Exposing the complexity of user affinity and unleashing its full power

Affinity modeling as we know it is simple and effective, but product metadata can be better leveraged to predict the most relevant experience to users.

Head of Content, Dynamic Yield

Every interaction a visitor has with a brand provides a valuable signal about their interests, preferences, and intent – whether viewing a specific product, adding an item to the cart, or making a purchase. As marketers, our job is to analyze these behaviors and the products themselves to gain a better understanding of how it’s all connected so we can build experiences that truly resonate. 

However, despite our best efforts, as visitors become exposed to a wider range of brands, colors, styles, prices, and more, it gets harder to assign meaning to a particular interaction. And oftentimes, the true nature of an engagement ends up lost in a web of complex data. Replaced by a guessing game in which different combinations of targeting conditions are tested in hopes of matching the right experience to the right audience, this challenge is the reason affinity profiling came to be. Adding a score to a user based on the metadata connected to a product they’ve interacted with meant teams could better predict the most relevant items to them in the future. 

But even in all of its glory, the industry is still naive in its application and implementation. In this post, I’ll explain where affinity is today, touch on the enormous untapped potential, and how we’re working to unleash its full power. 

Calculating affinity 

User affinity profiling starts by capturing each engagement with a digital property and then collecting all of the product attribute values associated with the interaction. 

Quick Reference: Example User Affinity Signals

Engagement type: Tthe interaction the user has with the product

Product attribute values: Tthe characteristics describing the product

Product view
Add to cart
Purchase
Filter product grid
Add to wishlist
Custom interaction 
Color: red
Gender: men
Style: T
Category: running

For example, let’s take a look at a sample journey:

Affinity Timeline

In the context of a single visitor, we might see four pageviews generated in which the product color was green, six in which the product was red, and two blue. Additionally, all of the items viewed consisted of Men’s t-shirt under $50. The visitor then places two items in their cart, a red and a green t-shirt, but ultimately purchases the red one. 

Attribute Color Style Gender Price Range
Value: Green Red Blue T Tank Cami Men Women Boys Girls <$50 >$50
Product view 4 6 2 12 0 0 12 0 0 0 12 0
Add to cart 1 1 0 2 0 0 2 0 0 0 2 0
Purchase 0 1 0 1 0 0 1 0 0 0 1 0
Attribute: Color
Value Green Red Blue
Product view 4 6 2
Add to cart 1 1 0
Purchase 0 1 0
Attribute: Style
Value T Tank Cami
Product view 12 0 0
Add to cart 2 0 0
Purchase 1 0 0
Attribute: Gender
Value Men Women Boys Girls
Product view 12 0 0 0
Add to cart 0 0 0 0
Purchase 0 0 0 0
Attribute: Price Range
Value <$50 >$50
Product view 12 0
Add to cart 2 0
Purchase 1 0

The entire session might look something like this, with all of the attribute values and engagement types to be considered marked for scoring. 

Affinity score weights

Every time a user interacts with a product, the score of each attribute value is updated by the type of interaction with it as well as when it occurred.

Engagement

In order for the data above to become meaningful, a correlation between the user interaction and an attribute value must first be determined. Today, this is done by assigning a weight to each engagement based on its assumed level of intent and then summing up the total number of engagements per each value.

Depending on the affinity algorithm, defaults weights might be quantified as follows. 

  • Purchase: 4X
  • Add to cart: 3X
  • Add to wishlist: 2X
  • Product view: 1X
  • And more

Understandably, the higher level of intent associated, the greater the weight given to the interaction. For example, making a purchase would receive more weight than adding an item to the cart, which itself, outweighs a simple product view. From there, the occurrence of each interaction is multiplied by the weight. 

Engagement score = interaction type weight x attribute value count

So if purchase weight is 4, add to cart is 2, and product view is 1, the total score for the value “color:red” is (1×6) + (2×1) + (4×1) = 12. 

Weight Color
Green Red Blue
Product view 1 4 6 2
Add to cart 2 1 1 0
Purchase 4 0 1 0
Score 0 0 0

These scores can now be used to build an affinity profile, which essentially ranks a list of preferred product colors, categories, brands, and so forth based on the customer’s activity. 

Recency

As user behavior and preferences change over time, affinity profiles must be refined with every new pageview, action, and event in real time. Therefore, the recency of each interaction is also taken into account, ensuring the algorithm favors newer activity over older data. 

Three time periods are typically used in scoring: 

  • Real-time or current browsing session: 8X
  • Recent history or in the last month: 2X
  • All-time or the last six months 1X

When layering this information into the attribute value score, in addition to multiplying the interaction weight, the recency weight is also multiplied. 

Recency score = recency weight (interaction weight x attribute value count)

Assuming the same user referenced above comes back to the site two weeks later and purchases the green t-shirt previously abandoned in their cart, the total score for “color: green” would be 2[(1×4) + (2×1) + (4×0)] + 8(4×1) = 44.

Attribute Color
Recency Value: Weight: Green Red Blue
Recent History
(weight = 2)
Product view 1 4 6 2
Add to cart 2 1 1 0
Purchase 4 0 1 0
Score 6x2=12 12x2=24 2x2=4
Real Time
(weight = 8)
Purchase 4 1 0 0
Score 4x8=32 0 0
Total Score 44 24 4
Recency: Recent History (weight = 2)
Attribute: Color:
Green Red Blue
Product View
Weight: 1
4 6 2
Add to Cart
Weight: 1
1 1 0
Purchase
Weight: 1
0 1 0
Score 6x2=12 12x2=24 2x2=4
Recency: Real Time (weight = 4)
Attribute: Color
Green Red Blue
Purchase
Weight: 4
0 1 0
Score 4x8=32 0 0
Total Score 44 24 4

Now the affinity profile reflects two strong scores (engagement and recency), enabling us to showcase products the user is more likely to engage with at any given moment in time. And through the normalization of site data, can even uncover “long tail” attributes correlated with a purchase, ensuring results aren’t skewed by demoting frequently-seen or always-popular attributes.

Releasing affinity into the wild

The straightforward method of affinity calculation described above is simple and effective – it leverages available product metadata to better uncover relationships between attributes from different product fields based on the two light formulas above. This makes it easy to communicate to stakeholders and thus emphasize the importance of good-quality attributes in the product feed – it also incorporates newly released products naturally, without the need for any additional setup or configuration. But the future of affinities is a lot more flexible, thanks to the help of machine learning, which introduces opportunities for increased ranking power and more complex correlations to be uncovered. 

These smart affinities can adjust the variables used in calculating scores, as opposed to leveraging the same formula with a static weighting system for all visitors. For example, by adjusting the weight of one attribute to focus on long-term behavior while another is geared toward the current session. Trained using past customer sessions which culminated in conversion, Deep Learning models can assign weights to different interactions and time scopes for each attribute as well as surface correlations between attributes. Then, based on a real-time user session, it can rank an attribute value, say color, by the probability of being added to the cart or purchased, even for colors or brands the individual hasn’t been exposed to before. And through a greater degree of data normalization, critical contextual nuances such as the type of site and product offering, date, and time, will all be taken into consideration during calculations. 

You could say, affinity modeling is becoming significantly more personalized. And with such much variability, teams need a lot more visibility into how affinities are generated, analyzed, and actioned. Today, most affinity data exists in a black box, meaning those leveraging it to optimize their personalization efforts have very little insight into how the algorithm at work calculated an affinity score. But “this is your affinity score, take it or leave it” isn’t good enough anymore. All of the data around affinities and attributes should be exposed, customizable, and personalized depending on the unique needs of the business

Dynamic Yield’s approach to affinity

We’ve been in the game for a while now, and our affinity-based personalization capabilities are growing to support all of the necessary applications listed above as well as those that may emerge during the continued evolution of affinity modeling. 

Our recommendation engine taps into direct engagement with product feed attributes, using data aggregated through rich affinity profiles to serve the most appropriate experience variations and products to each visitor. We also provide multiple methods and algorithms to power how these recommendations are personalized, allowing teams to employ deep learning and neural networks as well as continuously test and optimize a mix of affinity and popularity scores to best fit their needs.

And the affinity scores powering these recommendations don’t live behind gated walls. With our Unified Customer Profile API, this history of online activity can be fed through Dynamic Yield, or into any other system, such as a chatbot, smart TV, clienteling app, and more for the creation of other highly-tailored experiences. 

For example, using our Affinity Profile tool, a new user who lands on a site for the very first time would have an empty affinity state.

An affinity score populating as the user interacts with different products

Over the course of the session, the user’s affinity score would begin to populate, with important actions and behaviors like visiting a product page or adding an item to the cart contributing to the calculation. Top values are then bucketed according to the various signals of intent for a high-level view depicting which interactions held the greatest significance. 

Top attribute values ranked by significance

Furthermore, marketers can choose to drill into the user’s full profile, analyzing each of the values and attributes captured to better understand the nature of interactions made over time. For instance, what an individual’s score looks like for women’s products versus those for the home.

Diving into a user's affinity scores

And because affinity profiling is tied to a product feed, we make it easy to explore a brand’s catalog, offering helpful suggestions for recommendations based on either a personalized algorithm a popularity-based approach. 

Exploring recommendation strategies based on affinity vs popularity

And the ability to action these valuable user attributes doesn’t end there. 

Making affinity targetable

Up until now, if a marketer wanted to target an audience of people who had a high affinity to, say, the category ‘running,’ they would have to arbitrarily define the makeup of that segment through a series of targeting conditions. Based on historical interactions, this might be visitors who viewed 3 items or running products in the past week and who’s average order value (AOV) was above $75.

With Affinity-based conditions, visitors with preferences to specific brands, colors, categories, and other attributes from the product catalog can be segmented and automatically targeted without having to build the audience manually. 

Targeting users by affinity with Dynamic Yield

A few use cases of how teams leverage affinity profiles for the creation of highly-targeted experiences include:

  • Promoting relevant products in the homepage banner for users with an affinity to brown boots
  • Reordering menu items within the navigation based on the user’s affinity to Men’s pants
  • Moving product rows up on the shoe category page to surface Nike before any other brand
  • Serving targeted discount notifications to users with an affinity to women and/or accessories.

Affinity-based conditions open up an entirely new world of potential, giving eCommerce brands the power to capitalize on what they know about their customers but weren’t previously able to.

Turning complexity into actionability 

The beauty of user affinity lies in all of its variability – and the only way to unleash its full power to predict what an individual wants is through exposing and providing as much control over that information as well as how it can be used as possible. Dynamic Yield has built a system that will allow us to continue taking steps towards more efficient modeling, and we’re excited about the seemingly endless opportunities to help brands make more meaningful connections with their customers. 

If you want to learn more about our affinity-based personalization capabilities, contact us here

Rate this post