Note: The algorithm experiment feature is available only for Artificial Intelligence Recommendation (AIRec) instances of Standard Edition.
AIRec provides several types of algorithm models based on different industries. You can enable or disable algorithm models as needed and optimize specific algorithms on the experimental platform.
1. Item-based collaborative filtering algorithm
Collaborative filtering algorithms are divided into item-based collaborative filtering algorithms and user-based collaborative filtering algorithms.
AIRec uses item-based collaborative filtering algorithms.
An item-based collaborative filtering algorithm performs the following steps:
(1) Calculates the similarity between items.
(2) Generates a list of recommendations for users based on the similarity between items and the previous behavior of users.
Collaborative filtering algorithms
Traditional item-based collaborative filtering
The traditional item-based collaborative filtering algorithm is based on correlation rules.
The Swing algorithm is used to find similar items and the same items. In contrast, the traditional item-based collaborative filtering algorithm pays more attention to the discovery and diversity of items.
For example, when users provide feedback on a small variety of recommended items on the homepage, you can consider raising the priority of this algorithm to improve the diversity of recommended items.
Swing
The Swing algorithm calculates the similarity between commodities based on graph structures. It can extend to two-hop nodes with a high-dimensional network structure. The Swing algorithm is anti-noise and provides more accurate results than the traditional item-based collaborative filtering algorithm.
As mentioned earlier, the Swing algorithm is used to find similar items and the same items. Compared with the traditional item-based collaborative filtering algorithm, the Swing algorithm focuses more on correlation and requires more complex calculation.
Subcategory convergence optimization
To limit leaf categories, the subcategory convergence optimization algorithm recalls only similar items whose leaf category is the same as the leaf category of the current trigger item.
On the basis of item-based collaborative filtering, the limits on leaf categories ensure that the leaf categories of recommended items are consistent with those of previous behavior, such as clicked items.
For example, if a user clicks a lipstick and expects real-time recommendations for lipsticks next time, you can use this algorithm to recommend lipsticks to this user.
Parent category convergence optimization
To limit level 1 categories, the parent category convergence optimization algorithm recalls only similar items whose level 1 category is the same as the level 1 category of the current trigger item.
On the basis of item-based collaborative filtering, the limits on level 1 categories ensure that the level 1 categories of recommended items are consistent with those of previous behavior, such as clicked items.
For example, if a user clicks a lipstick and expects real-time recommendations for other makeup items such as foundation next time, you can use this algorithm to recommend other makeup items to this user.
Channel convergence optimization
To limit channels, the channel convergence optimization algorithm recalls only similar items whose channel is the same as the channel of the current trigger item.
On the basis of item-based collaborative filtering, the limits on channels ensure that the channels of recommended items are consistent with those of previous behavior, such as clicked items.
For example, if a user clicks a piece of military news and expects real-time recommendations for military news next time, you can use this algorithm to recommend military news to this user.
2. Content-based second-order transfer algorithm
Content-based second-order transfer algorithms
Second-order transfer based on user-preferred categories
For more information, see the category_path field of the item table in Data specifications.
The second-order transfer algorithm based on user-preferred categories calculates the preferred categories of users based on their previous behavior to find similar items.
Compared with collaborative filtering, content-based second-order transfer focuses more on users' preferences for item categories. This algorithm relies on information such as behavioral data and item categories that you uploaded. You must upload such information based on data specifications.
Second-order transfer based on user-preferred tags
For more information, see the tags field of the item table in Data specifications.
The second-order transfer algorithm based on user-preferred tags calculates the preferred tags of users based on their previous behavior to find similar items.
Compared with collaborative filtering, content-based second-order transfer focuses more on users' preferences for item tags. This algorithm relies on information such as behavioral data and item tags that you uploaded. You must upload such information based on data specifications.
Second-order transfer based on user-preferred channels
For more information, see the channel field of the item table in Data specifications.
The second-order transfer algorithm based on user-preferred channels calculates the preferred channels of users based on their previous behavior to find similar items.
Compared with collaborative filtering, content-based second-order transfer focuses more on users' preferences for channels. This algorithm relies on information such as behavioral data and item categories that you uploaded. You must upload such information based on data specifications.
Second-order transfer based on user-preferred authors
For more information, see the author field of the item table in Data specifications.
The second-order transfer algorithm based on user-preferred authors calculates the preferred authors of users based on their previous behavior to find similar items.
Compared with collaborative filtering, content-based second-order transfer focuses more on users' preferences for item categories. This algorithm relies on information such as behavioral data and item categories that you uploaded. You must upload such information based on data specifications.
Second-order transfer based on user-preferred stores
For more information, see the shop_id field of the item table in Data specifications.
The second-order transfer algorithm based on user-preferred stores calculates the preferred stores of users based on their previous behavior to find similar items.
Compared with collaborative filtering, content-based second-order transfer focuses more on users' preferences for stores. This algorithm relies on information such as behavioral data and stores that you uploaded. You must upload such information based on data specifications.
Second-order transfer based on user-preferred brands
For more information, see the brand_id field of the item table in Data specifications.
The second-order transfer algorithm based on user-preferred brands calculates the preferred brands of users based on their previous behavior to find similar items.
Compared with collaborative filtering, content-based second-order transfer focuses more on users' preferences for brands. This algorithm relies on information such as behavioral data and brands that you uploaded. You must upload such information based on data specifications.
Second-order transfer based on user-preferred organizations
For more information, see the organization field of the item table in Data specifications.
The second-order transfer algorithm based on user-preferred organizations calculates the preferred organizations of users based on their previous behavior to find similar items.
Compared with collaborative filtering, content-based second-order transfer focuses more on users' preferences for organizations. This algorithm relies on information such as behavioral data and organizations that you uploaded. You must upload such information based on data specifications.
3. Popular item recall algorithm
Popular item recall based on the popularity of global behavior
The popular item recall algorithm aggregates, analyzes, and calculates the behavioral data of sites such as apps, mini programs, and web pages to find popular items. By default, the popular item recall algorithm is preferentially used for new users.
You can use this algorithm to recall popular items in different scenarios and supplement insufficient personalized items.
4. New item recall algorithm
AIRec can use a small amount of traffic to detect and personalize the potential of new items. To do so, AIRec analyzes the behavior and interests of users, in combination with the characteristics and attributes of new items. Then, AIRec can decide whether to support or suppress the recommendations for new items.
New item recall algorithms
New item support based on user-preferred categories
The new item recall algorithm recalls personalized new items based on user-preferred categories that are specified by the category_path field of the item table. Before you use this algorithm, you must upload the release time and categories of items based on data specifications.
New item support based on user-preferred brands
The new item recall algorithm recalls personalized new items based on user-preferred brands that are specified by the brand_id field of the item table. Before you use this algorithm, you must upload the release time and brands of items based on data specifications.
New item support based on user-preferred stores
The new item recall algorithm recalls personalized new items based on user-preferred stores that are specified by the shop_id field of the item table. Before you use this algorithm, you must upload the release time and stores of items based on data specifications.
New item support based on user-preferred tags
The new item recall algorithm recalls personalized new items based on user-preferred tags that are specified by the tags field of the item table. Before you use this algorithm, you must upload the release time and tags of items based on data specifications.
New item support based on user-preferred channels
The new item recall algorithm recalls personalized new items based on user-preferred channels that are specified by the channel field of the item table. Before you use this algorithm, you must upload the release time and channels of items based on data specifications.
New item support based on the popularity of global behavior
The new item recall algorithm recalls popular new items based on the release time and behavior of new items. The release time is specified by the pub_time field of the item table. Before you use this algorithm, you must upload the release time and behavioral data of items based on data specifications.
New item support based on the latest release time
The new item recall algorithm sorts items in reverse order based on the release time that is specified by the pub_time field of the item table. The items with the later release time rank higher.
The new item recall algorithm recalls new items based on the release time of items. Before you use this algorithm, you must upload the release time of items based on data specifications.
5. Vector recall algorithm
Vector recall based on the sequence of user behavior
The vector recall algorithm uses the sequence of user clicks on items as the input to calculate the vectors of items and calculate the similarity between items or the interests of users in items. This algorithm has better generalization performance than the traditional item-based collaborative filtering algorithm.
This algorithm is based on the word2vec algorithm. More extensive common behavior of users indicates the more similar vectors between items. Compared with collaborative filtering, vector recall can build data models with high-order similarity to improve recall coverage. This algorithm is suitable for scenarios with extensive item behavior. Compared with the item-based collaborative filtering algorithm, this algorithm requires more extensive behavior and is more suitable for business scenarios with extensive behavior.
Tag-based vector recall
The tag-based vector recall algorithm calculates the similarity between items or the interests of users in items based on the vectors of item tags. This algorithm has better generalization performance than the traditional item-based collaborative filtering algorithm.
This algorithm calculates the feature vectors of items based on the semantic content of item tags. In this case, this algorithm requires high-quality tags. If items have high-quality tags, you can consider raising the priority of this algorithm.
Title-based vector recall
The title-based vector recall algorithm calculates the similarity between items or the interests of users in items based on the vectors of item titles. This algorithm has better generalization performance than the traditional item-based collaborative filtering algorithm.
This algorithm calculates the feature vectors of items based on the semantic content of item titles. In this case, this algorithm requires high-quality titles. If items have high-quality titles, you can consider raising the priority of this algorithm.