A vector processing method, apparatus, device, and medium

By introducing a TD×TD hybrid matrix and residual transformation processing into the recommendation system, the representation collapse problem of feature interaction matrices with smaller matrix dimensions is solved, thereby improving the feature information richness and prediction accuracy of the recommendation model.

CN122451207APending Publication Date: 2026-07-24TENCENT TECHNOLOGY (SHENZHEN) CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
TENCENT TECHNOLOGY (SHENZHEN) CO LTD
Filing Date
2026-04-08
Publication Date
2026-07-24

AI Technical Summary

Technical Problem

When existing recommendation systems extract interaction features, the effective rank of the feature interaction matrix with a smaller matrix dimension is affected by the dimension of the original matrix, resulting in insufficient interaction feature information, which in turn reduces the representation accuracy and prediction ability of the recommendation model.

Method used

By introducing a first mixing matrix with a matrix dimension of TD×TD to perform feature mixing processing on the feature concatenation vector, and combining residual transformation and feature extraction, the representation collapse phenomenon is alleviated and the richness of feature information is improved.

Benefits of technology

It effectively mitigates representation collapse and improves the accuracy of interaction representation vectors and the predictive ability of recommendation models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122451207A_ABST
    Figure CN122451207A_ABST
Patent Text Reader

Abstract

The application provides a vector processing method, device, equipment and medium, the method comprises the following steps: a feature input vector related to a business object and a to-be-recommended resource is converted into T first feature conversion vectors by a recommendation model; the T first feature conversion vectors are spliced to obtain a first feature splicing vector; the first feature splicing vector is mixed by a first mixing matrix in the recommendation model to obtain a first feature mixed vector; the first feature splicing vector and the first feature mixed vector are subjected to residual transformation to obtain T first residual transformation vectors; and T first feature extraction vectors are obtained by extracting features from the T first residual transformation vectors; and an interaction feature vector representing an interaction probability distribution is output by the recommendation model based on the T first feature extraction vectors. The application can output an interaction feature vector representing a more accurate interaction by using the recommendation model, thereby improving the prediction ability of the recommendation model.
Need to check novelty before this filing date? Find Prior Art