Artificial Intelligence Recommendation:Policy configuration

Overview

On the Policy Configuration page, you can view and configure global and scenario-based recommendation policies. Recommendation policies include deduplication rules, diversity rules, business metric definitions, and experience optimization rules. You can configure recommendation policies based on your business requirements. This allows the recommendation system to implement the expected recommendation logic.

Deduplication rules

Introduction

In some business scenarios, algorithms repeatedly recommend the same items. This can cause cognitive fatigue for users and eventually impair the user experience.

For example, if a user browses a face cream but does not click it to view details, the user may not be interested in the face cream. In this case, you want the system to stop recommending this item to the user for a period of time. You can configure deduplication rules to meet such requirements.

Operation guide

• Filter items by exposure behavior: After you enable this feature, you can configure a time window for this deduplication rule. For example, if you set the time window to 7 days, exposed items are not recommended to the same user for 7 days after the deduplication rule takes effect.

• Filter items by click behavior: After you enable this feature, you can configure a time window for this deduplication rule. For example, if you set the time window to 7 days, clicked items are not recommended to the same user for 7 days after the deduplication rule takes effect.

• Ensure personalized recommendation results for frequent access: After you enable this feature, if all items are exposed due to frequent access by users, the system is allowed to repeatedly recommend the exposed items.

Diversity rules

Introduction

In some business scenarios, homogeneous items may be recommended in a centralized manner. For example, if the system recognizes that a user is interested in fruits, the system increases recommendations for fruits so that more fruit items are displayed on the screen. If the system recognizes that a user is more interested in short video items, the system returns short video items that account for even as much as 80% of the recommended items in a single request. This may affect the diversity of recommendation results and user experience. To avoid the centralized recommendation of homogeneous items, we recommend that you configure diversity rules.

Operation guide

• Ensure a fixed ratio among different types of recommended items (mixed ranking): After you enable this feature, you can set item types and their corresponding proportions. Make sure that the sum of the proportions of all types is 100%. After the diversity rule takes effect, the recommendation results of a single request are returned based on the specified proportions of the item types.

• Improve the diversity of recommended items in a specific dimension (discretization): For the content and news industries, you can specify proportions for the author, category, type, and channel dimensions of recommended items. For the e-commerce industry, you can specify proportions for the store, category, type, and brand dimensions of recommended items. After you enable this feature, you can set window values. For example, you can set the number of consecutive items to 5. After the diversity rule takes effect, in the returned results of a recommendation request, the properties of five consecutive items do not share the same dimension.

Business metric definitions

Introduction

In some business scenarios, some item properties strongly correlate with user interest. For example, in some vertical content communities, the system can determine the interest of users in items based on the author dimension. Therefore, if you want the system to recommend items to users based on their preferred authors, you can define relevance factors to meet such requirements.

For industries such as news and short videos, the common requirement is to preferentially recommend new items. In this case, you must define the timeliness of new items first. This helps the system identify and recommend new items.

Operation guide

• Define relevance factors: You can configure up to three relevance factors and rank the factors. The system preferentially recommends items to users based on the relevance factors.

• Define new items: You can set a window value to define new items. The system uses the window value to determine new items when you configure experience optimization rules to recommend new items.

Experience optimization rules

Introduction

The default industry algorithm templates can be problematic for user experience. For example, content with similar titles is recommended in a centralized manner after the emergence of a hot topic, and outdated content without value is still frequently recommended. Experience optimization rules can help provide a solution to such issues.

Operation guide

Recommend new items: After you enable this feature, the system recommends new items to users first, and then recommends old items in scenarios where experience optimization rules take effect. New items are defined based on the window value for New Item Definition of Business Metric Definition in Global Policy.

Avoid a great deal of content with similar titles: After you enable this feature, the system identifies similar titles based on the built-in Natural Language Processing (NLP) models in scenarios where experience optimization rules take effect. This prevents a great deal of content with similar titles from appearing in the results of each request.