Last update on .

In this blog, we will see how we can perform in-depth customer analytics using publicly available inputs from the customers on company Facebook pages.  For this showcase, we will focus on the media sector and more precisely on the RTL group  (leading TV & Radio on the French speaking side of Belgium).  We analyzed the behavior of people acting on the Facebook pages of the RTL group and aggregated all available information to perform per-user analytics and predictions.  We are reusing the techniques detailed in previously published blog posts on Facebook Mining and Sentiment Analysis.  The techniques described in this post can be very useful for all major B2C companies involved in the media, telecoms, retails sectors.For this benchmark example, we will focus on two TV channels (RTL-TVI and Plug RTL) and two radio channels (Bel RTL and Radio Contact). All these channels are owned by the RTL group and target a specific audience.  For instance,  plug RTL channel is targeting your adults (age from 15 to 34) while Bel RTL is more focused toward senior people.

In a first step, we collect data for Facebook pages of these four channels.  To do so, we repeat the procedure described in the  Facebook Mining blog post for each of the channels.  We collected two months of data in the period ranging from the 22nd of September 2016 to the 22nd of November 2016.  We quickly see that there have been many posts from RTL over that period with about 100 likes per posts.  Plug-RTL being the exception with a limited number of posts and likes in that period.

[gallery ids="1120,1121,1122,1123" columns="2" size="large"]


Now that we have the data, we can identify who likes each individual Facebook posts and build a collection of Facebook users interacting with these four RTL pages.  With just two months of data, we collected about 125 thousand individual Facebook users.  Among which about 10% have been acting on at least two of the four pages. When you think that the large majority of Facebook users are using their name as Facebook pseudo, it means that we could actually collect about 125K real people names with just a few lines of code.  This is impressive.  If you are scared... stop reading here, because we are actually going to do much more than this...

At this point, we have enough information to start building a per-user dashboard (one per individual users that we have identified).  To do so, we need to keep track of what actions each user made on every single post.  This adds a little bit of complexity to our "user" database, but actually not that much...

Many user-centric metrics can then be built out of this user-action-post dataset.  A simple set of metrics that we can use is user activity over time and channel, overall user activity per channel, passive (likes) vs active (messages) user actions per channel.  Bellow, you can see the user-dashboard result for one random (anonymized) Facebook user ("Dany") out of our 125K known users.


We can also construct more complex metrics by, for instance, analyzing the type of posts the user likes or comments on.  We classified the posts in various topic categories and checked on which topics the user is performing actions.  We subdivided Facebook posts in 9 categories related to channel programs, channel presenters, miscellaneous, music related, politic related, television series, weather related, movie related and humor related. Then, we counted the number of messages of a given type a specific user liked for each channel, and use that to build the DNA profile of a specific customer.  Obviously, categories can be customized depending on your business needs and type of questions you are willing to answer.  On the figure bellow, extracted from Dany's dashboard, the numbers in parenthesis correspond to the number of posts that Dany liked among the total number of RTL posts in that category over the considered period.  The color bars represent the ratio of this two numbers (ranging from 0 to 1).


Now, it should be clear, that we can perform user segmentation or user 360° analysis by comparing Dany's DNA profile to the other users we identified.  The simplest method would be to perform a K-means clustering in view of grouping users similar interests.

We can also analyze the messages posted on Facebook by the customers.  The sentimental analysis algorithm that we developed in a previous post is particularly interesting as it allows us to spot negative messages from unsatisfied customers.  Identifying such messages allow us to engage communication with these unsatisfied customers before they change channel/brand (churn prevention) or before they spread their unsatisfactoriness with other customers.  This could potentially have a significant impact on the image of your company and on your communication strategy.  The faster we react on the Facebook page, the better we preserve the image of the company.

The figure bellow shows the 5 latest reactions from Dany to posts on RTL Facebook channel pages.  We can see that overall, Dany clearly likes the programs on RTL channels.  He is a fully satisfied customer.


In comparison, we can take a look at the similar metric for another user.   Anne commented 71 times over a period of just two months.  We see that she is a much less satisfied user and many of her messages are actually questions that would require an answer from RTL side.  Most of these questions are very easy to answer and would help to turn Anne into a satisfied user.  Note that in the dashboard, we require a confidence level on the sentence polarity of more than 75%.  Messages bellow this threshold are marked as "unclear" polarity.


They are several other things that we can do with such cross-channels Facebook data analysis.  We could, for instance, prevent further churn by comparing the activity of the customers on competitor (RTBF, vivacité, etc.) Facebook pages.  We could enhance user segmentation by looking at pages of other companies:  We can identify the cooking lovers by checking their activity on leader cooking pages, etc.  We could perform the study over a much longer time period in order to identify trends and provide feedback/suggestions for a better communication with the customers.  We could cross Facebook user database with other databases in order to find the mail or physical address of the customers and possibly engage communication, sending discounts, etc.  Possibilities are basically infinite...

Have you already faced similar type of issues ?  Feel free to contact us, we'd love talking to you…

If you enjoyed reading this post, please like it. It doesn't cost you anything, but matters for me!


Pingbacks are closed.



Comments are closed.