At InMobi daily we receive data about 800 million unique users and in-house Hadoop and spark platforms help us get meaningful insights from this data to drive a better targeted advertising model. This particular talk presents the architecture, data insights and the visualization of a ‘Location Based Social Graph’ we have built, pivoting around bssid, lat-long, essid of ad requests. Spark GraphX has enabled us to generate interesting network statistics at scale like degree distribution and connected components per geography and filter out the noisy edges. We are in the process of developing a home/work/public classification model to better understand nature of social cliques and the trending apps in each clique. In future we hope to overlay the graph with user attributes like demographics, app-affinity and segment-membership to derive meaningful relationships between users like Friends, Families or Colleagues and enable better CVR prediction based on look-alike modeling.
Seinjuti Chatterjee has a Bachelors in CS from NIT Durgapur, and a Masters in CS from University Of California, Santa Cruz. She had the opportunity of working in domains like Advertising, E-commerce Search, Content Optimization in companies spanning Adchemy Inc, Yahoo Inc, TheFind Inc, InMobi Pvt Ltd. She has built platforms that can ingest data of the scale of petabytes and derive insights by applying machine learnt models on user search patterns, visitation patterns etc. Currently she is working in the Sciences team at InMobi where the charter is to improve Ad Experience by building a location based social graph.
Ian Anderson obtained a doctorate in Computer Science from the University of Bristol in 2009. During this time he founded Overlay Media to develop technology to sense aspects of context on mobile devices. After supplying and working with Tier 1 mobile device manufacturers Overlay Media was acquired by InMobi in 2012. Within InMobi he took up the role of Principal Research Scientist and his current research interests include inferences that can be made with partial, discontinuous sensor data collected via mobile devices. This includes inferring user behavior traits such as social circles, user behavior patterns and inferences about predicted user.