Thesys Group The Data Mining Company - Social network analysis

Typical sponsor

Head of CRM
Chief Marketing Officer
Chief Sales Officer
Chief Strategy Officer
Board Member

Typical participant

Marketing Manager
Business Analysts
CRM Director
Retention Manager
Data Warehouse specialist
IT Manager

To whom do we recommend?

Telecom/Online companies who want to achieve improvements on traditional data mining churn prediction/acquisition models
Telecom companies who have a large prepaid portfolio where other rich segmentation data is not available, but aggregated CDRs can be extracted
MVNOs who don’t have resources to implement large data warehouse projects and -similarly to prepaid dominated operators- have less data available on customers in structured format other than CDR
Telecom companies who try to develop and diffuse socially contagious Value Added Services, or boost mobile advertising efficiency and click-through
Banks to estimate risk of loss of SME universe when big corporations churn

Aim of the project

Mining the warehouse of user calling and other communication and product adoption behavior offers the ability to improve targeting, retention and acquisition activities. This is new to an approach based not only on traditional measures of monthly spend, length of loyalty contract, handset subsidies but also combining these measures with clustering calling behavior, centrality, telecom adjusted PageRank© type influence score, spreading of VAS usage and affinity. With these measures we will help you reduce churn, identify false rotational churn, decrease your retention costs, and achieve double digit percentage increase in social service diffusion.

Methodology

The following example is given for a product diffusion project. During data understanding process and pre-analysis, building the social graphs and situating product diffusion around nodes at points in time are done. We are using the data of past and current call data records and other communication patterns to build the social networks in points in time and identify the past and current graph measures of the persons around whom service adoption accelerated. To weigh the links between individuals, we use intensity of communications between members. We will cut the graph according to business objectives if necessary, to get evening/weekend graphs for example. To improve the network information, we will also combine the service adoption data with demographic and other transactional and product affinity data. Each decision a customer made regarding a certain product as well as demographics characteristics will be included in a vector summarizing what we know about the customer. Using these vectors, we will assign similarity scores to pairs of consumers, and obtain a different set of weights, which, combined with the previous weights, will give us a more complete picture of interactions between customers. Then we will apply our clique-percolation algorithm to this network to identify communities and opinion leaders. This method allows us to build up closely connected group of members, who are likely to influence each other’s product adoption decisions. The building blocks of these “communities” are the so-called cliques. In a clique every member is connected to another one by link, not only based on social distance, but also based on service adoption history. Communities are obtained by identifying sets of cliques, which have large overlaps. In this way, every two member of a community is connected through a series of cliques. Combining these measures, we will determine each member’s potential to diffuse services/products. By analyzing the past growth of the network, we can identify the percentage each member has brought in from his viral potential and focus on the ones with adaptive campaign, who still have more potential.

Results

The main findings of social network analysis projects are:

Products and services within your portfolio that are affected by viral mechanisms
Optimal selection of target group for multi-wave acquisition or retention campaigns
Lifetime values adjusted by customer’s role in their communities
Long term risk assessment and expected cost of sustainability of various segments of your customer portfolio
Evaluating value of customer touch-points for third part advertisers
Likelihoods of pairs of calling cards representing rotational churn

DOs and DON’Ts

Involve innovation leaders and get support of a board member from your company
Go back in time up to 2 years for an initial service diffusion history around viral points
Select complete graphs in the sample dataset (city/island)
Start with an acquisition project for fast proof of concept
Customize algorithms based on dynamic definition of communities

Don’t

Don't try to engineer use case for a product sell before early diagnostic analysis
Don't buy a packaged software developed on other graphs in other regions
Don't mix calling circle with advanced graph derived variables

Example on fingerprinting and rotational churn: This figure comes from an SNA dataset, represents the comparison of weighted calling patterns of 2 different IDs within Operator's network in 2 different points in time, indicating rotational churn as the first ID stopped communication, the second gradually increased. Calling graphs in points of time can be also analyzed to uncover internal and external tariff migration patterns and help designing plans for segment protection, price competition.

Why we are different

Experience with multi-million node calling graphs, can handle large, computation intensive datasets
Unlike prepackaged software, our community definition is clustering members into overlapping, context specific communities, a node can be influential in spreading ringback tones in 1 community, but not viral for in churn in another one
Our variables are derived by local characteristics of the graph structure and communication intensity as customers in different demographies, geographies, social groups have different communication patterns
Have credible track record, doubled simple data mining lift measures

Social Network Analysis