CN-122020340-A - Model classification method, apparatus, device, storage medium, and computer program product

CN122020340ACN 122020340 ACN122020340 ACN 122020340ACN-122020340-A

Abstract

The application discloses a model classification method, a device, equipment, a storage medium and a computer program product, which are used for solving the problems that the existing model classification scheme cannot deeply analyze the internal structure of a model only according to the name, the creation time or a few surface parameters of the model, so that the classification result is coarse, the precision is insufficient, and the requirements of fine model management and scene call are difficult to support. The method comprises the steps of obtaining model structure parameters of a model to be classified, determining parameter coefficients to be classified corresponding to the model to be classified according to the model structure parameters, obtaining parameter domain reference coefficients corresponding to predetermined classification parameter domains, determining deviation values between the parameter coefficients to be classified and the parameter domain reference coefficients, and determining target classification parameter domains corresponding to the model to be classified according to the deviation values, wherein the parameter domain reference coefficients are determined based on the structure parameter coefficients of the corresponding classification parameter domains and the preset experimental model.

Inventors

LI CHUNPENG
LIU DEHONG
ZHU LE
WANG XINHUA
WANG MINGFEI
WU HUIPING

Assignees

中移在线服务有限公司
中国移动通信集团有限公司

Dates

Publication Date: 20260512
Application Date: 20251224

Claims (10)

1. A method of model classification, comprising: obtaining model structure parameters of a model to be classified, wherein the model structure parameters comprise at least one of total network layers, total nodes in series and total nodes in parallel; Determining a parameter coefficient to be classified corresponding to the model to be classified according to the model structure parameters; Acquiring a predetermined parameter domain reference coefficient corresponding to each classification parameter domain, wherein each parameter domain reference coefficient is determined based on a structural parameter coefficient of a preset experimental model of the corresponding classification parameter domain; determining deviation values between the parameter coefficients to be classified and the parameter domain reference coefficients; and determining a target classification parameter domain corresponding to the model to be classified according to the deviation value.
2. The method according to claim 1, wherein the parameter domain reference coefficients corresponding to the classification parameter domains are predetermined, specifically comprising: Obtaining a second model structure parameter of a preset experimental model; determining a structural parameter coefficient corresponding to each experimental model based on the second model structural parameters; determining a plurality of classification parameter fields corresponding to the structural parameter value fields according to a preset classification rule; determining a classification parameter domain matched with the structural parameter coefficient of each experimental model; And determining parameter domain reference coefficients corresponding to the classification parameter domains according to the structural parameter coefficients of the experimental models matched with the classification parameter domains for each classification parameter domain.
3. The method according to any one of claims 1 or 2, wherein the determining, according to the model structure parameter, a coefficient of a parameter to be classified corresponding to the model to be classified specifically includes: And determining the parameter coefficients to be classified corresponding to the model to be classified according to the total network layer number, the total node number, the total serial node number and the total parallel node number based on the preset serial structure weight coefficient and the preset parallel structure weight coefficient.
4. The method according to claim 2, wherein the determining, for each classification parameter domain, the parameter domain reference coefficient matched with the classification parameter domain according to the structural parameter coefficient of each experimental model matched with the classification parameter domain specifically comprises: Determining the dispersion degree of the structural parameter coefficients of each experimental model matched with the classification parameter domain; and when the dispersion is smaller than a preset dispersion threshold value, determining the coefficient mean value of the structural parameter coefficients of each experimental model matched with the classification parameter domain as the parameter domain reference coefficient of the classification parameter domain.
5. The method as recited in claim 4, further comprising: when the degree of dispersion is greater than or equal to the preset dispersion threshold value, respectively determining the difference value between each structural parameter coefficient and the coefficient mean value; Determining the sequence of the structural parameter coefficients according to the difference value corresponding to the structural parameter coefficients; And deleting all the structural parameter coefficients of each experimental model corresponding to the classification parameter domain according to the sorting, so as to obtain a second structural parameter coefficient set, and determining the parameter domain reference coefficient according to the second structural parameter coefficient set.
6. The method according to claim 5, wherein said determining said parameter domain reference coefficients from said second set of structural parameter coefficients comprises: If the deleting process is executed, determining the coefficient mean value as the parameter domain reference coefficient if the number of the deleted structural parameter coefficients is smaller than or equal to a preset deleting number threshold value; If the deleting process is executed, the number of the deleted structural parameter coefficients is larger than the preset deleting number threshold, a second average value is determined according to the maximum structural parameter coefficient and the minimum structural parameter coefficient in the second structural parameter coefficient set, and the second average value is determined to be the parameter domain reference coefficient.
7. A model classification apparatus, comprising: The system comprises a parameter acquisition unit, a classification unit and a classification unit, wherein the parameter acquisition unit is used for acquiring model structure parameters of a model to be classified, and the model structure parameters comprise at least one of total network layers, total nodes in series and total nodes in parallel; The parameter coefficient determining unit is used for determining a parameter coefficient to be classified corresponding to the model to be classified according to the model structure parameters; The reference coefficient acquisition unit is used for acquiring a predetermined parameter domain reference coefficient corresponding to each classification parameter domain, wherein each parameter domain reference coefficient is determined based on a structural parameter coefficient of a preset experimental model corresponding to the classification parameter domain; the deviation determining unit is used for determining deviation values between the parameter coefficients to be classified and the parameter domain reference coefficients; and the classification unit is used for determining a target classification parameter domain corresponding to the model to be classified according to the deviation value.
8. A model classification apparatus comprising: Processor, and A memory arranged to store computer executable instructions that, when executed, cause the processor to: obtaining model structure parameters of a model to be classified, wherein the model structure parameters comprise at least one of total network layers, total nodes in series and total nodes in parallel; Determining a parameter coefficient to be classified corresponding to the model to be classified according to the model structure parameters; Acquiring a predetermined parameter domain reference coefficient corresponding to each classification parameter domain, wherein each parameter domain reference coefficient is determined based on a structural parameter coefficient of a preset experimental model of the corresponding classification parameter domain; determining deviation values between the parameter coefficients to be classified and the parameter domain reference coefficients; and determining a target classification parameter domain corresponding to the model to be classified according to the deviation value.
9. A computer readable storage medium storing one or more programs, which when executed by an electronic device comprising a plurality of application programs, cause the electronic device to perform the model classification method of any of claims 1-6.
10. A computer program product comprising a computer program which, when executed by a processor, implements the model classification method according to any of claims 1-6.

Description

Model classification method, apparatus, device, storage medium, and computer program product Technical Field The present application relates to the field of wireless communications technologies, and in particular, to a model classification method, apparatus, device, storage medium, and computer program product. Background In the data-driven era, various algorithm models such as machine learning models, deep learning models and the like are widely applied to various fields such as financial wind control, intelligent customer service, electronic commerce recommendation, medical diagnosis and the like. With the diversification of business demands and the acceleration of technology iteration, the number of models required to be maintained and managed by enterprises or organizations is rapidly increased, and various models are increasingly diversified in terms of structure, function and complexity. In order to realize efficient organization, accurate calling and performance tracing of mass models, the existing scheme needs automatic and intelligent classification management of the models. Conventional automated classification methods of models typically rely on a single or limited basis of classification, e.g., based solely on the name of the model, time of creation, or a few surface parameters, such as input-output dimensions. However, because different models may be similar in surface parameters, but there are significant differences among complex features such as network structures, layers, node connection modes (serial connection and parallel connection) and the like in the models, the essential differences of the models cannot be deeply and accurately reflected only according to simple parameters. This results in coarse classification results, insufficient accuracy, difficulty in supporting the requirements of fine model management and scenerising calls. For example, in an intelligent customer service system, a lightweight model for processing simple business queries and a deep model for conducting complex multi-round conversations, if classified by basic parameters only, may be classified into the same class, thereby affecting the accuracy and efficiency of the system in assigning appropriate models for different user needs. Therefore, how to realize a model classification scheme capable of deeply analyzing the complexity of the internal structure of the model and realizing high-precision and automatic classification based on the model classification scheme is a technical problem to be solved in the field. Disclosure of Invention According to the method, the accuracy and the robustness of model classification are remarkably improved by introducing an outlier filtering mechanism and a dynamic reference value calculation strategy. The embodiment of the application provides a model classification method, which is used for solving the problems that the existing model classification scheme cannot deeply analyze the internal structure of a model only according to the name, creation time or a few surface parameters of the model, so that classification results are coarse, the precision is insufficient, and the requirements of fine model management and scenerization calling are difficult to support. The embodiment of the application also provides a model classification device which is used for solving the problems that the existing model classification scheme cannot deeply analyze the internal structure of a model only according to the name, creation time or a few surface parameters of the model, so that classification results are coarse, precision is insufficient, and the requirements of fine model management and scenerization calling are difficult to support. The embodiment of the application also provides model classification equipment which is used for solving the problems that the existing model classification scheme cannot deeply analyze the internal structure of a model only according to the name, creation time or a few surface parameters of the model, so that classification results are coarse, precision is insufficient, and the requirements of fine model management and scenerization calling are difficult to support. The embodiment of the application also provides a computer readable storage medium which is used for solving the problems that the existing model classification scheme cannot deeply analyze the internal structure of the model only according to the name, creation time or a few surface parameters of the model, so that classification results are coarse, precision is insufficient, and the requirements of fine model management and scenerization calling are difficult to support. A computer program product is used for solving the problems that the existing model classification scheme cannot deeply analyze the internal structure of a model only according to the name, creation time or a few surface parameters of the model, so that classification results are coarse, the precision is insufficient, and the