Method and device for evaluating business line online feasibility, medium and terminal

Abstract

The application relates to the fields of artificial intelligence and finance, and discloses a business online feasibility evaluation method and device, a medium and a terminal, which comprise the following steps: collecting offline business system data and online business system data, performing standardized mapping processing, and obtaining a standardized fusion data set; using a contrast learning encoder to encode the standardized fusion data set to obtain business embedding vectors corresponding to various business types, and obtaining business similarity features and sequence features of the various business types according to the business embedding vectors; based on the standardized fusion data set, user behavior aggregation features corresponding to the various business types are respectively constructed, and a business online feasibility scoring model is used to calculate the business online feasibility scores of the various business types based on the user behavior aggregation features, the business embedding vectors, the business similarity features and the sequence features. The application can be applied to the evaluation of the business online feasibility of financial technology and improves the evaluation accuracy.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence technology and finance, and in particular to a method, apparatus, medium, and terminal for assessing the feasibility of online business operations. Background Technology

[0002] With the rapid development of the digital economy, various industries are actively migrating their businesses from offline to online to reduce operating costs and improve service efficiency, with the financial industry being the most prominent example. However, the feasibility and user acceptance of different types of businesses in the online process vary significantly, and blindly pushing forward may lead to problems such as a decline in user experience and low business conversion rates. Therefore, there is an urgent need for an evaluation method for business online transformation to assist financial enterprises in formulating digital transformation strategies.

[0003] Currently, existing evaluation methods for business online transformation typically involve subjective assessment of the business through questionnaires or expert scoring, followed by statistical analysis to assist in decision-making.

[0004] However, because existing methods rely heavily on human experience or statistical indicators, the evaluation results are highly subjective and cannot accurately guide the digital transformation of financial businesses, thereby reducing the transaction rate of financial businesses after going online. Summary of the Invention

[0005] This application provides a method, apparatus, medium, and terminal for assessing the feasibility of online business operations, in order to solve the technical problem that existing assessment methods rely heavily on human experience or statistical indicators, resulting in highly subjective assessment results that cannot accurately guide the digital transformation of financial businesses, thereby reducing the transaction rate of online financial businesses.

[0006] Firstly, it provides a method for assessing the feasibility of online business operations, including: Data from offline business systems and online business systems are collected separately. The offline business system data and the online business system data are then subjected to standardized mapping processing to obtain a standardized fusion dataset. The standardized fusion dataset contains online user behavior records and offline user behavior records for different business types. Using a contrastive learning encoder that has completed model training, each user behavior record in the standardized fusion dataset is encoded to obtain a user behavior embedding vector corresponding to each user behavior record. All user behavior embedding vectors are aggregated according to business type to obtain a business embedding vector corresponding to each business type. The similarity between each business embedding vector and the anchor business embedding vector is calculated as a business similarity feature for each business type. The similarity scores are sorted in descending order to obtain a similarity sequence. Based on the similarity sequence, the ordinal feature of each business type is calculated. Based on the standardized fusion dataset, user behavior aggregation features corresponding to each of the aforementioned business types are constructed respectively. Using the business online feasibility scoring model that has been trained, the business online feasibility score for each of the aforementioned business types is calculated based on the user behavior aggregation features, business embedding vectors, business similarity features, and ordinal features corresponding to each of the aforementioned business types.

[0007] Secondly, a device for assessing the feasibility of online business operations is provided, including: The data acquisition module is used to collect data from offline business systems and online business systems respectively, and to perform standardized mapping processing on the offline business system data and the online business system data to obtain a standardized fusion dataset. The standardized fusion dataset contains online user behavior records and offline user behavior records of different business types. The contrastive learning module is used to encode each user behavior record in the standardized fusion dataset using a contrastive learning encoder that has completed model training, to obtain user behavior embedding vectors corresponding to each user behavior record, and to aggregate all user behavior embedding vectors according to business type to obtain business embedding vectors corresponding to each business type. The module also calculates the similarity between each business embedding vector and the anchored business embedding vector as the business similarity feature of each business type, sorts the similarity in descending order to obtain a similarity sequence, and calculates the ordinal feature of each business type based on the similarity sequence. The online feasibility assessment module is used to construct user behavior aggregation features corresponding to each of the business types based on the standardized fusion dataset, and to calculate the online feasibility score of each of the business types based on the user behavior aggregation features, business embedding vectors, business similarity features and ordinal features, using the business online feasibility scoring model that has been trained.

[0008] Thirdly, a storage medium is provided, which stores a computer program, and when the computer program is executed by a processor, it implements the steps of the above-mentioned feasibility assessment method for online business operations.

[0009] Fourthly, a terminal is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-mentioned method for assessing the feasibility of online business operations.

[0010] By employing the above technical solutions, the technical solutions provided in the embodiments of this application have at least the following advantages: This application provides a method, apparatus, medium, and terminal for assessing the feasibility of online business operations. First, offline and online business system data are collected separately. The offline and online business system data are then standardized and mapped to obtain a standardized fusion dataset. This standardized fusion dataset includes online and offline user behavior records for different business types. Second, a contrastive learning encoder with pre-trained models is used to encode each user behavior record in the standardized fusion dataset, obtaining a user behavior embedding vector corresponding to each user behavior record. All user behavior embedding vectors are then aggregated according to business type to obtain the data for each business type. The system calculates the business embedding vectors corresponding to each business type and the similarity between each business embedding vector and the anchor business embedding vector, using these similarities as business similarity features for each business type. It then sorts the similarities in descending order to obtain a similarity sequence and calculates the ordinal features for each business type based on this sequence. Finally, based on the standardized fusion dataset, it constructs user behavior aggregation features corresponding to each business type and uses the already trained business online feasibility scoring model to calculate the business online feasibility score for each business type based on the user behavior aggregation features, business embedding vectors, business similarity features, and ordinal features. Compared with existing technologies, the embodiments of this application fuse heterogeneous data from multiple sources, both offline and online, into a standardized fusion dataset through standardized mapping processing, eliminating the problem of inconsistent data definitions and laying a data foundation for subsequent cross-channel comparative analysis. Furthermore, by using a contrastive learning encoder to extract low-dimensional embedding representations of user behavior and aggregating them into business embedding vectors, it is possible to mine potential similarities between businesses from a data-driven perspective, avoiding biases caused by subjective human weighting. Finally, based on the business embedding vectors, business similarity features and ordinal features are calculated and combined with the constructed aggregated user behavior features, which are then input into the feasibility scoring model. This allows the model to simultaneously utilize the statistical features of the business itself, the similarity relationships between businesses, and ranking information, significantly improving the ability to distinguish and predict the feasibility of online operations for different businesses. Moreover, the entire evaluation process requires no human intervention and can automatically calculate the online feasibility score for each business, reducing reliance on human experience and effectively improving evaluation efficiency and accuracy. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A schematic diagram illustrating the application environment of the business online feasibility assessment method provided in the embodiments of this application is shown. Figure 2 A flowchart illustrating a method for assessing the feasibility of online business operations, as provided in an embodiment of this application, is shown. Figure 3 A flowchart illustrating the generation process of the guidance strategy provided in an embodiment of this application is shown; Figure 4 A flowchart illustrating the training process of the contrastive learning encoder provided in an embodiment of this application is shown. Figure 5 This document illustrates a flowchart of the construction process for user behavior aggregation features provided in an embodiment of this application. Figure 6 The training flowchart of the business online feasibility scoring model provided in the embodiments of this application is shown; Figure 7 This paper illustrates a block diagram of a device for evaluating the feasibility of online business operations, as provided in an embodiment of this application. Figure 8 A schematic diagram of the structure of a terminal provided in an embodiment of this application is shown. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] The method for assessing the feasibility of online business operations provided in this application embodiment can be applied to, for example... Figure 1In this application environment, the client communicates with the server via a network. The server can receive instructions from the client regarding the feasibility assessment of online business operations. It then collects data from both offline and online business systems, performs standardized mapping on these data to obtain a standardized fusion dataset. This standardized fusion dataset contains online and offline user behavior records for different business types. Using a contrastive learning encoder with pre-trained models, it encodes each user behavior record in the standardized fusion dataset to obtain a user behavior embedding vector corresponding to each record. All user behavior embedding vectors are then aggregated by business type to obtain business embedding vectors for each business type. The similarity between each business embedding vector and its anchored business embedding vector is calculated as a business similarity feature for each business type. The similarities are then sorted in descending order to obtain a similarity sequence, and the ordinal feature for each business type is calculated based on this sequence. Based on the standardized fusion dataset, user behavior aggregation features corresponding to each business type are constructed, and the standardized fusion dataset is used to construct the feasibility assessment of online business operations. The feasibility scoring model calculates the online feasibility score for each business type based on user behavior aggregation features, business embedding vectors, business similarity features, and ordinal features. The model then controls the client to output the score. In this embodiment, standardized mapping is used to fuse heterogeneous offline and online multi-source data into a standardized fusion dataset, eliminating inconsistencies in data definitions and laying a data foundation for subsequent cross-channel comparative analysis. Furthermore, a contrastive learning encoder extracts low-dimensional embedding representations of user behavior and aggregates them into business embedding vectors. This allows for data-driven analysis of potential similarities between businesses, avoiding biases caused by subjective human weighting. Finally, business similarity features and ordinal features are calculated based on the business embedding vectors and combined with the constructed user behavior aggregation features, inputting them into the feasibility scoring model. This enables the model to simultaneously utilize the statistical features of the business itself, the similarity relationships between businesses, and ranking information, significantly improving the ability to distinguish and predict the online feasibility of different businesses. Moreover, the entire evaluation process requires no manual intervention and can automatically calculate the online feasibility score for each business, reducing reliance on human experience and effectively improving evaluation efficiency and accuracy. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The following detailed description uses specific embodiments to illustrate this application.

[0015] This application provides a method for assessing the feasibility of online business operations, such as... Figure 2 As shown, the method includes: 101. Collect data from offline business systems and online business systems respectively, and perform standardized mapping processing on the offline business system data and online business system data to obtain a standardized fusion dataset.

[0016] Offline business systems include, but are not limited to, POS, CRM systems, branch monitoring systems, and voice recognition systems; offline business system data includes, but is not limited to, transaction serial number, transaction time, branch code, business type code, processing time, processing device number, operator employee number, transaction amount, transaction status, customer traffic in each area of ​​the branch, average user dwell time, peak hours, conversation start time, conversation duration, business keywords, sentiment score, and customer intent classification. Online business system data includes, but is not limited to, tracking systems, user profiling systems, and log systems; online business system data includes, but is not limited to, page browsing events, click events, business operation events, basic attribute tags, behavioral preference tags, and value stratification tags; page browsing events include, but are not limited to, page URL, page title, entry time, exit time, and page source; click events include, but are not limited to, element ID, element type, click time, and page; business operation events include, but are not limited to, business type, operation steps, operation result, and operation duration; basic attribute tags include, but are not limited to, age, gender, region, and device type; behavioral preference tags include, but are not limited to, browsing preferences, purchase preferences, and active time periods; value stratification tags include three levels: high value, medium value, and low value, which can be obtained by scoring based on the most recent business processing time, processing frequency, and amount involved using an RFM model. Standardized mapping processing can be implemented based on a pre-established standardized mapping dictionary for offline and online business types. This dictionary maintains mapping rules, with fields including: `source_type` (representing the source type, either offline or online), `source_code` (representing the source system code), `source_name` (representing the source system name), `standard_code` (representing the standard code), `standard_name` (representing the standard name), and `mapping_rule` (representing the mapping rule description). Mapping methods include one-to-one exact matching, many-to-one aggregation matching, and one-to-many split matching. Preferably, after completing standardized mapping processing, consistency verification can be performed to ensure 100% mapping coverage (number of mapped businesses / total number of businesses = 100%) and a 0% mapping conflict rate (the proportion of the same source code corresponding to multiple standard codes is 0). The standardized fusion dataset contains online and offline user behavior records for different business types.

[0017] In this embodiment, the current execution end can be the online business feasibility assessment module of the financial enterprise business management system.

[0018] Preferably, before performing standardized mapping, the collected data can be preprocessed, including but not limited to data deduplication, missing value imputation, and outlier removal. Specifically, data deduplication can employ a combined deduplication strategy based on primary keys and composite fields. Specifically, a composite primary key is formed using user ID, business type, and processing time to identify completely duplicate records and retain the latest one. For suspected duplicate records, such as multiple transactions for the same user ID within 5 minutes, a time window threshold is set, retaining only one record within the window, prioritizing records with a successful processing status. Missing value imputation can be performed using a hierarchical strategy. For example, for fields with a missing rate of less than 5%, such as processing time and satisfaction rating, random forest imputation can be used for prediction and imputation. For fields with a missing rate of more than 30%, such as detailed geographical location information, if the field is not a core feature, it is directly deleted. If it is a core feature, it is imputed with the mean or median and a missing column is marked so that the model can learn the information contained in the missing pattern itself. It should be noted that the data must be standardized before all imputation operations and then inverse standardized restoration should be performed after imputation. Outlier removal can employ a multi-method combined anomaly detection strategy. Specifically, for continuous numerical fields such as processing time and transaction amount, box plots can be used; for fields with approximately normal distributions, Z-scores can be used; and for anomalies in high-dimensional feature vectors, the isolation forest method can be used. Records identified as outliers by all three methods are marked. Outliers that can be explained by business logic, such as large transactions by VIP customers, are retained, while those that cannot be explained are deleted. The proportion of outliers handled is controlled within 5% of the total data volume; exceeding this proportion triggers a data quality review process. Furthermore, when standardizing the data, Min-Max normalization is used for continuous numerical fields to scale the data to the [0,1] interval; and RobustScaler is used for robust normalization of fields with extreme outliers.

[0019] 102. Using the contrastive learning encoder that has completed model training, encode each user behavior record in the standardized fusion dataset to obtain the user behavior embedding vector corresponding to each user behavior record. Aggregate all user behavior embedding vectors according to business type to obtain the business embedding vector corresponding to each business type. Calculate the similarity between each business embedding vector and the anchored business embedding vector to serve as the business similarity feature for each business type. Sort each similarity in descending order to obtain a similarity sequence. Calculate the ordinal feature for each business type based on the similarity sequence.

[0020] In this embodiment, firstly, a contrastive learning encoder that has completed model training is used to encode each user behavior record in the standardized fusion dataset, obtaining the user behavior embedding vector corresponding to each user behavior record. Further, all user behavior embedding vectors are grouped according to business type, with each group containing user behavior embedding vectors belonging to the same business type. The average of all user behavior embedding vectors under each business type is then taken as the business embedding vector corresponding to that business type, thus achieving aggregation of all user behavior embedding vectors by business type. Further, anchor industries are pre-selected. For services, which are known to have high feasibility for online deployment (e.g., services that have been successfully migrated and have a score > 0.8), the corresponding anchor service embedding vector is obtained, and the similarity between each service embedding vector and the anchor service embedding vector is calculated as the service similarity feature for each service type. Further, the similarity scores are sorted in descending order to obtain a similarity sequence, and the percentile of each service type is calculated, which is the ratio of the order to the total number of service types. For example, if there are 100 service types, the percentile of the 5th ranked service type is 5 / 100 = 0.05, and this percentile value is used as the order feature of the service type.

[0021] 103. Based on the standardized fusion dataset, construct user behavior aggregation features corresponding to each business type, and use the business online feasibility scoring model that has been trained to calculate the business online feasibility score for each business type based on the user behavior aggregation features, business embedding vector, business similarity features and ordinal features corresponding to each business type.

[0022] The user behavior aggregation features include at least the following types: average monthly number of transactions, online transaction rate, transaction duration, transaction time preference, transaction success rate, and repurchase rate within 7 days after transaction. The average monthly number of transactions type includes time windows such as the past 3 months, past 6 months, and past 12 months. The transaction duration type includes the mean, median, and standard deviation. The transaction time preference type includes the proportion of transactions in the morning, afternoon, and evening. After feature cross-referencing and transformation, it should be noted that in actual operation, the number of features is huge. Based on this, the user behavior aggregation features corresponding to each business type are obtained through multiple screenings in this embodiment. The screening methods include, but are not limited to, filtering, wrapping, and significance testing.

[0023] Compared with existing technologies, the embodiments of this application fuse heterogeneous data from multiple sources, both offline and online, into a standardized fusion dataset through standardized mapping processing, eliminating the problem of inconsistent data definitions and laying a data foundation for subsequent cross-channel comparative analysis. Furthermore, by using a contrastive learning encoder to extract low-dimensional embedding representations of user behavior and aggregating them into business embedding vectors, it is possible to mine potential similarities between businesses from a data-driven perspective, avoiding biases caused by subjective human weighting. Finally, based on the business embedding vectors, business similarity features and ordinal features are calculated and combined with the constructed aggregated user behavior features, which are then input into the feasibility scoring model. This allows the model to simultaneously utilize the statistical features of the business itself, the similarity relationships between businesses, and ranking information, significantly improving the ability to distinguish and predict the feasibility of online operations for different businesses. Moreover, the entire evaluation process requires no human intervention and can automatically calculate the online feasibility score for each business, reducing reliance on expert experience and effectively improving evaluation efficiency and accuracy.

[0024] In one embodiment of this application, for further definition and explanation, such as Figure 3 As shown, the method in the embodiment further includes: 201. Construct and train an online transfer probability prediction model that has completed model training.

[0025] Accordingly, step 201 of the embodiment specifically includes: constructing an initial online migration probability prediction model based on a long short-term memory network; obtaining an online migration training sample set, wherein each online migration training sample in the online migration training sample set includes a historical user behavior sequence and a real historical next channel type label; using the initial online migration probability prediction model, predicting the corresponding historical next user behavior online migration probability based on each historical user behavior sequence; calculating the online migration loss value between the historical next user behavior online migration probability and the real historical next channel type label; adjusting the model parameters of the initial online migration probability prediction model with reference to the online migration loss value, and continuing to train the model and calculate the online migration loss value until the online migration loss value converges, thereby obtaining an online migration probability prediction model with completed model training, and predicting the next user behavior online migration probability based on the online migration probability prediction model with completed model training.

[0026] The initial online migration probability prediction model includes a first long short-term memory layer, a second long short-term memory layer, a fully connected layer, and an output layer. The online migration loss function can be a binary cross-entropy function.

[0027] 202. For each user, extract all user behavior records from the standardized fusion dataset based on the current user's identity identifier, and sort all user behavior records in ascending order by timestamp to obtain the current user's user behavior sequence.

[0028] The user behavior sequence contains multiple user behavior nodes; each user behavior node includes at least a timestamp, channel type, business type code, processing time, processing result, and user profile summary; the user profile summary includes at least age segmentation, value stratification, and digital literacy score.

[0029] It should be noted that the user behavior sequence must be of a fixed length. If the length is insufficient, zero vectors can be padded at the beginning. If the length exceeds the limit, the nearest time step can be truncated.

[0030] 203. Using the online migration probability prediction model that has been trained, predict the probability of the next user behavior online migration based on the user behavior sequence of each user.

[0031] 204. Based on the preset migration probability level classification rules, the migration probability level of the corresponding users is classified according to the online migration probability of each user behavior in the next step, resulting in multiple probability level migration groups.

[0032] In this embodiment of the application, the preset migration probability level classification rule is, for example, as follows: For the high-probability migration group, the probability of the next user behavior migrating online is >0.7; For the medium probability migration group, 0.4 ≤ the probability of the next user behavior migrating online ≤ 0.7; Low-probability migration group: 0.1 ≤ probability of online migration for the next user behavior < 0.4; For the extremely low probability migration group, the probability of the next user behavior migrating online is <0.1.

[0033] Based on the above rules, users are divided into high-probability migration group, medium-probability migration group, low-probability migration group, and very low-probability migration group according to each user's online migration probability.

[0034] 205. Develop corresponding guidance strategies for each probability level migration group.

[0035] In this embodiment, the guidance strategy for the high-probability migration group includes pushing online application coupons, simplifying the online application process, and providing dedicated customer service for first-time online applications; the guidance strategy for the medium-probability migration group includes sending online application tutorial links, showcasing successful cases, and providing a hybrid application service that combines online appointment with offline assistance; and the guidance strategy for the low-probability migration group includes optimizing the offline service experience and collecting reasons for not migrating.

[0036] In addition, the effectiveness of the onboarding strategy can be verified through A / B testing. Specifically, target users are randomly divided into an experimental group and a control group. The experimental group receives the onboarding strategy, while the control group does not receive the onboarding strategy. The actual migration rate is tracked over 30 days.

[0037] In this embodiment, an online migration probability prediction model based on a long short-term memory network is used to predict the probability of a user's next online migration behavior. By capturing long-term dependencies in the user behavior sequence, the accuracy of predicting whether the user's next behavior will occur on an online channel is improved. Furthermore, corresponding guidance strategies are specified according to different migration probability levels, realizing refined user segmentation and personalized guidance strategy formulation, improving the overall migration conversion efficiency while avoiding the problem of aversion caused by blind push.

[0038] In one embodiment of this application, for further definition and explanation, such as Figure 4 As shown, before step 102 of the embodiment uses the contrastive learning encoder that has completed model training to encode each user behavior record in the standardized fusion dataset, the method of the embodiment further includes: 301. Based on the multilayer perceptron architecture, construct an initial contrastive learning encoder.

[0039] 302. Obtain the comparative learning training sample set.

[0040] 303. Using the initial contrastive learning encoder, encode the two historical user behavior records contained in the contrastive learning training samples to obtain two historical user behavior embedding vectors, and calculate the cosine similarity between the two historical user behavior embedding vectors.

[0041] 304. Construct a contrastive learning loss function with the optimization objective of the cosine similarity of positive sample pairs approaching 1 and the cosine similarity of negative sample pairs approaching -1, and calculate the contrastive learning loss value based on the cosine similarity.

[0042] 305. Adjust the model parameters of the initial contrastive learning encoder with reference to the contrastive learning loss value, and continue to train the model and calculate the contrastive learning loss value until the contrastive learning loss value converges, thus obtaining the contrastive learning encoder with completed model training, and perform encoding processing based on the contrastive learning encoder with completed model training.

[0043] The initial contrastive learning encoder comprises an input layer, a first hidden layer, a second hidden layer, and a projector head. The input layer takes user behavior records as input, with the dimension matching the number of features in the user behavior records. The first hidden layer contains 128 neurons, using a linear rectified function as the activation function, followed by batch normalization (BatchNorm) and a dropout layer (Dropout=0.3). The second hidden layer contains 64 neurons, also using a linear rectified function followed by batch normalization (BatchNorm). The projector head contains 32 neurons and outputs user behavior embedding vectors. Each contrastive learning training sample in the training sample set is a pair of samples, either positive or negative. When a pair is positive, the two historical user behavior records it contains consist of one historical online user behavior record and one historical offline user behavior record of the same business type. When a pair is negative, the two historical user behavior records it contains consist of two historical user behavior records of different business types.

[0044] In this embodiment, a contrastive learning encoder based on a multilayer perceptron architecture is used to encode user behavior records, compressing high-dimensional user behavior records into low-dimensional dense embedding vectors, thus reducing the computational complexity of subsequent processing. Furthermore, during the training process of the contrastive learning encoder, a contrastive learning framework is used, based on positive and negative sample pairs, to enable the encoder to learn that the behavioral features of the same business in online and offline channels are close to each other in the embedding space, while the behavioral features of different businesses are far apart. This allows the user behavior embedding vectors output by the trained contrastive learning encoder to automatically capture the essential similarities between businesses. Simultaneously, through positive sample pair constraints, the contrastive learning encoder is forced to map the behavior of the same business in different channels to similar positions in the embedding space, thereby learning shared semantic representations across channels and solving the channel heterogeneity problem.

[0045] In one embodiment of this application, for further definition and explanation, such as Figure 5 As shown, step 103 of the embodiment constructs user behavior aggregation features corresponding to each business type based on the standardized fusion dataset, including: 401. Divide the standardized fusion data in the standardized fusion dataset into online user behavior data subsets and offline user behavior data subsets according to channel type.

[0046] Optionally, in this embodiment of the application, before constructing the user behavior aggregation features, user behavior clustering can be performed in advance based on the standardized fusion dataset to obtain different types of user groups, and each user can be labeled with a corresponding user type label to complete the expansion of the standardized fusion dataset, so as to construct the user behavior aggregation features based on the expanded standardized fusion dataset.

[0047] 402. The first feature selection process yields multiple first-selection features.

[0048] Accordingly, step 402 of the embodiment specifically includes: for each continuous variable type feature, obtaining online user behavior data corresponding to the current continuous variable type feature from the online user behavior data subset, and obtaining offline user behavior data corresponding to the current continuous variable type feature from the offline user behavior data subset; calculating the statistic corresponding to the current continuous variable type feature based on the continuous variable type significance test formula, according to the online user behavior data and offline user behavior data corresponding to the current continuous variable type feature, and determining the significance difference index between online user behavior and offline user behavior of the current continuous variable type feature based on the statistic corresponding to the current continuous variable type feature; and for each categorical variable type feature, obtaining online user behavior data corresponding to the current categorical variable type feature from the online user behavior data subset, and obtaining offline user behavior data corresponding to the current continuous variable type feature from the online user behavior data subset. The system obtains offline user behavior data corresponding to the current categorical variable type feature from the user behavior data subset. Based on the categorical variable type significance test formula, it calculates the statistics corresponding to the current categorical variable type feature according to the online and offline user behavior data. Based on the statistics corresponding to the current categorical variable type feature, it determines the significance difference index between online and offline user behavior for the current categorical variable type feature. Based on the significance difference index between online and offline user behavior for each continuous variable type feature and the significance difference index between online and offline user behavior for each categorical variable type feature, it performs the first feature screening process on all features in the standardized fusion dataset, retaining features that show significant differences between different channels as the first screening features.

[0049] Continuous variables include, for example, processing time and satisfaction ratings. Categorized variables include, for example, service types and processing results. The significance test formula for continuous variables can be the independent samples t-test, used to compare the significance difference in the means between online and offline user behavior data corresponding to continuous variable characteristics. The significance test formula for categorical variables can be the chi-square test, used to compare the significance difference in the distributions between online and offline user behavior data corresponding to categorical variable characteristics.

[0050] In this embodiment, for each feature, online user behavior data and offline user behavior data are obtained respectively, and the corresponding significance test formula is substituted to calculate the statistic corresponding to the feature. The significance difference index, i.e., p-value, is obtained by consulting the degree of freedom table or by software automatically calculating the statistic. Finally, the significance difference index is compared with the preset significance level threshold to determine whether the difference is significant, and the features with significant differences are retained as the first screening features.

[0051] 403. Based on each first screening feature, the standardized fusion dataset is filtered to obtain the standardized fusion dataset after the first feature screening. The standardized fusion dataset after the first feature screening is divided into multiple standardized fusion dataset subsets of business types according to business type. The statistics of each first screening feature in each standardized fusion dataset subset of business type are calculated respectively.

[0052] In this embodiment, firstly, a filtering process is performed on the standardized fusion dataset, retaining only the user behavior records corresponding to each first filtering feature, resulting in a standardized fusion dataset after the first feature filtering process. Further, the standardized fusion dataset after the first feature filtering process is divided according to business type, with each business type corresponding to a business type standardized fusion dataset subset. Finally, within each business type standardized fusion dataset subset, the statistics for each feature are calculated; these can be either the mean or the median.

[0053] For example, suppose that the standardized fusion dataset after the first feature screening process contains 3 user behavior records, namely (1) credit card business, online, processing time 5 minutes, successful processing; (2) credit card business, offline, processing time 30 minutes, successful processing; (3) transfer business, online, processing time 2 minutes, successful processing. The standardized fusion data is divided into a credit card business subset, including (1) and (2), and a transfer business subset, including (3). If the first screening feature is the processing time, the statistical value of the processing time in the standardized fusion data subset of the credit card business is, average processing time = (5+30) / 2 = 17.5; the statistical value of the processing time in the standardized fusion data subset of the transfer business is, average processing time = 2.

[0054] 404. For each first screening feature, calculate the variance of the current first screening feature across all business types, and filter out the current first screening feature when the variance is lower than the preset variance threshold to complete the second feature screening process and obtain multiple second screening features.

[0055] In this embodiment of the application, continuing the example above, firstly, the mean of the average processing time for credit card business type and transfer business type is calculated as (17.5+2) / 2=9.75. Then, the variance of the average processing time for credit card business type and transfer business type is calculated as [(17.5-9.75)²+(2-9.75)²] / 2=60. Finally, it is determined whether the feature of average processing time can be retained based on the preset variance threshold.

[0056] 405. For each second screening feature, calculate the Pearson correlation coefficient between the current second screening feature and the preset target variable, and filter out the current second screening feature when the Pearson correlation coefficient is lower than the preset Pearson correlation coefficient threshold, so as to complete the third feature screening process and obtain multiple third screening features.

[0057] Among them, the preset target variable can be the feasibility score of business online transformation for business types that have completed online migration.

[0058] 406. Based on the recursive feature elimination method, a fourth feature filtering process is performed on each third filtering feature to construct the user behavior aggregate features corresponding to each business type.

[0059] Among them, the recursive feature elimination method can be based on the XGBoost base learner, which removes the least important third screening feature each time, and retrains the XGBoost base learner using the remaining third screening features. The model performance is evaluated through cross-validation until the number of features is reduced to a preset feature number threshold, so as to complete the construction of user behavior aggregated features.

[0060] In this embodiment, features irrelevant to the channel are eliminated based on significant differences, ensuring that the subsequent model only focuses on features that can truly distinguish user channel preferences, thereby improving the effectiveness of the features; redundant features that cannot distinguish the online feasibility of different business types are eliminated based on the variance threshold method; features most relevant to the online feasibility score are retained based on the correlation coefficient method; and suboptimal selection caused by ignoring feature synergy is avoided based on the recursive feature elimination method.

[0061] In one embodiment of this application, for further definition and explanation, such as Figure 6 As shown, before step 103 of the embodiment utilizes the business online feasibility scoring model that has completed model training to calculate the business online feasibility score for each business type based on the user behavior aggregation features, business embedding vectors, business similarity features, and ordinal features corresponding to each business type, the method of the embodiment further includes: 501. Based on the extreme gradient boosting model, construct an initial business online feasibility scoring model.

[0062] 502. Obtain the feasibility score training sample set.

[0063] 503. Divide the feasibility scoring training sample set into a preset number of parts, and select one part as the validation subset and the rest as the training subset.

[0064] 504. For each validation subset, use the corresponding training subset to train the initial business online feasibility scoring model in the first iteration round. Calculate the root mean square error (RMSE) of the current validation subset after the first iteration round. Repeat this process for multiple iteration rounds of model training and calculation of the corresponding RMSE of the current validation subset to construct the RMSE loss curve of the validation subset. Stop iterating when the RMSE loss value of the validation subset does not decrease in multiple consecutive iteration rounds to complete the current folded cross-validation. Take the iteration round number corresponding to the lowest point in the RMSE loss curve of the validation subset as the optimal iteration round number for the current folded cross-validation. Repeat this process for multiple folded cross-validation and determine the corresponding optimal iteration round number.

[0065] 505. Calculate the average of all optimal iteration rounds as the global optimal iteration rounds, and use the feasibility score training sample set to retrain the initial business online feasibility score model until the global optimal iteration rounds are reached, thus obtaining the business online feasibility score model with completed model training. Calculate the business online feasibility score based on the business online feasibility score model with completed model training.

[0066] The initial feasibility scoring model for online business implementation consists of multiple decision trees, each of which is a regression tree. The output is the sum of the outputs of all decision trees. Each feasibility scoring training sample in the training sample set includes historical user behavior aggregation features, historical business embedding vectors, historical business similarity features, historical ordinal features, and real historical business online feasibility scores. The number of folds is consistent with the preset number of samples.

[0067] In this embodiment of the application, for example, the preset number of samples is 5. First, the feasibility scoring training sample set can be divided into 5 subsets, namely subset 1, subset 2, subset 3, subset 4, and subset 5. Then, subset 1 is used as the validation subset, and the other 4 subsets are combined as the training subset. Furthermore, the initial business online feasibility scoring model is trained in the first iteration using this training subset, resulting in a business online feasibility scoring model that has completed the first iteration. The predicted historical business online feasibility score for each sample in the validation subset is calculated using this model, and the root mean square error (RMSE) between the predicted historical business online feasibility score and the corresponding true historical business online feasibility score is calculated. This RMSE is the current validation subset's RMSE after the first iteration. This process is repeated for multiple iterations, and the RMSE of the current validation subset is calculated. A RMSE loss curve for the validation subset is constructed. Iteration stops when the RMSE loss value of the validation subset does not decrease within 50 consecutive iterations, completing the first fold cross-validation. The iteration number corresponding to the lowest point in the RMSE loss curve of the validation subset is taken as the optimal iteration number for the first fold cross-validation. Similarly, subset 2 is used as the validation subset, and the remaining four subsets are combined as the training subset. Second-fold cross-validation is then performed to determine the optimal number of iterations for this second-fold cross-validation. This process is repeated for five-fold cross-validation. Furthermore, the average of the five optimal iteration counts is calculated as the globally optimal number of iterations. Using the feasibility score training sample set, the initial business online feasibility scoring model is retrained until the globally optimal number of iterations is reached, resulting in a fully trained business online feasibility scoring model.

[0068] Preferably, a multi-indicator comprehensive evaluation system can also be used to evaluate the feasibility scoring model for business onlineization, including but not limited to root mean square error, mean absolute error, coefficient of determination, mean absolute percentage error, etc.

[0069] In this embodiment, by dividing the feasibility score training sample set into multiple parts for multi-fold cross-validation, the evaluation bias caused by a single data partition is effectively avoided, and the generalization ability of the model is improved. Furthermore, during each fold cross-validation process, the root mean square error of the validation subset is continuously recorded. When the root mean square error loss value of the validation subset does not decrease in multiple consecutive iterations, training is stopped, and the iteration number corresponding to the lowest point in the root mean square error loss value curve of the validation subset is taken as the optimal value of that fold, preventing overfitting. At the same time, the average of the optimal iteration number of multiple folds is taken as the global optimal iteration number, and the model is retrained to that number using all training samples. This integrates the evaluation results of multiple data partitions, enhancing the stability and reliability of model training.

[0070] In specific application scenarios, based on the feasibility score of online business for each business type and the probability of the next online migration of user behavior corresponding to each user's user behavior sequence, a heat map to show the activity of user channels during a certain period, a sorted distribution map to show the online business scores of each business, a Sankey diagram to show the user migration path, and a strategy report can be generated and visualized.

[0071] The process of creating the heatmap involves aggregating and statistically analyzing user behavior data by hour and channel, calculating the number of active users per channel per hour, and performing logarithmic transformation to compress the long-tail distribution, making the heatmap color distribution more uniform. The heatmap supports mouse hover interaction, displaying specific values ​​and percentage information when hovering, and supports drill-down filtering by business type and user group. The process of creating the ranking distribution chart is as follows: First, obtain the feasibility score for online business implementation for each business type. Then, display the feasibility score for online business implementation for each business type using a horizontal bar chart, arranged from highest to lowest score. The specific score value is labeled at the top of each bar, and the business name is displayed to the right of the bar. Clicking on any bar brings up a details panel, which includes feature analysis for that business type, a SHAP feature contribution waterfall chart, and a radar chart comparing similar businesses. Feature analysis includes average processing time, success rate, and user satisfaction. The SHAP feature contribution waterfall chart displays the top 5 positive features and bottom 5 negative features that contribute the most to the score. The radar chart comparing similar businesses displays the comparison results of the 5 businesses with the closest scores for that business. The charts support filtering by business category, such as account opening, transaction, and query. The nodes of the Sankey diagram include offline business outlets, online APP modules, and churn nodes, i.e., users who have not migrated. Directed edges represent user traffic, and the width of the edge is proportional to the number of migrating users. The drawing process is to obtain the probability of the next online migration of each user's behavior sequence, bin the users according to the migration probability, and count the number of people migrating from each starting channel to the target channel. It can also dynamically switch according to user groups and time periods. The strategy report structure includes an execution summary, data overview, a business online feasibility ranking table, user migration forecast, risks and challenges, and an implementation roadmap. The execution summary includes the top 10 businesses in terms of online feasibility score and an overall migration trend forecast. The data overview includes data collection period, sample size, and coverage description. The business online feasibility ranking table includes all businesses sorted by score, including business name, score, recommended strategy, and priority level. The user migration forecast includes the number and percentage of users at each migration probability level and recommended guidance strategies. The risks and challenges include a list of businesses with an online feasibility score <0.3 and an analysis of the reasons. The implementation roadmap can be implemented in three phases: Phase 1 (1-3 months) focuses on promoting the top 10 high-feasibility businesses; Phase 2 (4-6 months) focuses on nurturing medium-feasibility businesses; and Phase 3 (7-12 months) focuses on tackling low-feasibility businesses.

[0072] This application provides a method for assessing the feasibility of online business operations. First, data from both offline and online business systems are collected. The offline and online business system data are then standardized and mapped to obtain a standardized fusion dataset. This standardized fusion dataset includes online and offline user behavior records for different business types. Second, a contrastive learning encoder with pre-trained models is used to encode each user behavior record in the standardized fusion dataset, obtaining a user behavior embedding vector corresponding to each record. All user behavior embedding vectors are then aggregated according to business type to obtain the corresponding embedding vector for each business type. The system calculates the business embedding vectors and the similarity between each business embedding vector and the anchor business embedding vector, using these similarities as business similarity features for each business type. It then sorts the similarities in descending order to obtain a similarity sequence, and calculates the ordinal features for each business type based on the similarity sequence. Finally, based on the standardized fusion dataset, it constructs user behavior aggregation features corresponding to each business type, and uses the already trained business online feasibility scoring model to calculate the business online feasibility score for each business type based on the user behavior aggregation features, business embedding vectors, business similarity features, and ordinal features. Compared with existing technologies, the embodiments of this application fuse heterogeneous data from multiple sources, both offline and online, into a standardized fusion dataset through standardized mapping processing, eliminating the problem of inconsistent data definitions and laying a data foundation for subsequent cross-channel comparative analysis. Furthermore, by using a contrastive learning encoder to extract low-dimensional embedding representations of user behavior and aggregating them into business embedding vectors, it is possible to mine potential similarities between businesses from a data-driven perspective, avoiding biases caused by subjective human weighting. Finally, based on the business embedding vectors, business similarity features and ordinal features are calculated and combined with the constructed aggregated user behavior features, which are then input into the feasibility scoring model. This allows the model to simultaneously utilize the statistical features of the business itself, the similarity relationships between businesses, and ranking information, significantly improving the ability to distinguish and predict the feasibility of online operations for different businesses. Moreover, the entire evaluation process requires no human intervention and can automatically calculate the online feasibility score for each business, reducing reliance on expert experience and effectively improving evaluation efficiency and accuracy.

[0073] Furthermore, as a response to the above Figure 2 The implementation of the method shown in this application provides an assessment device for the feasibility of online business operations, such as... Figure 7 As shown, the device includes: Data acquisition module 61, comparative learning module 62, online feasibility assessment module 63; The data acquisition module 61 is used to collect offline business system data and online business system data respectively, and to perform standardized mapping processing on the offline business system data and the online business system data to obtain a standardized fusion dataset, wherein the standardized fusion dataset contains online user behavior records and offline user behavior records of different business types. The contrastive learning module 62 is used to encode each user behavior record in the standardized fusion dataset using a contrastive learning encoder that has completed model training, to obtain user behavior embedding vectors corresponding to each user behavior record, and to aggregate all user behavior embedding vectors according to business type to obtain business embedding vectors corresponding to each business type. It also calculates the similarity between each business embedding vector and the anchored business embedding vector as the business similarity feature of each business type, sorts the similarity in descending order to obtain a similarity sequence, and calculates the ordinal feature of each business type based on the similarity sequence. The online feasibility assessment module 63 is used to construct user behavior aggregation features corresponding to each of the business types based on the standardized fusion dataset, and to calculate the online feasibility score of each of the business types based on the user behavior aggregation features, business embedding vectors, business similarity features and ordinal features corresponding to each of the business types using the business online feasibility scoring model that has been trained.

[0074] In specific application scenarios, the device further includes an online migration probability prediction module, used for: For each user, all user behavior records of the current user are extracted from the standardized fusion dataset based on the current user's identity identifier, and the all user behavior records are sorted in ascending order by timestamp to obtain the user behavior sequence of the current user, wherein the user behavior sequence contains multiple user behavior nodes; Using the online migration probability prediction model that has been trained, the probability of the next user behavior online migration is predicted according to the user behavior sequence of each user. Based on the preset migration probability level classification rules, the corresponding users are classified into multiple probability level migration groups according to the online migration probability of each user behavior in the next step. Develop appropriate guidance strategies for each of the aforementioned probability level migration groups.

[0075] In specific application scenarios, before the contrastive learning module, the device further includes a contrastive learning encoder training module, used for: An initial contrastive learning encoder is constructed based on a multilayer perceptron architecture. Obtain a contrastive learning training sample set, wherein each contrastive learning training sample in the contrastive learning training sample set is a sample pair, and the sample pair is either a positive sample pair or a negative sample pair. When the sample pair is a positive sample pair, the two historical user behavior records contained in the positive sample pair consist of one historical online user behavior record and one historical offline user behavior record of the same business type. When the sample pair is a negative sample pair, the two historical user behavior records contained in the negative sample pair consist of two historical user behavior records of different business types. Using the initial contrastive learning encoder, the two historical user behavior records contained in the contrastive learning training samples are encoded to obtain two historical user behavior embedding vectors, and the cosine similarity between the two historical user behavior embedding vectors is calculated. A contrastive learning loss function is constructed with the optimization objective of the cosine similarity of positive sample pairs approaching 1 and the cosine similarity of negative sample pairs approaching -1, and the contrastive learning loss value is calculated based on the cosine similarity. The model parameters of the initial contrastive learning encoder are adjusted according to the contrastive learning loss value, and the model is trained and the contrastive learning loss value is calculated until the contrastive learning loss value converges, thus obtaining a contrastive learning encoder with completed model training, and encoding processing is performed based on the contrastive learning encoder with completed model training.

[0076] In specific application scenarios, the online feasibility assessment module includes a user behavior aggregation feature construction unit, used for: The standardized fusion data in the standardized fusion dataset is divided into an online user behavior data subset and an offline user behavior data subset according to channel type; For each continuous variable type feature, online user behavior data corresponding to the current continuous variable type feature is obtained from the online user behavior data subset, and offline user behavior data corresponding to the current continuous variable type feature is obtained from the offline user behavior data subset. Based on the continuous variable type significance test formula, a statistic corresponding to the current continuous variable type feature is calculated according to the online and offline user behavior data. Based on the statistic corresponding to the current continuous variable type feature, a significant difference index between online and offline user behavior for the current continuous variable type feature is determined. For each categorical variable type feature, online user behavior data corresponding to the current categorical variable type feature is obtained from the online user behavior data subset, and offline user behavior data is obtained from the offline user behavior data subset. The system obtains offline user behavior data corresponding to the current categorical variable type feature from the subset, and calculates the statistic corresponding to the current categorical variable type feature based on the categorical variable type significance test formula, according to the online and offline user behavior data corresponding to the current categorical variable type feature. Based on the statistic corresponding to the current categorical variable type feature, it determines the significance difference index between online and offline user behavior of the current categorical variable type feature. Based on the significance difference index between online and offline user behavior of each continuous variable type feature and the significance difference index between online and offline user behavior of each categorical variable type feature, it performs a first feature screening process on all features in the standardized fusion dataset, retaining features with significant differences between different channels as the first screening features. Based on each of the first screening features, the standardized fusion dataset is filtered to obtain a standardized fusion dataset after the first feature screening process. The standardized fusion dataset after the first feature screening process is then divided into multiple standardized fusion dataset subsets for different business types according to business type. The statistics of each first screening feature in each of the standardized fusion dataset subsets for different business types are calculated. For each first screening feature, the variance of the current first screening feature across all business types is calculated, and the current first screening feature is filtered out when the variance is lower than a preset variance threshold, so as to complete the second feature screening process and obtain multiple second screening features. For each second screening feature, calculate the Pearson correlation coefficient between the current second screening feature and the preset target variable, and filter out the current second screening feature when the Pearson correlation coefficient is lower than the preset Pearson correlation coefficient threshold, so as to complete the third feature screening process and obtain multiple third screening features; Based on the recursive feature elimination method, a fourth feature filtering process is performed on each of the third filtering features to construct the user behavior aggregate features corresponding to each of the business types.

[0077] In specific application scenarios, prior to the online feasibility assessment module, the device further includes a user behavior clustering module, used for: Based on the standardized fusion dataset, user behavior clustering is performed to obtain different types of user groups, and each user is labeled with a corresponding user type label to expand the standardized fusion dataset. Based on the expanded standardized fusion dataset, user behavior aggregation features are constructed.

[0078] In specific application scenarios, prior to the online feasibility assessment module, the device further includes a business online feasibility scoring model training module, used for: Based on the extreme gradient boosting model, an initial business online feasibility scoring model is constructed. Obtain a feasibility score training sample set, wherein each feasibility score training sample in the feasibility score training sample set includes historical user behavior aggregation features, historical business embedding vectors, historical business similarity features, historical ordinal features, and real historical business online feasibility scores. The feasibility score training sample set is divided into a preset number of parts, and one part is selected as the verification subset, while the remaining parts are used as the training subset. For each validation subset, the initial business online feasibility scoring model is trained in the first iteration using the corresponding training subset. The root mean square error (RMSE) of the current validation subset is calculated after the first iteration. This process is repeated for multiple iterations, and the RMSE of the current validation subset is calculated to construct the RMSE loss curve of the validation subset. If the RMSE loss of the validation subset does not decrease in multiple iterations, the iteration is stopped to complete the current fold cross-validation. The iteration number corresponding to the lowest point in the RMSE loss curve of the validation subset is taken as the optimal iteration number of the current fold cross-validation. This process is repeated for multiple fold cross-validation and the corresponding optimal iteration number is determined. The number of folds is consistent with the preset number of parts. The average of all optimal iteration rounds is calculated as the global optimal iteration round number. Using the feasibility score training sample set, the initial business online feasibility score model is retrained until the global optimal iteration round number is reached, resulting in a business online feasibility score model with completed model training. The business online feasibility score is then calculated based on the business online feasibility score model with completed model training.

[0079] In specific application scenarios, prior to the online migration probability prediction module, the device further includes an online migration probability prediction model training module, used for: An initial online migration probability prediction model was constructed based on a long short-term memory network. Obtain an online migration training sample set, wherein each online migration training sample in the online migration training sample set includes a historical user behavior sequence and a real historical channel type label; Using the initial online migration probability prediction model, the probability of the next online migration of the user behavior is predicted according to each of the historical user behavior sequences. Calculate the online migration loss value between the probability of the next online migration of the user behavior in the past and the channel type tag in the real history; The model parameters of the initial online migration probability prediction model are adjusted according to the online migration loss value, and the model is trained and the online migration loss value is calculated until the online migration loss value converges, so as to obtain the online migration probability prediction model with completed model training, and predict the probability of the next user behavior online migration based on the online migration probability prediction model with completed model training.

[0080] This application provides an assessment device for the feasibility of online business operations. First, it collects data from both offline and online business systems, and performs standardized mapping processing on the offline and online data to obtain a standardized fusion dataset. This standardized fusion dataset includes online and offline user behavior records for different business types. Second, using a contrastive learning encoder that has completed model training, it encodes each user behavior record in the standardized fusion dataset to obtain a user behavior embedding vector corresponding to each user behavior record. Then, it aggregates all the user behavior embedding vectors according to business type to obtain the corresponding embedding vector for each business type. The system calculates the business embedding vectors and the similarity between each business embedding vector and the anchor business embedding vector, using these similarities as business similarity features for each business type. It then sorts the similarities in descending order to obtain a similarity sequence, and calculates the ordinal features for each business type based on the similarity sequence. Finally, based on the standardized fusion dataset, it constructs user behavior aggregation features corresponding to each business type, and uses the already trained business online feasibility scoring model to calculate the business online feasibility score for each business type based on the user behavior aggregation features, business embedding vectors, business similarity features, and ordinal features. Compared with existing technologies, the embodiments of this application fuse heterogeneous data from multiple sources, both offline and online, into a standardized fusion dataset through standardized mapping processing, eliminating the problem of inconsistent data definitions and laying a data foundation for subsequent cross-channel comparative analysis. Furthermore, by using a contrastive learning encoder to extract low-dimensional embedding representations of user behavior and aggregating them into business embedding vectors, it is possible to mine potential similarities between businesses from a data-driven perspective, avoiding biases caused by subjective human weighting. Finally, based on the business embedding vectors, business similarity features and ordinal features are calculated and combined with the constructed aggregated user behavior features, which are then input into the feasibility scoring model. This allows the model to simultaneously utilize the statistical features of the business itself, the similarity relationships between businesses, and ranking information, significantly improving the ability to distinguish and predict the feasibility of online operations for different businesses. Moreover, the entire evaluation process requires no human intervention and can automatically calculate the online feasibility score for each business, reducing reliance on expert experience and effectively improving evaluation efficiency and accuracy.

[0081] According to one embodiment of this application, a storage medium is provided, the storage medium storing at least one executable instruction, which can execute the business online feasibility assessment method in any of the above method embodiments.

[0082] Based on this understanding, the technical solution of this application can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, or portable hard drive), and includes several instructions to cause a computer device (such as a personal computer, server, or network device) to execute the methods described in the various implementation scenarios of this application.

[0083] Figure 8 The diagram shows a structural schematic of a terminal according to one embodiment of the present application. The specific embodiments of the present application do not limit the specific implementation of the terminal.

[0084] like Figure 8 As shown, the terminal may include: a processor 702, a communications interface 704, a memory 706, and a communications bus 708.

[0085] The processor 702, communication interface 704, and memory 706 communicate with each other via communication bus 708.

[0086] The communication interface 704 is used to communicate with other network elements such as clients or other servers.

[0087] The processor 702 is used to execute program 710, specifically to perform the relevant steps in the above-mentioned method embodiment for assessing the feasibility of online business operations.

[0088] Specifically, program 710 may include program code that includes computer operation instructions.

[0089] The processor 702 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The computer device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0090] Memory 706 is used to store program 710. Memory 706 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0091] Specifically, program 710 can be used to cause processor 702 to perform the following operations: Data from offline business systems and online business systems are collected separately. The offline business system data and the online business system data are then subjected to standardized mapping processing to obtain a standardized fusion dataset. The standardized fusion dataset contains online user behavior records and offline user behavior records for different business types. Using a contrastive learning encoder that has completed model training, each user behavior record in the standardized fusion dataset is encoded to obtain a user behavior embedding vector corresponding to each user behavior record. All user behavior embedding vectors are aggregated according to business type to obtain a business embedding vector corresponding to each business type. The similarity between each business embedding vector and the anchor business embedding vector is calculated as a business similarity feature for each business type. The similarity scores are sorted in descending order to obtain a similarity sequence. Based on the similarity sequence, the ordinal feature of each business type is calculated. Based on the standardized fusion dataset, user behavior aggregation features corresponding to each of the aforementioned business types are constructed respectively. Using the business online feasibility scoring model that has been trained, the business online feasibility score for each of the aforementioned business types is calculated based on the user behavior aggregation features, business embedding vectors, business similarity features, and ordinal features corresponding to each of the aforementioned business types.

[0092] It should be noted that the functions or steps that the storage medium or terminal can achieve are described in the relevant descriptions of the server side and client side in the aforementioned method embodiments. To avoid repetition, they will not be described one by one here.

[0093] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0094] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0095] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0096] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for assessing the feasibility of online business operations, characterized in that, include: Data from offline business systems and online business systems are collected separately. The offline business system data and the online business system data are then subjected to standardized mapping processing to obtain a standardized fusion dataset. The standardized fusion dataset contains online user behavior records and offline user behavior records for different business types. Using a contrastive learning encoder that has completed model training, each user behavior record in the standardized fusion dataset is encoded to obtain a user behavior embedding vector corresponding to each user behavior record. All user behavior embedding vectors are aggregated according to business type to obtain a business embedding vector corresponding to each business type. The similarity between each business embedding vector and the anchor business embedding vector is calculated as a business similarity feature for each business type. The similarity scores are sorted in descending order to obtain a similarity sequence. Based on the similarity sequence, the ordinal feature of each business type is calculated. Based on the standardized fusion dataset, user behavior aggregation features corresponding to each of the aforementioned business types are constructed respectively. Using the business online feasibility scoring model that has been trained, the business online feasibility score for each of the aforementioned business types is calculated based on the user behavior aggregation features, business embedding vectors, business similarity features, and ordinal features corresponding to each of the aforementioned business types.

2. The method according to claim 1, characterized in that, The method further includes: For each user, all user behavior records of the current user are extracted from the standardized fusion dataset based on the current user's identity identifier, and the all user behavior records are sorted in ascending order by timestamp to obtain the user behavior sequence of the current user, wherein the user behavior sequence contains multiple user behavior nodes; Using the online migration probability prediction model that has been trained, the probability of the next user behavior online migration is predicted according to the user behavior sequence of each user. Based on the preset migration probability level classification rules, the corresponding users are classified into multiple probability level migration groups according to the online migration probability of each user behavior in the next step. Develop appropriate guidance strategies for each of the aforementioned probability level migration groups.

3. The method according to claim 1, characterized in that, Before the contrastive learning encoder, which has already been trained on a model, encodes each user behavior record in the standardized fusion dataset, the method further includes: An initial contrastive learning encoder is constructed based on a multilayer perceptron architecture. Obtain a contrastive learning training sample set, wherein each contrastive learning training sample in the contrastive learning training sample set is a sample pair, and the sample pair is either a positive sample pair or a negative sample pair. When the sample pair is a positive sample pair, the two historical user behavior records contained in the positive sample pair consist of one historical online user behavior record and one historical offline user behavior record of the same business type. When the sample pair is a negative sample pair, the two historical user behavior records contained in the negative sample pair consist of two historical user behavior records of different business types. Using the initial contrastive learning encoder, the two historical user behavior records contained in the contrastive learning training samples are encoded to obtain two historical user behavior embedding vectors, and the cosine similarity between the two historical user behavior embedding vectors is calculated. A contrastive learning loss function is constructed with the optimization objective of the cosine similarity of positive sample pairs approaching 1 and the cosine similarity of negative sample pairs approaching -1, and the contrastive learning loss value is calculated based on the cosine similarity. The model parameters of the initial contrastive learning encoder are adjusted according to the contrastive learning loss value, and the model is trained and the contrastive learning loss value is calculated until the contrastive learning loss value converges, thus obtaining a contrastive learning encoder with completed model training, and encoding processing is performed based on the contrastive learning encoder with completed model training.

4. The method according to claim 1, characterized in that, Based on the standardized fusion dataset, the user behavior aggregation features corresponding to each of the aforementioned business types are constructed, including: The standardized fusion data in the standardized fusion dataset is divided into an online user behavior data subset and an offline user behavior data subset according to channel type; For each continuous variable type feature, online user behavior data corresponding to the current continuous variable type feature is obtained from the online user behavior data subset, and offline user behavior data corresponding to the current continuous variable type feature is obtained from the offline user behavior data subset. Based on the continuous variable type significance test formula, a statistic corresponding to the current continuous variable type feature is calculated according to the online and offline user behavior data. Based on the statistic corresponding to the current continuous variable type feature, a significant difference index between online and offline user behavior for the current continuous variable type feature is determined. For each categorical variable type feature, online user behavior data corresponding to the current categorical variable type feature is obtained from the online user behavior data subset, and offline user behavior data is obtained from the offline user behavior data subset. The system obtains offline user behavior data corresponding to the current categorical variable type feature from the subset, and calculates the statistic corresponding to the current categorical variable type feature based on the categorical variable type significance test formula, according to the online and offline user behavior data corresponding to the current categorical variable type feature. Based on the statistic corresponding to the current categorical variable type feature, it determines the significance difference index between online and offline user behavior of the current categorical variable type feature. Based on the significance difference index between online and offline user behavior of each continuous variable type feature and the significance difference index between online and offline user behavior of each categorical variable type feature, it performs a first feature screening process on all features in the standardized fusion dataset, retaining features with significant differences between different channels as the first screening features. Based on each of the first screening features, the standardized fusion dataset is filtered to obtain a standardized fusion dataset after the first feature screening process. The standardized fusion dataset after the first feature screening process is then divided into multiple standardized fusion dataset subsets for different business types according to business type. The statistics of each first screening feature in each of the standardized fusion dataset subsets for different business types are calculated. For each first screening feature, the variance of the current first screening feature across all business types is calculated, and the current first screening feature is filtered out when the variance is lower than a preset variance threshold, so as to complete the second feature screening process and obtain multiple second screening features. For each second screening feature, calculate the Pearson correlation coefficient between the current second screening feature and the preset target variable, and filter out the current second screening feature when the Pearson correlation coefficient is lower than the preset Pearson correlation coefficient threshold, so as to complete the third feature screening process and obtain multiple third screening features; Based on the recursive feature elimination method, a fourth feature filtering process is performed on each of the third filtering features to construct the user behavior aggregate features corresponding to each of the business types.

5. The method according to claim 1, characterized in that, Before constructing the user behavior aggregation features corresponding to each of the business types based on the standardized fusion dataset, the method further includes: Based on the standardized fusion dataset, user behavior clustering is performed to obtain different types of user groups, and each user is labeled with a corresponding user type label to expand the standardized fusion dataset. Based on the expanded standardized fusion dataset, user behavior aggregation features are constructed.

6. The method according to claim 1, characterized in that, Before calculating the feasibility score for online business development for each business type using the pre-trained business online feasibility scoring model, based on user behavior aggregation features, business embedding vectors, business similarity features, and ordinal features corresponding to each business type, the method further includes: Based on the extreme gradient boosting model, an initial business online feasibility scoring model is constructed. Obtain a feasibility score training sample set, wherein each feasibility score training sample in the feasibility score training sample set includes historical user behavior aggregation features, historical business embedding vectors, historical business similarity features, historical ordinal features, and real historical business online feasibility scores. The feasibility score training sample set is divided into a preset number of parts, and one part is selected as the verification subset, while the remaining parts are used as the training subset. For each validation subset, the initial business online feasibility scoring model is trained in the first iteration using the corresponding training subset. The root mean square error (RMSE) of the current validation subset is calculated after the first iteration. This process is repeated for multiple iterations, and the RMSE of the current validation subset is calculated to construct the RMSE loss curve of the validation subset. If the RMSE loss of the validation subset does not decrease in multiple iterations, the iteration is stopped to complete the current fold cross-validation. The iteration number corresponding to the lowest point in the RMSE loss curve of the validation subset is taken as the optimal iteration number of the current fold cross-validation. This process is repeated for multiple fold cross-validation and the corresponding optimal iteration number is determined. The number of folds is consistent with the preset number of parts. The average of all optimal iteration rounds is calculated as the global optimal iteration round number. Using the feasibility score training sample set, the initial business online feasibility score model is retrained until the global optimal iteration round number is reached, resulting in a business online feasibility score model with completed model training. The business online feasibility score is then calculated based on the business online feasibility score model with completed model training.

7. The method according to claim 2, characterized in that, Before using the online migration probability prediction model that has already been trained to predict the probability of the next user behavior online migration based on the user behavior sequence of each user, the method further includes: An initial online migration probability prediction model was constructed based on a long short-term memory network. Obtain an online migration training sample set, wherein each online migration training sample in the online migration training sample set includes a historical user behavior sequence and a real historical channel type label; Using the initial online migration probability prediction model, the probability of the next online migration of the user behavior is predicted according to each of the historical user behavior sequences. Calculate the online migration loss value between the probability of the next online migration of the user behavior in the past and the channel type tag in the real history; The model parameters of the initial online migration probability prediction model are adjusted according to the online migration loss value, and the model is trained and the online migration loss value is calculated until the online migration loss value converges, so as to obtain the online migration probability prediction model with completed model training, and predict the probability of the next user behavior online migration based on the online migration probability prediction model with completed model training.

8. A device for evaluating the feasibility of online business operations, characterized in that, include: The data acquisition module is used to collect data from offline business systems and online business systems respectively, and to perform standardized mapping processing on the offline business system data and the online business system data to obtain a standardized fusion dataset. The standardized fusion dataset contains online user behavior records and offline user behavior records of different business types. The contrastive learning module is used to encode each user behavior record in the standardized fusion dataset using a contrastive learning encoder that has completed model training, to obtain user behavior embedding vectors corresponding to each user behavior record, and to aggregate all user behavior embedding vectors according to business type to obtain business embedding vectors corresponding to each business type. The module also calculates the similarity between each business embedding vector and the anchored business embedding vector as the business similarity feature of each business type, sorts the similarity in descending order to obtain a similarity sequence, and calculates the ordinal feature of each business type based on the similarity sequence. The online feasibility assessment module is used to construct user behavior aggregation features corresponding to each of the business types based on the standardized fusion dataset, and to calculate the online feasibility score of each of the business types based on the user behavior aggregation features, business embedding vectors, business similarity features and ordinal features, using the business online feasibility scoring model that has been trained.

9. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for assessing the feasibility of business onlineization as described in any one of claims 1 to 7.

10. A terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for assessing the feasibility of online business operations as described in any one of claims 1 to 7.

Images