Frequently Asked Questions: Evergage Decisions™ Contextual Bandit

Evergage Blog

Ideas and Strategies for Real-Time Personalization
Frequently Asked Questions: Evergage Decisions™ Contextual Bandit

November 27, 2018 by

We recently announced a new Evergage module called Evergage Decisions, which includes a ground-breaking algorithm called “Contextual Bandit” that allows companies to deliver the optimal promotion, offer, hero image – or even complete experience — to each individual site visitor.

To explain the algorithm in more detail, Cliff Lyon (VP of Engineering at Evergage) and Jordan Bentley (Senior Data Scientist at Evergage) presented a webinar earlier this month in which they discussed and demonstrated the power, benefits and uses of the new algorithm.

I highly recommend you watch the full webinar replay for an explanation of the methodology and primary use cases of the algorithm, but I’ve pulled some of the highlights into an FAQ in this blog post.

What is the Contextual Bandit?

The Contextual Bandit is an algorithm that uses predictive machine learning to evaluate the probability of engagement with a promotion, image, offer or experience and compare it with its potential business value to select the optimal content for each person. Essentially, you give Contextual Bandit several different experiences to display in a specific area of your website, and it selects the most ideal one for each individual visitor.

The important thing to note about the Contextual Bandit algorithm is that it is always learning and optimizing. The more it learns about how visitors interact with each promotion, the better it will be at picking the right one for each individual. This kind of personalization can offer greater lift than A/B testing alone, where picking the right experience for each user beats choosing a universal winner.

What do I use it for?

While many of Evergage’s existing algorithms are ideal for product and content recommendations, the Contextual Bandit is ideal for selecting specific experiences or promotions that you would ordinarily need to target manually to different segments via rules. Using machine learning offers greater accuracy in targeting users and requires far less effort.

Take, for instance, your homepage hero experience — the main image, copy and CTA you include above the fold on your homepage. No matter your industry, you probably have a number of different ideas for what to feature there. B2B marketers may want to promote their latest white papers, announce upcoming webinars, or ask visitors to take a free trial. E-commerce sites may want to promote upcoming sales, announce new products in specific categories, or ask visitors to download their mobile apps. Ordinarily, you either need to pick one single promotion, or set manual rules to show different promotions to different segments of visitors.

But what you really want to do is show each person the promotion that is right for him or her. That’s where the Contextual Bandit comes in. You set up your multiple different promotions and then you let the algorithm do the heavy lifting.

Can I prioritize promotions that have a higher value?

Contextual Bandit will pick the experience that is best for each person, but marketers may become wary about losing some control over that selection. In particular, marketers may want to prioritize promotions that have a higher potential value to the business.

Let’s say, as an e-commerce marketer, you have a few different promotions running (for simplicity, let’s say there are just two promotions). One announces the launch of a new line from a well-known designer brand, and the other is a promotion for a sale on certain items. You want to ensure that the designer brand promotion gets enough attention, because the expected average order value from that promotion is much higher than that from the sale items.

Fortunately, Contextual Bandit can account for these value differences. To use Contextual Bandit, you associate a dollar value reward with each promotion. The dollar value is based on what something is worth to your business, and allows you to accurately compare the value of a conversion for one promotion over another. That way, the algorithm can calculate not just a person’s likelihood to click on each promotion, but also what the value for the business is if he does click. So even if the algorithm decides that there is a slightly higher probability that a person will click the sale promotion, it may still decide to show the designer brand promotion if the potential value is higher. Like the rest of the features in Evergage, the Contextual Bandit calculates all of this in real time using the most up-to-date visitor information available.

Can I avoid showing the same person the same promotion several times?

Marketers may also want to account for promotion fatigue if their visitors tend to return to their site many times. The Contextual Bandit takes into account how many times a person has seen a single promotion without acting on it, and learns the right time to try showing them something new.

Can I modify the promotions after I’ve launched it?

Most websites do not keep the same promotions running forever. Marketers frequently add and remove promotions. With the Contextual Bandit, you can add or remove promotions from consideration after you’ve launched the algorithm without affecting the overall effectiveness. This is possible because the models for each promotion operate independently, so you can swap out promotions without losing the learning from the rest of the promotions. The rewards for each promotion can also be changed at any time without any negative effects.

Can I still use Contextual Bandit on low traffic volume pages?

Contextual Bandit can still be used for low traffic sites or lower traffic pages. One of the benefits about our Contextual Bandit algorithm is that it starts personalizing immediately, it doesn’t wait until it has collected a certain amount of data. So there is no waiting period where personalization isn’t taking place.

The Contextual Bandit uses a two-phase approach where a simpler model is used at first to get higher performance out of small data sets. Once there is enough data for the more advanced machine learning model to operate effectively, it takes over. At any given time, the Contextual Bandit is using the most effective approach for the size of your data set.

Is Contextual Bandit effective for new visitors or visitors with limited history?

Even sites that have a heavy ratio of new visitors to returning visitors can receive significant benefit from Contextual Bandit. The algorithm can easily be used for new visitors, because we already know a lot about a person even when they’re new to the site. At the very least, we know that they’re new visitors, and that’s useful information for the algorithm to leverage. And we can also immediately identify his geolocation, referral source, browser, device type, and firmographic details if applicable (such as company, industry, company size, other technologies used etc.).

Final Thoughts

Contextual Bandit is the most advanced approach to automated experience optimization on the market today. Jordan, Cliff and the rest of the Evergage engineering and data science teams have put an immense amount of work into researching, designing and evolving this algorithm — testing it against real data sets and vetting it with real marketers — to make it as effective as possible at driving real business value. This article covered just a few of the common questions they’ve received; they answer these and more in greater detail in the webinar. They’ve also authored a technical white paper on Contextual Bandit which is available upon request from your customer success or account manager.

Please watch the webinar replay for all the details, and request a demo today to learn more about how Contextual Bandit could work for your business.

Share This
testingpage2