VisualDNA technology is a full stack of scalable personalization technology that results in outstanding performance for increasing user engagement and revenue through online personalization:
1. The Chassis: Scalable cloud based infrastructure:
Our technology is built in Amazon’s Web Services (cloud computing) using Hadoop (Map Reduce) and Cassandra (non-relational database) to enable us to service millions of users for both our front end user preference collection and back end recommendation, inference, ad targeting and analytics.
2. The Fuel: User profile collection using image-based surveys:
We use image-based surveys as they have two significant benefits: 1. Very high user engagement rates, 2. Rich, accurate, first party user data.
We achieve start rates of 90% of users who are shown the survey starting the survey and complete rates of 93% of users who start the survey completing the survey, resulting in an overall completion rate of 84%. Our surveys are presented to users in a frictionless way as an integral feature that helps the user get more relevant content, products or deals on a site.
Our surveys predict user tags in our 3000 node taxonomy covering the following six areas: 1. Personality traits (e.g. extroversion), 2. Tastes and preferences (e.g. luxury lover), 3. Interests (e.g. golf), 4. Purchase intent (e.g. car buyer), 5. Long term goals (e.g. marriage), 6. Demographics (age, gender, lifestage and children). We are unique in our ability to identify valuable information about the user important for recommendation that is not easily identifiable by other means online. Indeed, personality traits, tastes and preferences and offline interests that are not expressed online are vital to ensuring the right content or product is put in front of the right user.
3. Oil: User preference inference from online behaviour:
Our inference algorithms predict profiles of users who have not taken our surveys based on their online behaviour. This allows us to scale profile acquisition to all of a site’s users who have cookies supported in their browser (approx 95% of users).
4. Engine 1: Recommendation engine with automatic learning:
We have successfully applied our recommendation engine to a number of different class of items including news article recommendation, video recommendation and deal recommendation. We typically achieve increases in click through rates of 80-90% when our recommendations are A/B tested against a control group shown random recommendations.
5. Engine 2: Personalized ad targeting:
Our ad targeting engine ensures that users get relevant ads and advertisers get a higher return on their ad spend. There are two components to our personalized ad targeting technology; Ad targeting, showing an ad to only the users who are most likely to click and convert, and Ad personalisation, customizing the creative tto the individual to increase the likelihood they will click and convert. Our technology allows us to deliver personalized ads to users we have profiled (either through surveys or inference) wherever they go online. We achieve this through our integrations with DoubleClick Ad Exchange, Invite Media and AppNexus.
6. Dashboard: Audience analytics and dashboard presentation:
Our analytics and dashboard presents aggregated audience data enabling publishers, advertisers and merchants to gain unparalleled insight into their audience or customer base. We have a number of different views that we present both in image space and profile space.
Image ranking report A highly visual representation that shows, for a particular survey, the proportion of users in the audience who selected each image and how this audience differs from a baseline audience (typically the average Internet audience). More detailed analysis is possible by using individual images as pivots to show the proportion of users in the audience for each second image, who also selected a specific image.
Tag cloud report
A similar purpose to the image ranking report, but represents the audience profile in words.
Cluster analysis report
When we have a sufficiently large set of users (1000+) we perform cluster analysis to identify the 3-7 most significant clusters in an audience or customer base. This provides a clearer representation of the key distinct groups within the audience and enables coarse grain customization of a product to the key audience groups.
Performance analytics report
Our performance analytics give interaction and conversion statistics for different types of users.
Examples
Recommendation Example: Personalized deal of the day emails
Our recommendation engine has been used to create personalized emails for a “deal of the day” site that has multiple deals each day. In this case an email containing the most relevant deals for each user needs to be sent out each morning. Thus we maintain a recommender that learns the relationship between user profiles and the 1000 deal categories (rather than the specific deals themselves) and preclassify each new deal into one of these 1000 categories.
For each deal for each individual user we are able to compute the probability they will click and purchase a deal in that category and can then create an email that contains the two or three deals that the user is most likely to purchase. The subject line of the email can be similarly customized for each user. The result is higher email open rates, click rates and ultimately deal purchases.
Ad Technology Example: Personalized deal of the day ads
In this example we deliver ads to users containing the most relevant deal of the day for that user. Our flash ad unit reads the VisualDNA user cookie to extract their unique id and sends a request back to our server to ask for the deal that should be presented to that user. The server conducts a look up from a precomputed recommendation of the most relevant deal for that user and returns the deal id back to the flash ad unit. The flash ad unit then fetches the text and image for the specified deal and renders it into the unit. Our precomputed recommendation can update in real-time according to user interactions. For example, if a user had already visited the deal of the day site and viewed a particular deal, the deal id lookup can be updated immediately to show the same deal, a form of retargeting.
We have been granted two patents protecting our methodology with two more patents pending.