Adobe Employee

Adobe Target Experts on Target High Performance Recommendations

Forum|Forum|2 years ago
January 24, 2024
14 replies
8585 views

February 28th, 2024

We are happy to welcome Cristinel Anastasoaie, Brent Kostak, and Timothy Furlow for an AMA session about Adobe Target High Performance Recommendations.

Our experts:

Brent Kostak is a Product Marketing Manager at Adobe. Brent brings data-driven personalization solutions to market that connect global brands to their customers. Adobe enables enterprise marketing and engineering teams to streamline data and create a single, personalized view of their customers to deliver exceptional digital experiences. Brent has over 10 years of experience helping organizations leverage software technologies and services to better compete and serve their mission.
Cristinel Anastasoaie is a Senior Product Manager for Adobe Target. Cristinel is focused on machine learning personalization in Adobe Target, Analytics for Target integration, and the product user interface. Before his current Adobe role, Cristinel held product management positions on various content management and digital marketing products focused on consumers.
Tim Furlow is a Senior Manager of Personalization & Optimization at Lenovo. Tim's team is responsible for driving the test and learn program worldwide. Tim's approach to Target places emphasis on CRO, and is informed by his career in web analytics, marketing technology, and personalization at Fortune 500 companies.

How this AMA works:

This thread will open on Wednesday, February 28, for you to start submitting your questions.
Reply to this post with any questions you have for our Experts. They will reply to as many of your questions as possible.
After the AMA is over, the thread will be locked for new replies, but it will remain visible as a resource.

This post is no longer active and is closed to new replies. Need help? Start a new post to ask your question.

TF_L

Level 1

Hey all,

Look forward to your questions on recommendations. Cant wait to hear from you all!

markcol

Amelia_Waliany

Author

Adobe Employee

Hi @markcol, you're in the right place! Feel free to post any of your Target Recommendations related questions here in the thread and our team of experts will follow up with you by replying to your question here in the thread.

cristinel

Adobe Employee

Hello everyone, thank you for joining in this Coffee Break. Looking forward to your questions!

JyotiSharmaV

Community Advisor

For the recommendation activity, we have to Feed data to Target, which, if fed in the CSV format, has the first column named "entity.Id" https://experienceleague.adobe.com/docs/target/using/recommendations/entities/entity-attributes.html?lang=en

What if we don't have that unique code for us? What can go in that field?

bkostak

Adobe Employee

Hi Jyoti! These would be unique single value parameters which are classified internally with an organization and mapped to not just products but could be content components such as movies, articles, etc and predefined for each 'entity.id' attribute value (e.g. unique numeric values)

J

jingle77

For popularity-based recommendations, can we use a custom metric? For example, could we build a recommendation algorithm based on a sold-to-view ratio? The metric would be something like product orders/product views.

TF_L

Level 1

@jingle77 you would need to build out a custom algorithm and send it into Target. This is quite possible, but nothing in the box in Adobe Target.

J

jingle77

For "Bring your own" algorithms, is there a template, particularly in Python, we could refer to if we wanted to try out a custom algorithm?

bkostak

Adobe Employee

Hi @jingle77 Thank you for your question! Based on custom templates / scripts, you could use your python algorithm and upload the output in custom criteria configuration as a .csv. You can customized all the OOTB options in the criteria menu to create your own criteria configuration. Note - Custom criteria replace the “offline training” portion of Item-Based recommendations, but behave similarly to Item-Based recommendation algorithms during the online content delivery phase, in that a single key is used for retrieval of recommendations and business rules/filters are then applied.

https://experienceleague.adobe.com/docs/target/using/recommendations/criteria/recommendations-csv.html?lang=en

B

BryanSt3

Hi team!

Is there a product and traffic minimum to use recommendations? Can you layer in user segments into a recipe, such as a “best sellers for those who did X”? Looking to see how granular these recommendations can get, and what other data layers can be added on top.

cristinel

Adobe Employee

One way to tackle this would to filter out recommendations based on profile attributes assuming that the profile stores information you need to filter recommendations on. Some more details about this type of filtering can be found on experience league page: https://experienceleague.adobe.com/docs/target/using/recommendations/criteria/dynamic-static/profile-attribute-matching.html?lang=en

TF_L

Level 1

Thank you @bryanst3 for your question, traffic for a recommendation would depend on your test time to confidence using a statistical calculator. If you have low traffic and cannot meet confidence, I would look at prior data of recommendation interaction and make a trade off as long as your analytics team agrees. Your best bet to understand granularity would be to store profile attributes and to create more of an affinity approach. In addition, you can layer several stacking segments in a recommendation test depending on your use case.

H

hagarwal

I'd like to hear suggestions on using AI activities when there's a publishing schedule that's shorter than the learning may take to build the model.

bkostak

Adobe Employee

Hi @hagarwal thank you for your question! Models to building times within AI activities will correlate to traffic to selected activity locations and conversion rates associated with those success metrics.

For Auto-Target specifically, the personalization activity does not attempt to build a personalized model for a given experience until there are least 50 conversions for that experience. If the model built is of insufficient quality (as determined by offline evaluation on hold-out “test” data, using a metric known as AUC), the model is not used to serve traffic in a personalized manner.

Additional points on model building:

After an activity is live, Auto-Target considers up to the last 45 days of randomly served data when attempting to build models. For example, control traffic, plus some extra randomly served data held out by the algorithm.
When Revenue per Visit is your success metric, these activities typically require more data to build models due to the higher data variance that typically exists in visit-revenue compared to conversion rate.
Because models are built on a per-experience basis, replacing one experience with another experience means that sufficient traffic (at least 50 conversions) must be collected for the new experience before personalized models can be rebuilt.

A

AmsaPa

Is it better to optimize recommended for you with content analysis techniques using product description, user reviews, etc and introduce users to new/unexpected items that align with their interests beyond the browsing history?

TF_L

Level 1

@amsapa thank you for your question I highly suggest a read through of https://experienceleague.adobe.com/docs/target/using/recommendations/introduction-to-recommendations.html?lang=en.

Adobe Target recommendations can be used for so many capabilities than products - Travel, Gaming, Videos, Publish, B2B Sales, and more. It will open up so many possibilities on how to utilize recommendations.

Now how you optimize, you have to be willing to take risks on different UI, featuring reviews, and also testing by different locations WW. You may find what works for one country does not work for another.

S

Sharma_Manu

Hi Team,
Is there any way to dynamically personalize the recommendation algorithm based on user's behaviour and journey or switch the recommendation recipe If existing recipe is not returning any products?

cristinel

Adobe Employee

Thank you @sharma_manu . Recommendation algorithms could be configured to use dynamic inclusion rules based on user profile attributes or on page context similar to category affinity filtering. Using dynamic rules brings more engagement. More details about available filtering rules could be found on Experience League: https://experienceleague.adobe.com/docs/target/using/recommendations/criteria/dynamic-static/use-dynamic-and-static-inclusion-rules.html?lang=en

J

jingle77

Have any users of Adobe Target shown insights regarding how Adobe Target recommendations have impacted product relationships and/or the customer journey? For example, if a site has implemented rec for you, did it cause a significant shift in traffic viewing/ordering a particular set of products?

cristinel

Adobe Employee

Thank you for your great question and that would be a great new feature of Target and / or Target + Analytics. So far, the Target reporting data as well as the algorithms co-occurrence scores are not available. Please use the Adobe Target Community Ideas to post this as a feature request and have other community members vote for this feature.

Show more replies

Sign up

Login with SSO

Login to the community

Login with SSO

Scanning file for viruses.

This file cannot be downloaded