Social Network Mathematical Epidemiology Given population ?, and infectious disease ?, ??(?)- Subset of population that is currently susceptible to ? at time ? ??(?)- Subset of population that is currently infected by ? at time ?

Mathematical Epidemiology Compartmental Models The interaction between these subsets an be modeled in the following way: ? ?:Once individuals get infected they remain infected and infectious forever ? ? ?:Individuals can get reinfected by the same virus

Diffusion Model 2 Stages Decide who gets infected by whom independently Stage 1 Decide the time of infection (observed symptoms) Stage 2

Latent Social Network Inference 2 8 1 4 3 6 10

Inference Steps Occurs in two stages: Stage 1: Given the transaction fit the inter-arrivals to a power law Stage 2: Given the estimated power law, and the transactions, find the Influence matrix that maximizes the likelihood of the observed transactions Maximum Likelihood Estimation

back to caching We now have the social graph over the network of users We need a video relevance function to assign relevance scores to videos Rank the videos according to relevance scores and store the top K videos in the cache

Video Relevance Calculate a temporal score(v) proportional to its frequency, and decaying in the time since last arrival. ????? (? ??) social score(v) is calculated according to the sum of the probabilities of new infections ? ? ? ???? (? ?? ?) Combine temporal score with social score.

Results Comparison % Improve Inter-Arrival/CRF 11.6 Combined/Inter-Arrrival 13.2 Table 1: Percentage Improvements of algorithms using all users Fig. 1: Average Hitrate for all approaches over different cache sizes Fig.2: Cache Size comparison between purely social and baseline using all users

Results Connected Users Comparison % Improve Temporal/CRF 15.6 Combined/CRF 21.1 Table 2: Percentage Improvements of algorithms without isolated nodes Fig. 3: Average Hitrate for all approaches over different cache sizes without isolated nodes Fig.4: Cache Size comparison between purely social and baseline without isolated nodes

Future Directions Explore other epidemic models On the fly update of nodes and edges Graph clustering into different communities and influence groups User recommendations using the social graph Object detection and tagging via twitter