Visual modeling method based on intelligent semantic understanding

Abstract

The present invention relates to a visual modeling method based on intelligent semantic understanding, comprising the following steps: a user inputs a modeling task goal, performs word segmentation processing, and obtains a word segmentation result; a user inputs a source data set, and obtains the characteristics of the source data set; a test model range is determined based on the word segmentation result; m test data are selected from the source data set to test the test model range, and a model test result is obtained; n preset algorithm models are selected from the preset algorithm models as alternative solutions, and corresponding n groups of algorithm model hyperparameters are determined; the n preset algorithm models and their corresponding n groups of algorithm model hyperparameters and the characteristics of the source data set are subjected to neural network learning; after learning, the n preset algorithm models are sorted and recommended according to model evaluation indicators. The present invention provides a form of recommended algorithm models in the form of intelligent recommendation for the modeling task goal given by the user, and provides users with a variety of model solutions that can directly solve their task goals.

Description

Technical Field

[0001] The present invention relates to the technical fields of intelligent recommendation and artificial visual modeling, and in particular to a visual modeling method based on intelligent semantic understanding. Background Art

[0002] With the advent of the cloud era, big data has attracted increasing attention. Big data generally describes the vast amounts of unstructured and semi-structured data generated by a company, which would be prohibitively time-consuming and expensive to download into relational databases for analysis. To unlock the deeper value of massive amounts of enterprise data, machine learning models are being applied to various industries and intelligent systems.

[0003] Currently, machine learning still has high theoretical and programming barriers for users. A growing number of visual modeling platforms are attempting to lower the barrier to entry for using machine learning algorithms and models. For example, drag-and-drop visual modeling methods, while lowering the programming barrier, still impose strong theoretical requirements. The reasons are: ① Visual modeling platforms encapsulate algorithms and models into readily usable components at the smallest granularity of algorithm and model functionality. This implicitly requires users to master machine learning theory and understand the data processing pipeline, including the need for different data preprocessing and cleaning techniques for different types of data. Users also need to have a solid understanding of feature processing engineering and the strengths and weaknesses of various neural networks. ② Building machine learning models serves specific application scenarios and achieves specific goals. While component-based visual modeling methods may seem to lower the barrier to entry, the interconnectedness of data processing and component execution flows after dragging and dropping results in a longer debugging cycle for models built through visualization. ③ There are numerous types and variations of neural networks, and different neural network combinations and network parameters directly impact the generation of the optimal model. Even with visual modeling methods, extensive experimentation is required to verify whether the modeling objectives are met. Summary of the Invention

[0004] The purpose of the present invention is to provide a variety of model solutions that can directly solve the task goals of users in the form of intelligent recommendations and recommendation algorithm models for the modeling task goals given by users, so that users can model the set task goals and then analyze and sort out the big data to finally obtain the task results. For example, intelligent semantic analysis and processing based on big data can be used to obtain customers' interest in products, so that stores can recommend products for sale based on their interests. Therefore, a visual modeling method based on intelligent semantic understanding is provided.

[0005] In order to achieve the above-mentioned object of the invention, the embodiment of the present invention provides the following technical solutions:

[0006] A visualization modeling method based on intelligent semantic understanding includes the following steps:

[0007] Step S1, the user inputs the modeling task objective, performs word segmentation on the modeling task objective to obtain the word segmentation result; the user inputs the source data set and obtains the features of the source data set;

[0008] Step S2, determine the test model range based on the word segmentation result; select m test data from the source data set to test the test model range, and obtain the model test result, where the model test result includes N pre-set algorithm models;

[0009] Step S3, select n pre-set algorithm models from the N pre-set algorithm models as alternative solutions, and determine the corresponding n sets of algorithm model hyperparameters, 1 < n ≤ N;

[0010] Step S4, perform neural network learning on the n pre-set algorithm models, their corresponding n sets of algorithm model hyperparameters, and the features of the source data set; after learning, sort the n pre-set algorithm models according to the model evaluation index and recommend.

[0011] Furthermore, the step of determining the test model range based on the word segmentation result includes: calculating the similarity of the model labels in the algorithm model library based on the word segmentation result to determine the test model range; the algorithm model library includes model ID, model label, model name, and model description fields, and the model label includes classification, clustering, prediction, and regression.

[0012] Furthermore, the step of selecting m test data from the source data set to test the test model range and obtaining the model test result includes:

[0013] Based on the features of the source data set, obtain the hyperparameter list corresponding to each model ID in the test model range from the model hyperparameter library;

[0014] Select m test data from the source data set to test the test model range, 1 < m < 0.1M, where M is the total number of the source data set, and obtain the model test result; sort the pre-set algorithm models obtained in the model test result according to the size of the model F1 value, and visually recommend the top N pre-set algorithm models, N > 1.

[0015] Furthermore, the step of determining the corresponding n sets of algorithm model hyperparameters includes: screening out the corresponding n sets of algorithm model hyperparameters in the model hyperparameter library according to the model ID of the test model range, and the pre-set initial value is set for the algorithm model hyperparameters of each alternative solution, and the user modifies and resets the pre-set initial value through the visual interface method.

[0016] Furthermore, the step of sorting and recommending n preset algorithm models according to model evaluation indicators after learning includes: using visual icons to sort the n preset algorithm models according to different evaluation indicators, and the evaluation indicators include accuracy, recall rate, and F1 value; the user selects an optimal solution from the n preset algorithm models, or uses the recommended first solution as the final solution to the goal of this modeling task.

[0017] Furthermore, the step of inputting a source data set by the user and obtaining the features of the source data set includes: calculating the value distribution and value missing rate of each feature according to the input source data set, and determining the feature selection result according to the set threshold x, 0.5 <x<1。

[0018] Furthermore, the step of performing neural network learning on n preset algorithm models and their corresponding n sets of algorithm model hyperparameters and the features of the source data set includes:

[0019] The n preset algorithm models of this modeling task and their corresponding n sets of algorithm model hyperparameters, as well as the characteristics of the source data set, are batch iteratively learned, and the BP algorithm is used to optimize the neural network weights.

[0020] The loss function for neural network weight optimization is:

[0021]

[0022] Among them, M represents the total number of source data sets, i represents the i-th test data, j represents the j-th test data, and x i Represents any feature in the i-th test data; represents the characteristic division of the i-th test data and the j-th test data in the cosine space; n represents the number of preset algorithm models, and k represents the k-th preset algorithm model; Represents the weight of the k-th preset algorithm model; Represents the weight bias of the feature in the i-th test data; Represents the weight of the i-th test data; represents the total features in the i-th test data, Represents the total features in the j-th test data; Represents the algorithm model hyperparameters of the k-th preset algorithm model.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] (1) The present invention first uses intelligent semantic understanding, guided by the modeling task goal, to ensure model quality. Secondly, for the data in the structured data, the correlation analysis between the modeling task target field value and other feature values ​​can be performed to filter out features with weak correlation. It can also be used to filter out features with high missing feature fields by calculating the missing rate of feature field values, thereby achieving automatic feature selection. This process does not require manual intervention, reduces the user's workload, and further improves the quality of the model. Finally, because the use of preset algorithm models has good data flow specifications and standards, the possibility of errors in the modeling task flow is reduced, and modeling efficiency is improved.

[0025] (2) In the process of obtaining the optimal model, the user only needs to submit the target of the modeling task through the visual interface, input the source data set, configure the parameters of the algorithm model, and set the storage location of the model results to complete the modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0027] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0029] It should be noted that similar reference numerals and letters represent similar items in the following figures. Therefore, once an item is defined in one figure, it does not need to be further defined or explained in subsequent figures. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used to distinguish the description and are not to be understood as indicating or implying relative importance, or implying any actual relationship or order between these entities or operations.

[0030] Example 1:

[0031] The present invention is achieved through the following technical solutions: Figure 1 As shown in FIG, the visual modeling method based on intelligent semantic understanding includes the following steps:

[0032] In step S1, the user inputs a modeling task target, performs word segmentation processing on the modeling task target, and obtains a word segmentation result; the user inputs a source data set, and obtains the characteristics of the source data set.

[0033] Users input modeling task objectives based on their own business needs. For example, if a user needs to predict a customer's interest in a product in order to make recommendations, the user can input the modeling task objective "Task 1 - Predicting customer interest in a product." For another example, if a user needs to predict a customer's default risk in order to consider whether to cooperate, the user can input the modeling task objective "Task 2 - Predicting customer default risk."

[0034] Use the word segmentation algorithm to segment the input modeling task target and obtain the word segmentation results. For example, after the word segmentation of Task 1, the word segmentation results are "prediction, user, right, product, favorite", and after the word segmentation of Task 2, the word segmentation results are "prediction, customer, default, risk".

[0035] The user then enters the source dataset, which is the big data related to the modeling task. Two methods are supported for selecting source datasets: local upload and database connection. Source datasets can include multimodal data types such as structured tabular data, text data, and image data.

[0036] According to the input source data set, the value distribution and value missing rate of each feature are calculated and statistically analyzed, and the feature selection result is determined according to the set threshold x, 0.5 <x<1。

[0037] Step S2, determining the test model range based on the word segmentation result; selecting m test data from the source data set to test the test model range, and obtaining a model test result, which includes N preset algorithm models.

[0038] Based on the word segmentation results, similarity calculation is performed on the model labels in the algorithm model library to determine the scope of the test model; the algorithm model library includes model ID, model label, model name, and model description field, and the model label includes but is not limited to classification, clustering, prediction, and regression.

[0039] Based on the characteristics of the source dataset, obtain the hyperparameter list corresponding to each model ID in the test model range from the model hyperparameter library. Select m test data from the source dataset to test the test model range, where 1 < m < 0.1M and M is the total number of data in the source dataset, and obtain the model test results. Sort the preset algorithm models obtained in the model test results according to the size of the model F1 value, and visually recommend the top N preset algorithm models, where N > 1.

[0040] Step S3, select n preset algorithm models from the N preset algorithm models as alternative solutions, and determine the corresponding n groups of algorithm model hyperparameters, where 1 < n ≤ N.

[0041] The user selects n preset algorithm models from the recommended N preset algorithm models as alternative solutions for the current modeling task objective, where 1 < n ≤ N. Filter out the corresponding n groups of algorithm model hyperparameters in the model hyperparameter library according to the model ID in the test model range. The algorithm model hyperparameters of each alternative solution are set with preset initial values, and the user modifies and resets the preset initial values through the visual interface.

[0042] Step S4, perform neural network learning on the n preset algorithm models, their corresponding n groups of algorithm model hyperparameters, and the characteristics of the source dataset; after learning, sort the n preset algorithm models according to the model evaluation index and recommend them.

[0043] Perform batch iterative learning on the n preset algorithm models for the current modeling task objective, their corresponding n groups of algorithm model hyperparameters, and the characteristics of the source dataset, and use the BP algorithm to optimize the neural network weights. <...>The loss function for neural network weight optimization is:

[0045]

[0046] where, M represents the total number of data in the source dataset, i represents the i-th test data, j represents the j-th test data, x i represents any feature in the i-th test data; represents the feature分度 in the cosine space between the i-th test data and the j-th test data; n represents the number of preset algorithm models, and k represents the k-th preset algorithm model; represents the weight of the k-th preset algorithm model; represents the weight bias of the feature in the i-th test data; represents the weight of the i-th test data; represents the total feature in the i-th test data, represents the total feature in the j-th test data; represents the algorithm model hyperparameters of the k-th preset algorithm model.

[0047] Using visual icons, n preset algorithm models are sorted according to different evaluation indicators, including accuracy, recall rate, and F1 value. The user selects the best solution from the n preset algorithm models, or uses the first recommended solution as the final solution for this modeling task.

[0048] The core process of traditional AI visual modeling is: create a task → select a dataset → select feature engineering components → select model components → select evaluation metrics → submit training → obtain model results. Within this process, feature engineering includes components for data splitting, row and column processing, SQL, and data cleaning, while model building involves components such as logistic regression, decision trees, CNNs, and RNNs. While this may seem to lower the barrier to entry for AI visual modeling, allowing for zero-code model building, it actually still requires a high level of knowledge in feature engineering, data preprocessing, and neural network theory. Furthermore, if any component in the task fails, the model will not function properly, resulting in extremely low model efficiency.

[0049] The core process of AI visual modeling of the present invention is: input modeling task objectives → select data sets → model selection (automatically select features, preset hyperparameter list) → submit training → obtain model results → select the optimal model. First, use intelligent semantic understanding to guide the modeling task objectives to ensure model quality. Secondly, for the data in the structured data, you can perform correlation analysis on the modeling task target field values ​​and other feature values ​​to filter out features with weak correlations. You can also calculate the missing rate of feature field values ​​to filter out features with high missing feature fields, thereby achieving automatic feature selection. This process does not require manual intervention, reduces the user's workload, and further improves the quality of the model. Finally, because the preset algorithm model has good data flow specifications and standards, the possibility of errors in the modeling task flow is reduced, and the modeling efficiency is improved.

[0050] This invention not only realizes zero-code AI visual modeling, but also takes the modeling task goal as its original intention, providing users with a variety of model solutions that can directly solve their task goals, and promotes better implementation and application of AI models.

[0051] Example 2:

[0052] This embodiment is illustrated by the technical solution of embodiment 1. The user's demand is to predict the customer's interest in the product based on the product sales data and make product recommendations.

[0053] Step 1: Enter the modeling task objective "predict customer interest in products".

[0054] Step 2: Segment the modeling task objective to obtain the segmentation result "prediction, customer, towards, product, preference".

[0055] Step 3: Input the source dataset, which is the customer product order table. This table includes the following 19 fields: product name, manufacturer, shipping date, country / region, region, city, subcategory, customer name, province / autonomous region, category, segment, order ID, order date, shipping method, profit, discount, quantity, sales amount.

[0056] Step 4: (1) According to the input source dataset, count metrics such as the value distribution and value missing rate of each feature, and set a threshold x, where 0.5 < x < 1. If the missing rate is greater than the set threshold x, discard the feature; if the feature value distribution is more uniform, the weaker the impact of the feature on the modeling task objective, then discard the feature, thus performing automatic feature selection.

[0057] (2) Map according to the segmentation result of the modeling task objective and the model labels in the algorithm model library to obtain the model with the model label "prediction" as the test model range for this task objective. Based on the features of the source dataset, obtain the hyperparameter list corresponding to each model ID in the test model range from the model hyperparameter library, as shown in Table 1. The algorithm model library includes fields such as model ID, model label, model name, and model description. The model labels include classification, clustering, prediction, and regression.

[0058]

[0059] Table 1

[0060] (3) Select m test model test data from the source dataset to test the test model range, where 1 < m < 0.1M and M is the total number of records in the source dataset, to obtain the model test results. Sort the pre-set algorithm models obtained in the model test results according to the size of the model F1 value, and visually recommend the top N pre-set algorithm models, where N > 1. See the example data in Table 2. The pre-set algorithm models with the F1 value ranking in the top 3 (N = 3) in the training results are RCNN (F1 value equal to 92%), GRU (F1 value equal to 89%), and LSTM (F1 value equal to 85%). Then recommend these three pre-set algorithm models to the user for selection.

[0061]

[0062] Table 2

[0063] Step 5: The user selects 2 (n = 2) pre-set algorithm models, RCNN and GRU, from the 3 (N = 3) pre-set algorithm models as alternative solutions for this modeling task objective, where 1 < n ≤ N.

[0064] Step 6: Based on the model IDs of the two alternative solutions selected by the user, the corresponding two sets of algorithm model hyperparameters are selected from the model hyperparameter library. The algorithm model hyperparameters of each alternative solution have preset initial values, and the user can modify and reset the preset initial values ​​through the visual interface.

[0065] Step 7: The two alternative solutions selected for this modeling task and the corresponding two sets of algorithm model hyperparameters, as well as the characteristics of the source data set, are batch-trained for iterative learning, and the BP algorithm is used to optimize the neural network weights.

[0066] Step 8: Use visual charts to present the two alternative solutions based on different evaluation metrics, such as accuracy, recall, and F1 value. In this example, the pre-built RCNN model has an accuracy of 94%, a recall of 88%, and an F1 value of 91%, while the pre-built GRU model has an accuracy of 91%, a recall of 90%, and an F1 value of 90.5%. Users can select the optimal solution from the two alternatives based on their modeling task objectives. For example, if high accuracy is required, the user can choose the RCNN model, while if high recall is required, the user can choose the GRU model. Alternatively, the system recommends the optimal solution, which is ranked based on the weighted average of the evaluation metrics. This recommends the RCNN model, which results in the final solution.

[0067] The above is a modeling approach for the task objective of "predicting customer interest in products." In the realm of enterprise big data processing, this approach can also be used to model monitoring big data from internet finance industries, such as internet payment, internet insurance, and internet consumer finance, thereby obtaining directly usable models based on customer needs. Based on intelligent semantic understanding, this solution uses monitoring big data as the source dataset and, guided by modeling task objectives, implements risk monitoring, prevention, and identification for the internet finance industry. This essentially establishes the ability to monitor, prevent, and identify financial risks, including business risks and internet risks. Users simply submit the modeling task objective through a visual interface, input the source dataset, configure the algorithm model parameters, and set the storage location for the model results to complete the modeling process.

[0068] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A visual modeling method based on intelligent semantic understanding, characterized by: The steps include: Step S1: The user inputs the modeling task objective, performs word segmentation on the modeling task objective to obtain the word segmentation result; The user inputs the source dataset and obtains the features of the source dataset; Step S2: Determine the test model scope based on the word segmentation result; Select m test data from the source dataset to test the test model scope, and obtain the model test result, where the model test result includes N preset algorithm models; The step of determining the test model scope based on the word segmentation result includes: calculating the similarity of the model labels in the algorithm model library based on the word segmentation result to determine the test model scope; the algorithm model library includes model ID, model label, model name, and model description fields, and the model labels include classification, prediction, and regression; Step S3: Select n preset algorithm models from the N preset algorithm models as alternative solutions, and determine the corresponding n sets of algorithm model hyperparameters, where 1 < n ≤ N; Step S4: Perform neural network learning on the n preset algorithm models, their corresponding n sets of algorithm model hyperparameters, and the features of the source dataset; after learning, sort the n preset algorithm models according to the model evaluation metrics and then recommend.

2. The visual modeling method based on intelligent semantic understanding according to claim 1, characterized in that: The step of selecting m test data from the source dataset to test the test model scope and obtaining the model test result includes: Based on the features of the source dataset, obtain the hyperparameter list corresponding to each model ID in the test model scope from the model hyperparameter library; Select m test data from the source dataset to test the test model scope, where 1 < m < 0.1M and M is the total number of the source dataset, and obtain the model test result; sort the preset algorithm models obtained in the model test result according to the size of the model F1 value, and visually recommend the first N preset algorithm models, where N > 1.

3. The visual modeling method based on intelligent semantic understanding according to claim 1, characterized in that: The step of determining the corresponding n sets of algorithm model hyperparameters includes: screening out the corresponding n sets of algorithm model hyperparameters from the model hyperparameter library according to the model ID of the test model scope, and the preset initial values are set for the algorithm model hyperparameters of each alternative solution, and the user modifies and resets the preset initial values through the visual interface method.

4. The visual modeling method based on intelligent semantic understanding according to claim 1, characterized in that: The step of sorting and recommending the n preset algorithm models according to the model evaluation metrics after learning includes: using the display form of visual icons to sort the n preset algorithm models according to different evaluation metrics, and the evaluation metrics include accuracy, recall rate, and F1 value; the user selects an optimal solution from the n preset algorithm models, or uses the recommended first solution as the final solution for the current modeling task objective.

5. The visual modeling method based on intelligent semantic understanding according to claim 1, characterized in that: The step of the user inputting the source dataset and obtaining the features of the source dataset includes: calculating and statistically analyzing the value distribution and value missing rate of each feature according to the input source dataset, and determining the feature selection result according to the set threshold x, where 0.5 < x < 1.

Images