CN-120706964-B - Information generation method and device

CN120706964BCN 120706964 BCN120706964 BCN 120706964BCN-120706964-B

Abstract

The invention discloses an information generation method and device, and relates to the technical field of information processing, wherein the method comprises the steps of responding to a user performance prediction request, and extracting working data of a user to be predicted from different data sources; the method comprises the steps of inputting working data of users to be predicted into a prediction model, wherein in the prediction model, a corresponding branch network is activated based on a position type corresponding to the users to be predicted, specified characteristics are determined by the branch network, the outputted specified characteristics are processed and then input into a shared coding network to extract general characteristics, prediction results under different performance indexes of all positions are output through an output head based on the general characteristics, scores under different performance indexes of the users at different positions are determined based on the prediction results, and a performance comprehensive value is determined based on the scores. And intelligent prediction of performance is realized.

Inventors

LIU LI
ZHANG XINJIE
ZHAO YAN
TANG MIAO
MA XUEJING

Assignees

兵器装备集团财务有限责任公司

Dates

Publication Date: 20260512
Application Date: 20250613

Claims (8)

1. An information generation method, comprising: extracting working data of a user to be predicted from different data sources; The working data of the user to be predicted is input into a prediction model, wherein in the prediction model, a corresponding branch network is activated based on the position type corresponding to the user to be predicted, and specified characteristics are determined by the branch network; based on the prediction result, determining scores of different performance indexes of users at different positions, and determining a performance comprehensive value based on the scores; When the output designated feature is processed, the method comprises processing the post type to obtain a post type embedded vector, wherein E j =e.onehot (j), E is a post embedded matrix, Splicing the appointed characteristic h branch and the post type embedded vector to obtain a spliced characteristic ; The step of outputting the prediction results under the different performance indexes of each post based on the general feature through the output head comprises the steps of splicing the splicing feature with the general feature to obtain a target splicing feature, and outputting the prediction results based on the target splicing feature , wherein, For the purpose of stitching the features, In order to activate the function, Is a bias term; In order for the dynamic weights to be given, , wherein, In order to share the basic weights of the coding network, To dynamically adjust tensors.
2. The information generating method according to claim 1, wherein when training the predictive model, the method includes collecting historical multi-source heterogeneous data of the user under different post types, wherein the different multi-source heterogeneous data are associated with the post types and are marked with different performance index information as training samples; determining a corresponding branch network based on the post type, taking a training sample associated with the post type as input of the corresponding branch network, and taking corresponding performance index information as output to train the prediction model; Wherein each branch network parameter is updated based on the loss function and the shared code network parameter is updated based on the loss function during training.
3. The information generating method according to claim 2, wherein the task-specific loss corresponding to the different post types j includes a classification loss when determining the loss function Regression loss ; Determining regularization loss based on parameters of the shared coding network; Determining a loss function based on the task-specific loss and the regularization loss: , wherein, , The weights of the regression task and the classification task are respectively.
4. The information generating method according to claim 3, wherein the prediction model activates a branch network based on a post type corresponding to a user to be predicted, and outputting the specified feature by the branch network comprises determining data types included in training samples corresponding to different post types; if the training samples corresponding to the same post type simultaneously contain the numerical value type and the text type, fusing the numerical value characteristics and the text characteristics extracted by each branch network to obtain the appointed characteristics: , as a numerical feature, h text is a text feature; , belonging to the weight corresponding to different post types, Attention weights corresponding to the numerical features, Attention weights corresponding to text features; Wherein, during training, the vector is embedded based on the post type And (3) adjusting the modal fusion weight: ; and determining fusion features, namely the designated features, based on the modal fusion weights.
5. The information generating method according to claim 4, wherein the prediction model activates a branch network based on a post type corresponding to a user to be predicted, and outputting a specified feature by the branch network comprises: And if the training samples corresponding to the same post type are of a numerical value type or a text type, a first branch network is activated for the training samples of which the data types are of a numerical value type, wherein the training samples are normalized through the first branch network and then are input into a full-connection layer to extract numerical characteristics, or a second branch network is activated for the training samples of which the data types are of a text type, wherein the training samples are extracted into CLS vectors through the second branch network and are input into the full-connection layer to perform dimension reduction processing, so that the text characteristics are obtained.
6. The information generating method according to claim 1, characterized in that the method further comprises: Generating an initial development path randomly based on the development action of each performance index in a development action library, wherein the development action comprises a development action name, required time, required cost, any performance index lifting amplitude after the development action is completed, and coverage of the development action on each requirement of an enterprise; and screening the initial development paths based on the objective function to obtain development paths corresponding to different users in different posts, wherein each development path comprises three stages of short-term, medium-term and long-term, and each stage comprises m development actions.
7. The information generating method as defined in claim 6, wherein in constructing the objective function, the method includes determining a difference between a score value at any one of different performance indicators of any one of the stations and a target threshold value of each of the performance indicators corresponding to the any one of the stations, calculating a corresponding short-plate weight based on the difference , wherein, Difference value corresponding to each performance index/total difference value corresponding to all performance indexes; based on short plate weights Any performance index improvement amplitude after development action is completed Determining a first objective function ; Set j weights of each requirement of enterprise Coverage of development actions in development path to enterprise requirements j Determining a second objective function Wherein the sum of the contributions of all development actions in the development path to the enterprise requirement j is coverage ; A third objective function is determined based on the total time cost T required by the user to complete all of the development actions in the development path and the direct funds cost K paid by the company for the user development path, , wherein, , Is a cost coefficient; And optimizing the objective function consisting of the first objective function, the second objective function and the third objective function based on the preset total time and total cost as constraint conditions.
8. An intelligent prediction device for user performance, comprising: the data acquisition unit is used for extracting working data of a user to be predicted from different data sources; The prediction unit inputs the working data of the user to be predicted into a prediction model, wherein in the prediction model, a corresponding branch network is activated based on a position type corresponding to the user to be predicted, and specified characteristics are determined by the branch network; The output unit is used for determining scores of different post users under different performance indexes based on the prediction result and determining a performance comprehensive value based on the scores; When the output designated feature is processed, the method comprises processing the post type to obtain a post type embedded vector, wherein E j =e.onehot (j), E is a post embedded matrix, Splicing the appointed characteristic h branch and the post type embedded vector to obtain a spliced characteristic ; The step of outputting the prediction results under the different performance indexes of each post based on the general feature through the output head comprises the steps of splicing the splicing feature with the general feature to obtain a target splicing feature, and outputting the prediction results based on the target splicing feature , wherein, For the purpose of stitching the features, In order to activate the function, Is a bias term; In order for the dynamic weights to be given, , wherein, In order to share the basic weights of the coding network, To dynamically adjust tensors.

Description

Information generation method and device Technical Field The present invention relates to the field of information processing technologies, and in particular, to an information generating method and apparatus. Background The calculation of user performance is usually performed periodically according to performance criteria, for example, different enterprises set up performance assessment rules for different users, and then performance calculation is performed manually on performance calculation days. The method cannot judge the performance of the user in advance, so that the user cannot be intervened in advance, and therefore, the intelligent prediction of the performance of the user is a technical problem to be solved. Disclosure of Invention The invention mainly aims to provide a data processing method and device for solving the defects in the related art. In order to achieve the above object, according to a first aspect of the present invention, there is provided an information generating method, including extracting work data of a user to be predicted from different data sources, inputting the work data of the user to be predicted into a prediction model, wherein in the prediction model, a corresponding branch network is activated based on a post type corresponding to the user to be predicted, and a specified feature is determined by the branch network, the outputted specified feature is processed and then input into a shared coding network to extract a general feature, prediction results under different performance indexes of different posts are outputted based on the general feature through an output head, and scores under different performance indexes of different posts are determined based on the prediction results, and a performance integrated value is determined based on the scores. Optionally, outputting the prediction result under different performance indexes of each post based on the general feature through an output head comprises processing a post type to obtain a post type embedded vector, wherein E j =e.onehot (j), E is a post embedded matrix, j is {1,2, and N j } is a post type tag, and splicing the specified feature h branch with the post type embedded vector to obtain a spliced feature h cond＝[hbranch;ej ]. Optionally, outputting the prediction results under the different performance indexes of each post based on the general feature through the output head comprises splicing the appointed feature with the general feature to obtain a target spliced feature, and outputting the prediction results based on the target spliced featureWhere h is the target feature, σ=relu is the activation function,And W is a dynamic weight, W=W base+ΔW·ej, wherein W base is a basic weight of the shared coding network, and DeltaW is a dynamic adjustment tensor. Optionally, when the prediction model is trained, the method comprises the steps of collecting historical multi-source heterogeneous data of a user under different post types, wherein the different multi-source heterogeneous data are associated with the post types and are marked with different performance index information as training samples, determining a branch network corresponding to the data based on the post types, taking the training samples associated with the post types as input of the corresponding branch network, and taking the corresponding performance index information as output to train the prediction model, and updating each branch network parameter based on a loss function and updating the shared coding network parameter based on the loss function during training. Optionally, in determining the penalty function, the task-specific penalty corresponding to the different job type j includes a classification penaltyRegression lossDetermining a regularization loss based on parameters of a shared coding network, determining a loss function based on the task total loss and the regularization loss: wherein, alpha and beta are the weights of regression task and classification task respectively. Optionally, the prediction model activates a branch network based on the post type corresponding to the user to be predicted, the branch network outputs specified features including determining data types contained in training samples corresponding to different post types, if the training samples corresponding to the same post type contain numerical value types and text types at the same time, fusing the numerical value features and the text features extracted by each branch network to obtain specified features, wherein h fused＝wnum·hnum+wtext·htext,hnum is the numerical value features, h text is the text features, w num,wtext belongs to weights corresponding to different post types, w num is the attention weights corresponding to the numerical value features, w text is the attention weights corresponding to the text features, and during training, adjusting the modal fusion weights based on the post type embedded vector e j, and determining the fusion fea