A model fusion-based index calculation method, device and medium

Abstract

The application discloses a model fusion-based index calculation method and device and a medium, wherein the method comprises the following steps: acquiring mixed data used for calculating a target index and a user calculation demand; determining a data proportion corresponding to the mixed data; determining a structured data calculation model and an unstructured data calculation model in a preset model library based on the data proportion and / or the user calculation demand; determining result weights corresponding to the structured data calculation model and the unstructured data calculation model respectively based on the data proportion and / or the user calculation demand; and determining an index calculation result based on the result weights, the structured data calculation model and the unstructured data calculation model. Based on the proportions of structured data and unstructured data in the mixed data, the combination of a traditional algorithm and a large model calculation can be optimized, so that the advantages of the traditional algorithm in structured data processing and interpretability can be fully utilized.

Description

Technical Field

[0001] This application relates to the field of indicator calculation, specifically to an indicator calculation method, device, and medium based on model fusion. Background Technology

[0002] Metric calculation refers to the process of transforming raw data into meaningful quantitative indicators using specific mathematical formulas or statistical methods. These indicators can more intuitively and concisely reflect data characteristics, business conditions, or system performance. In recent years, with the explosive growth of data volume and the increasing diversification of data types, traditional single-technology approaches have become insufficient to meet practical needs.

[0003] Currently, the industry is increasingly showing a trend of combining multiple technologies to calculate complex indicators. For example, some data analysis platforms are beginning to explore combining traditional statistical methods with simple machine learning models. However, this combination is often a simple superposition, lacking systematicity and adaptability, and failing to fully leverage the advantages of each technology. Summary of the Invention

[0004] To address the aforementioned issues, this application proposes a method, device, and medium for calculating indicators based on model fusion, wherein the method includes: In response to a user-sent instruction to calculate an indicator, the system acquires mixed data for calculating the target indicator, as well as the user's calculation requirements, including real-time performance and / or interpretability. It then determines the proportion of mixed data, including the proportion of structured data and the proportion of unstructured data. Based on the data proportions and / or the user's calculation requirements, it determines a structured data calculation model and an unstructured data calculation model from a pre-defined model library. Based on the data proportions and / or the user's calculation requirements, it determines the result weights corresponding to the structured data calculation model and the unstructured data calculation model, respectively. Finally, based on the result weights, the structured data calculation model, and the unstructured data calculation model, it determines the indicator calculation result.

[0005] In one example, after obtaining the mixed data used to calculate the target indicator, the method further includes: cleaning the structured data and unstructured data in the mixed data according to the data type; and extracting features from the cleaned data to obtain feature data of the corresponding type.

[0006] In one example, determining the index calculation result based on the result weight, the structured data calculation model, and the unstructured data calculation model specifically includes: obtaining the first feature anchor point of the structured data in the mixed data; when the unstructured data calculation model processes the corresponding unstructured data, determining the anchor point data related to the first feature anchor point; and increasing the corresponding weight of the anchor point data.

[0007] In one example, determining the index calculation result based on the result weight, the structured data calculation model, and the unstructured data calculation model specifically includes: obtaining a second feature anchor point in the unstructured data of the mixed data; generating a structured label based on the second feature anchor point; and incorporating the structured label into the structured data calculation model.

[0008] In one example, determining the structured data calculation model and the unstructured data calculation model in a preset model library based on the data proportion and / or the user's calculation needs specifically includes: determining the data calculation bias based on the data proportion and / or the user's calculation needs; and determining the structured data calculation model and the unstructured data calculation model in the preset model library based on the data calculation bias.

[0009] In one example, determining the index calculation result based on the result weights, the structured data calculation model, and the unstructured data calculation model specifically includes: calculating a first intermediate result corresponding to the structured data in the mixed data based on the structured data calculation model; calculating a second intermediate result corresponding to the unstructured data in the mixed data based on the unstructured data calculation model; and performing a weighted calculation on the first intermediate result and the second intermediate result based on the result weights to obtain the index calculation result.

[0010] In one example, after determining the indicator calculation result based on the result weight, the structured data calculation model, and the unstructured data calculation model, the method includes: obtaining the historical indicator calculation results of the same user's historical indicator calculation tasks; determining the calculation error between the historical indicator calculation results and the indicator calculation results; and when the calculation error is higher than a preset error threshold, adjusting the result weight and the model weight of the unstructured data calculation model until the calculation error is lower than the preset error threshold.

[0011] In one example, adjusting the result weights and the model weights of the unstructured data calculation model until the calculation error is lower than a preset error threshold specifically includes: counting the number of adjustments to the result weights and the model weights of the unstructured data calculation model; when the number of adjustments exceeds a preset number threshold, re-selecting the structured data calculation model and the unstructured data calculation model from the preset model library.

[0012] This application also provides a model fusion-based index calculation device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the method as described in any of the above examples.

[0013] This application also provides a non-volatile computer storage medium storing computer-executable instructions configured to perform the steps of the method as described in any of the above examples.

[0014] The method proposed in this application offers the following advantages: Based on the ratio of structured to unstructured data in mixed data, this application optimizes the combination of traditional algorithms and large-scale model computation. This fully leverages the advantages of traditional algorithms in structured data processing and interpretability, as well as the strengths of large-scale models in handling unstructured data and capturing nonlinear relationships, achieving complementary advantages. It solves the problems of insufficient computational power of traditional algorithms in complex scenarios and the high computational resource consumption, slow inference speed, and poor interpretability of large-scale models, making the indicator calculation process more in line with user needs and improving the user experience. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating an index calculation method based on model fusion in an embodiment of this application. Figure 2 This is a schematic diagram of the structure of an index calculation device based on model fusion in an embodiment of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] Traditional algorithms excel at calculating simple, rule-based indicators, boasting clear computational logic and high execution efficiency. For instance, basic financial indicators like the debt-to-equity ratio (total liabilities / total assets × 100%) can be quickly derived through simple numerical calculations. However, when faced with complex, ambiguous, or data-driven formulas, traditional algorithms often become inefficient. For example, when predicting systemic risk indicators based on multiple complex factors, traditional algorithms struggle to effectively handle massive, variable data and fail to accurately capture the complex relationships between data points, resulting in low accuracy and inefficiency.

[0018] Large-scale models, with their powerful learning capabilities and ability to uncover complex data patterns, demonstrate certain advantages in handling complex indicator calculation tasks. In sentiment analysis indicator calculation within natural language processing, large language models can accurately determine the sentiment tendency of text and provide corresponding indicator values ​​by learning from large amounts of text data. However, the drawbacks are that training and running large models consumes significant computational resources and time, resulting in high costs. Furthermore, when dealing with simple indicator calculations with clear mathematical logic and rules, they can exhibit over-complexity, complicating simple problems. Due to their black-box nature, the interpretability of the calculation process and results is poor, making it difficult for users to understand and verify the rationality of the calculations.

[0019] Existing technologies lack specificity in the interpretable output of fusion models, either being overly complex (unsuitable for scenarios with low interpretability requirements) or too simplistic (unable to meet the traceability requirements of scenarios with high interpretability).

[0020] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0021] Figure 1This document illustrates a flowchart of a model fusion-based indicator calculation method provided in one or more embodiments. This method can be applied to indicator calculation in various business sectors, such as internet finance, e-commerce, instant messaging, gaming, and government services. The process can be executed by computing devices specific to the sector (e.g., risk control servers or smart mobile terminals for payment services). Certain input parameters (such as result weights) or intermediate results can be manually adjusted to improve accuracy.

[0022] The analysis method involved in the embodiments of this application can be implemented by a terminal device or a server, and this application does not impose any special limitations on it. For ease of understanding and description, the following embodiments are all described in detail using a server as an example.

[0023] It should be noted that the server can be a single device or a system composed of multiple devices, i.e., a distributed server. This application does not make any specific limitations in this regard.

[0024] like Figure 1 As shown in the embodiments of this application, a method for calculating indicators based on model fusion is provided, including: S101: In response to the indicator calculation instruction sent by the user, obtain the mixed data used to calculate the target indicator, as well as the user's calculation requirements, which include real-time performance and / or interpretability.

[0025] When a user needs to calculate a certain indicator based on data, they send an indicator calculation command to the server and upload mixed data to the server. Upon receiving the indicator calculation command from the user, the server receives the mixed data from the user. Mixed data refers to the data used to calculate the target indicator, including both structured and unstructured data. For example, when calculating a basic financial indicator such as the debt-to-equity ratio (total liabilities / total assets × 100%), the mixed data can be data related to both total liabilities and total assets.

[0026] When acquiring mixed data, in addition to users uploading mixed data themselves, the server can also collect raw data from various data sources through interfaces, database connections, file reading, etc., including but not limited to structured data (such as table data in relational databases), semi-structured data (such as log files in JSON format), and unstructured data (such as customer reviews and logs in text format), thereby obtaining mixed data.

[0027] When the server receives the metric calculation instruction, it also retrieves the user's calculation requirements from the instruction. These requirements include the real-time nature and interpretability of the metric calculation. Understandably, different calculation requirements will lead to variations in subsequent data processing to better align with the user's needs.

[0028] In one embodiment, after obtaining the mixed data, it is necessary to perform data cleaning and feature extraction to facilitate subsequent data processing. Different cleaning and feature extraction methods should be used depending on the data type. Specifically, for structured data, SQL statements can be used to remove duplicate data, such as `select distinct * from table_name`, and missing values ​​in numeric fields can be filled using the mean, e.g., `update table_name set column_name = avg(column_name) where column_name is null`. For unstructured data, regular expressions can be used to remove special characters and garbled text. For example, the `java.util.regex` package and the `Pattern.compile` method can be used to compile the regular expression `[^a-zA-Z0-9\\u4e00-\\u9fa5]`, which matches content that is not alphanumeric, numeric, or Chinese characters. The `Matcher.replaceAll` method replaces the matched content with an empty string. When extracting data features, for structured data, traditional feature engineering methods such as Principal Component Analysis (PCA) and correlation analysis can be used to extract key features. The PCA algorithm is implemented using the open-source mathematical library Apache CommonsMath. For unstructured data, large models can be used for feature extraction. A pre-trained model is loaded using a deep learning framework (Deeplearning4j). Convolutional Neural Networks (CNNs) are used to extract visual features from image data, while Recurrent Neural Networks (RNNs) or Transformer models are used to extract semantic features from text data. Extracting data features before inputting them into the model improves model performance, addresses data heterogeneity issues, and optimizes computational efficiency.

[0029] S102: Determine the data proportion corresponding to the mixed data, wherein the data proportion includes the proportion of structured data and the proportion of unstructured data.

[0030] After obtaining the mixed data, it is necessary to determine the proportion of different data types within the mixed data, namely, the proportion of structured data and the proportion of unstructured data. Once the data proportions are determined, subsequent data processing procedures can be adjusted accordingly.

[0031] S103: Based on the data proportion and / or the user's computing needs, determine the structured data computing model and the unstructured data computing model in the preset model library.

[0032] When selecting a data computation model, the type of computation model needs to be determined based on the data proportion and user computational needs. These computation models include structured data computation models for calculating structured data and unstructured data computation models for calculating unstructured data, and these models are pre-stored in a preset model library.

[0033] When considering factors that include both data proportion and user computing needs, it is necessary to quantify user computing needs before selecting a model. For example, when obtaining the user's computing need level, users should be required to select the real-time requirement level and the interpretability requirement level themselves, so that the specific level of user computing needs can be used as a quantitative indicator.

[0034] It is understandable that different computing models have different computing speeds and interpretability. Therefore, different computing models can be set with corresponding user computing requirement levels. That is, a computing model corresponds to an interpretability level and a real-time level. When selecting a computing model, the user computing requirement level can be matched to obtain the corresponding computing model.

[0035] Simultaneously, structured data calculation models can be bound to unstructured data calculation models. Based on the data proportion and user calculation needs, the data calculation bias for this indicator calculation task can be determined, and then the corresponding calculation model group can be selected in the database based on this bias. When determining the data calculation bias, the bias can be based on the data proportion and the quantified user calculation needs. For example, a weighted average can be calculated using preset coefficients, taking the data proportion and the user calculation need value corresponding to the user calculation need level (e.g., 0.8 for a high-level user calculation need). Finally, based on the specific value of the data calculation bias, the calculation model group corresponding to that value can be selected from the model database.

[0036] Specifically, when the mixed data contains only structured data or only unstructured data, there is no need to consider the selection of another calculation model. In this case, the corresponding data calculation model can be directly determined from the preset model library according to the user's calculation needs, and the result weight of the data calculation model is 1.

[0037] Users can also specify the computational model they want to choose. For example, the following are examples of selecting a computational model for different interpretability requirements: For scenarios with a high interpretability level, when selecting algorithms, logically transparent structured data models (such as linear regression models and decision tree models) should be prioritized for user recommendation; and when selecting unstructured data computation models, computation models with higher interpretability (such as the distilled BERT model) should be prioritized, and a feature contribution output module (which quantifies the influence weight of keywords such as "overdue" and "debt" in the text through SHAP values) should be embedded as an explanation. The result weight can be determined based on the data proportion; for example, if the proportion of structured data is 80%, the weight can be set to [0.8, 0.2].

[0038] For scenarios with a medium interpretability level, when selecting algorithms, prioritize recommending models that balance accuracy and interpretability (such as gradient boosting trees) to users, and output the core formula and key parameters (such as "conversion rate = 0.6 × clicks + 0.4 × dwell time") as explanations; for unstructured data models, recommend users to choose conventional pre-trained models (such as RNNs), and output the top 3 influencing features (such as "the positive contribution of 'high cost-performance ratio' in product reviews is 0.2") as explanations.

[0039] For scenarios with a low interpretability level, when selecting algorithms, simple statistical models (such as mean and percentage calculations) should be recommended to users first; for structured data models, high-performance large models (such as Transformer) can be recommended to users, without the need for additional interpretability processing.

[0040] S104: Based on the data proportion and / or the user's computing needs, determine the result weights corresponding to the structured data computing model and the unstructured data computing model, respectively.

[0041] After selecting a suitable computational model, if both structured and unstructured data computational models exist, it is necessary to determine the weights of the results from each model. These weights are used to evaluate the importance of the computational results from each model to the final indicator calculation result. The larger the weight of the computational model's result, the more the calculation results of this model are relied upon when calculating the indicator.

[0042] S105: Determine the index calculation results based on the result weights, the structured data calculation model, and the unstructured data calculation model.

[0043] After determining the result weights of the two calculation models, the calculation results of the two calculation models can be weighted and averaged to determine the final index calculation result.

[0044] Specifically, when determining the indicator calculation results, it is necessary to calculate the first intermediate result corresponding to the structured data in the mixed data based on the structured data calculation model, and calculate the second intermediate result corresponding to the unstructured data in the mixed data based on the unstructured data calculation model. Finally, based on the result weights, the first intermediate result and the second intermediate result are weighted and calculated to obtain the indicator calculation result.

[0045] When the output dimensions of the first intermediate result and the second intermediate result are different, the first intermediate result and the second intermediate result can be normalized by L2 and then weighted averaged.

[0046] For example, for a certain indicator calculation task, the first intermediate result is 0.75, and the second intermediate result is a 5-dimensional vector [0.2, 0.5, 0.1, 0.1, 0.1]. Then, through L2 normalization, the second intermediate result can be transformed into [0.348, 0.871, 0.174, 0.174, 0.174]. This transformed result is then compared with the first intermediate result [1, 0, 0, 0] after L2 normalization and expansion to a 5-dimensional vector. We perform a weighted average of [0.7, 0.3]. If the weight of the result is [0.7, 0.3], then the fusion result is: [1.0×0.7+0.348×0.3,0×0.7+0.871×0.3,0×0.7+0.174×0.3,0×0.7+0.174×0.3,0×0.7+0.174×0.3]=[0.804,0.261,0.052,0.052,0.052]. The final risk score is 0.804 for the first dimension.

[0047] In one embodiment, during the computation process, the structured data computation model and the unstructured data computation model can dynamically interact with each other, thereby enabling the structured data computation model and the unstructured data computation model to complement each other in terms of features.

[0048] Specifically, for the structured data computation model, the first feature anchor point of the structured data in the mixed data can be obtained during the computation, and when the unstructured data computation model processes the corresponding unstructured data, the anchor point data related to the first feature anchor point can be determined, and the corresponding weight of the anchor point data can be increased.

[0049] Specifically, for the structural data calculation model, when processing structured data, the first diagnostic anchor point can be output (such as the structured feature of "excessive rebar spacing" in the engineering quality risk score); and when the non-structured data calculation model processes unstructured data (such as acceptance report text), it should prioritize focusing on content related to the anchor point (such as keywords in the text) and strengthen the weight of these features through an attention mechanism (such as increasing the attention weight of related keywords by 20%).

[0050] For unstructured data computation models, second feature anchor points can be obtained from the unstructured data in the mixed data, and structured labels can be generated based on these anchor points. These structured labels are then incorporated into the structured data computation model. For example, the unstructured data computation model extracts "potential impact features" (such as the text "inadequate qualifications of construction personnel") from the unstructured data, generating structured labels (such as "personnel qualification risk = 0.8"). Traditional algorithms incorporate this label into the computation logic (e.g., updating the formula to "engineering quality risk score = 0.4 × concrete strength + 0.3 × rebar spacing + 0.2 × schedule deviation + 0.1 × personnel qualification risk"), achieving feature complementarity.

[0051] In one embodiment, after obtaining the indicator calculation result, the result weights and model parameters can be adjusted based on the calculation error value. This error value can be obtained by calculating the mean squared error between historical indicator calculation results from historical indicator calculation tasks and the current indicator calculation result. Specifically, the calculation error between the historical indicator calculation results and the current indicator calculation result can be determined by obtaining the historical indicator calculation results from historical indicator calculation tasks of the same user. When the calculation error exceeds a preset error threshold, the result weights and the model weights of the unstructured data calculation model are adjusted until the calculation error falls below the preset error threshold.

[0052] Furthermore, if the number of adjustments to the result weights and the weight parameters in the model is too many but the error is always higher than the preset error threshold, it is considered that the selected calculation model does not meet the requirements of the index calculation. The number of adjustments to the result weights and the model weights of the unstructured data calculation model can be counted, and when the number of adjustments exceeds the preset number threshold, the structured data calculation model and the unstructured data calculation model can be reselected from the preset model library.

[0053] It should be noted that when selecting a model and determining the result weights, one can choose a computational model and determine the result weights for different models solely based on data proportion; alternatively, one can choose a computational model and determine the result weights for different models solely based on user computational requirements; or one can choose a computational model and determine the result weights simultaneously based on both data proportion and user computational requirements. Furthermore, user computational requirements can include both real-time performance and interpretability, or only one of these. Understandably, when users have higher requirements for real-time performance or interpretability, they tend to choose structured data computational models for computation.

[0054] In one embodiment, a real-time monitoring mechanism can be introduced, utilizing the performance monitoring tool JavaMission Control and a custom monitoring class to monitor changes in data characteristics and computational requirements during the computation process. When the data volume suddenly increases and real-time requirements rise, the monitoring mechanism can detect this change promptly and feed the information back to the fusion strategy generation module. Based on the results of real-time monitoring, the fusion strategy is dynamically optimized and adjusted. The dynamic updating of the fusion strategy is achieved using concurrent programming ThreadPoolExecutor and an event-driven mechanism. Automatic adjustment of the fusion strategy, i.e., adjusting weight allocation, is implemented. A gradient descent algorithm (implemented through the WeightOptimizer class) is used to automatically adjust the fusion weights based on the computational error value after result verification feedback, i.e., adjusting the value of the result weights. A reinforcement learning algorithm is employed, using the reinforcement learning library RL4J, allowing the system to autonomously learn the optimal fusion strategy based on historical computational results and feedback information. Through continuous training and optimization, the system can more accurately select the fusion method and adjust the fusion weights under different data scenarios and computational requirements. When a surge in data volume is detected, the JVM resource status is obtained through ManagementFactory, and the memory allocation of large models is dynamically adjusted to reduce model caching and prioritize computation threads.

[0055] like Figure 2 As shown in the illustration, this application also provides a model fusion-based index calculation device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to: In response to a user-sent instruction to calculate an indicator, the system acquires mixed data for calculating the target indicator, as well as the user's calculation requirements, including real-time performance and / or interpretability. It then determines the proportion of mixed data, including the proportion of structured data and the proportion of unstructured data. Based on the data proportions and / or the user's calculation requirements, it determines a structured data calculation model and an unstructured data calculation model from a pre-defined model library. Based on the data proportions and / or the user's calculation requirements, it determines the result weights corresponding to the structured data calculation model and the unstructured data calculation model, respectively. Finally, based on the result weights, the structured data calculation model, and the unstructured data calculation model, it determines the indicator calculation result.

[0056] This application embodiment also provides a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured as follows: In response to a user-sent instruction to calculate an indicator, the system acquires mixed data for calculating the target indicator, as well as the user's calculation requirements, including real-time performance and / or interpretability. It then determines the proportion of mixed data, including the proportion of structured data and the proportion of unstructured data. Based on the data proportions and / or the user's calculation requirements, it determines a structured data calculation model and an unstructured data calculation model from a pre-defined model library. Based on the data proportions and / or the user's calculation requirements, it determines the result weights corresponding to the structured data calculation model and the unstructured data calculation model, respectively. Finally, based on the result weights, the structured data calculation model, and the unstructured data calculation model, it determines the indicator calculation result.

[0057] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0058] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0059] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0060] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0062] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0064] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0065] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0066] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0067] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for calculating indicators based on model fusion, characterized in that, include: In response to a user-sent instruction to calculate an indicator, the system acquires mixed data for calculating the target indicator, as well as the user's calculation requirements, which include real-time performance and / or interpretability. Determine the data proportions corresponding to the mixed data, wherein the data proportions include the proportion of structured data and the proportion of unstructured data; Based on the data proportion and / or the user's computing needs, a structured data computing model and an unstructured data computing model are determined in a preset model library. Based on the data proportion and / or the user's computing needs, determine the result weights corresponding to the structured data computing model and the unstructured data computing model, respectively. Based on the result weights, the structured data calculation model, and the unstructured data calculation model, the index calculation results are determined. The determination of the indicator calculation result based on the result weight, the structured data calculation model, and the unstructured data calculation model specifically includes: Obtain the first feature anchor point of the structured data in the mixed data; When the unstructured data calculation model processes the corresponding unstructured data, it determines the anchor point data related to the first feature anchor point. Increase the corresponding weight of the anchor point data; The determination of the indicator calculation result based on the result weight, the structured data calculation model, and the unstructured data calculation model specifically includes: Obtain the second feature anchor point from the unstructured data in the mixed data; Structured labels are generated based on the second feature anchor point; The structured tags are incorporated into the structured data calculation model.

2. The method according to claim 1, characterized in that, After acquiring the mixed data used to calculate the target indicator, the method further includes: Based on the data type, data cleaning is performed on the structured data and unstructured data in the mixed data respectively; The cleaned data is then subjected to feature extraction to obtain feature data of the corresponding type.

3. The method according to claim 1, characterized in that, The step of determining structured data calculation models and unstructured data calculation models from a preset model library based on the data proportion and / or the user's calculation needs specifically includes: Based on the data proportion and / or the user's computing needs, determine the data computing bias; Based on the data calculation bias, the structured data calculation model and the unstructured data calculation model are determined in the preset model library.

4. The method according to claim 1, characterized in that, The determination of the indicator calculation result based on the result weight, the structured data calculation model, and the unstructured data calculation model specifically includes: The first intermediate result corresponding to the structural data in the mixed data is calculated based on the structural data calculation model. The second intermediate result corresponding to the unstructured data in the mixed data is calculated based on the unstructured data calculation model. Based on the result weights, the first intermediate result and the second intermediate result are weighted and calculated to obtain the index calculation result.

5. The method according to claim 1, characterized in that, After determining the index calculation result based on the result weight, the structured data calculation model, and the unstructured data calculation model, the method includes: Obtain the historical metric calculation results of the same user's historical metric calculation task; Determine the calculation error between the historical indicator calculation result and the indicator calculation result; When the calculation error is higher than a preset error threshold, the result weights and the model weights of the unstructured data calculation model are adjusted until the calculation error is lower than the preset error threshold.

6. The method according to claim 5, characterized in that, The step of adjusting the weights of the results and the model weights of the unstructured data calculation model until the calculation error is lower than a preset error threshold specifically includes: The number of adjustments made to the weights of the results and the model weights of the unstructured data calculation model is counted. When the number of adjustments exceeds a preset threshold, a new structured data calculation model and an unstructured data calculation model are selected from the preset model library.

7. An index calculation device based on model fusion, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the steps of the method as claimed in any one of claims 1-6.

8. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are configured to perform the steps of the method as claimed in any one of claims 1-6.

Images