Posts Tagged ‘Marketing’

Integration with Facebook custom audiences

Thursday, August 9th, 2018 by Nick

We have enabled integration with Facebook custom audiences. This feature allows you to synchronize Sonamine predictions with Facebook custom audiences.

How might this be useful? For example, you can now retarget Facebook users who are likely to churn out of your game. With lookalike campaigns, you can acquire FB users who are most likely to convert to paying users or have the highest predicted LTV. The list goes on.

Showing promotions to mobile game players

Monday, February 6th, 2012 by Nick

When a player is in your game and ready to purchase something with real money for the first time, how do you show a promotion to her?  After having spoken to dozens and dozens of developers, we are convinced that this capability must be baked into the game design, and also budgeted up front.  In this post, we cover some of the considerations.


Why do you need a promotion system?  Economics.  In the free to play world, you need to use promotional messages to improve the conversion rate.  Email marketing might work, but the response rates are abysmal and most games do not collect emails.  This is not a cure all for badly designed games, but a complementary facet of the game business.


The person using this could be a marketing promotions manager, a revenue optimization or monetization analyst.  In all cases, they want to improve monetization by pushing specific offers and content to specific users.  The key is to give these users a simple UI to manage the promotions and which users to target them to (see targeting below).


Promotions is a well understood tool in many direct-to-consumer industries like mobile phone service companies or financial services.  Even retail businesses are getting into promotions and offers.  A promotion calendar is usually drawn up at the beginning of each year / quarter.  There could be festival oriented promotions (eg. Valentines day) or just regular weekly conversion promotions.  The key is to avoid spamming your user base, and you do that by cleverly managing who you show promotions to (see Sonamine offerings) and when.


Mobile games, especially native OS games, face a particularly tough situation with promotions.  Not only must the game retrieve live dynamic promotions, but the click-through action must be to a section of the game that is appropriate for the promotion.  For example, if the promotion is to get 50% discount off a virtual item, the click through must be through to the buy page for that item, with discount applied.  Since every game is unique with different set up, it is not possible to use a generic mobile advertising system to insert “interstitial” ads into the game, especially since most interstitial ads are either video oriented or open up a browser, both terrible user experiences.

Target players for promotion

Another reason why you cannot easily use a mobile ad provider to insert “interstitial” or “banner” ads into the game and apply “house ads” is that these systems do not allow you to target individual users differently for different promotions.  In practice what you want is a way to link different player lists to specific promotions.  Again another reason why a better option may be to build this inhouse.  One simple option is to create a user_promotions data structure (hash_map, key value store or what ever suits you) with user_id and promo_id.  Then your game can access this and retrieve the promo as needed.

Click through action

Finally, in the mobile game architecture, it is best to ensure that promos can easily have a deep-link into separate parts of the games, passing along appropriate parameters associated with the promos.   For example, the product purchase page might take parameters to filter out the virtual item list.

Some example promos placements

Common places to show promos : splash screen, level up screens, the buy-pages of course.

Comments from the field…

Investigative SNA or large scale SNA? Part 1

Wednesday, January 19th, 2011 by Nick

In the recent Predictive Analytics World show, Karl Rexer was presenting the highlights from the annual data miner survey.  735 data miners from 60 countries (mostly North America and Europe) responded.  You can contact Karl directly for inquiries on this survey (

In terms of algorithms used, it was interesting to see that

  • 12% of the data miners mentioned using Social Network Analysis
  • 9% of the data miners mentioned using Link Analysis.

This means there are at least 88 data miners out there using social network analysis!  This is quite a large number and we are glad to see this.

However this finding also raises some questions:

  • how are they using SNA?
  • what tools are they using?
  • are the models using SNA in production?
  • are data miners exploring individual communities or looking at large scale modeling?

Unfortunately the data does not provide any clues to the answers.  But it does bring up some key differences between various SNA tools.  We’ll touch briefly on one key difference and expand on it in the next post.

There are two distinct flavors of SNA.  One is investigative in nature and involves an analyst discovering and analyzing individual social networks.  In this flavor of SNA, if the analyst finds fraud or criminal activity, it is more than sufficient to make up the cost of the software and the analyst time.   Financial services, law enforcement companies use this type of SNA.  Vendors include i2 (see and SAS (see

The second is a large scale backend version of SNA.  In this case, you cannot have analysts looking at individual users, for example telephone customers.  Rather you are trying to add SNA variables into predictive analytics and data mining.  In many cases, this large scale SNA can be used as a filter to help narrow down the cases for investigative SNA.  Sonamine is the only vendor in this space, providing large scale SNA for up to billions of nodes in the social network.

It is likely that the data miners in the survey were probably using the large scale SNA to add SNA insights into their predictive models.  Next post, we will cover the differences between these two types and how customers should evaluate the key requirements.

Viral churn analysis presentation at IIR conference

Friday, October 29th, 2010 by Nick

Pursway was at IIR Customer Segmentation and Intelligence show this week in London.  As usual Ron gave a fantastic presentation regarding influencer marketing.

One interesting twist in their marketing spin is that they are now careful to position themselves as not replacing existing systems, but existing alongside current marketing practices.  They must have received some sales objections to this effect.

One thing is clear, Pursway has been performing many POCs with different operators.  By receiving and analyzing the CDR data that these operators send them, they shared their findings at this conference.  I wonder what are the operators’ perspective about this…See point 20 below.

Here were their main points:

1) customers have greater control and this trend will only increase.
2) who are your most valuable customers?  Are they the ones who spend the most or the ones that influence others to spend on your brand?
3) current marketing looks only at individuals
4) customers like to copy one another
5) word of mouth is a well known and proven phenonmenon
6) how do you identify these influencers?  By influencers, we do not mean celebrities or the most connected subscribers, but the everyday customers and community based marketing
7) each influencer can cause about 40 other people to purchase a product, not 100,000 people.
8) about 8-15% of population are influencers
9) high spending customers are snobs.  They spend a lot of money and actually do not associate with people enough to be influencers
10) most influencers are currently under the radar of most marketing departments
11) first step is the understand who is a friend of who.
12) it is important to distinguish between a link and a relationship.  They are often not the same.
13) To identify relationships, you need to look at the pattern of calls, including frequency, length, location.
14) To identify influencers, you need to look for what is being influenced and when.
15) Pursway has identified between 5 to 7 types of influencers spanning churn, data services and other types.
16) to act on influencers, you should provide them experience based marketing, such as free data plan for 30 days.
17) influencers will not work for marketers, because their credibility and ability to influence might hinge on them not working for you.
18) influence is measured by how many of his/her friends had the same behavior (purchase, churn etc) after the influencer (purchase, churn etc)
19) you can create a metric called viral effect for each subscriber that measures the number of friends that exhibited the same behavior after that particular subscriber
20) after receiving and analyzing the call detail records (CDR) data from multiple operators in the european mobile industry, Pursway calculated the viral churn effect of 2 different groups – influencers and non influencers.  There were significant differences (I created graph for easy consumption of the numbers, MNO stands for mobile network operator) :

viral churn21) when an influencer churns, the damage to their community is already done, so you need to treat the influencers different to prevent them from churning.

One case study of targeting influentials was provided by them.  This was 12 month long program of “be the first to know” for the 5M postpaid base of a European operator.  The influencers were given sneak previews of new products with the goal of increasing number of distinct products being used per subscriber.  The goal was to study the effect on the friends of these influencers versus a control group.

The influencer group had an average of 1.6 products while their friends (presumably non influencers) averaged 0.9 products.  This 0.9 was also the average of the control group.  After 12 months, of the program, the influencers increased their product average to 2.17 (35% increase) and their friends increased their product average to 1.37 (52% increase).  The control group increased their product average to 1.09 (21% increase).

During Q&A, someone asked what type of data was used to identify the influencers.  Ron answered that they looked at the the time sequence of events (purchase churn etc) and looked to see how many of the friends of each subscriber “followed” after the initial event.

Shameless plug here: Sonamine software’s cascade scorer function produces three different types of such viral scores : initiators, followers, sinks.  These are essentially the combinations of succeptibility to being influenced (yes/no) and ability to influence (yes/no).

Also see last two posts on how to create a predictive model for who the influencers are!

Influencer marketing for data miners (2)

Tuesday, September 21st, 2010 by Nick

Recall in the previous post that we have found a way to quantify the value of each person based on his/her effect on their friends.   The simple approach we discussed previously is that for each person that bought a product, you counted the number of her calling circle that purchased the same product within a certain timeframe.

Yes, there are objections that we covered such a timeframe and really knowing whether influence took place.  Another common objection is how to take into account double counting.  If both person A and B bought the product and one of their COMMON friends C also bought the product later, do you count C in both A and B, or do you divide C’s purchase between A and B?

Regardless of which method you choose, you will end up with a “network value” for each purchaser.  So the next step is to use traditional data mining techniques to model this network value, and to predict this network value.

Unlike churn or propensity modeling, you are not trying to score each user on their probability of churning or responding to a campaign, rather you are trying to predict influence size of each subscriber.  In this case, you would probably use a f unction estimation technique.  The dependent variable will be the influence size metric – the number of purchases within each subscribers network that occurred after the subscriber’s purchase date.  The same techniques apply in terms of estimating goodness of fit, by looking at mean-square-errors.  You could also use train/test concept if you have sufficient purchase data.

If you are going to send a marketing promotion to the influencers, bear in mind that you need to consider each person’s probability of responding to a campaign.  If an influencer does not respond to a campaign, then it is useless to promote to them.  So in order to combine the probability of response to the influence score, you can perform a simple calculation

impact of influencer in campaign = (probability of responding)  x  (influence size)

For example, if an influencer Jane has a very low probability of responding, but has a high influence size, you would still want to promote the product to her!  Alternatively, if influencer Peter has low influence but has a high probability of responding, you will still want to promote the product to him. Obvious you can segment the campaign list further such as high propensity-high influence, high-propensity-low influence etc…

How to measure results of campaign accurately

In the campaign above, the goal is to (1) get the target campaign list to buy a product (2) get the purchaser to influence their friends to buy the same product within a specific period of time, say 1 month.  Hence the measurement of campaign results must start with the initial purchases of the target campaign list, and then include their friends’ purchase within one month.

One key question is how would you construct a control group?  This part is most often overlooked in influencer marketing campaigns!

In order to perform an apples-to-apples comparison, the campaign control group must ensure that there is no overlap between the entire 2nd level network and the entire 2nd level network of the treatment group.  The reason is to make sure that the friends of the control group are not affected by the promotion through other friends!

To be more specific imagine the following :

The treatment group 2nd level network = campaign target list + all the friends of the campaign target list.

The control group 2nd level network = control target list + all the friends of the control target list.

The key is to ensure that there is no overlap between the “all the friends of the campaign target list” and the “all the friends of the control target list” In fact there should be no overlap at all between the treatment and control group 2nd level networks.

How does Sonamine help you?

Sonamine provides software to help you implement these influence data mining models and marketing campaigns.

To create the influence metric of each subscriber – use Sonamine software cascade scorer function.

To create the different social network variables to add to your influence and propensity model – use Sonamine software to generate these variables.