CN-122019885-A - Recommendation method and recommendation system

CN122019885ACN 122019885 ACN122019885 ACN 122019885ACN-122019885-A

Abstract

The invention discloses a recommendation method and a recommendation system, wherein the recommendation method comprises the steps of S1, obtaining unbiased features and auxiliary features of users, respectively carrying out embedding processing on the unbiased features and the auxiliary features to obtain feature embedding and other feature embedding, S2, carrying out clustering on all users through a clustering module, S3, distributing target clusters for the users based on the distance from the users to each cluster, obtaining group characterization of the target clusters, S4, generating complete user characterization comprising a clustering center and personalized offset features, S5, inputting the complete user characterization into a main recommendation model to output recommendation results, and S6, constructing inter-cluster correlation regular loss and main model basic loss to finish cold start recommendation of low-activity users. According to the invention, clustering of the users is finished firstly through unbiased characteristics, and interest migration from high activity to low activity is realized, so that cold start recommendation for the low activity users is realized.

Inventors

WANG JUN
LU SIYU
TANG XINTAO

Assignees

广州天宸健康科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260205

Claims (9)

1. A recommendation method, comprising the steps of: S1, acquiring unbiased features and auxiliary features of a user, and respectively performing embedding processing on the unbiased features and the auxiliary features to obtain feature embedding and other feature embedding; s2, taking the unbiased characteristics as input, clustering all users by a clustering module to generate N clusters, and performing cluster mapping and cluster decoupling to obtain cluster embedding; S3, distributing a target cluster for the user based on the distance between the user and each cluster, and acquiring a cluster characterization of the target cluster, wherein the cluster characterization is calculated by high active user characteristics in the clusters: s4, inputting the feature embedding, other feature embedding and cluster embedding into a feature fusion module to generate a complete user representation comprising a cluster center and personalized offset features; s5, inputting the complete user characterization into a main recommendation model, carrying out user interest prediction and recommendation calculation, and outputting a recommendation result; And S6, constructing inter-cluster correlation regular loss and main model basic loss, and optimizing a clustering module and a main recommendation model through joint loss to finish cold start recommendation of the low-activity user.
2. The recommendation method of claim 1, wherein in step S1, the unbiased feature is a user inherent feature with high coverage, including one or more of a user' S gender, age, region, and model number of a mobile phone.
3. The recommendation method of claim 1, wherein in step S2, the clustering module updates by NTM, adjusts update strengths of different clusters by a cluster attention mechanism, and the clustering module decouples from a main recommendation model training process, and the gradient does not return.
4. The recommendation method of claim 1, wherein in step S3, the group characterization is generated by calculating the distance between each user and N clusters, screening the cluster closest to the N clusters as a target cluster, extracting the characteristics of all high active users in the target cluster, and obtaining the group characterization of the cluster through mean calculation or weighted summation.
5. The recommendation method of claim 1, wherein the radius of the personalized offset feature in step S4 is adjustable, and smaller radius personalized feature offsets can improve compactness in clusters and reduce generalization errors of clusters.
6. The recommendation method of claim 1, wherein in step S5, the construction process of the inter-cluster correlation regular loss is to construct a covariance matrix, express the correlation between clusters by covariance, calculate a matrix norm of the covariance matrix, and use the matrix norm as the regular loss.
7. The recommendation method of claim 6, wherein the matrix norm is one of an L1 norm, an L2 norm, or a Frobenius norm for measuring correlation integration between two different clusters.
8. The recommendation method of claim 1, wherein in step S6, the joint loss function is: L=Lbase+λLcluster; Wherein Lbase is a basic loss, lcluster is a clustering loss, and lambda is a super parameter.
9. The recommendation system is characterized by comprising a feature acquisition and embedding processing unit, a user clustering unit, a group characterization generating unit, a feature fusion unit, a recommendation unit and a joint loss optimizing unit, wherein: The feature acquisition and embedding processing unit is used for acquiring unbiased features and auxiliary features of a user, and respectively carrying out embedding processing on the unbiased features and the auxiliary features to obtain feature embedding and other feature embedding; the user clustering unit is used for taking the unbiased characteristics as input, clustering all users through the clustering module to generate N clusters, and performing cluster mapping and cluster decoupling to obtain cluster embedding; the group representation generating unit is used for distributing target clusters for users based on the distance between the users and each cluster, and acquiring group representations of the target clusters, wherein the group representations are calculated by high active user features in the clusters: The feature fusion unit is used for inputting the feature embedding, other feature embedding and cluster embedding into the feature fusion module to generate a complete user representation comprising a cluster center and personalized offset features; the recommendation unit is used for inputting the complete user representation into a main recommendation model, carrying out user interest prediction and recommendation calculation, and outputting a recommendation result; the combined loss optimization unit is used for constructing inter-cluster correlation regular loss and main model basic loss, and cold start recommendation of the low-activity user is completed through the combined loss optimization clustering module and the main recommendation model.

Description

Recommendation method and recommendation system Technical Field The invention relates to the technical field of information mining, in particular to a recommendation method and a recommendation system. Background In the traditional recommended service scenario, the user group presents obvious two-stage differentiation characteristics that the low active user accounts for a very high proportion, becoming a 'silent majority', while the medium and high active users account for only a small proportion. The differentiation causes the recommendation model to face a severe cold start problem, and the method is characterized in that only sparse behavior data of the low-activity user is easy to be ignored by the model to become a model blind point, and the training process of the recommendation model is more dependent on behavior sequence data of the medium-high-activity user, so that the recommendation effect of the model on the low-activity user is extremely poor, and a local failure phenomenon occurs. The existing thinking for solving the cold start of the user comprises independent modeling, transfer learning, hot list recommendation and the like, but the methods have obvious defects that the independent modeling is based on strong assumption conditions, the applicable scene is limited, the effect of the transfer learning is uncontrollable, the negative transfer problem is easy to occur, the hot list recommendation cannot meet the personalized requirements of the user, and the core problem that the low-activity user behavior is submerged cannot be effectively solved. Therefore, there is a need for a cold start recommendation scheme that can leverage user basic features to achieve cross-activity interest migration. Disclosure of Invention The invention provides a recommendation method for solving the technical problems in the prior art, which comprises the following steps: S1, acquiring unbiased features and auxiliary features of a user, and respectively performing embedding processing on the unbiased features and the auxiliary features to obtain feature embedding and other feature embedding; s2, taking the unbiased characteristics as input, clustering all users by a clustering module to generate N clusters, and performing cluster mapping and cluster decoupling to obtain cluster embedding; S3, distributing a target cluster for the user based on the distance between the user and each cluster, and acquiring a cluster characterization of the target cluster, wherein the cluster characterization is calculated by high active user characteristics in the clusters: s4, inputting the feature embedding, other feature embedding and cluster embedding into a feature fusion module to generate a complete user representation comprising a cluster center and personalized offset features; s5, inputting the complete user characterization into a main recommendation model, carrying out user interest prediction and recommendation calculation, and outputting a recommendation result; And S6, constructing inter-cluster correlation regular loss and main model basic loss, and optimizing a clustering module and a main recommendation model through joint loss to finish cold start recommendation of the low-activity user. Further, in step S1, the unbiased feature is a user inherent feature with high coverage rate, including one or more combinations of user gender, age, region, and mobile phone model. Further, in step S2, the clustering module updates by adopting the NTM method, adjusts the update strength of different clusters through the cluster attention mechanism, and decouples the clustering module from the main recommendation model training process, so that the gradient is not returned. Further, in step S3, the group characterization is generated by calculating the distance between each user and N clusters, screening the cluster closest to the N clusters as a target cluster, extracting the characteristics of all the high active users in the target cluster, and obtaining the group characterization of the cluster through mean calculation or weighted summation. Further, the radius of the personalized offset feature in step S4 is adjustable, and the personalized feature offset with smaller radius can improve the compactness in the cluster and reduce the generalization error of the cluster. Further, in step S5, the construction process of the inter-cluster correlation regular loss is to construct a covariance matrix, express the correlation among clusters through covariance, calculate the matrix norm of the covariance matrix, and take the matrix norm as the regular loss. Further, the matrix norm is one of an L1 norm, an L2 norm or a Frobenius norm, and is used for measuring correlation integration between two different clusters. Further, in step S6, the joint loss function is: L=Lbase+λLcluster Wherein Lbase is a basic loss, lcluster is a clustering loss, and lambda is a super parameter. The application also provides a recommendation system, whic