Training method of after-sales auditing model and after-sales auditing method and device

Abstract

The invention discloses an after-sales auditing model training method, an after-sales auditing method and an after-sales auditing device. Relates to the technical field of computers. A specific embodiment of the training method of the after-sales auditing model comprises the steps of extracting implicit features from an obtained original feature set based on an implicit interaction strategy; wherein the original feature set comprises user features and after-sales features; performing feature cross processing on the original feature set by using a predefined rule to obtain explicit features; splicing the implicit features and the explicit features to generate multi-dimensional cross features, and training based on the multi-dimensional cross features to obtain an after-sales auditing model; the dimensionality and diversity of the features can be enriched, and the feature construction efficiency is improved. A specific embodiment of the after-sales auditing method comprises the steps of obtaining to-be-processed after-sales data, and determining an auditing result of the to-be-processed after-sales data according to an after-sales auditing model; displaying the audit result of the to-be-processed after-sales data to the user; the accuracy of the auditing result can be improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a training method for a post-sales audit model, a post-sales audit method and a device. Background Art

[0002] At present, business features are generally constructed through manual analysis or by simply using crude means to construct massive features, and then the usable features are selected to train the model.

[0003] In the process of implementing the present invention, the inventors found that there are at least the following problems in the related art:

[0004] The existing model data features are insufficient, the feature construction is relatively simple and inefficient, and the audit results are less accurate. Summary of the invention

[0005] In view of this, the embodiments of the present invention provide a training method, an after-sales audit method and a device for an after-sales audit model. By performing multi-dimensional feature cross-pollination on the original data, implicit features and explicit features are constructed to obtain higher-order multi-dimensional cross-features, which can enrich the dimensions and diversity of features, improve the efficiency of feature construction, and improve the accuracy of audit results.

[0006] To achieve the above object, according to one aspect of an embodiment of the present invention, a training method for a post-sales review model is provided, comprising:

[0007] Extracting implicit features from the acquired original feature set based on an implicit interaction strategy; wherein the original feature set includes user features and after-sales features;

[0008] Use predefined rules to perform feature cross processing on the original feature set to obtain explicit features;

[0009] Implicit features and explicit features are concatenated to generate multi-dimensional cross features, and the after-sales review model is obtained based on multi-dimensional cross feature training.

[0010] Optionally, implicit features are extracted from the acquired original feature set based on an implicit interaction strategy, including:

[0011] Using the bilinear mechanism, any two original features in the acquired original feature set are subjected to outer product operation to generate a feature interaction matrix;

[0012] The original feature set is extracted based on the feature interaction matrix to obtain implicit features.

[0013] Optionally, implicit features are extracted from the acquired original feature set based on an implicit interaction strategy, including:

[0014] The self-attention mechanism is used to determine the attention weights between any original features in the acquired original feature set to obtain the weight distribution information;

[0015] The original features are divided into multiple subspaces, and feature interactions are performed in multiple subspaces based on weight distribution information to generate implicit features.

[0016] Optionally, implicit features are extracted from the acquired original feature set based on an implicit interaction strategy, including:

[0017] Using the feature selection mechanism, each original feature in the acquired original feature set is globally averaged and pooled to obtain a global description vector.

[0018] Input the global description vector into the fully connected layer to obtain the weight coefficient of each original feature;

[0019] The importance of the original features is adjusted based on the weight coefficient of each original feature, and the original feature set is extracted according to the importance to obtain implicit features.

[0020] Optionally, a post-sales review model is obtained based on multi-dimensional cross-feature training, including:

[0021] Label the multi-dimensional cross-features, build a labeled sample set, and train the after-sales review model based on the labeled sample set.

[0022] Optionally, after the after-sales review model is obtained through training based on the labeled sample set, the following steps are further included:

[0023] Generate an unlabeled sample set according to the multi-dimensional cross-features; wherein the number of samples in the unlabeled sample set is greater than the number of samples in the labeled sample set;

[0024] Input the unlabeled sample set into the after-sales review model to obtain the output results of each unlabeled sample in the unlabeled sample set;

[0025] Based on the output results of each unlabeled sample, unlabeled samples whose sample confidence exceeds a preset threshold are selected from the unlabeled sample set and marked as credible samples;

[0026] Add trusted samples to the labeled sample set to generate a training sample set, and iteratively train the after-sales review model based on the training sample set.

[0027] Optionally, after obtaining the after-sales review model based on multi-dimensional cross-feature training, the following is further included:

[0028] Use pre-configured business indicator detection mechanisms to monitor the output of the post-sales audit model in real time;

[0029] In response to the output result being less than the current business setting indicator, the after-sales review model is updated according to the acquired new user information and multi-dimensional after-sales related information by utilizing the pre-configured self-learning mechanism.

[0030] Optionally, before extracting implicit features from the acquired original feature set based on the implicit interaction strategy, the following steps are further included:

[0031] Obtain user information and multi-dimensional after-sales related information; wherein the multi-dimensional after-sales related information includes after-sales record information, product information and logistics information;

[0032] Determine the feature types of user information and multi-dimensional after-sales related information; wherein the feature types include discrete features, continuous features, and text features;

[0033] Based on the feature type, the user information and the multi-dimensional after-sales related information are encoded and converted using a preset feature conversion strategy to generate an original feature set according to the user information and the multi-dimensional after-sales related information after the encoding conversion.

[0034] According to a second aspect of an embodiment of the present invention, a post-sale review method is provided, comprising:

[0035] Obtaining after-sales data to be processed, and determining the audit result of the after-sales data to be processed according to a after-sales audit model; wherein the after-sales audit model is obtained by any method of claims 1 to 5;

[0036] The review results of the pending after-sales data are displayed to the user.

[0037] Optionally, after determining the audit result of the to-be-processed after-sales data according to the after-sales audit model, the method further includes:

[0038] The audit result is adjusted based on a pre-configured correction factor to display the adjusted audit result to the user.

[0039] According to a third aspect of an embodiment of the present invention, there is provided a training device for a post-sales audit model, comprising:

[0040] An implicit processing module is used to extract implicit features from the acquired original feature set based on an implicit interaction strategy; wherein the original data features include user features and after-sales features;

[0041] The explicit processing module is used to perform feature cross-processing on the original feature set using predefined rules to obtain explicit features.

[0042] The model training module is used to concatenate implicit features and explicit features to generate multi-dimensional cross-features, and obtain the after-sales review model based on multi-dimensional cross-feature training.

[0043] According to a fourth aspect of an embodiment of the present invention, there is provided a post-sale review device, including:

[0044] A model prediction module, used to obtain after-sales data to be processed, and determine the audit result of the after-sales data to be processed according to the after-sales audit model; wherein the after-sales audit model is obtained by any method in the first aspect of the embodiment of the present invention;

[0045] The result display module is used to display the review results of the after-sales data to be processed to the user.

[0046] According to a fifth aspect of an embodiment of the present invention, there is provided an electronic device, including:

[0047] one or more processors;

[0048] a storage device for storing one or more programs,

[0049] When one or more programs are executed by one or more processors, the one or more processors implement the method of any of the above embodiments.

[0050] According to a sixth aspect of an embodiment of the present invention, there is provided a computer-readable medium on which a computer program is stored, and when the program is executed by a processor, the method of any of the above embodiments is implemented.

[0051] According to a seventh aspect of an embodiment of the present invention, there is provided a computer program product, including a computer program, which implements the method of any of the above embodiments when the computer program is executed by a processor.

[0052] An embodiment of the above invention has the following advantages or beneficial effects: implicit features are extracted from the acquired original feature set based on an implicit interaction strategy; wherein the original feature set includes user features and after-sales features; the original feature set is subjected to feature cross-processing using predefined rules to obtain explicit features; implicit features and explicit features are spliced ​​to generate multi-dimensional cross-features, and an after-sales audit model is obtained based on multi-dimensional cross-feature training; the embodiment of the present invention constructs implicit features and explicit features by performing multi-dimensional feature cross-features on the original data to obtain higher-order multi-dimensional cross-features, which can enrich the dimension and diversity of features and improve the efficiency of feature construction. After-sales data to be processed is obtained, and the audit results of the after-sales data to be processed are determined according to the after-sales audit model; wherein the after-sales audit model is obtained by any method in the first aspect of the present invention; the audit results of the after-sales data to be processed are displayed to the user; and the accuracy of the audit results can be improved.

[0053] The further effects of the above-mentioned non-conventional optional manner will be described below in conjunction with specific implementation examples. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings are used to better understand the present invention and do not constitute an improper limitation of the present invention.

[0055] Figure 1 is a schematic diagram of the main process of the training method of the after-sales review model according to an embodiment of the present invention;

[0056] Figure 2 is an overall architecture diagram of a post-sales audit model according to a preferred embodiment of the present invention;

[0057] Figure 3 It is a flow chart of sample self-construction of a training method of a post-sales audit model according to a preferred embodiment of the present invention;

[0058] Figure 4 is a schematic diagram of a post-sale review method according to an embodiment of the present invention;

[0059] Figure 5 is a schematic diagram of main modules of a training device for a training model of a post-sales audit model according to an embodiment of the present invention;

[0060] Figure 6 is a schematic diagram of main modules of a post-sale audit device according to an embodiment of the present invention;

[0061] Figure 7 is an exemplary system architecture diagram to which embodiments of the present invention may be applied;

[0062] Figure 8 It is a schematic diagram of the structure of a computer system of a terminal device or a server suitable for implementing an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The following is a description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0064] It should be noted that the acquisition, storage and application of personal information etc. involved in the embodiments of the present invention are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0065] At present, business features are generally constructed through manual analysis or by crude means, and then the usable features are selected to train the model. That is, the existing model data features are insufficient, the feature construction is relatively simple and inefficient, and the accuracy of the audit results is poor.

[0066] In view of this, according to a first aspect of an embodiment of the present invention, a method for training an order arrival time prediction model is provided.

[0067] Figure 1 FIG. 1 is a schematic diagram of the main process of the training method of the after-sales audit model according to an embodiment of the present invention. Figure 1 As shown, the training method of the after-sales review model according to an embodiment of the present invention mainly includes the following steps S101 to S103.

[0068] Step S101, extracting implicit features from the acquired original feature set based on an implicit interaction strategy; wherein the original feature set includes user features and after-sales features.

[0069] The original feature set is obtained from the database or related system. The original feature set includes user features, after-sales features, order features, etc. Among them, user features include user personal information, historical behavior, preference data, etc., and after-sales features include customer feedback, service records, return or exchange information, etc. In order to further improve the prediction ability of the model, it is necessary to mine the implicit relationships and patterns in these original features, that is, to extract implicit features. Implicit features are implicit features and patterns extracted from the original feature set. Implicit features can be extracted through complex implicit interaction strategies or deep learning models. They are usually not easy to observe or quantify directly, but they can effectively capture potential user behavior patterns or implicit laws in after-sales services.

[0070] Implicit interaction strategies can be implemented through technologies such as deep neural networks, matrix decomposition, or self-attention mechanisms. User features and after-sales features are embedded and converted into low-dimensional vector representations. Then, through multiple nonlinear transformations of the network layer or through cross operations between features, implicit features of different dimensions are extracted; these implicit features can capture the potential relationship between users and after-sales services. For example, a deep learning model is used to analyze product features, and the model automatically finds that there is a complex relationship between "color" and "material", which may affect the popularity or sales of the product. This relationship is implicit and cannot be directly represented by a simple feature combination.

[0071] Step S102, performing feature cross processing on the original feature set using predefined rules to obtain explicit features.

[0072] Explicit features are obtained by combining or interacting existing original features. These explicit features can show the relationship between the original features, so as to better describe the internal structure of the data. Specifically, rules can be defined. For example, assuming that there are two features in the original feature set, namely user age and purchase frequency, these two features can be combined through predefined rules to generate a new feature, which represents the behavior patterns of users of different age groups at different purchase frequencies. Such cross-features can help the model identify the behavioral differences of specific user groups. In actual operation, feature crossover can be achieved through simple mathematical operations (such as multiplication and addition), logical operations (such as classification combination) or more complex functions (such as polynomial combination); for example, for a categorical feature and a numerical feature, a new categorical feature can be generated by combination, or a new discrete feature can be generated by the range of the numerical feature. This crossover processing can capture complex relationships that cannot be described by a single feature. The generated explicit features can significantly improve the prediction performance of the model and introduce important feature interaction information hidden in the original data.

[0073] Step S103, concatenating implicit features and explicit features to generate multi-dimensional cross features, and obtaining an after-sales review model based on multi-dimensional cross feature training.

[0074] In the above steps, implicit features are automatically extracted from the original data through deep learning or other complex models, and contain implicit patterns between user and after-sales features. Explicit features are obtained by cross-processing the original features through predefined rules. These explicit features can explicitly express the relationship and interaction between features. Then, the implicit features and explicit features are spliced ​​to form a comprehensive feature vector, which contains multi-dimensional feature information, including both explicit business logic and rules, and potential patterns mined from the original data. This spliced ​​multi-dimensional cross-feature more comprehensively describes the complex relationship between user behavior and after-sales service, allowing the model to better capture important information in the data. Based on these multi-dimensional cross-features, the after-sales audit model is trained. These multi-dimensional cross-features can be input into a machine learning model, such as a decision tree, support vector machine or deep neural network. By learning the mapping relationship between these multi-dimensional cross-features and the after-sales audit results, the model can automatically generate the audit results. In this way, the model can not only handle complex data relationships, but also effectively improve the accuracy and efficiency of the audit. The splicing of implicit features and explicit features not only enriches the input information of the model, but also enhances the model's ability to capture potential patterns and explicit rules in after-sales audits, making the final generated after-sales audit model have higher expressiveness and predictive ability.

[0075] The embodiment of the present invention extracts implicit features from the acquired original feature set based on an implicit interaction strategy; wherein the original feature set includes user features and after-sales features; performs feature cross-processing on the original feature set using predefined rules to obtain explicit features; concatenates implicit features and explicit features to generate multi-dimensional cross-features, and obtains a after-sales review model based on multi-dimensional cross-feature training; the embodiment of the present invention performs multi-dimensional feature cross-processing on original data to construct implicit features and explicit features to obtain higher-order multi-dimensional cross-features, which can enrich the dimension and diversity of features and improve the efficiency of feature construction.

[0076] Optionally, implicit features are extracted from the acquired original feature set based on an implicit interaction strategy, including: using a bilinear mechanism to perform an outer product operation on any two original features in the acquired original feature set to generate a feature interaction matrix; and extracting the original feature set based on the feature interaction matrix to obtain implicit features.

[0077] The bilinear mechanism generates a feature interaction matrix by calculating the outer product of two features. This matrix can capture the high-order interaction information between the two features. For example, there are two original features, namely user behavior features and product features. These two features can be represented as vectors respectively. Through the bilinear mechanism, these two vectors are outer-producted to obtain a feature interaction matrix. The generated feature interaction matrix contains a large amount of high-dimensional information about the relationship between the original features. By further processing this feature interaction matrix, such as using deep neural networks or other dimensionality reduction techniques, useful information can be extracted and marked as implicit features. The bilinear mechanism significantly enhances the learning ability of the model, enabling the model to better adapt to and process complex relationships in the original data set.

[0078] Optionally, implicit features are extracted from the acquired original feature set based on an implicit interaction strategy, including: using a self-attention mechanism to determine the attention weights between any original features in the acquired original feature set to obtain weight distribution information; dividing the original features into multiple subspaces, and performing feature interactions in multiple subspaces based on the weight distribution information to generate implicit features.

[0079] The self-attention mechanism analyzes each feature in the original feature set pair by pair to determine the attention weights between them. These weights reflect the importance of a feature to other features in the current context, that is, which features have a greater impact on the final decision of the model. By calculating these attention weights, a weight distribution information can be obtained, which describes the relationship structure within the original feature set. Divide the original features into multiple subspaces to analyze the interaction relationship between the original features in a more fine-grained manner. In each subspace, the original features may exhibit different interaction patterns. Therefore, by dispersing the original features into multiple subspaces, the implicit information in the data can be captured more comprehensively.

[0080] Optionally, implicit features are extracted from the acquired original feature set based on an implicit interaction strategy, including: using a feature selection mechanism to perform global average pooling on each original feature in the acquired original feature set to obtain a global description vector; inputting the global description vector into a fully connected layer to obtain a weight coefficient for each original feature; adjusting the importance of the original feature based on the weight coefficient of each original feature, and extracting the original feature set according to the importance to obtain implicit features.

[0081] In the feature selection mechanism, each original feature in the original feature set is first subjected to global average pooling. Global average pooling is to compress and summarize the information of each original feature on the entire data set to generate a global description vector. This vector can represent the performance of the feature in the global scope and capture the overall trend and importance of the feature. This global description vector is input into the fully connected layer, which performs a linear transformation on the input vector and outputs the weight coefficient of each original feature. These weight coefficients reflect the importance of each original feature relative to other original features in the entire original feature set. The larger the weight, the greater the contribution of the original feature to the final decision of the model. Based on these weight coefficients, the importance of the original features is adjusted. Specifically, each original feature is weighted according to its weight coefficient to increase or decrease its influence in the original feature set. This adjustment process can effectively highlight the features that play a key role in model prediction, while reducing or ignoring those relatively unimportant features.

[0082] Optionally, the after-sales review model is obtained based on multi-dimensional cross-feature training, including: adding labels to the multi-dimensional cross-features, constructing a labeled sample set, and obtaining the after-sales review model based on the labeled sample set training.

[0083] Appropriate labels are added to the obtained multidimensional cross features. The labels represent the results or categories that the model needs to predict in the end. For example, in the after-sales review scenario, the labels represent the predicted probability of a transaction or service request, whether there is a risk, or whether it meets specific review standards. The process of adding labels can be based on existing business data or manual annotation. For example, based on historical after-sales data, the correct labels can be added to each sample according to the known review results or expert judgments. These labels help the model learn how to infer the correct results from the input features. The multidimensional cross features with labels are combined into a labeled sample set. The labeled sample set contains all the information required for model training, that is, the input features of each labeled sample and its corresponding correct label. The labeled sample set is input into the machine learning algorithm. By adjusting the model parameters, the model can predict the labels as accurately as possible. During the training process, the model is continuously optimized to minimize the difference between the predicted results and the true labels. Finally, the fully trained model can make accurate predictions on new data and help automate the after-sales review task. The embodiment of the present invention utilizes the after-sales audit model trained with the labeled sample set to effectively identify potential risks and problems, improve the quality and efficiency of after-sales service, and not only reduce the burden of manual audit, but also significantly improve the accuracy and consistency of the audit.

[0084] Optionally, after obtaining the after-sales audit model based on the labeled sample set training, it also includes: generating an unlabeled sample set based on multidimensional cross-features; wherein the number of samples in the unlabeled sample set is greater than the number of samples in the labeled sample set; inputting the unlabeled sample set into the after-sales audit model to obtain the output results of each unlabeled sample in the unlabeled sample set; based on the output results of each unlabeled sample, screening out unlabeled samples whose sample confidence exceeds a preset threshold from the unlabeled sample set, and marking them as credible samples; adding the credible samples to the labeled sample set to generate a training sample set, and iteratively training the after-sales audit model based on the training sample set.

[0085] Specifically, an unlabeled sample set is generated based on multi-dimensional cross features, and the number of samples in the unlabeled sample set is much larger than that in the labeled sample set. These unlabeled samples are input into the after-sales audit model, and the model predicts each unlabeled sample and outputs the corresponding result. Each output result is accompanied by a confidence value, which indicates the confidence of the model in the prediction result. The higher the confidence, the higher the accuracy of the prediction result. Based on these output results, the unlabeled sample set is screened. Specifically, unlabeled samples with confidence exceeding the preset threshold are selected. Although these high-confidence samples have not been manually labeled, they can be regarded as "automatically labeled" samples with high credibility because the model is very confident in their prediction results. These credible samples are added to the original labeled sample set to form a new training sample set. This new training sample set contains more information, including not only the originally labeled samples, but also the automatically labeled samples that the model believes to have high credibility. The after-sales review model is iteratively trained using this expanded new training sample set. Through this iterative training, the model can learn from more data and continuously optimize its own prediction capabilities. Each iteration further improves the generalization ability of the model, allowing more accurate judgments and predictions when processing new data.

[0086] To address the problem of insufficient data samples, an embodiment of the present invention constructs an intelligent audit model through a sample quality improvement method based on self-construction of samples. It can address problems such as difficulty in labeling data for after-sales audits and the small number of labeled samples due to high time costs. By utilizing the model's own learning ability, based on a small number of labeled samples and combined with the predicted sample confidence, the learning model is continuously iterated and optimized, so that the model can learn from a small number of labeled samples to obtain reasonable after-sales audit results, greatly reducing the cost and time of labeling samples.

[0087] Optionally, after obtaining the after-sales review model based on multi-dimensional cross-feature training, it also includes: using a pre-configured business indicator detection mechanism to monitor the output results of the after-sales review model in real time; in response to the output result being less than the current business setting indicator, using a pre-configured self-learning mechanism to update the after-sales review model based on the acquired new user information and multi-dimensional after-sales related information.

[0088] The results of real-time detection of core business monitoring indicators are used as self-learning capability driving signals. Among them, core business monitoring indicators are the market indicators of after-sales model review, including model accuracy, manual sampling accuracy, model review ratio, etc. When the business indicator detection mechanism or business indicator detection model detects that it is lower than the current business setting indicator, self-learning is triggered, and model learning and updating are performed according to new business data containing new business logic. For example, recent new data processing and processing and data set division are performed, and the newly trained model is verified by the test set before the model is launched and updated. This process is an automated process without manual intervention to achieve self-update and learning of the model. In addition, scheduled updates can be performed, so that when the business indicators are relatively normal, the model will still be learned and updated to try to ensure that the model has new business knowledge and business capabilities. Finally, the self-supervised learning ability of the after-sales review prediction model is realized, which greatly improves the review ability of the after-sales review model.

[0089] Optionally, before extracting implicit features from the acquired original feature set based on the implicit interaction strategy, it also includes: acquiring user information and multi-dimensional after-sales related information; wherein the multi-dimensional after-sales related information includes after-sales record information, product information and logistics information; determining feature types of user information and multi-dimensional after-sales related information; wherein the feature types include discrete features, continuous features and text features; based on the feature type, encoding and converting the user information and multi-dimensional after-sales related information using a pre-set feature conversion strategy, so as to generate an original feature set based on the user information and multi-dimensional after-sales related information after the encoding and conversion processing.

[0090] User information includes the user's personal information, purchase history, behavior records, etc. Multi-dimensional after-sales related information includes after-sales record information, product information, and logistics information. Specifically, the above user information and multi-dimensional after-sales related information can be obtained from the enterprise's database or real-time system. Then, the feature types of this information are determined. The feature types include discrete features, continuous features, and text features; discrete features include categorical variables, such as product category, user gender, etc.; continuous features are numerical variables, such as purchase amount, logistics time, etc.; text features are descriptive text data, such as customer feedback or product descriptions, etc. Based on these feature types, the information is encoded and converted using a pre-set feature conversion strategy. For discrete features, unique hot encoding or label encoding can be used; continuous features can be normalized or standardized; and text features can be encoded using word embedding or other natural language processing techniques. Through these conversions, the data is converted into a numerical form that can be directly used by the model. The user information and multi-dimensional after-sales related information that have been processed by encoding conversion are combined together to generate the original feature set, which will be used as the input of the model for subsequent analysis and prediction tasks. In this way, the diversity of the data is retained, while the efficiency and accuracy of the model in processing data are improved.

[0091] Figure 2 1 is an overall architecture diagram of the after-sales audit model according to a preferred embodiment of the present invention. Figure 2As shown in the figure, the architecture uses multiple feature processing methods to generate the final output. Specifically, the input features are divided into four categories: sparse features (sparse_sparse), dense features (dense), text features (textfeature), and sequence features (seq_feature). Each feature type is processed by a different embedding layer, such as sparse features through "sparse_embedding", dense features through "dense_embedding", text features through "text_embedding", and sequence features through "seq_embedding", converting the original features into low-dimensional vector representations suitable for model use. The embedded features are processed by different implicit interaction strategies. "sparse_embedding" passes through a bilinear interaction layer called "BilinearInteraction" or an attention mechanism interaction layer called "InteractingLayer", combined with "SENET" (Squeeze-and-Excitation, feature selection) to adjust the importance weights of the features. In addition, "dense" is processed by "Deeplayer" to generate feature representation; "textfeature" is processed by a "Textcnn" to capture the contextual relationship in the text; "seq_feature" is processed by "Reduce_mean" to reduce the dimension or average the sequence features. After these different processing layers, all features are "concat" together to form a unified representation. These concatenated features will be input into the two sub-networks of "deep layer" and "CIN" (Compressed Interaction Network, feature interaction network structure), which are used to extract high-order features and cross-features respectively. Finally, the output of "deep layer" and "CIN" passes through the "dense" layer to output the final prediction result of the model.

[0092] The input data of the after-sales audit model includes multiple features in multiple dimensions. The embodiment of the present invention embeds the feature structure into the model to achieve efficient cross-feature and extract rich relevant information in more dimensions. Feature cross-feature mainly occurs in Figure 2The bilinear part of BilinearInteraction in the above method changes the previous inner product or Hadamard product feature interaction mode, combines the advantages of both to jointly maintain a set of matrices to achieve feature interaction; while the self-attention part mainly realizes feature interaction in multiple different subspaces, pays attention to the information of features in different subspaces, and enriches the diversity of feature information. At the same time, this paper introduces the SENET network model to dynamically learn the importance of each feature and pay more attention to important features. In addition, the input data can also realize explicit feature interaction through the CIN model, realizing feature crossover from another perspective. Finally, the feature data is spliced ​​to further increase the diversity of features, making the information learned by the model richer, thereby improving the model effect.

[0093] Figure 3 1 is a flow chart of sample self-construction of the training method of the after-sales audit model according to the preferred embodiment of the present invention. The method of sample self-construction in the embodiment of the present invention realizes the model effect that meets the requirements of after-sales intelligent audit from a small number of samples. Sample self-construction is similar to sample enhancement, except that its "enhanced samples" are all from actual real data that has been judged by the model, such as Figure 3 As shown in the figure, Labeled Data is a small amount of labeled sample data. First, this part is used to train the after-sales audit model. Then the trained model M (Model M) is used to predict a large amount of unlabeled data (Unlabeled Data), and the samples with higher confidence in the prediction results are marked as trusted samples (Trust Part) data sample sets, and then mixed with the labeled real samples to form a new training set (New Mixture Data), retrain the model M' (Model M'), and then re-predict the labeled data as model M, select samples with higher confidence, and re-mix with the previous training set to form a new training set iterative model. Repeat this process until the model performance no longer improves, indicating that the sample construction process is completed. Using this sample construction method, you can save a lot of tedious and inefficient labeling processes, and still get a more ideal model to handle after-sales business.

[0094] According to a second aspect of an embodiment of the present invention, a post-sale review method is provided.

[0095] Figure 4 Schematic diagram of the main process of the order scheduling method according to an embodiment of the present invention; Figure 4 As shown, the order scheduling method according to the embodiment of the present invention mainly includes the following steps S401 to S402.

[0096] Step S401, obtaining after-sales data to be processed, and determining the audit result of the after-sales data to be processed according to a after-sales audit model; wherein the after-sales audit model is obtained by any method in the first aspect of the embodiment of the present invention.

[0097] Step S402: Display the review result of the after-sales data to be processed to the user.

[0098] During the after-sales audit, the after-sales data to be processed is first obtained from the enterprise's database or after-sales management system. The after-sales data includes the user's after-sales request, product return information, and related logistics records. The pre-trained after-sales audit model is used to audit these pending data. The after-sales audit model has been trained with a large amount of historical data and can make judgments based on the input features to determine the audit results of each after-sales request. This result can be audit pass, audit failure, or further manual audit, or after-sales prediction probability, etc. The system displays these audit results to the user. The display method can intuitively present the audit conclusions through the user interface (UI) or notification system. Users can view the detailed audit information of each after-sales request in the system, including the basis for the model to make the conclusion. This display method allows users to quickly understand the processing status of each after-sales request and make corresponding decisions or take further actions. Through this automated audit and result display method, the efficiency and accuracy of after-sales audit can be greatly improved.

[0099] Optionally, after determining the audit result of the after-sales data to be processed according to the after-sales audit model, it also includes: adjusting the audit result based on a pre-configured correction coefficient to display the adjusted audit result to the user.

[0100] Correction coefficients can be pre-set according to business rules or policies to fine-tune the output results of the model. For example, some high-risk categories of goods may require stricter audit standards, and correction coefficients can be used to reduce the probability of these goods passing the audit. On the contrary, for low-risk or high-quality customers, correction coefficients can increase the probability of passing the audit, thereby optimizing the user experience. After the model obtains preliminary audit results, these results can be adjusted by the correction coefficients, and the results output by the model are combined with the correction coefficients to generate the final audit results. For example, if the model outputs a certain after-sales request with a pass probability of 80%, and the correction coefficient is set to 0.9, the final pass probability is adjusted to 72%. This adjusted result can not only improve the accuracy of the audit, but also better support the business strategy of the enterprise, making the audit process more effective and flexible in actual applications.

[0101] According to a third aspect of an embodiment of the present invention, a training device for a post-sales review model is provided.

[0102] Figure 5is a schematic diagram of the main modules of the training device of the after-sales audit model according to an embodiment of the present invention, such as Figure 5 As shown, a training device 500 for a post-sales audit model includes:

[0103] The implicit processing module 501 is used to extract implicit features from the acquired original feature set based on the implicit interaction strategy; wherein the original data features include user features and after-sales features;

[0104] The explicit processing module 502 is used to perform feature cross processing on the original feature set using predefined rules to obtain explicit features.

[0105] The model training module 503 is used to concatenate implicit features and explicit features to generate multi-dimensional cross features, and obtain the after-sales review model based on the multi-dimensional cross feature training.

[0106] Optionally, the implicit processing module 501 is further used for:

[0107] Use the bilinear mechanism to perform outer product operation on any two original features in the acquired original feature set to generate a feature interaction matrix;

[0108] The original feature set is extracted based on the feature interaction matrix to obtain implicit features.

[0109] Optionally, the implicit processing module 501 is further used for:

[0110] The self-attention mechanism is used to determine the attention weights between any original features in the acquired original feature set to obtain weight distribution information;

[0111] The original features are divided into multiple subspaces, and feature interactions are performed in multiple subspaces based on weight distribution information to generate implicit features.

[0112] Optionally, the implicit processing module 501 is further used for:

[0113] Using the feature selection mechanism, each original feature in the acquired original feature set is globally averaged and pooled to obtain a global description vector.

[0114] Input the global description vector into the fully connected layer to obtain the weight coefficient of each original feature;

[0115] The importance of the original feature is adjusted based on the weight coefficient of each original feature, and the original feature set is extracted according to the importance to obtain implicit features.

[0116] Optionally, the model training module 503 is further used to:

[0117] Label the multi-dimensional cross-features, build a labeled sample set, and train the after-sales review model based on the labeled sample set.

[0118] Optionally, the training device 500 further includes a sample self-building module, and the sample self-building module is used to:

[0119] Generate an unlabeled sample set according to the multi-dimensional cross-features; wherein the number of samples in the unlabeled sample set is greater than the number of samples in the labeled sample set;

[0120] Input the unlabeled sample set into the after-sales review model to obtain the output results of each unlabeled sample in the unlabeled sample set;

[0121] Based on the output results of each unlabeled sample, unlabeled samples whose sample confidence exceeds a preset threshold are selected from the unlabeled sample set and marked as credible samples;

[0122] Add trusted samples to the labeled sample set to generate a training sample set, and iteratively train the after-sales review model based on the training sample set.

[0123] Optionally, the training device 500 further includes a self-learning module, and the self-learning module is used to:

[0124] Use pre-configured business indicator detection mechanisms to monitor the output of the post-sales audit model in real time;

[0125] In response to the output result being less than the current business setting indicator, the after-sales review model is updated according to the acquired new user information and multi-dimensional after-sales related information by utilizing the pre-configured self-learning mechanism.

[0126] Optionally, the training device 500 further includes a feature set generation module, and the feature set generation module is used to:

[0127] Obtain user information and multi-dimensional after-sales related information; wherein the multi-dimensional after-sales related information includes after-sales record information, product information and logistics information;

[0128] Determine the feature types of user information and multi-dimensional after-sales related information; wherein the feature types include discrete features, continuous features, and text features;

[0129] Based on the feature type, the user information and the multi-dimensional after-sales related information are encoded and converted using a preset feature conversion strategy to generate an original feature set according to the user information and the multi-dimensional after-sales related information after the encoding conversion.

[0130] It should be noted that the specific implementation content of the training device for the after-sales audit model in the embodiment of the present invention has been described in detail in the above training method for the after-sales audit model, so the repeated content will not be described here.

[0131] According to a fourth aspect of an embodiment of the present invention, a post-sale review device is provided.

[0132] Figure 6 is a schematic diagram of the main modules of the after-sales audit device according to an embodiment of the present invention, such as Figure 6 As shown, a post-sale audit device 600 mainly includes:

[0133] The model prediction module 601 is used to obtain the after-sales data to be processed, and determine the audit result of the after-sales data to be processed according to the after-sales audit model; wherein the after-sales audit model is obtained by any method in the first aspect of the embodiment of the present invention;

[0134] The result display module 602 is used to display the review results of the after-sales data to be processed to the user.

[0135] Optionally, the auditing device 600 further includes a result correction module, which is used to:

[0136] The audit result is adjusted based on a pre-configured correction factor to display the adjusted audit result to the user.

[0137] It should be noted that the specific implementation content of the after-sales audit device in the embodiment of the present invention has been described in detail in the above after-sales audit method, so the repeated content will not be described again here.

[0138] According to a fifth aspect of an embodiment of the present invention, there is provided an electronic device, including:

[0139] one or more processors;

[0140] a storage device for storing one or more programs,

[0141] When one or more programs are executed by one or more processors, the one or more processors implement the method provided by the first aspect of the embodiment of the present invention and / or the second aspect of the embodiment of the present invention.

[0142] According to a sixth aspect of an embodiment of the present invention, a computer-readable medium is provided, on which a computer program is stored. When the program is executed by a processor, the method provided by the first aspect of the embodiment of the present invention and / or the second aspect of the embodiment of the present invention is implemented.

[0143] According to a seventh aspect of an embodiment of the present invention, a computer program product is provided, including a computer program, which, when executed by a processor, implements the method provided by the first aspect of the embodiment of the present invention and / or the second aspect of the embodiment of the present invention.

[0144] Figure 7An exemplary system architecture 700 is shown to which the training method for a post-sale audit model or the training apparatus for a post-sale audit model according to an embodiment of the present invention can be applied.

[0145] like Figure 7 As shown, system architecture 700 may include terminal devices 701, 702, 703, network 704 and server 705. Network 704 is used to provide a medium for communication links between terminal devices 701, 702, 703 and server 705. Network 704 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0146] The user can use the terminal devices 701, 702, 703 to interact with the server 705 through the network 704 to receive or send messages, etc. Various communication client applications can be installed on the terminal devices 701, 702, 703, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc. (only for example).

[0147] The terminal devices 701 , 702 , and 703 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, and desktop computers.

[0148] The server 705 may be a server that provides various services, such as a backend management server (only an example) that provides support for shopping websites browsed by users using the terminal devices 701, 702, and 703. The backend management server may analyze and process the received data such as the model training request, and feed back the processing result (after-sales review model - only an example) to the terminal device.

[0149] It should be noted that the training method of the order arrival time prediction model provided in the embodiment of the present invention is generally run by the server 705 , and accordingly, the training device of the order arrival time prediction model is generally set in the server 705 .

[0150] It should be noted that the after-sales review method provided in the embodiment of the present invention is generally run by the server 705 , and accordingly, the after-sales review device is generally disposed in the server 705 .

[0151] It should be understood that Figure 7 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0152] Reference below Figure 8 , which shows a schematic diagram of the structure of a computer system 800 of a terminal device suitable for implementing an embodiment of the present invention. Figure 8The terminal device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0153] like Figure 8 As shown, the computer system 800 includes a central processing unit (CPU) 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage part 808 to a random access memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the system 800 are also stored. The CPU 801, the ROM 802, and the RAM 803 are connected to each other through a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0154] The following components are connected to the I / O interface 805: an input section 806 including a keyboard, a mouse, etc.; an output section 807 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, a modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the I / O interface 805 as needed. A removable medium 811, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 810 as needed, so that a computer program read therefrom is installed into the storage section 808 as needed.

[0155] In particular, according to the embodiments disclosed in the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program contains program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 809, and / or installed from the removable medium 811. When the computer program is executed by the central processing unit (CPU) 801, the above-mentioned functions defined in the system of the present invention are executed.

[0156] It should be noted that the computer-readable medium shown in the present invention may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present invention, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0157] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0158] The modules involved in the embodiments of the present invention may be implemented in software or hardware. The modules described may also be arranged in a processor, for example, may be described as: a processor including an implicit processing module, an explicit processing module and a model training module. The names of these modules do not, in some cases, constitute a limitation on the module itself. For example, the implicit processing module may also be described as "a module for extracting implicit features from the acquired original feature set based on an implicit interaction strategy". Alternatively, a processor includes a model prediction module and a result display module, wherein the names of these modules do not, in some cases, constitute a limitation on the module itself. For example, the result display module may also be described as "a module for displaying the audit results of the after-sales data to be processed to the user".

[0159] As another aspect, the present invention further provides a computer-readable medium, which may be included in the device described in the above embodiment; or it may exist independently and not be assembled into the device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by a device, the device implements the following method: extracting implicit features from the acquired original feature set based on implicit interaction strategies; wherein the original feature set includes user features and after-sales features; performing feature cross-processing on the original feature set using predefined rules to obtain explicit features; splicing the implicit features and the explicit features to generate multi-dimensional cross-features, and obtaining an after-sales review model based on multi-dimensional cross-feature training. Alternatively, the device implements the following method: obtaining after-sales data to be processed, and determining the review results of the after-sales data to be processed according to the after-sales review model; wherein the after-sales review model is obtained by any method in the first aspect of the present invention; and displaying the review results of the after-sales data to be processed to the user

[0160] The computer program product provided by an embodiment of the present invention includes a computer program, which, when executed by a processor, implements the training method of the after-sales review model in the embodiment of the first aspect of the present invention and / or the after-sales review method in the embodiment of the second aspect of the present invention.

[0161] According to the technical solution of the embodiment of the present invention, it has the following advantages or beneficial effects: implicit features are extracted from the acquired original feature set based on implicit interaction strategy; wherein the original feature set includes user features and after-sales features; the original feature set is subjected to feature cross-processing using predefined rules to obtain explicit features; implicit features and explicit features are spliced ​​to generate multi-dimensional cross-features, and a after-sales audit model is obtained based on multi-dimensional cross-feature training; the embodiment of the present invention constructs implicit features and explicit features by performing multi-dimensional feature cross-features on the original data to obtain higher-order multi-dimensional cross-features, which can enrich the dimension and diversity of features and improve the efficiency of feature construction. After-sales data to be processed are obtained, and the audit results of the after-sales data to be processed are determined according to the after-sales audit model; wherein the after-sales audit model is obtained by any method in the first aspect of the present invention; the audit results of the after-sales data to be processed are displayed to the user; and the accuracy of the audit results can be improved.

[0162] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0163] It should be noted that the acquisition, storage and application of user personal information involved in the technical solution of the present disclosure are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.

Claims

1. A training method for a post-sales audit model, characterized in that: include: Extracting implicit features from the acquired original feature set based on an implicit interaction strategy; wherein the original feature set includes user features and after-sales features; Performing feature cross processing on the original feature set using predefined rules to obtain explicit features; The implicit features and explicit features are concatenated to generate multi-dimensional cross features, and an after-sales review model is obtained by training based on the multi-dimensional cross features.

2. The method according to claim 1, characterized in that Based on the implicit interaction strategy, implicit features are extracted from the original feature set, including: Using the bilinear mechanism, any two original features in the acquired original feature set are subjected to outer product operation to generate a feature interaction matrix; The original feature set is extracted based on the feature interaction matrix to obtain implicit features.

3. The method according to claim 1, characterized in that Based on the implicit interaction strategy, implicit features are extracted from the original feature set, including: The self-attention mechanism is used to determine the attention weights between any original features in the acquired original feature set to obtain the weight distribution information; The original features are divided into a plurality of subspaces, and feature interactions are performed in the plurality of subspaces based on the weight distribution information to generate implicit features.

4. The method according to claim 1, characterized in that: Based on the implicit interaction strategy, implicit features are extracted from the original feature set, including: Using the feature selection mechanism, each original feature in the acquired original feature set is globally averaged and pooled to obtain a global description vector. Inputting the global description vector into a fully connected layer to obtain a weight coefficient for each original feature; The importance of each original feature is adjusted based on a weight coefficient of each original feature, and the original feature set is extracted according to the importance to obtain an implicit feature.

5. The method according to claim 1, characterized in that The after-sales review model is obtained based on the multi-dimensional cross-feature training, including: Labels are added to the multi-dimensional cross features to construct a labeled sample set, and a post-sales review model is obtained by training based on the labeled sample set.

6. The method according to claim 5, characterized in that After the after-sales review model is obtained through training based on the labeled sample set, the method further includes: Generate an unlabeled sample set according to the multi-dimensional cross-features; wherein the number of samples in the unlabeled sample set is greater than the number of samples in the labeled sample set; Inputting the unlabeled sample set into the after-sales review model to obtain output results of each unlabeled sample in the unlabeled sample set; Based on the output results of the respective unlabeled samples, unlabeled samples whose sample confidence exceeds a preset threshold are screened out from the unlabeled sample set, and marked as credible samples; The credible samples are added to the annotated sample set to generate a training sample set, and the after-sales audit model is iteratively trained according to the training sample set.

7. The method according to claim 1, characterized in that After the after-sales review model is obtained based on the multi-dimensional cross-feature training, it also includes: Using a pre-configured business indicator detection mechanism to monitor the output results of the after-sales review model in real time; In response to the output result being less than the current business setting indicator, the after-sales review model is updated according to the acquired new user information and multi-dimensional after-sales related information by using a pre-configured self-learning mechanism.

8. The method according to claim 1, characterized in that Before extracting implicit features from the acquired original feature set based on the implicit interaction strategy, it also includes: Acquire user information and multi-dimensional after-sales related information; wherein the multi-dimensional after-sales related information includes after-sales record information, product information and logistics information; Determine the feature types of the user information and multi-dimensional after-sales related information; wherein the feature types include discrete features, continuous features, and text features; The user information and multi-dimensional after-sales related information are transcoded based on the feature type using a preset feature conversion strategy to generate an original feature set based on the transcoded user information and multi-dimensional after-sales related information.

9. A post-sale audit method, characterized in that: include: Obtaining after-sales data to be processed, and determining the audit result of the after-sales data to be processed according to a after-sales audit model; wherein the after-sales audit model is obtained by any method described in claims 1 to 5; The review results of the after-sales data to be processed are displayed to the user.

10. The method according to claim 9, characterized in that After determining the audit result of the after-sales data to be processed according to the after-sales audit model, the method further includes: The audit result is adjusted based on a preconfigured correction factor to display the adjusted audit result to the user.

11. A training device for an after-sales audit model, characterized in that: include: An implicit processing module, used to extract implicit features from the acquired original feature set based on an implicit interaction strategy; wherein the original data features include user features and after-sales features; The explicit processing module is used to perform feature cross processing on the original feature set using predefined rules to obtain explicit features. The model training module is used to splice the implicit features and explicit features to generate multi-dimensional cross features, and obtain the after-sales review model based on the multi-dimensional cross feature training.

12. A post-sale audit device, characterized in that: include: A model prediction module, used for obtaining after-sales data to be processed, and determining the audit result of the after-sales data to be processed according to the after-sales audit model; wherein the after-sales audit model is obtained by any method described in claims 1 to 8; The result display module is used to display the review result of the after-sales data to be processed to the user.

13. An electronic device, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 10.

14. A computer readable medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 10 is implemented.

15. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 10 is implemented.

Images