The Fifth Elephant 2014

A conference on big data and analytics

'Know Your Customer!' - Advanced Data Science for Audience Segmentation

Submitted by prabhakar srinivasan (@prabhacar7) on Monday, 21 April 2014

videocam_off

Technical level

Advanced

Section

Full talk

Status

Confirmed & Scheduled

View proposal in schedule

Vote on this proposal

Login to vote

Total votes:  +17

Objective

Have you ever wondered how Cisco does Customer Segmentation? What is Cisco's technology stack to deal with Big Data? What tools and technologies are adopted to bring best-of-breed algorithms from data science to inform on the problem of identifying segments in the audience. How does supervised and semi-supervised machine-learning along with Bayesian predictive analytics combine to produce a very interesting cocktail of technology solutions?

What is the system design and lambda architecture that we use to successfully deploy the solution to scale?

What is the practical business use of audience segmentation? What applications can derive benefit from Audience Segmentation?

I will provide answers to the above questions and many more in this demo plus talk on Cisco's Audience Segmentation solution.

Description

Audience Segmentation is a very important practical necessity in pretty much every field. Whether it is internet subscribers or paytv subscribers there is an intense need from the advertisers and service providers to know who is living in a household and what their demography profiles are and what their interests are?

An ensemble of techniques which include advanced linear and non-linear dimensionality reduction, unsupervised learning algorithms, bayesian predictive analytics and expert systems come together to form a compelling pragmatic data science solution stack for the big data environment.

Requirements

Some background knowledge of Big Data tools and data science algorithms like Bayes theorem, machine-learning algorithms, knowledge-based systems. Some background knowledge about the business requirements of Audience Segmentation. Some exposure to Mahout would be helpful.

Speaker bio

I am Prabhakar Srinivasan. I work as a data scientist at Cisco. I have invented a unique technique to do customer segmentation which works for the PayTv domain of Cisco products running on Big data infrastructure. I wish to share this success story with the community with the hope that others who are trying to solve a similar problem can gain some practical insight on doing data science on big data stack that really gives business value.

I have successfully deployed my invention in Europe for some Telecom giants. I have presented this technique during talks in Monaco during Cisco's internal conferences. It is time to share this success story with the wider Big data community.

Comments

  • 1
    Govind Kanshi (@govindsk) 4 years ago

    Thanks Prabhakar - Do you think It will be possible to share the library/framework/toolset ? That will be super-helpful for attendees to go ahead and use the technique with proven toolset.

  • 1
    prabhakar srinivasan (@prabhacar7) Proposer 4 years ago

    Hello Govind,
    Thanks for your comment. The details you are looking for are given below:

    Library/tools: Mahout, Python-scikit-learn, Apache Commons Math, Numpy, D3
    (where the libraries were found lacking in functionality the code was developed in-house and this work is in process of being contributed back to open source)

    Framework: There is no generic framework for doing Audience Segmentation but a generic methodology is proposed which combines Bayesian inference, Unsupervised Machine-Learning and Expert Systems. Since this is a hard problem and needs adaptation to each domain, it is very hard to find a turn-key solution. Cisco's approach has many key take-aways in terms of the development process from prototyping to deployment to scale. This comprises of a machine-learning workflow pipeline which can be adapted to Audience/Customer segmentation with other domains as well.
    The concepts that cut across all domains in a generic manner are the concepts of feature extraction, dimensionality reduction, modeling the problem as a data-science problem, identifying the need for supervised and unsupervised learning, evaluating the models, feeding bigdata as input in lambda architecture and building engineering dashboards and higher-level features built on top of customer segmentatation. Using this generic approach, one can adapt to a particular domain data model like eCommerce, Media, Travel, etc.,

  • 1
    Sandeep Mukhopadhyay (@sandeepmukho) 4 years ago

    Hi Prabhakar
    Could you share slides beforehand. Also could you share links of videos, if you presented this before ?

Login with Twitter or Google to leave a comment