Jaa


Personalized recommendations lists

Intelligent Recommendations provides businesses with a toolkit of relevant scenarios. This article describes User picks, or personalized recommendations based on full user history, recent activity, or session activity.

Types of personalized recommendations

The User picks scenario is a style of personalized recommendations that focuses on capturing the user’s tastes, or preferences, and positions a user in unique locations in the item space.

This scenario creates highly personalized recommendations in the context of a large catalog of items. The distance between a user and a particular item decides its strength of relationship. Vectors that 's closer together represent a stronger connection.

User picks can be exposed to users in diverse ways, meaning that there are multiple flavors of "picks for you."

Depending on the business strategy and user intent, the models can:

  • Learn from everything known about users.

  • Add weight to events, inferring more from recent actions or events.

  • Consider only recent events.

Example of personalized recommendations.

Picks based on entire user history

When the picks are based on user history, this scenario takes into consideration all user-consumed items from the catalog in the past over multiple interactions with the system. In the retail vertical, the picks include the entire purchase history of a user. The models use collaborative filtering techniques to recommend the next set of items a user might enjoy, based entirely on their history of consumption, such as purchase history.

Examples based on entire user history:

  • For gamers, Xbox games related to what they play most often

  • Recommending similar films

  • Suggesting a new TV series

  • Documentation or training activities that a user might have interest in reading or pursuing

Picks based on recent activity

Sometimes, recent interactions with the system matter more and represent a better signal for personalized recommendations. In this case, the models can either weight recent signals or use only the most recent interactions as a seed (starting point).

Examples based on recent activity:

  • Frequent and longtime grocery shopper (changing needs)

  • Games or movies recommendations (changing relevance)

  • Music playlist generation (changing or evolving taste)

Picks based on real-time activity

When the picks are based on real-time signals, with the current session as input, it's a scenario with the shortest time range. The signals come as real-time events and, together with pre-trained models, can help serve instant recommendations relevant to the current situation.

Examples based on real-time activity:

  • User recently viewed items. Display a list of related items.

  • User just finished playing a game. What upsell opportunities are available?

  • User downloaded content. What additional content to choose?

  • User is reading articles. What should be read next?

Best practices when using personalized recommendations

With personalized recommendations, note best practices for specific situations.

Full list personalization

Often, retailers don’t need personalized recommendations for an entire catalog. Instead, they have a specific catalog subset to choose from for exposing to consumers. It might be already ordered by priority. In this case, an editor or curator might apply a slight reshuffle to push the items of user interest higher in the list. Intelligent Recommendations can support this experience "on the fly," based on existing complete models. Retailers can require a complete reshuffling of the list, or only remove items a user has already purchased. By removing user history from a list, any set of items can be slightly personalized, ensuring that there's no wasted real estate for product placement.

Examples of full list personalizations:

  • Personalized trending lists

  • Personalized deals, choosing from discounted products or new deals

Diversifying taste in personalized recommendations

Understanding user actions is an important role for personalizing recommendations lists. Similarly, the "like" action can potentially have a plethora of different meanings. Hence, not every like action is the same. A parent may like listening to blues music videos and their child on the same account enjoys watching spaceship launch videos. Our models use the multiple persona algorithm in Intelligent Recommendations and recognize that a user’s taste can vary. User interactions are split into clusters (separate groups) to diversify results. The split provides suggestions from two separate clusters and interleaves them in the results returned to users. This feature protects the changing taste of users and ensures that user interests can't overpower one another.

Examples of diversifying tastes:

  • Diverse taste in music or movies

  • Shopping various categories (shoes, jewelry, cleaning supplies)

  • Family accounts with more than one person

See also

Fine-tune results
Provide item-based recommendations lists
Provide trending recommendations lists