Tech Blog: Tuning with Thresholds & Business Scenarios

Today I would like to talk to you about tuning Chatterbox Labs classifiers to suit your business use case. To recap, Chatterbox Labs build machine learning and natural language processing products that monetize real-time social data by pinpointing consumers on their social path to purchase; we deploy our technology within a partner software platform or stack. To read more about Path 2 Purchase, head here: http://chatterbox.co/products.

Path 2 Purchase classifies data using a cascading decision tree. At the top of the tree are important utility classifications that cut through the noise (chatter that is not monetizable), and remove advertising, news and corporate messaging, whilst at the bottom of the tree are the classifiers that deliver the Path 2 Purchase segments: Interest, Consideration, Lead Validation and Post Purchase.

Importantly, each time the algorithms perform a classification (at each stage of the decision tree) we produce a score associated with that classification – it is called a confidence score.

Using the tree and the confidence scores, Chatterbox have implemented a system within Path 2 Purchase which enables you to control how you would like to tune this traversal across the tree. This system is based on the premise that different industry use cases will have different quality vs quantity requirements. Let's take two example opposing business scenarios:

  • Purchase Intent for Automotive: Let's say our partner is focusing purely on the high value Lead Validation stage for an automotive client, funneling imminent purchase signals into a CRM or engagement system to get buyers into the dealership and close the deal. This client wants to ensure maximum quality and, given the high sale value of each item, are not interested in the higher quantity available in the earlier Interest and Consideration stages. The partner would use a Conservative setting for the thresholds meaning more messages are held higher up the decision tree, so that only very high confidence messages are funneled into the Lead Validation stage. This means the client receives very high quality and can focus the time needed on closing the deals.

  • Social Advertising for Pay Per View TV: Let's take the opposite end of the spectrum here. In this instance the partner is focusing on mass, targeted advertising so that they can deliver the right message to the right audience in real-time. The premise here is that those buyers early on their consumer journey in the Interest stage need a different creative message to those in the Consideration stage and again to those in the final Lead Validation stage. This automatic segmentation is based on real public social data, so users will be matched 100% for advertising. For this client quantity and scale are key. The partner would use a Liberal setting for the thresholds, effectively opening the gates and facilitating maximum coverage within the Interest, Consideration and Lead Validation stages. This means the client is able to deliver their targeted to creatives to the relevant segment of potential buyers, at huge scale in real-time.

Now for the tech part. Chatterbox's real-time classifications are applied to the incoming real-time feed from the social network (in fact, to achieve significant scale, 148k classifications can be performed every second using 9 Java threads benchmarked on a simple Mac Mini!). This is great and hits massive scale – however you may not know at the time the data arrives which use case your client is going to be using. That's fine and Chatterbox is designed to allow for this - on the incoming feed Chatterbox algorithms traverse the whole decision tree, classifying everything. You can write this simple dictionary of classification scores to your own scalable datastore within your firewall or infrastructure.

When you pull your data out of the datastore to use in your platform or client facing environment, you apply the thresholds – configured together as a Business Scenario. Want to interpret the data using a conservative setting? No problem. Change your mind and decide you want to use more liberal settings? Simple, just apply a different Business Scenario to the classification stores in your datastore. Of course, all of this functionality along with defaults, is provided by Chatterbox; you do not need to implement this!

So there we have it - with Chatterbox's real-time Path 2 Purchase technology you can not only monetize your data in a live stream, but you can very simply tune it using thresholds and scenarios to different industry, brand and product use cases.

If you would like to know more about Chatterbox Labs' technology please contact Andrew Watson – VP Strategic Alliances (andrew@chatterbox.co).