A page operation disorder identification method and device
By constructing a page operation obstacle identification method using neural networks and GBDT models, the problem of untimely response to user operation obstacles in online financial products has been solved, achieving accurate identification and timely assistance, and reducing customer churn.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SICHUAN XW BANK CO LTD
- Filing Date
- 2022-11-04
- Publication Date
- 2026-05-19
AI Technical Summary
In existing technologies, when users of online financial products encounter operational obstacles, there is no timely and accurate response, leading to customer churn. Existing solutions are inefficient, costly, or provide a poor customer experience.
A method for identifying page operation obstacles is constructed. By collecting user operation data and page link graph relationships, a machine learning model is trained to identify whether the user has encountered an operation obstacle, and a threshold is set to determine whether the user needs assistance.
It enables timely and accurate identification of user operational obstacles, improving identification accuracy and efficiency, reducing customer churn, and enhancing customer experience.
Smart Images
Figure CN115687047B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of Internet technology, specifically to a method and apparatus for identifying page operation obstacles. Background Technology
[0002] With the development of internet technology, more and more industries, such as commerce and financial management, have launched online products, including apps or H5 websites. Previously, offline services typically had dedicated staff available to assist customers with any issues. However, many online products now require self-service operation. If customers encounter difficulties or obstacles during this process, they often have to abandon the transaction. For example, in traditional bank branches, a customer manager is present to guide customers through banking transactions and answer any questions promptly. But when using financial products offered online, customers are often self-service, and there are few ways to resolve operational obstacles. For instance, some users may experience interruptions when using financial products offered by a bank's app or H5 website. For example, incomplete information when applying for a loan may prevent them from accessing the next page, or a page freeze may cause a transaction failure when purchasing wealth management products. These interruptions can have various causes, including system errors / bugs, customer operational obstacles, lack of customer willingness to trade, and other reasons. Such interruptions can lead to customer churn.
[0003] The current industry practice regarding this issue involves two solutions when a customer's actions are interrupted, such as a loan application. One is to passively wait for the customer to initiate contact, and the other is for customer service personnel to make outbound calls. Customer service personnel manually determine if the customer is encountering an obstacle by asking questions and reviewing their system activity records, and then answer their questions. However, the current approach has shortcomings: Firstly, only a minority of customers proactively contact the customer, making passive waiting inefficient. Secondly, proactively contacting all customers in an attempt to recover those genuinely experiencing difficulties not only wastes significant resources but also alienates customers and damages their experience. Summary of the Invention
[0004] In order to solve the problems existing in the prior art, the present invention provides a method and apparatus for identifying page operation obstacles, which aims to solve the problem that existing applications and related web pages cannot respond promptly and accurately to provide assistance when users encounter operation obstacles.
[0005] A method for identifying page operation obstacles includes the following steps:
[0006] Step 1: Collect and process a certain amount of data related to users' product usage process; the amount of data can be customized.
[0007] Step 2: Construct a recognition model to output the probability of the customer encountering operational obstacles based on the data processed in Step 1. The recognition model includes three sub-models: a first neural network, a second neural network, and a GBDT model.
[0008] Step 3: Distribution optimization. The three sub-models in the recognition model obtained in step 2 make the probability output by the recognition model more accurate.
[0009] Optimize the first neural network: Use the data processed in step 1 to train the first neural network in a supervised manner to obtain the optimal first neural network;
[0010] Optimize the second neural network: Use the data processed in step 1 to train the second neural network in a supervised manner to obtain the optimal second neural network;
[0011] Optimize the GBDT model: Use the features extracted by the optimized first and second neural networks and combine them with the user's basic information to train the GBDT model to obtain the optimal GBDT model;
[0012] Step 4: Based on the probability of user having operational obstacles output by the recognition model obtained in Step 3, and combined with the threshold of the probability of user having operational obstacles, determine whether the user has encountered obstacles in the process of using the product.
[0013] Preferably, in step 1, the relevant data of the user's product usage process includes the user's page operation history data, the graph network structure data of the page links of the user's product usage, the results of customer service personnel's outbound calls to reach users, and the user's own basic information data.
[0014] The user's page operation history data includes the link relationship of the user's browsing pages (such as jumping from page A to page C and then to page B), the time spent on the page, the frequency of clicking each function button, the usage duration, device type data, etc.
[0015] The user's product page network structure data includes the link relationships between pages, link function tag data, etc.
[0016] The results of customer service personnel's outbound calls to customers are collected. By analyzing the feedback data, historical customers are divided into two categories: those with operational obstacles and those without operational obstacles. These categories are used to define labels for training machine learning models.
[0017] Preferably, in step 1, the data processing procedure includes encoding, binning, truncation, standardization, and normalization.
[0018] Preferably, step 2 includes the following steps, wherein the order of steps 2.1 and 2.2 is interchangeable:
[0019] Step 2.1: Construct a first neural network to extract features of user page operation behavior. The first neural network includes an input layer, a feature embedding layer, and an output layer. The feature embedding layer includes a converter network layer. Both the input layer and the output layer are fully connected networks.
[0020] Step 2.2: Construct a second neural network to extract the page graph network structure features of the customer's use of the product. The second neural network includes a graph coding layer and an output layer, wherein the graph coding layer includes a GCN network layer and the output layer is a fully connected network.
[0021] Step 2.3: Based on the features extracted by the first neural network described in Step 2.1 and the features extracted by the second neural network described in Step 2.2, as well as the user's own information, construct a GBDT model to output the probability that the user has operational difficulties.
[0022] Preferably, step 3, optimizing the first neural network, includes the following steps:
[0023] Step 3.1.1: Initialize the parameters of the first neural network to satisfy a uniform distribution between [-1, 1];
[0024] Step 3.1.2: Calculate the cross-entropy loss using the output1 of the initialized first neural network and the user's obstacle label y:
[0025] Loss1=Crossentropy(y,output1);
[0026] In the formula, Loss1 is the loss function value of the first neural network; Crossentropy is the cross-entropy loss function; output1 is the output of the first neural network; and y is the label indicating whether the user encountered an obstacle.
[0027] Step 3.1.3: Optimize the model parameters using gradient descent to obtain a smaller loss function, namely: MinLoss1;
[0028] Step 3.1.4: Repeat step 3.1.3 for multiple iterations until the value of Loss1 no longer decreases, thus obtaining the optimal first neural network.
[0029] Preferably, step 3, optimizing the second neural network, includes the following steps:
[0030] Step 3.2.1: Initialize the parameters of the second neural network to satisfy a standard normal distribution;
[0031] Step 3.2.2: Calculate the cross-entropy loss using the output of the initialized second neural network and the label indicating whether the user encountered an obstacle:
[0032] Loss2=Crossentropy(y,output2);
[0033] In the formula, Loss2 is the loss function value of the second neural network; Crossentropy is the cross-entropy loss function; output2 is the output of the second neural network; and y is the label indicating whether the user encountered an obstacle.
[0034] Step 3.2.3: Optimize the model parameters using gradient descent to obtain a smaller loss function, namely: MinLoss2;
[0035] Step 3.2.4: Repeat step 3.2.3 for multiple iterations until the value of Loss2 no longer decreases, thus obtaining the optimal second neural network.
[0036] Preferably, step 3, optimizing the GBDT model, includes the following steps:
[0037] Step 3.3.1: Generate the first CART decision tree D1(x,θ1) based on the features extracted by the optimal first neural network, the features extracted by the second neural network, and the user's basic information, to fit the label y indicating whether the customer has encountered an obstacle;
[0038] Step 3.3.2: Use the output value of the CART decision tree (output) gbdt1 Calculate the cross-entropy loss based on the label y indicating whether the customer encountered an obstacle:
[0039] Loss gbdt1 =Crossentropy(y,output) gbdt1 );
[0040] Loss in the formula gbdt1 The output is the loss function value of the first CART decision tree, D1(x,θ1), where Crossentropy is the cross-entropy loss function. gbdt1 y is the output of the first CART decision tree D1(x,θ1); y is the label indicating whether the user encountered an obstacle.
[0041] Step 3.3.3: Calculate the gradient of the loss function:
[0042]
[0043] In the formula, grad1 is the gradient of the loss function of the first CART decision tree D1(x,θ1), θ1 is the parameter of the first CART decision tree D1, and Loss... gbdt1 The loss function value of the first CART decision tree, D1(x,θ1);
[0044] Step 3.3.4: Based on the features extracted by the optimal first neural network, the features extracted by the optimal second neural network, and the user's basic information, generate a second CART decision tree D2(x,θ2) to fit the gradient grad1 of the loss function from the previous step. This process continues for the i-th tree (where i>=2), where D... i (x,θ i ), D i Let be the i-th decision tree, used to fit the gradient grad of the loss function from the previous step. i-1 ;
[0045] Step 3.3.5: assemble the newly generated CART decision tree D. i (x,θ i Multiply the learning rate ε by the model and add it to the model to form a new model;
[0046] Step 3.3.6: Calculate the gradient grad of the loss function of the new model. i :
[0047]
[0048] In the formula grad i The gradient of the loss function for the new model; Loss gbdti The loss function value for the new model; θ i Let D be the i-th decision tree. i parameters
[0049] Step 3.3.7: Repeat steps 3.3.4 to 3.3.6 until the number of decision trees reaches the preset number M, which is the optimal GBDT model, i.e. the optimal recognition model. M can be customized.
[0050] Preferably, in step 4, based on the recognition model obtained in step 3, the input of relevant user operation product data and user information processed in step 1 can output the probability that the user has an operational obstacle. By analyzing historical data, a threshold probability is determined. When the probability of the customer's operational obstacle is greater than the threshold probability, it is considered that the customer has encountered an operational obstacle.
[0051] A preferred page operation obstacle recognition device includes a data processing module, an recognition module, and a customer service system module. The data processing module obtains relevant data about the user's product usage process from the customer service system module, preprocesses the data, and transmits it to the recognition module. The recognition module extracts features from the data, calculates the probability that the user has encountered an operation obstacle, determines whether the user has encountered an obstacle based on a set threshold, and finally transmits the determination result to the customer service system module to notify relevant customer service personnel to provide timely assistance to the user.
[0052] Preferably, the identification module includes a first neural network module, a second neural network module, and a GBDT module; wherein the first neural network module is used to extract features of user page operation behavior, the second neural network module is used to extract page graph network structure features of customer using the product, and the GBDT module combines the features extracted by the two neural network modules and the user's basic information to output the probability that the customer has an operation obstacle. A threshold probability is determined by analyzing historical data, and when the probability of the customer's operation obstacle is greater than the threshold probability, it is considered that the customer has encountered an operation obstacle.
[0053] The beneficial effects of this invention include:
[0054] By utilizing customer action data and page link graph data, a machine learning model is built. The model is trained using historical operation obstacle tags collected and reported by customer service personnel. The dual neural network model structure can effectively extract and match the characteristics of customer action behavior and the link relationship structure of the page itself, thereby enabling timely and accurate identification of whether customers have encountered operation obstacles.
[0055] This invention is an operational obstacle recognition method constructed from a completely new perspective and data dimension, which not only achieves automatic recognition but also improves recognition accuracy. Attached Figure Description
[0056] Figure 1 This is a schematic diagram of obtaining the operation obstacle label in Example 1.
[0057] Figure 2 This is a schematic diagram illustrating the implementation of a page operation obstacle identification method in Example 1.
[0058] Figure 3 This is a schematic diagram of a page operation obstacle recognition device according to Embodiment 2. Detailed Implementation
[0059] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, 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 a part of the embodiments of this application, and not all of the embodiments. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0060] Example 1
[0061] The following is in conjunction with the appendix Figure 2 Specific embodiments of the present invention will be described in detail;
[0062] A method for identifying page operation obstacles includes the following steps:
[0063] Step 1: Collect and process a certain amount of data related to users' product usage process; the amount of data can be customized.
[0064] The data related to the user's product usage process includes the user's page operation history data, the graph network structure data of the page links used by the user, the results of customer service personnel's outbound calls to reach users, and the user's own basic information data.
[0065] The user's page operation history data includes the link relationship of the user's browsing pages (such as jumping from page A to page C and then to page B), the time spent on the page, the frequency of clicking each function button, the usage duration, device type data, etc.
[0066] The user's product page network structure data includes the link relationships between pages, link function tag data, etc.
[0067] The results of customer service personnel's outbound calls to customers are collected. By analyzing the feedback data, historical customers are divided into two categories: those with operational obstacles and those without operational obstacles. These categories are used to define labels for training machine learning models.
[0068] The data processing steps include encoding, binning, truncation, standardization, and normalization.
[0069] Step 2: Construct a recognition model to output the probability of a customer encountering operational obstacles based on the data processed in Step 1. The recognition model includes three sub-models: a first neural network, a second neural network, and a GBDT model; the order of steps 2.1 and 2.2 is interchangeable.
[0070] Step 2.1: Construct a first neural network to extract features of user page operation behavior. The first neural network includes an input layer, a feature embedding layer, and an output layer. The feature embedding layer includes a converter network layer. Both the input layer and the output layer are fully connected networks.
[0071] Step 2.2: Construct a second neural network to extract the page graph network structure features of the customer's use of the product. The second neural network includes a graph coding layer and an output layer, wherein the graph coding layer includes a GCN network layer and the output layer is a fully connected network.
[0072] Step 2.3: Based on the features extracted by the first neural network described in Step 2.1 and the features extracted by the second neural network described in Step 2.2, as well as the user's own information, construct a GBDT model to output the probability that the user has operational difficulties.
[0073] Step 3: Distribution optimization. The three sub-models in the recognition model obtained in step 2 make the probability output by the recognition model more accurate.
[0074] Optimize the first neural network: Use the data processed in step 1 to train the first neural network in a supervised manner to obtain the optimal first neural network;
[0075] Step 3.1.1: Initialize the parameters of the first neural network to satisfy a uniform distribution between [-1, 1];
[0076] Step 3.1.2: Calculate the cross-entropy loss using the output1 of the initialized first neural network and the user's obstacle label y:
[0077] Loss1=Crossentropy(y,output1);
[0078] In the formula, Loss1 is the loss function value of the first neural network; Crossentropy is the cross-entropy loss function; output1 is the output of the first neural network; and y is the label indicating whether the user encountered an obstacle.
[0079] Step 3.1.3: Optimize the model parameters using gradient descent to obtain a smaller loss function, namely: MinLoss1;
[0080] Step 3.1.4: Repeat step 3.1.3 for multiple iterations until the value of Loss1 no longer decreases, thus obtaining the optimal first neural network.
[0081] Optimize the second neural network: Use the data processed in step 1 to train the second neural network in a supervised manner to obtain the optimal second neural network;
[0082] Step 3.2.1: Initialize the parameters of the second neural network to satisfy a standard normal distribution;
[0083] Step 3.2.2: Calculate the cross-entropy loss using the output of the initialized second neural network and the label indicating whether the user encountered an obstacle:
[0084] Loss2=Crossentropy(y,output2);
[0085] In the formula, Loss2 is the loss function value of the second neural network; Crossentropy is the cross-entropy loss function; output2 is the output of the second neural network; and y is the label indicating whether the user encountered an obstacle.
[0086] Step 3.2.3: Optimize the model parameters using gradient descent to obtain a smaller loss function, namely: MinLoss2;
[0087] Step 3.2.4: Repeat step 3.2.3 for multiple iterations until the value of Loss2 no longer decreases, thus obtaining the optimal second neural network.
[0088] Optimize the GBDT model: Use the features extracted by the optimized first and second neural networks and combine them with the user's basic information to train the GBDT model to obtain the optimal GBDT model;
[0089] Step 3.3.1: Generate the first CART decision tree D1(x,θ1) based on the features extracted by the optimal first neural network, the features extracted by the second neural network, and the user's basic information, to fit the label y indicating whether the customer has encountered an obstacle;
[0090] Step 3.3.2: Use the output value of the CART decision tree gbdt1 Calculate the cross-entropy loss based on the label y indicating whether the customer encountered an obstacle:
[0091] Loss gbdt1 =Crossentropy(y,output) gbdt1 );
[0092] Loss in the formula gbdt1 The output is the loss function value of the first CART decision tree, D1(x,θ1), where Crossentropy is the cross-entropy loss function. gbdt1 y is the output of the first CART decision tree D1(x,θ1); y is the label indicating whether the user encountered an obstacle; Step 3.3.3: Calculate the gradient of the loss function:
[0093]
[0094] In the formula, grad1 is the gradient of the loss function of the first CART decision tree D1(x,θ1), θ1 is the parameter of the first CART decision tree D1, and Loss... gbdt1 The loss function value of the first CART decision tree, D1(x,θ1);
[0095] Step 3.3.4: Based on the features extracted by the optimal first neural network, the features extracted by the optimal second neural network, and the user's basic information, generate a second CART decision tree D2(x,θ2) to fit the gradient grad1 of the loss function from the previous step. This process continues for the i-th tree (where i>=2), where D... i (x,θ i ), D i Let be the i-th decision tree, used to fit the gradient grad of the loss function from the previous step. i-1 ;
[0096] Step 3.3.5: assemble the newly generated CART decision tree D.i (x,θ i Multiply the learning rate ε by the model and add it to the model to form a new model;
[0097] Step 3.3.6: Calculate the gradient grad of the loss function of the new model. i :
[0098]
[0099] In the formula grad i The gradient of the loss function for the new model; Loss gbdti The loss function value for the new model; θ i Let D be the i-th decision tree. i parameters
[0100] Step 3.3.7: Repeat steps 3.3.4 to 3.3.6 until the number of decision trees reaches the preset number M, which is the optimal GBDT model, i.e. the optimal recognition model. M can be customized.
[0101] Step 4: Based on the probability of user having operational difficulties output by the recognition model obtained in Step 3, and combined with the threshold of the probability of user having operational difficulties, determine whether the user has encountered difficulties in using the product.
[0102] Based on the recognition model obtained in step 3, inputting the relevant user operation product data and user information processed in step 1 will output the probability that the user has an operational obstacle. By analyzing historical data, a threshold probability is determined. When the probability of a customer's operational obstacle is greater than the threshold probability, the customer is considered to have encountered an operational obstacle.
[0103] Corresponding to step 1 above:
[0104] The data collected from the customer service system for model training and inference is mainly divided into four parts: The first part consists of user page operation history data, including the user's page browsing path (e.g., jumping from page A to page C and then to page B), time spent on each page, frequency of clicks on various function buttons, usage duration, device type data, etc., serving as the data source for the first neural network to extract features. The second part consists of user product page graph network structure data, including the link relationships between pages and the function label data of those links, serving as the data source for the second neural network to extract features. The third part consists of feedback from customer service personnel's outbound calls, used to obtain user operation obstacle tags; a specific illustration of this process is attached. Figure 1The fourth part uses the user's basic information data as input to the GBDT model. The data collected in these four parts includes unstructured data (text, language, etc.) and structured data (tabular data). Therefore, preprocessing such as encoding, binning, truncation, standardization, and normalization is required. For structured data such as gender, age, and income, preprocessing mainly involves filling missing values (e.g., filling missing gender values with the mode), truncation (e.g., annual income greater than 1 million is rounded to 1 million, less than 50,000 is rounded to 50,000), and binning (e.g., age segments 0-20, 20-30, 30-50, 50-100, etc.). For unstructured data such as text, images, and speech, preprocessing mainly involves encoding, standardization, and normalization. For example, a bachelor's degree or higher is encoded as 3, less than a bachelor's degree as 2, and no degree as 1; RGB values of images are encoded as arrays; and language text is encoded as word vectors.
[0105] Corresponding to step 2 above:
[0106] The recognition model is constructed as follows: First, a first neural network is built, consisting of an input layer, a feature embedding layer, and an output layer. The input layer comprises multiple fully connected network layers, whose main function is to map the features of the input page operation history data to specific feature dimensions for input to the feature embedding layer. The feature embedding layer mainly consists of a transformer network layer, whose main function is to process and calculate the results of the input layer and extract sequence features. The output layer consists of multiple fully connected layers, whose main function is to map the features input to the feature embedding layer to the final output. Next, a second neural network is constructed, mainly consisting of a graph coding layer and an output layer. The graph coding layer mainly consists of a GCN (Graph Convolution Neural Network). The GBDT model is composed of a network layer, whose main function is to encode the graph network relationship between page links and extract effective features from the page link relationship. The output layer structure consists of multiple fully connected layers, whose main function is to map the features input to the graph encoding layer to the final output. Finally, the feature outputs of the first neural network feature embedding layer and the graph encoding layer of the second neural network are combined with the user's basic information as input to the GBDT model. The GBDT model outputs an obstacle probability, which is then combined with a probability threshold given by expert experience to obtain the result of whether the customer has encountered an operational obstacle.
[0107] Corresponding to step 3 above:
[0108] Optimize the recognition model; since the recognition model uses a multi-sub-model structure, a step-by-step optimization approach is adopted.
[0109] First, optimize the first neural network, hereinafter referred to as model_nn1;
[0110] a) Randomly initialize the parameters of model_nn1 so that they satisfy a uniform distribution between [-1, 1];
[0111] b) Calculate the cross-entropy loss using the initialized model output1 and the label y indicating whether the customer encountered an obstacle:
[0112] Loss1=Crossentropy(y,output1);
[0113] In the formula, Loss1 is the value of the model_nn1 loss function; Crossentropy is the cross-entropy loss function; output1 is the output of model_nn1; and y is the label indicating whether the user encountered an obstacle.
[0114] c) Use gradient descent to optimize the model parameters. The optimization goal is to minimize the loss function.
[0115] Min Loss1;
[0116] d) Repeat step c for multiple iterations until the value of Loss1 no longer decreases, thus obtaining the optimal model_nn2;
[0117] Then optimize the second neural network, hereinafter referred to as model_nn2, similar to optimizing the first neural network;
[0118] d) Randomly initialize the parameters of model_nn2 to make them conform to a standard normal distribution;
[0119] e) Calculate the cross-entropy loss using the initialized model output2 and the customer's obstacle label y:
[0120] Loss2=Crossentropy(y,output2);
[0121] In the formula, Loss2 is the value of the model_nn2 loss function; Crossentropy is the cross-entropy loss function; output2 is the output of model_nn2; and y is the label indicating whether the user encountered an obstacle.
[0122] f) Use gradient descent to optimize the model parameters. The optimization objective is:
[0123] Min Loss2;
[0124] g) Repeat step g for multiple iterations until the value of Loss2 no longer decreases, thus obtaining the optimal model_nn2;
[0125] Finally, optimize the GBDT model, hereinafter referred to as model_gbdt;
[0126] g) Use the intermediate features output by model_nn1 and model_nn2, along with basic user information, to generate a CART decision tree D1(x,θ1) to fit the customer's label y for whether or not they like the product.
[0127] h) Use the output value of the CART decision tree gbdt1 Calculate the cross-entropy loss with y:
[0128] Loss gbdt1 =Crossentropy(y,output) gbdt1 );
[0129] Loss in the formula gbdt1 The output is the loss function value of the first CART decision tree D1(x,θ1), where Crossentropy is the cross-entropy loss function. gbdt1 y is the output of the first CART decision tree D1(x,θ1); y is the label indicating whether the user encountered an obstacle.
[0130] Calculate the gradient of the loss function:
[0131]
[0132] In the formula, grad1 is the gradient of the loss function of the first CART decision tree D1(x,θ1), θ1 is the parameter of the first CART decision tree D1(x,θ1), and Loss is... gbdt1 The loss function value of the first CART decision tree D1(x,θ1) is used; the following steps are repeated until the number of trees reaches the preset number M, which is the optimal GBDT model, i.e., the optimal recognition model.
[0133] i) Use the intermediate features output by model_nn1 and model_nn2, along with the user's basic information, to generate a second CART decision tree D2(x,θ2), which is used to fit the gradient grad1 of the loss function from the previous step. This process is repeated for each tree, with D... i (x,θ i ), D i Let be the i-th decision tree, used to fit the gradient grad of the loss function from the previous step. i-1 ;
[0134] j) The newly generated CART decision tree D i (x,θ i Multiplying the learning rate ε by the model and adding it to the model creates a new model.
[0135] k) Calculate the gradient of the loss function of the new model;
[0136]
[0137] In the formula grad i The gradient of the loss function for the new model; Loss gbdti The loss function value for the new model; θ i For the new model parameters;
[0138] Corresponding to step 4 above:
[0139] Data related to the user's product usage process that triggered the interruption of key transaction operations is collected and processed. The processed data is then transmitted to the recognition model trained in step 3 to determine the probability that the user has encountered an operational obstacle. A threshold probability is determined by analyzing historical data. When the probability of the user's operational obstacle is greater than the threshold probability, the user is considered to have encountered an operational obstacle. Finally, the determination result of whether the user has encountered an operational obstacle is transmitted to the customer service system.
[0140] This invention utilizes customer operation behavior data and page link graph relationship data to construct a machine learning model. It is trained using historical operation obstacle tags collected and reported by customer service personnel. The dual neural network model structure can effectively extract and match the characteristics of customer operation behavior and the link relationship structure of the page itself, thereby enabling timely and accurate identification of whether customers have encountered operation obstacles. This solves the problem that existing applications and related web pages cannot respond and provide assistance in a timely and accurate manner when users encounter operation obstacles.
[0141] Example 2
[0142] See attached document Figure 3 , attached Figure 3 The diagram shows a structural schematic of a page operation obstacle recognition device. This device can be a module, program segment, or code on an electronic device. It should be understood that this device is related to the aforementioned appendix. Figure 2 Corresponding to the method implementation examples, it is capable of executing the attached... Figure 2 The various steps involved in the method embodiments, and the specific functions of the device, can be found in the description above. To avoid repetition, detailed descriptions are appropriately omitted here. (See attached...) Figure 3The device shown includes a data processing module, a recognition module, and a customer service system module. In a specific implementation, for example, a customer applies for a loan using a bank's online application. After completing the basic information, the next step, facial recognition, fails repeatedly. After several attempts, the customer gives up, and the loan application is interrupted, triggering a critical transaction interruption. At this point, the customer service system module transmits relevant data collected during the user's product usage, including the user's page operation history, the graph network structure data of the page links used by the user, and the user's basic information, to the data processing module. The data processing module preprocesses the data and then transmits the preprocessed data to the recognition module. The customer's operation data is input into the first neural network module, and the page link relationship data is input into the second neural network module. The intermediate layer features of the two neural networks and the customer's basic information are then input into the GBDT module. The GBDT module outputs the recognition result indicating whether the customer encountered an operational obstacle to the customer service system module, allowing customer service to contact the user promptly for assistance.
[0143] This invention utilizes customer operation behavior data and page link graph relationship data to construct a machine learning model. It is trained using historical operation obstacle tags collected and reported by customer service personnel. The dual neural network model structure can effectively extract and match the characteristics of customer operation behavior and the link relationship structure of the page itself, thereby enabling timely and accurate identification of whether customers have encountered operation obstacles. This solves the problem that existing applications and related web pages cannot respond and provide assistance in a timely and accurate manner when users encounter operation obstacles.
[0144] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.
Claims
1. A method for identifying page operation obstacles, characterized in that, Includes the following steps: Step 1: Collect and process a certain amount of data related to users' product usage process; the amount of data can be customized. Step 2: Construct a recognition model to output the probability of the customer encountering operational obstacles based on the data processed in Step 1. The recognition model includes three sub-models: a first neural network, a second neural network, and a GBDT model. Step 3: Distribution optimization. The three sub-models in the recognition model obtained in step 2 make the probability output by the recognition model more accurate. Optimize the first neural network: Use the data processed in step 1 to train the first neural network in a supervised manner to obtain the optimal first neural network; Optimize the second neural network: Use the data processed in step 1 to train the second neural network in a supervised manner to obtain the optimal second neural network; Optimize the GBDT model: Use the features extracted by the optimized first and second neural networks and combine them with the user's basic information to train the GBDT model to obtain the optimal GBDT model; Step 4: Based on the probability of user having operational difficulties output by the recognition model obtained in Step 3, and combined with the threshold of the probability of user having operational difficulties, determine whether the user has encountered difficulties in using the product. In step 1, the relevant data of the user's product usage process includes the user's page operation history data, the graph network structure data of the page links of the user's product usage, the results of customer service personnel's outbound calls to reach users, and the user's own basic information data. Step 2 includes the following steps: The order of steps 2.1 and 2.2 can be interchanged: Step 2.1: Construct a first neural network to extract features of user page operation behavior. The first neural network includes an input layer, a feature embedding layer, and an output layer. The feature embedding layer includes a converter network layer. Both the input layer and the output layer are fully connected networks. Step 2.2: Construct a second neural network to extract the page graph network structure features of the customer's use of the product. The second neural network includes a graph coding layer and an output layer, wherein the graph coding layer includes a GCN network layer and the output layer is a fully connected network. Step 2.3: Based on the features extracted by the first neural network described in Step 2.1 and the features extracted by the second neural network described in Step 2.2, as well as the user's own information, construct a GBDT model to output the probability that the user has operational difficulties.
2. The method for identifying page operation obstacles according to claim 1, characterized in that, In step 1, the data processing procedure includes encoding, binning, truncation, standardization, and normalization.
3. The method for identifying page operation obstacles according to claim 1, characterized in that, Step 3, optimizing the first neural network, includes the following steps: Step 3.1.1: Initialize the parameters of the first neural network to satisfy a uniform distribution between [-1, 1]; Step 3.1.2: Calculate the cross-entropy loss using the output1 of the initialized first neural network and the user's obstacle label y: ； In the formula The value of the loss function for the first neural network; The cross-entropy loss function is used; output1 is the output of the first neural network; y is the label indicating whether the user encountered an obstacle. Step 3.1.3: Optimize the model parameters using gradient descent to obtain a smaller loss function: ; Step 3.1.4: Repeat step 3.1.3, iterating multiple times, until... When the value no longer decreases, the optimal first neural network is obtained.
4. The method for identifying page operation obstacles according to claim 1, characterized in that, Step 3, optimizing the second neural network, includes the following steps: Step 3.2.1: Initialize the parameters of the second neural network to satisfy a standard normal distribution; Step 3.2.2: Calculate the cross-entropy loss using the output of the initialized second neural network and the label indicating whether the user encountered an obstacle: ; In the formula The value of the loss function for the second neural network; y is the cross-entropy loss function; output2 is the output of the second neural network; y is the label indicating whether the user encountered an obstacle. Step 3.2.3: Optimize the model parameters using gradient descent to obtain a smaller loss function: ; Step 3.2.4: Repeat step 3.2.3, iterating multiple times, until... The value of no longer decreases, thus obtaining the optimal second neural network.
5. The method for identifying page operation obstacles according to claim 1, characterized in that, Step 3, optimizing the GBDT model, includes the following steps: Step 3.3.1: Generate the first CART decision tree D1 based on the features extracted by the optimal first neural network, the features extracted by the second neural network, and the user's basic information. , used to fit the label y indicating whether the customer encountered an obstacle; Step 3.3.2: Use the output values of the CART decision tree Calculate the cross-entropy loss based on the label y indicating whether the customer encountered an obstacle: Loss gbdt1 = Crossentropy(y,output gbdt1 ); Loss in the formula gbdt1 The first CART decision tree D1 Loss function value, The cross-entropy loss function; output gbdt1 The first CART decision tree D1 The output is y, which is a label indicating whether the user encountered an obstacle. Step 3.3.3: Calculate the gradient of the loss function: ； In the formula The first CART decision tree D1 loss function gradient, For the parameters of the first CART decision tree D1, Loss gbdt1 The first CART decision tree D1 Loss function value; Step 3.3.4: Generate the second CART decision tree D2 based on the features extracted by the optimal first neural network, the features extracted by the optimal second neural network, and the user's basic information. This is used to fit the gradient grad1 of the loss function from the previous step, and so on. For the i-th tree (where i >= 2), we have D. i D i Let be the i-th decision tree, used to fit the gradient grad of the loss function from the previous step. i-1 ; Step 3.3.5: assemble the newly generated CART decision tree. Multiply by the learning rate These are accumulated into the model to form a new model; Step 3.3.6: Calculate the gradient grad of the loss function of the new model. i : ； In the formula The gradient of the loss function for the new model; Loss gbdti The loss function value for the new model; For the i-th decision tree parameters Step 3.3.7: Repeat steps 3.3.4 to 3.3.6 until the number of decision trees reaches the preset number M, which is the optimal GBDT model, i.e. the optimal recognition model. M can be customized.
6. The method for identifying page operation obstacles according to claim 1, characterized in that, In step 4, based on the recognition model obtained in step 3, the probability of a user having an operational obstacle is output by inputting the relevant user operation product data and user information processed in step 1. A threshold probability is determined by analyzing historical data. When the probability of a customer having an operational obstacle is greater than the threshold probability, the customer is considered to have encountered an operational obstacle.
7. A page operation obstacle recognition device, characterized in that, It includes a data processing module, an identification module, and a customer service system module. The data processing module obtains relevant data about the user's product usage process from the customer service system module, preprocesses the data, and then passes it to the identification module. The identification module extracts features from the data, calculates the probability that the user has encountered an operational obstacle, determines whether the user has encountered an obstacle based on a set threshold, and finally transmits the determination result to the customer service system module to notify the relevant customer service personnel to provide timely assistance to the user. The identification module includes a first neural network module, a second neural network module, and a GBDT module. The first neural network module is used to extract features of user page operation behavior, the second neural network module is used to extract page graph network structure features of customer using the product, and the GBDT module combines the features extracted by the two neural network modules and the user's basic information to output the probability that the customer has operational difficulties.
8. The page operation obstacle recognition device according to claim 7, characterized in that, The GBDT module determines a threshold probability by analyzing historical data. When the probability of a customer encountering an operational obstacle is greater than the threshold probability, the customer is considered to have encountered an operational obstacle.