Cost evaluation model generation method and device and cost evaluation method and device

Abstract

The invention discloses a cost evaluation model generation method and device and a cost evaluation method and device, and the method comprises the steps: carrying out the data enhancement and generation processing of data in each field in the cost evaluation of a government affair informatization project, and obtaining heterogeneous data in each field; performing data matching and data alignment on the heterogeneous data in each field, and performing data fusion by adopting a weighted average algorithm to generate a cost evaluation data set; decomposing data in the cost evaluation data set into shared feature data and specific feature data, and performing data generalization by using a feature generalization network; dynamically fusing the shared feature data and the specific feature data after data generalization to obtain a feature representation data set; and constructing and training a cost evaluation model of the government affair informatization project by adopting a pre-training method to obtain a trained cost evaluation model. When the model is used for cost evaluation, objectivity, accuracy and credibility of cost evaluation of the government affair informatization project can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of cost assessment, and in particular to a cost assessment model generation method and device, and a cost assessment method and device. Background Art

[0002] The cost assessment of government information technology projects (also known as government information technology project measurement or government information technology project cost) refers to the process of evaluating, calculating and controlling the costs of various resources, equipment, manpower and services required for the project during the implementation of the government information technology project. However, due to objective factors such as the personalized customization needs of government information technology project construction, the rapid product iteration speed and the lack of a unified and applicable reference basis, the current data collection for the cost assessment of government information technology projects is difficult (the business processes and needs of each government department are highly personalized, resulting in the need to customize the corresponding data collection plan for each project during the data collection stage), the small amount of data retention (new functions and systems are constantly launched, while old systems and data may be eliminated soon, and there is a lack of sufficient historical data as a reference during the assessment) and poor data quality (lack of unified standards and specifications, uneven data quality in different regions or departments, and difficulty in ensuring the accuracy and reliability of the data), which in turn leads to the lack of objectivity, accuracy and credibility of the cost assessment of government information technology projects. Summary of the invention

[0003] The present invention provides a cost assessment model generation method and device, as well as a cost assessment method and device, which utilize data generalization, fusion and heterogeneity to construct and train a cost assessment model for a government information project, and use the model to perform cost assessment to improve the problems of government information project cost assessment in terms of difficulty in data collection, data retention, poor data quality, etc., thereby improving the objectivity, accuracy and credibility of government information project cost assessment.

[0004] In order to achieve the above object, an embodiment of the present invention provides a method for generating a cost assessment model, comprising: Perform data enhancement and generation processing on the data in each field of the cost assessment of government information projects to obtain heterogeneous data in each field; After matching and aligning the heterogeneous data in each field, a weighted average algorithm is used to fuse the aligned heterogeneous data to generate a cost assessment data set for the government information project; Decomposing the data in the cost assessment data set into shared feature data and specific feature data; performing data generalization on the shared feature data and the specific feature data using a feature generalization network; The shared feature data and the specific feature data after data generalization are dynamically fused to obtain a feature representation data set; based on the feature representation data set, a pre-training method is used to construct and train a cost assessment model for the government informationization project to obtain a trained cost assessment model.

[0005] As an improvement of the above solution, the method of generalizing the shared feature data and the specific feature data using a feature generalization network includes: Performing data generalization on the shared feature data and the specific feature data using a feature generalization network; Wherein, the feature generalization network includes a feature network layer, a task network layer and a feedback network layer; Using the feature network layer to extract features from the shared feature data and the specific feature data to obtain common features; Based on the task network layer, the general features are mapped to corresponding government information projects to obtain predicted mapping category probability distribution, and the difference between the mapping category probability distribution and the real category of the government information project is measured by using a supervised classification loss function; In the feedback network layer, an auxiliary loss function is used to evaluate the common features and the mapping category probability distribution to update the parameters of the feature network layer and the task network layer.

[0006] As an improvement of the above solution, the data in the cost assessment data set is decomposed into shared feature data and specific feature data, including: After the data in the cost assessment dataset is segmented using a word segmenter and a BPE algorithm, feature extraction is performed using a Word2Vec model to obtain vector features of each word; Using the self-attention mechanism, the vector features are associated and calculated to generate a feature matrix; the words in the cost assessment data set are weighted and summed according to the feature matrix to obtain a self-attention vector; the self-attention vector is decomposed into shared features and specific features; The shared features corresponding to the shared features whose variance is smaller than a preset variance threshold are deleted, and the specific features corresponding to the specific features whose variance is smaller than a preset variance threshold are deleted to obtain shared feature data and specific feature data.

[0007] As an improvement of the above scheme, after data matching and data alignment are performed on the heterogeneous data in each field, a weighted average algorithm is used to perform data fusion on the aligned heterogeneous data to generate a cost assessment data set for the government informationization project, including: Use similarity and rule matching methods to match heterogeneous data in each field and filter out low-quality data generated; An embedding model is used to map the heterogeneous data in each matched field into the same low-dimensional vector space for data alignment; A weighted average algorithm is used to perform weighted average calculation on all aligned heterogeneous data to obtain fused data, and the fused data is smoothed and denoised to generate a cost assessment data set for the government informationization project.

[0008] As an improvement of the above solution, the shared feature data and the specific feature data after data generalization are dynamically fused to obtain a feature representation data set, including: Extracting common shared features from the shared feature data after data generalization, and obtaining a weight vector for each field according to the common shared features; Extract the feature representation of each field in the specific feature data after data generalization; fuse the weight vector and feature representation of each field to obtain a feature representation data set.

[0009] As an improvement of the above solution, the cost assessment model of the government information project is constructed and trained by a pre-training method according to the feature representation data set to obtain a trained cost assessment model, including: Using the feature representation data set as input data, and using a pre-training method to construct a cost assessment model for the government informationization project; Based on the cross entropy loss function and the SGF optimization algorithm, the cost evaluation model is iteratively trained using the feature representation data set to update the parameters of the cost evaluation model; Using the cost assessment data set to verify the updated cost assessment model, if the verification result does not achieve the expected effect, adjusting the model parameters and hyperparameters according to the verification result; Then, based on the cross entropy loss function and the SGF optimization algorithm, the feature representation data set is used to iteratively train the adjusted cost assessment model, the parameters of the adjusted cost assessment model are updated, and the updated cost assessment model is verified using the cost assessment data set until the verification result achieves the expected effect, thereby obtaining a trained cost assessment model.

[0010] As an improvement of the above scheme, the data in each field of the cost assessment of the government information project is enhanced and generated to obtain heterogeneous data in each field, including: Collect and organize various assessment data that affect the cost of government information projects, and classify the assessment data by category to obtain data for each field; Data enhancement processing is performed on the data in each field, and based on the context information of the data in each field and the dependencies between the data, data generation processing is performed on the data in each field in an adversarial manner to obtain heterogeneous data in each field.

[0011] In order to achieve the above-mentioned purpose, an embodiment of the present invention provides a cost assessment model generation device, including: The heterogeneous data acquisition module is used to perform data enhancement and generation processing on the data in each field of the cost assessment of the government information project to obtain heterogeneous data in each field; An evaluation data generation module is used to perform data matching and data alignment on the heterogeneous data in each field, and then use a weighted average algorithm to perform data fusion on the aligned heterogeneous data to generate a cost evaluation data set for the government informationization project; A feature data generalization module is used to decompose the data in the cost assessment data set into shared feature data and specific feature data; and generalize the shared feature data and the specific feature data using a feature generalization network; The evaluation model construction module is used to dynamically fuse the shared feature data and the specific feature data after data generalization to obtain a feature representation data set; based on the feature representation data set, a pre-training method is used to construct and train the cost evaluation model of the government informationization project to obtain a trained cost evaluation model.

[0012] In order to achieve the above object, an embodiment of the present invention further provides a cost assessment method, comprising: Input the evaluation data of the target government informationization project into the trained cost evaluation model, divide the evaluation data into fields and regions, determine the cost evaluation standard of the evaluation data, and output the cost evaluation result of the target government informationization project; Wherein, the trained cost assessment model is obtained by adopting the above-mentioned cost assessment model generation method.

[0013] In order to achieve the above-mentioned purpose, an embodiment of the present invention provides a cost assessment device, including: An evaluation result acquisition module is used to input the evaluation data of the target government informationization project into the trained cost evaluation model, divide the evaluation data into fields and regions, determine the cost evaluation standard of the evaluation data, and output the cost evaluation result of the target government informationization project; Wherein, the trained cost assessment model is obtained by adopting the above-mentioned cost assessment model generation method.

[0014] Compared with the prior art, the embodiments of the present invention disclose a method and device for generating a cost assessment model and a method and device for generating a cost assessment model. The methods and devices enhance and generate data in each field of the cost assessment of a government information project to obtain heterogeneous data in each field. After data matching and data alignment are performed on the heterogeneous data in each field, a weighted average algorithm is used to fuse the aligned heterogeneous data to generate a cost assessment data set for the government information project. The data in the cost assessment data set is decomposed into shared feature data and specific feature data. A feature generalization network is used to generalize the shared feature data and the specific feature data. The shared feature data and the specific feature data after data generalization are dynamically fused to obtain a feature representation data set. According to the feature representation data set, a pre-training method is used to construct and train a cost assessment model for the government information project to obtain a trained cost assessment model. Using a trained cost assessment model to conduct cost assessment based on the assessment data of the target government information technology project can improve the problems of government information technology project cost assessment in terms of difficulty in data collection, data retention, and poor data quality, improve the objectivity, accuracy, and credibility of government information technology project cost assessment, and solve the problem that some regions are unable to conduct effective assessment of government information technology projects due to the lack of current reference basis. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of a method for generating a cost assessment model provided by an embodiment of the present invention; Figure 2 is a flowchart of an evaluation data processing process provided by an embodiment of the present invention; Figure 3 It is a structural schematic diagram of a cost assessment model generating device provided by an embodiment of the present invention; Figure 4 It is a flow chart of a cost assessment method provided by an embodiment of the present invention; Figure 5 It is a structural schematic diagram of a cost assessment device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] It should be noted that the terms "comprises" and "specifically" and any variations of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or inherent to these processes, methods, products or apparatuses.

[0018] See also Figure 1 , Figure 1 The present invention provides a flow chart of a method for generating a cost assessment model, which includes: S1, perform data enhancement and generation processing on the data in each field of the cost assessment of the government information project to obtain heterogeneous data in each field; S2, after performing data matching and data alignment on the heterogeneous data in each field, a weighted average algorithm is used to perform data fusion on the aligned heterogeneous data to generate a cost assessment data set for the government informationization project; S3, decomposing the data in the cost assessment data set into shared feature data and specific feature data; performing data generalization on the shared feature data and the specific feature data using a feature generalization network; S4, dynamically fuse the shared feature data and the specific feature data after data generalization to obtain a feature representation data set; based on the feature representation data set, use a pre-training method to build and train a cost assessment model for the government informationization project to obtain a trained cost assessment model.

[0019] Exemplarily, various assessment data that affect the cost of government information technology projects are collected and sorted, and data enhancement and generation are performed on the cost assessment data of government information technology projects in different fields (development, procurement, service and basis), so as to obtain high-quality and diverse data, and then expand the data set; then the data is matched and aligned, and the heterogeneous data of government information technology project cost assessment from different fields, formats and types are integrated by knowledge fusion to form a comprehensive and complete cost assessment data set for government information technology projects; with the help of feature decoupling, the features in the data set are decomposed into shared features and specific features, and the feature generalization network is used to generalize the shared features and specific features of the heterogeneous fusion data set, so that the data has generalization capabilities for the field of government information technology projects; finally, all shared features and different features are integrated by means of ensemble learning, and a pre-training method is used to construct and train the cost assessment model of the government information technology project, so that the model has generalization recognition capabilities, and ultimately improves the objectivity, accuracy and credibility of the current government information technology project cost assessment. And it will solve the problem that some regions are unable to effectively evaluate government information technology projects due to the lack of current reference basis.

[0020] Specifically, in step S3, generalizing the shared feature data and the specific feature data using a feature generalization network includes: Performing data generalization on the shared feature data and the specific feature data using a feature generalization network; Wherein, the feature generalization network includes a feature network layer, a task network layer and a feedback network layer; Using the feature network layer to extract features from the shared feature data and the specific feature data to obtain common features; Based on the task network layer, the general features are mapped to corresponding government information projects to obtain predicted mapping category probability distribution, and the difference between the mapping category probability distribution and the real category of the government information project is measured by using a supervised classification loss function; In the feedback network layer, an auxiliary loss function is used to evaluate the common features and the mapping category probability distribution to update the parameters of the feature network layer and the task network layer.

[0021] Exemplarily, the generalization capability of shared feature data and specific feature data to unknown targets or tasks is realized by constructing a feature generalization network, wherein the feature generalization network consists of three parts, including a feature network layer, which is responsible for extracting features from input data and using these features for subsequent tasks; a task network layer, which generalizes data based on the features extracted by the feature network. It provides a supervised classification loss to guide the training of the feature network; a feedback network layer, which learns an auxiliary loss function, which is designed to encourage the feedback network to extract enhanced losses and provide additional feedback to the feature network, thereby supporting more complex domain generalization tasks. For example, a pre-trained CNN is used as the input of the feature generalization network, and the first four convolution blocks are frozen, and only the remaining convolution blocks and the average pooling layer are updated to improve computational efficiency. This module is mainly responsible for extracting common features from the shared features and specific features of the input so that they can be applied to different government information projects. One or more fully connected layers are connected after the feature network layer as a task network layer, which is used to map the output of the feature network layer to a specific category of government information projects. At the same time, in the mapping process, a supervised classification loss function is set to measure the difference loss between the output and the actual government informationization project. By minimizing the difference loss, the accuracy of data mapping to the correct category can be improved. That is, for each training sample, first extract the common features from the shared features and the specific features in the feature network layer, and then map these features to the corresponding government informationization project category in the task network layer, and output the predicted probability probability, while the supervised classification loss is used to measure the difference between the predicted probability distribution and the real category. The feedback network layer uses a multi-layer perceptron to introduce an auxiliary evaluation loss function to evaluate the quality of the output features of the feature network layer and the task network layer, and then optimize the generalization ability of the features through continuous iteration. Specifically, the features processed by the feature network layer and the task network layer can be evaluated by comparing the use and non-use of the auxiliary judgment loss function, so as to generate a meta-loss, and then the meta-loss is used to update the parameters of the feature network layer and the task network layer, thereby guiding the feature network layer and the task network layer to generate features that are more conducive to generalization.

[0022] It is worth mentioning that domain generalization is an important concept in machine learning and computer vision. It refers to the ability of a model to successfully apply knowledge or patterns learned from one or more source domains to target domains that have not been seen during training, even if there are distribution differences between these domains. Simply put, domain generalization is to hope that the model can "learn from one example and apply it to other cases", not only limited to performing well in a specific data set or specific environment, but also able to maintain stable performance across different environments or data sets.

[0023] Specifically, in step S3, decomposing the data in the cost assessment data set into shared feature data and specific feature data includes: After the data in the cost assessment dataset is segmented using a word segmenter and a BPE algorithm, feature extraction is performed using a Word2Vec model to obtain vector features of each word; Using the self-attention mechanism, the vector features are associated and calculated to generate a feature matrix; the words in the cost assessment data set are weighted and summed according to the feature matrix to obtain a self-attention vector; the self-attention vector is decomposed into shared features and specific features; The shared features corresponding to the shared features whose variance is smaller than a preset variance threshold are deleted, and the specific features corresponding to the specific features whose variance is smaller than a preset variance threshold are deleted to obtain shared feature data and specific feature data.

[0024] Exemplarily, by learning a function, the feature vector is mapped into a set of parts with different functions, so that each part only contains information related to itself, that is, the data features in the cost assessment data set are decomposed into shared features and specific features. For example, SentencePiece is used as a customized word segmenter for feature extraction to segment the text, and the byte-level Byte Pair Encoding (BPE) algorithm is used to ensure that the information after segmentation is not lost. The Word2Vec model is used to extract features from the cost assessment data set to obtain the vector features of each word. The self-attention mechanism is used to process the extracted vector features, so that the data features in the field of government informationization projects are associated and calculated, and then a feature matrix is ​​generated. According to the feature matrix, the vocabulary in the cost assessment data set is weighted and summed to generate a self-attention vector, which contains all the information in the field of government informationization projects, that is, the data in the field of government informationization projects after heterogeneous fusion is decomposed into multiple sub-feature vectors. Finally, the UndoBias method is used to learn a universal weight for each field of all government informationization projects. And the incremental weight for each field based on this , then the domain weight of each domain is expressed as , by using the maximum margin classification to learn the combination of the above weights, the shared features and specific features can be decomposed. Screening is performed by calculating the variance of each feature. Features with too small variances indicate that the samples have basically no difference in this feature, and most of the values ​​may be the same or the values ​​of the entire feature are the same. Such features have little effect on subsequent steps. The embodiment of the present invention adopts the VarianceThreshold method, by setting a threshold, all features that are smaller than the threshold are deleted.

[0025] Specifically, the step S2 includes: S21, uses similarity and rule matching methods to match heterogeneous data in each field and filters out low-quality data generated; S22, uses an embedding model to map the heterogeneous data in each matched field into the same low-dimensional vector space for data alignment; S23, using a weighted average algorithm to perform weighted average calculation on all aligned heterogeneous data to obtain fused data, and performing smoothing and denoising on the fused data to generate a cost assessment data set for the government informationization project.

[0026] Exemplarily, data matching and data alignment are performed on the heterogeneous data in each field to ensure that they can be correctly associated and integrated during the fusion process. Then, the heterogeneous data in different fields are averaged and weighted by the weighted average algorithm to achieve the fusion of heterogeneous data in each field and form a comprehensive and complete cost assessment data set for government information projects. For example, on the one hand, a similarity-based matching method (cosine similarity) is adopted, that is, the semantic similarity of the text is compared by measuring the cosine value of the angle between two text vectors in multidimensional space, thereby realizing data alignment in the field. On the other hand, a rule-based matching method is adopted, that is, data is identified, classified and filtered through a pre-defined set of rules. Data matching is performed on the evaluation basis in the data set, such as management methods, implementation details and other file data, and low-quality data generated by data operations are filtered. Data alignment aims to correspond and integrate the data in the field after data operations to ensure data consistency and accuracy. This embodiment uses the TransE model to train the heterogeneous data in each field after data matching, so that the heterogeneous data in each field are mapped to the same low-dimensional vector space, and then a similarity threshold is set (which can be set as needed). When the similarity of the data exceeds the similarity threshold, it can be considered that they may represent the same object or content, thereby achieving data alignment in each field. This embodiment uses a weighted average algorithm to fuse the data, that is, by weighted averaging the data in different data sources or data sets according to certain weights, so as to obtain the fused data results. First, an adaptive algorithm is used to dynamically adjust the weights. This method can adjust the weights in real time according to the quality of the data itself or the fusion effect to achieve a better fusion effect. Secondly, a weighted average calculation is performed on all data sets in each field, and the formula is:

[0027] in, It is The weight of the data source, It is The data value of the data source, is the number of data sources, and the fused data results are obtained. Finally, the fused data is smoothed (noise and outliers in the fusion results are removed) and denoised (the noise level in the fusion results is further reduced through filtering technology) to improve the availability and accuracy of the data, and to integrate heterogeneous data from different fields to form a comprehensive and complete cost assessment data set for government information projects.

[0028] Specifically, in step S4, the shared feature data and the specific feature data after data generalization are dynamically fused to obtain a feature representation data set, including: Extracting common shared features from the shared feature data after data generalization, and obtaining a weight vector for each field according to the common shared features; Extract the feature representation of each field in the specific feature data after data generalization; fuse the weight vector and feature representation of each field to obtain a feature representation data set.

[0029] Exemplarily, by designing a reasonable feature layer and selecting a suitable extractor, a feature fusion integrated learning network is constructed; the integrated learning network used in this embodiment includes the following three parts: a shared CNN feature extractor is responsible for extracting common features from the input data; a specific feature extraction layer, for each source domain, there is a corresponding feature extraction layer, which is responsible for feature representation for unseen targets; a domain weight branch, which accepts the output of the CNN feature extractor and learns a weight vector , used to measure the contribution of each source domain target feature representation; using the weight output by the domain weight branch , the features from different specific branches are integrated. Specifically, each domain feature is multiplied by the corresponding weight and then aggregated (such as weighted average). Because it is based on the learned weights of shared features and specific features, it can automatically adjust the importance of features of different types of government information project data, so it can construct feature representations for unknown targets.

[0030] Specifically, in step S4, the cost assessment model of the government information project is constructed and trained by a pre-training method according to the feature representation data set to obtain a trained cost assessment model, including: Using the feature representation data set as input data, and using a pre-training method to construct a cost assessment model for the government informationization project; Based on the cross entropy loss function and the SGF optimization algorithm, the cost evaluation model is iteratively trained using the feature representation data set to update the parameters of the cost evaluation model; Using the cost assessment data set to verify the updated cost assessment model, if the verification result does not achieve the expected effect, adjusting the model parameters and hyperparameters according to the verification result; Then, based on the cross entropy loss function and the SGF optimization algorithm, the feature representation data set is used to iteratively train the adjusted cost assessment model, the parameters of the adjusted cost assessment model are updated, and the updated cost assessment model is verified using the cost assessment data set until the verification result achieves the expected effect, thereby obtaining a trained cost assessment model.

[0031] Exemplarily, the fused feature representation is used as input data and input into the Transformer architecture, and a feature fusion layer and an embedding layer are set therein to combine the received fused features with the output of the embedding layer; the cross-entropy loss function is selected as the loss function to measure the difference between the content generated by the model and the real content; then the cost assessment dataset is used for training, the gradient is calculated by the back propagation algorithm, and the SGF optimization algorithm is used to update the model parameters to minimize the loss function; finally, the heterogeneous data is used as the validation set to verify the model, and the model parameters and hyperparameters are adjusted according to the verification results, and the training and optimization process is iterated until the content generated by the model reaches the expected result.

[0032] In the embodiment of the present invention, the shared features and specific feature data after feature generalization are dynamically integrated by designing a network structure and adopting an integrated learning method, and the data set is applied to the training of the model to improve the adaptive evaluation capability of the model, so as to realize effective evaluation of the cost evaluation of government informationization projects when data collection is difficult, data retention and data quality are poor.

[0033] Specifically, the step S1 includes: S11, collect and organize various assessment data that affect the cost of government information projects, and classify the assessment data by category to obtain data for each field; S12, performing data enhancement processing on the data in each field, and performing data generation processing on the data in each field in an adversarial manner according to the context information of the data in each field and the dependency relationship between the data, to obtain heterogeneous data in each field.

[0034] For example, Figure 2 As shown, Figure 2It is a flowchart of an evaluation data processing process provided by an embodiment of the present invention, which collects and organizes various types of evaluation data (development, procurement, service, and basis) that affect the cost of government information projects, and classifies the evaluation data by category to obtain data in each field. The evaluation data used in this embodiment is the main factor affecting the cost evaluation of government information projects, including the content of information projects (software development, system operation and maintenance, project operation, hardware acquisition, customized equipment, etc.), the cost guidance of e-government projects in various provinces and cities, the wage guidance price of the human resources market in various regions over the years, the cost evaluation method, etc., and the existing government information project cost evaluation data is divided into fields. Government information projects usually cover multiple fields, such as e-government platform construction, smart city systems, public security management systems, etc. Each field has its specific project characteristics, cost structure and evaluation standards. Through field division, the data in these different fields can be classified so that data processing and analysis can be carried out more targetedly later. Data enhancement is performed on the data of each field to improve the randomness and variation of the field data; TF-IDF is used to extract keywords or key phrases (which may represent the theme or context of the data set) from the data set, and then the context information obtained is converted into a numerical vector through Word2Vec and added back to the original data set; GAN (Generative Adversarial Networks) is a deep learning model consisting of two neural networks - the generator and the discriminator. The two networks compete with each other during the training process, forming a competitive relationship to improve the quality of the generated data. In this example, SeqGAN is selected as the model for data generation. First, the enhanced data set is input into the generator, and the generator will generate simulated data (fake samples) for the adversarial network based on the input data; then the data set and the simulated data are input into the discriminator together, and the maximum and minimum alternating optimization strategy is used to iterate SqeGAN, so that the generator and the discriminator compete with each other, and finally a high-quality data set is generated.

[0035] More specifically, the data enhancement processing for each field includes: The data in each field are augmented by text simplification, text paraphrase and noise injection.

[0036] Exemplarily, the source data is processed by some open source text simplification models and manual editing methods, the original complex text is simplified into a more understandable version, and then added to the data set to increase the dimension of the data; by re-stating the content of the original text, a new text instance is generated, thereby increasing the corpus size and implicit information of a specific scene without adding additional information, thereby enriching the connotation and extension of the data set; noise injection, as a data enhancement method, has significant effects in enhancing the robustness of the model, preventing overfitting, improving generalization ability and improving model stability. In this embodiment, Gaussian noise is selected as the injected noise, and a random noise matrix matching the shape of the data set is generated by a tool (such as the random number generation library numpy in Python), and the generated noise matrix is ​​added to the data set.

[0037] The embodiment of the present invention discloses a method for generating a cost assessment model. The method performs data enhancement and generation processing on data in each field in the cost assessment of a government information project to obtain heterogeneous data in each field. After data matching and data alignment are performed on the heterogeneous data in each field, a weighted average algorithm is used to fuse the aligned heterogeneous data to generate a cost assessment data set for the government information project. The data in the cost assessment data set is decomposed into shared feature data and specific feature data. A feature generalization network is used to generalize the shared feature data and the specific feature data. The shared feature data and the specific feature data after data generalization are dynamically fused to obtain a feature representation data set. According to the feature representation data set, a pre-training method is used to construct and train a cost assessment model for the government information project to obtain a trained cost assessment model. Using a trained cost assessment model to conduct cost assessment based on the assessment data of the target government information technology project can improve the problems of government information technology project cost assessment in terms of difficulty in data collection, data retention, and poor data quality, improve the objectivity, accuracy, and credibility of government information technology project cost assessment, and solve the problem that some regions are unable to conduct effective assessment of government information technology projects due to the lack of current reference basis.

[0038] Figure 3 1 is a schematic diagram of the structure of a cost assessment model generation device provided by an embodiment of the present invention. The cost assessment model generation device 10 includes: The heterogeneous data acquisition module 11 is used to perform data enhancement and generation processing on the data in each field of the cost assessment of the government information project to obtain heterogeneous data in each field; An evaluation data generation module 12 is used to perform data matching and data alignment on the heterogeneous data in each field, and then use a weighted average algorithm to perform data fusion on the aligned heterogeneous data to generate a cost evaluation data set for the government informationization project; The feature data generalization module 13 is used to decompose the data in the cost assessment data set into shared feature data and specific feature data; and generalize the shared feature data and the specific feature data using a feature generalization network; The evaluation model construction module 14 is used to dynamically fuse the shared feature data and the specific feature data after data generalization to obtain a feature representation data set; based on the feature representation data set, a pre-training method is used to construct and train the cost evaluation model of the government information project to obtain a trained cost evaluation model.

[0039] Specifically, the using a feature generalization network to perform data generalization on the shared feature data and the specific feature data includes: The shared feature data and the specific feature data are generalized using a feature generalization network. Wherein, the feature generalization network includes a feature network layer, a task network layer and a feedback network layer; Using the feature network layer to extract features from the shared feature data and the specific feature data to obtain common features; Based on the task network layer, the general features are mapped to corresponding government information projects to obtain predicted mapping category probability distribution, and the difference between the mapping category probability distribution and the real category of the government information project is measured by using a supervised classification loss function; In the feedback network layer, an auxiliary loss function is used to evaluate the common features and the mapping category probability distribution to update the parameters of the feature network layer and the task network layer.

[0040] Specifically, decomposing the data in the cost assessment data set into shared feature data and specific feature data includes: After the data in the cost assessment dataset is segmented using a word segmenter and a BPE algorithm, feature extraction is performed using a Word2Vec model to obtain vector features of each word; Using the self-attention mechanism, the vector features are associated and calculated to generate a feature matrix; the words in the cost assessment data set are weighted and summed according to the feature matrix to obtain a self-attention vector; the self-attention vector is decomposed into shared features and specific features; The shared features corresponding to the shared features whose variance is smaller than a preset variance threshold are deleted, and the specific features corresponding to the specific features whose variance is smaller than a preset variance threshold are deleted to obtain shared feature data and specific feature data.

[0041] Specifically, the evaluation data generating module 12 is used to: Use similarity and rule matching methods to match heterogeneous data in each field and filter out low-quality data generated; An embedding model is used to map the heterogeneous data in each matched field into the same low-dimensional vector space for data alignment; A weighted average algorithm is used to perform weighted average calculation on all aligned heterogeneous data to obtain fused data, and the fused data is smoothed and denoised to generate a cost assessment data set for the government informationization project.

[0042] Specifically, the shared feature data and the specific feature data after data generalization are dynamically merged to obtain a feature representation data set, including: Extracting common shared features from the shared feature data after data generalization, and obtaining a weight vector for each field according to the common shared features; Extract the feature representation of each field in the specific feature data after data generalization; fuse the weight vector and feature representation of each field to obtain a feature representation data set.

[0043] Specifically, the cost assessment model of the government information project is constructed and trained by a pre-training method according to the feature representation data set to obtain a trained cost assessment model, including: Using the feature representation data set as input data, and using a pre-training method to construct a cost assessment model for the government informationization project; Based on the cross entropy loss function and the SGF optimization algorithm, the cost evaluation model is iteratively trained using the feature representation data set to update the parameters of the cost evaluation model; Using the cost assessment data set to verify the updated cost assessment model, if the verification result does not achieve the expected effect, adjusting the model parameters and hyperparameters according to the verification result; Then, based on the cross entropy loss function and the SGF optimization algorithm, the feature representation data set is used to iteratively train the adjusted cost assessment model, the parameters of the adjusted cost assessment model are updated, and the updated cost assessment model is verified using the cost assessment data set until the verification result achieves the expected effect, thereby obtaining a trained cost assessment model.

[0044] Specifically, the heterogeneous data acquisition module 11 is used to: Collect and organize various assessment data that affect the cost of government information projects, and classify the assessment data by category to obtain data for each field; Data enhancement processing is performed on the data in each field, and based on the context information of the data in each field and the dependencies between the data, data generation processing is performed on the data in each field in an adversarial manner to obtain heterogeneous data in each field.

[0045] A cost assessment model generating device provided by an embodiment of the present invention can implement all processes of the cost assessment model generating method of the above embodiment. The functions of each module in the device and the technical effects achieved are respectively the same as the functions and technical effects achieved by the cost assessment model generating method of the above embodiment, and will not be repeated here.

[0046] Figure 4 1 is a flow chart of a cost assessment method provided by an embodiment of the present invention, the cost assessment method comprising: S101, inputting the evaluation data of the target government informationization project into the trained cost evaluation model, dividing the evaluation data into fields and regions, determining the cost evaluation standard of the evaluation data, and outputting the cost evaluation result of the target government informationization project; The trained cost assessment model is obtained by using the cost assessment model generation method as described above.

[0047] For example, the trained cost assessment model has integrated the generalization ability of government information project data, so it can conduct cost assessment of government information projects when data collection is difficult, data retention is poor, and data quality is poor. The model is fed with the content of a city's information project, including software, operation and maintenance, operations, and personalized equipment, and the model is asked to conduct cost assessment. First, since the city has not set its own cost assessment standards, the model will refer to the cost standards of other cities in the province where the city is located to generalize the city's cost standards and conduct assessment based on them. Next, the software part will be evaluated based on the estimation method (prediction method or estimation method) adopted for the city's past software costs. If there is no past data for the city, it will be generalized to the estimation method commonly used in the province for evaluation; the cost of operation and maintenance will be evaluated based on the city's human resources market wage guidance price of the current year or the previous year and combined with the calculation method of the labor rate in the e-government engineering cost guidance book. If there is a lack of the city's human resources market wage guidance price in recent years, it will be generalized based on the fluctuation range of the wage guidance price of the position in the province in recent years to obtain the cost; the hardware cost will be evaluated based on the similarity of the equipment parameters, because the equipment models of different manufacturers are different, but the effectiveness of their equipment can be compared by comparing parameters, and the model will also give an assessment of the hardware costs according to the degree of parameter similarity; and for some personalized equipment, the model will be evaluated based on the content of some of its modules combined with equipment with similar functions in other provinces and cities, and then a complete and appropriate information project cost evaluation report will be obtained.

[0048] A cost assessment method disclosed in an embodiment of the present invention inputs the assessment data of a target government informationization project into a trained cost assessment model, divides the assessment data into fields and regions, determines the cost assessment standard of the assessment data, and outputs the cost assessment result of the target government informationization project; wherein the trained cost assessment model is obtained using the cost assessment model generation method as described above. Using a trained cost assessment model to perform cost assessment based on the assessment data of the target government informationization project can improve the problems of difficulty in data collection, data retention, and poor data quality in the cost assessment of government informationization projects, improve the objectivity, accuracy, and credibility of the cost assessment of government informationization projects, and solve the problem that some regions cannot effectively evaluate government informationization projects due to the lack of current reference basis.

[0049] Figure 5 1 is a schematic diagram of the structure of a cost assessment device provided by an embodiment of the present invention. The cost assessment device 100 includes: The evaluation result acquisition module 101 is used to input the evaluation data of the target government informationization project into the trained cost evaluation model, divide the evaluation data into fields and regions, determine the cost evaluation standard of the evaluation data, and output the cost evaluation result of the target government informationization project; The trained cost assessment model is obtained by using the cost assessment model generation method as described above.

[0050] A cost assessment device provided by an embodiment of the present invention can implement all processes of the cost assessment model generation method of the above-mentioned embodiment. The functions of each module in the device and the technical effects achieved are respectively the same as the functions and technical effects achieved by the cost assessment method of the above-mentioned embodiment, and will not be repeated here.

[0051] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for generating a cost assessment model, characterized in that: include: Perform data enhancement and generation processing on the data in each field of the cost assessment of government information projects to obtain heterogeneous data in each field; After matching and aligning the heterogeneous data in each field, a weighted average algorithm is used to fuse the aligned heterogeneous data to generate a cost assessment data set for the government information project; Decomposing the data in the cost assessment data set into shared feature data and specific feature data; Performing data generalization on the shared feature data and the specific feature data using a feature generalization network; The shared feature data and the specific feature data after data generalization are dynamically fused to obtain a feature representation data set; based on the feature representation data set, a pre-training method is used to construct and train a cost assessment model for the government informationization project to obtain a trained cost assessment model.

2. The method for generating a cost assessment model according to claim 1, wherein: The method of using a feature generalization network to generalize the shared feature data and the specific feature data includes: Performing data generalization on the shared feature data and the specific feature data using a feature generalization network; Wherein, the feature generalization network includes a feature network layer, a task network layer and a feedback network layer; Using the feature network layer to extract features from the shared feature data and the specific feature data to obtain common features; Based on the task network layer, the general features are mapped to corresponding government information projects to obtain predicted mapping category probability distribution, and the difference between the mapping category probability distribution and the real category of the government information project is measured by using a supervised classification loss function; In the feedback network layer, an auxiliary loss function is used to evaluate the common features and the mapping category probability distribution to update the parameters of the feature network layer and the task network layer.

3. The method for generating a cost assessment model according to claim 1, wherein: Decomposing the data in the cost assessment data set into shared feature data and specific feature data includes: After the data in the cost assessment dataset is segmented using a word segmenter and a BPE algorithm, feature extraction is performed using a Word2Vec model to obtain vector features of each word; Using the self-attention mechanism, the vector features are associated and calculated to generate a feature matrix; the words in the cost assessment data set are weighted and summed according to the feature matrix to obtain a self-attention vector; the self-attention vector is decomposed into shared features and specific features; The shared features corresponding to the shared features whose variance is smaller than a preset variance threshold are deleted, and the specific features corresponding to the specific features whose variance is smaller than a preset variance threshold are deleted to obtain shared feature data and specific feature data.

4. The method for generating a cost assessment model according to claim 1, wherein: After data matching and data alignment are performed on the heterogeneous data in each field, a weighted average algorithm is used to perform data fusion on the aligned heterogeneous data to generate a cost assessment data set for the government informationization project, including: Use similarity and rule matching methods to match heterogeneous data in each field and filter out low-quality data generated; An embedding model is used to map the heterogeneous data in each matched field into the same low-dimensional vector space for data alignment; A weighted average algorithm is used to perform weighted average calculation on all aligned heterogeneous data to obtain fused data, and the fused data is smoothed and denoised to generate a cost assessment data set for the government informationization project.

5. The method for generating a cost assessment model according to claim 1, wherein: The shared feature data and the specific feature data after data generalization are dynamically merged to obtain a feature representation data set, including: Extracting common shared features from the shared feature data after data generalization, and obtaining a weight vector for each field according to the common shared features; Extract the feature representation of each field in the specific feature data after data generalization; fuse the weight vector and feature representation of each field to obtain a feature representation data set.

6. The method for generating a cost assessment model according to claim 1, wherein: The method of constructing and training the cost assessment model of the government information project using a pre-training method according to the feature representation data set to obtain a trained cost assessment model includes: Using the feature representation data set as input data, and using a pre-training method to construct a cost assessment model for the government informationization project; Based on the cross entropy loss function and the SGF optimization algorithm, the cost evaluation model is iteratively trained using the feature representation data set to update the parameters of the cost evaluation model; Using the cost assessment data set to verify the updated cost assessment model, if the verification result does not achieve the expected effect, adjusting the model parameters and hyperparameters according to the verification result; Then, based on the cross entropy loss function and the SGF optimization algorithm, the feature representation data set is used to iteratively train the adjusted cost assessment model, the parameters of the adjusted cost assessment model are updated, and the updated cost assessment model is verified using the cost assessment data set until the verification result achieves the expected effect, thereby obtaining a trained cost assessment model.

7. The method for generating a cost assessment model according to claim 1, wherein: The data in each field of the cost assessment of the government information project is enhanced and generated to obtain heterogeneous data in each field, including: Collect and organize various assessment data that affect the cost of government information projects, and classify the assessment data by category to obtain data for each field; Data enhancement processing is performed on the data in each field, and based on the context information of the data in each field and the dependencies between the data, data generation processing is performed on the data in each field in an adversarial manner to obtain heterogeneous data in each field.

8. A cost assessment model generation device, characterized in that: include: The heterogeneous data acquisition module is used to perform data enhancement and generation processing on the data in each field of the cost assessment of the government information project to obtain heterogeneous data in each field; An evaluation data generation module is used to perform data matching and data alignment on the heterogeneous data in each field, and then use a weighted average algorithm to perform data fusion on the aligned heterogeneous data to generate a cost evaluation data set for the government information project; A feature data generalization module, used for decomposing the data in the cost assessment data set into shared feature data and specific feature data; Performing data generalization on the shared feature data and the specific feature data using a feature generalization network; The evaluation model construction module is used to dynamically fuse the shared feature data and the specific feature data after data generalization to obtain a feature representation data set; based on the feature representation data set, a pre-training method is used to construct and train the cost evaluation model of the government informationization project to obtain a trained cost evaluation model.

9. A cost assessment method, characterized in that: include: Input the evaluation data of the target government informationization project into the trained cost evaluation model, divide the evaluation data into fields and regions, determine the cost evaluation standard of the evaluation data, and output the cost evaluation result of the target government informationization project; Wherein, the trained cost assessment model is obtained by using the cost assessment model generation method as described in any one of claims 1-7.

10. A cost assessment device, characterized in that: include: An evaluation result acquisition module is used to input the evaluation data of the target government informationization project into the trained cost evaluation model, divide the evaluation data into fields and regions, determine the cost evaluation standard of the evaluation data, and output the cost evaluation result of the target government informationization project; Wherein, the trained cost assessment model is obtained by using the cost assessment model generation method as described in any one of claims 1-7.

Images