The Deep Learning strategy is the most successful personalized recommendation strategy in our toolkit. It defines the most likely item users will interact with next.
The strategy works as an exploratory and discovery algorithm, conveying recommendations that are likely to be titled “Recommended for you”. It is optimal for top-funnel experiences like email campaigns and on the site homepage. Both are the initial point of product discovery for most online shoppers.
How does the Deep Learning algorithm work?
To determine which item you are most likely to interact with next, Dynamic Yield uses a deep learning algorithm based on the word2vec model. To learn more about the word2vec model, see The rise of deep learning-based recommendations.
Every time a user interacts with a product, we note all the recent products they have interacted with. The closer in time a user interacts with two different products, the stronger their association (or vector). Based on all user interactions, a model is built to determine how likely a user is to interact with product A after interacting with product B.
As each user browses your site and interacts with different products, we use the model to calculate which product is most strongly associated with the set of products in their recent history. This represents the most likely product similar users interacted with. The results improve for users who have richer recent histories, but the algorithm is also effective for users who only have one product interaction, even during current sessions.
The algorithm is updated weekly, redefining all product associations based on your traffic. However, it takes into account the current user’s behavior in real time.
This algorithm is continuously improved and the logic is subject to change.
- Dynamic Yield has been collecting data from your site for at least 30 days (data is collected as soon as you add the Dynamic Yield script to your site). There is no limit on feed size. This algorithm works well for small and large feeds alike.
- This feature is part of AdaptML®, our self-training deep learning AI system. Contact your Customer Success Manager to learn more about AdaptML®.
When creating recommendation campaigns using the Deep Learning strategy, we strongly recommend using the following campaign settings:
- Variation stickiness: Sticky for the user
- Attribution window: Variation is served and clicked
- End attribution window after: 1 or 7 days
Custom filter rules and filters are applied after the results have been returned for a user. This means they are filtering an already-small set of products. We recommend to use few or no filtering rules to ensure that you don’t filter out all of the results, which would result in the user being served the fallback strategy (Deep Learning → User Affinity → Popularity).
How does this algorithm compare to collaborative Filtering and User Affinity?
These are all personalized recommendation algorithms that work with different technologies to provide items that are “Recommended for you”. Currently, the Deep Learning algorithm shows the best performance results in most cases when used on the homepage. It combines the best of aspects of the other algorithms: The option to work in real time on feeds of any size (like User Affinity) and the ability to continuously learn based on a predictive model (like Collaborative Filtering). We recommend starting off by testing this new algorithm in comparison to any other algorithm currently running on your site.