Home / Blogs

Rumblings for an In-Session Recommendation Engine at Email Service Providers, Part I

Email Campaign builders (marketers) are flying blind. I know ESPs are genuinely timely about rolling out new products for their marketers, but there is a colossal gap in adopting data science and MLops into the email campaign building workflow.  Even MailChimp does not seem to have the answer just yet, and half-baked attempts over the years to optimize the subject line haven't been inspiring.

This article will attempt to articulate why a machine learning in-session-based recommendation engine is urgently required in the workflow at ESPs and professional marketers who send many emails.  It will also hope to accomplish how the recommendation engine will optimize specific email metrics like open-rate, conversion rate, or click-through rate, and why industry-specific inputs and campaign type inputs are necessary to optimize every campaign module, perhaps within the email campaign editor.

To fully comprehend the campaign builders' arduous plight, these marketers have zero visibility into what their next campaign will deliver without real-time predictive analytics built into the workflow of every module of the buildout process. Currently, there is no way to truly optimize campaigns reliably without an in-session real-time recommendation engine on a per-module basis. Furthermore, this requires enormous compute power, given that these models are deep learning CNN models.

Firstly, when we begin building an email campaign from scratch, we must address the dreaded selection of the template.

As we enthusiastically attempt to select a template, we fully capitulate, not having the slightest idea if a one or two-column template will convert best.

Further, we have no idea if this template will optimize the campaign we are attempting to build. Subsequently, not knowing the optimal text length in an email (not subject line) is just the very beginning of a slippery downward trajectory of a campaign gone awry. Further, selecting a random image and not understanding the full extent of the damage that image can have on our click-through before hitting send is just plainly shooting in the dark. There is a better way, science. A Real-Time In-Session Recommendation Engine.  

What email campaign building currently requires is a recommendation engine at "each step" of the campaign building process, with real-time predictive analytics built-in at every module within the workflow to give the marketer confidence that the email will convert (if he/she has optimized for that variable).

To conduct this scientifically, the ESP or marketer needs to first build in fields for predictors or classifiers at the inception of the campaign building process, way before the templates are selected. These classifiers are important inputs that allow the recommendation engine to optimize certain chosen variables that the marketer wants to optimize for the campaign. Our model and examples to follow in parts II and III of this blog series will use 3-5 variables to optimize before each email campaign buildout starts.  

With 20M clients, why is MailChimp lagging? When you go to MailChimp, and start the campaign building process, you have to choose a random template. The marketer has no idea which template will convert best. Let us say one of your predictors was for "industry type," and the marketer happens to be in the automotive industry. We could ask the recommendation engine to select the templates that convert best for the automotive industry? Moreover, in this example, let us ask the marketer to optimize for CTR rate in this campaign? OK Great. That is two classifiers.

Then, let us now ask for a 3rd classifier, the marketer wants to optimize for a type of campaign. Let make this a promotional campaign. The ML recommendation engine is now armed with three classifiers to build out a series of predictions. That is it. Now the real efficiency takes place.

We are using historical data to predict the campaign's outcome, but we are also introducing new data sets to make the prediction more accurate. Now, we have:

Predictor 1: Industry: Automotive
Predictor 2: Campaign Type: Promotional
Predictor 3: Target Variable: Click-Through Rate

Given that our target variable is click-through rate, we might assume that a 2-column template will convert better than a 3-column template? But Who knows. That is the calamity marketers are facing now. Nevertheless, with the recommendation engine, a 1:2 Column width converts better. With this information, the marketer has made a scientific and informed decision armed with any real-time analytics to boot. Furthermore, it creates that dopamine effect every time a marketer looks at these analytics. We need to jump into the historical data and introduce new datasets so the predictive analytics algorithm can help us make ever more informed decisions.

Real-Time In Session Recommendation Engine for the Template Module Only, if optimized by specific industry and CTR

From a machine learning perspective, a deep learning model, such as this, takes enormous resources of computing power because accuracy is critical, and explainability lags slightly. Think about the computing resources needed at MailChimp to run these models in real-time for each of their 20M clients.

Ensembles. Also, we must introduce ensembles to harden our predictions. A good ensemble has better performance than any other independent contributing model. Ensemble diversity is a property of a good ensemble where contributing models make different errors with the same input.

These techniques underscore the importance of different models running for the same industry inputs. Seeking independent models and uncorrelated predictions provides a guide for thinking about and introducing diversity into ensemble models. Without a diversity of ensembles, one independent model can run into anomalies.

So to minimize risk in the predictions, the template considers a deep learning model with ensemble techniques and enormous computing power. However, the marketer is now much better informed about which template to choose for the highest CTR in the Automotive Industry for a Promotional type campaign. In part II of this blog series, we will discuss predicting for email text length, and in part III, we will optimize real-time predictions for images; if you want a demo, feel free to reach out.

By Fred Tabsharani, Loxz Digital Group-ML Modeling – Fred Tabsharani has spent the last two decades in IT as a recognized industry leader in Digital Marketing. He holds an MBA from John F. Kennedy University and has added four AI certifications, two from the UC Berkeley, School of Information, Google, and IBM. Fred is a 10 year veteran of M3AAWG. Visit Page

CircleID Newsletter The Weekly Wrap

More and more professionals are choosing to publish critical posts on CircleID from all corners of the Internet industry. If you find it hard to keep up daily, consider subscribing to our weekly digest. We will provide you a convenient summary report once a week sent directly to your inbox. It's a quick and easy read.

I make a point of reading CircleID. There is no getting around the utility of knowing what thoughtful people are thinking and saying about our industry.

VINTON CERF
Co-designer of the TCP/IP Protocols & the Architecture of the Internet

Comments

 Be the first to post a comment!

Add Your Comments

 To post your comments, please login or create an account.

Related

Topics

Brand Protection

Sponsored byAppdetex

Domain Names

Sponsored byVerisign

Cybersecurity

Sponsored byVerisign

IPv4 Markets

Sponsored byIPXO

Threat Intelligence

Sponsored byWhoisXML API

Domain Management

Sponsored byMarkMonitor