BlueConic recently launched AI Workbench for data scientists and marketers to bridge the gap between teams and better enable AI in marketing. But it’s not just about technology. Teams have to speak the same language to avoid miscommunications that will impact the time to value you see with machine learning initiatives.
We turned to one of BlueConic’s data scientists and solutions architect, Bart Leusink, to understand the terminology and create an AI glossary for marketers.
Bart, can you help us define what the following terms mean and how AI Workbench enables these things in BlueConic?
Sure thing! Let’s dive in.
First things first, what is a “model?”
A “model” is a data structure that represents a real-world process (e.g. the relation between visits, pageviews, and propensity to buy). You’re basically codifying your hypothesis of how (a small part of) the world works. Let’s say, for example, that your hypothesis is ‘the number of pageviews and visits of a customer determines how likely that a customer is to buy something.’ In that case, the relationship between pageviews, visits, and whether the customer bought something is the model.
What does it mean to build the model?
“Building the model” typically refers to writing code that represents these “real-world” relationships in such a way that they can be learned based on the data. Within AI workbench, you could build a custom model or import one from a Python library. We embedded Jupyter notebooks which allows data scientists to bring in their own models, or tweak any existing models within the BlueConic library. Those who aren’t familiar with Python can also use our built-in models for things like propensity scoring and customer lifetime value.
Processing the data, what does that mean?
Before the data can be used for training a model, it is processed. For example, numbers are normalized to increase the performance of the machine learning algorithm. One cool thing about doing this with BlueConic’s AI Workbench is that data scientists don’t have to spend nearly as much time processing the data as they would normally have because a lot of the data is already transformed. BlueConic gives you a single view of the customer so you don’t need to stitch data from different files together to try and create that view.
Since you can also use the full power of the Python programming ecosystem with AI Workbench, complex transformations are easy to apply to BlueConic profile data. There are instances where you will still have to spend time transforming data – like converting a date to the number of days since that date since that’s a little more complex.
So then, the next step is to train the model?
That’s right. Typically, you train a model by giving it a set of data and allowing the model to learn what types of outcomes you’re looking for. In the case of look-alike modeling, for example, you might train a model to look for attributes in a set of profiles that matches ones you’ve specified to the model as “high-value customers.” Now, traditionally, you’d have to pre-determine which attributes you want to put into the model to train it because you’re pulling lists of data from disparate marketing systems. With AI Workbench, you have more flexibility to experiment with different hypotheses because the data you work with isn’t restricted to the data in the file you exported.
That’s pretty neat! And then the fun begins, right? You get to run the model across the database?
Yes! Typically, because data is dispersed across different marketing technology systems, data scientists have a hard time getting a model into production across various external systems. To do so often requires you to build a connection to these systems- like, for example, you’d have to develop on-site personalization capabilities to adjust something on your site for a customer who has a high customer lifetime value. With our certified connections already plugged into BlueConic, you can easily get the model into production across these different activation channels by running it in AI Workbench.
What’s the difference between running and deploying a model?
Good question. Running the model means “pressing start” on the model, so to speak. Deploying the model means you’re sending the outcome of the model to your external marketing systems. Again, because models are being run in BlueConic, you can deploy the model across all your marketing system because BlueConic is already connected to all your activation channels. BlueConic AI Workbench will enable you to attach scores as profile properties to activate segments in real-time. It also enables you to do look-alike modeling based on changing behaviors, for example.
When will the models run or be deployed?
You’ll usually hear of “scheduling” a model, meaning, you can determine the time period in which the model is run. Currently, it’s difficult for data scientists to ensure predictions are kept up-to-date because the data the model is built on is static. By scheduling the notebook BlueConic will automatically ensure that the marketer can use the latest predictions in their segments.
Ok, then you want to update or optimize the model to make sure it’s the best it can be?
You got it. The beauty of machine learning is that it’s constantly learning and looking for new patterns. Iterating on models usually takes a while because data is coming from different sources – gathering, cleaning data takes up most of data scientists’ time. Using Jupyter within BlueConic gives you access to unified customer data – which updates as the customer’s attributes change. Your models will have access to the most up-to-date data because it’s pulling right from our persistent, dynamic profiles; and you can create smarter segments based on machine learning outputs that get attached to those profiles (like prediction scores).
This was great, Bart! Thanks for taking the time to explain this stuff.