Prediction of product feature data and training of prediction model

Abstract

Disclosed in the embodiments of the present specification are a product feature data prediction model training method and apparatus, and a product feature data prediction method and apparatus. The product feature data prediction model training method comprises: using a tree model structure to perform training by means of performing a branching operation on a tree model on the basis of product feature data for classification selected from sample data in a node and product prediction feature data, so as to obtain a product feature data prediction tree model. The product feature data prediction method comprises: using a product feature data prediction tree model obtained by means of training to predict a product feature data prediction value of a product to be predicted.

Description

Prediction of product feature data and training of prediction models Technical Field

[0001] Multiple embodiments of this specification relate to the field of information technology, and in particular, to product feature data prediction model training methods and devices, and prediction methods and devices. Background Art

[0002] In information technology, forecasting refers to the process of making decisions or predictions about future events, trends, or outcomes based on existing information, data, algorithms, and models. Forecasting has widespread applications across various fields, including economics, finance, marketing, and supply chain management, helping people better understand and respond to future changes and challenges.

[0003] Taking high-end wealth management products as an example, these products are complex and have specific sales windows. Currently, there are two potential sales challenges: First, when new products are released, it's difficult to determine the sales window, effectively preventing new product sales from being effectively guaranteed. Second, when reselling existing products, it's difficult to accurately identify optimal rescheduling plans, resulting in sales volumes falling short of expectations. Therefore, sales forecasting for high-end wealth management products is crucial. Accurate sales forecasts can better determine sales windows and rescheduling plans, helping financial institutions and sales teams better plan their sales strategies. Summary of the Invention

[0004] Multiple implementations of this specification describe product feature data prediction model training methods and devices, prediction methods and devices, and their technical solutions are as follows.

[0005] In the first aspect, an embodiment of the present specification provides a product feature data prediction model training method, comprising the following steps: obtaining a plurality of sample data containing historical feature data of a product, and selecting product prediction features and a number of product features for classification; presetting a root node of a tree model, and storing all sample data in the root node; performing branching operations on the tree model based on the product feature data for classification and the product prediction feature data corresponding to the sample data in the node to obtain a product feature data prediction tree model, wherein the product feature data prediction value corresponding to each node in the product feature data prediction tree model is calculated by using the product prediction feature data corresponding to the sample data in each node.

[0006] On the second aspect, an embodiment of this specification provides a product feature data prediction method, comprising the following steps: obtaining sample data of the product to be predicted, wherein the sample data of the product to be predicted includes product feature data for classification; inputting the sample data of the product to be predicted into a product feature data prediction tree model trained by the aforementioned training method to obtain a predicted value of the product feature data of the product to be predicted.

[0007] On the third aspect, an embodiment of this specification provides a product feature data prediction model training device, including: a data acquisition module, used to obtain multiple sample data containing product historical feature data; a feature selection module, used to select product prediction features and several classification product features in the sample data; an initialization storage module, used to preset the root node of the tree model and store all sample data in the root node; a training module, used to obtain a product feature data prediction tree model by performing branch operations on the tree model based on the classification product feature data and product prediction feature data corresponding to the sample data in the node, wherein the product feature data prediction value corresponding to each node in the product feature data prediction tree model is calculated by the product prediction feature data corresponding to the sample data in each node.

[0008] In a fourth aspect, an embodiment of the present specification provides a product feature data prediction device, comprising: an acquisition module for acquiring sample data of a product to be predicted, wherein the sample data of the product to be predicted includes product feature data for classification; a prediction module for inputting the sample data of the product to be predicted into a product feature data prediction tree model trained by the aforementioned training method to obtain a predicted value of the product feature data of the product to be predicted.

[0009] In a fifth aspect, an embodiment of this specification provides an electronic device, comprising a processor and a memory; the processor is connected to the memory; the memory is used to store executable program code; the processor runs a program corresponding to the executable program code by reading the executable program code stored in the memory, so as to execute the methods described in the first and second aspects of the above embodiments.

[0010] In a sixth aspect, an embodiment of this specification provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the methods described in the first and second aspects of the above embodiments.

[0011] The beneficial effects brought about by the technical solutions provided in some embodiments of this specification include at least: by performing branching operations on the tree model based on several product feature data for classification and product prediction feature data to obtain a product feature data prediction tree model, all possible branching schemes can be determined by considering the correlation between several product features for classification, which can improve the accuracy of the product feature data prediction model to a certain extent. At the same time, a multi-branch tree model structure is adopted to select multiple branches according to the actual situation of the product prediction feature data of the sample data and the product features for classification, so that the trained product feature data prediction tree model has greater flexibility and adaptability, and can clearly and intuitively represent the branch structure and decision path of the product feature data prediction model, thereby increasing the interpretability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] FIG1 is a global schematic diagram of an application scenario of a product feature data prediction model according to an embodiment of this specification.

[0013] FIG2 is a schematic diagram of an interaction in an application scenario of a product feature data prediction model according to an embodiment of this specification.

[0014] FIG3 is a schematic diagram of a scenario for deploying a product feature data prediction model according to an embodiment of this specification.

[0015] FIG4 is a flowchart of a product feature data prediction model training method according to an embodiment of the present specification.

[0016] FIG5 is a flowchart of a product feature data prediction model training method according to an embodiment of the present specification.

[0017] FIG6 is a flowchart of a product feature data prediction model training method according to an embodiment of the present specification.

[0018] FIG7 is a schematic diagram showing the principle of the product feature data prediction model training method according to an embodiment of this specification.

[0019] FIG8 is a flowchart of a product feature data prediction model training method according to an embodiment of this specification.

[0020] FIG9 is a flowchart of a product feature data prediction model training method according to an embodiment of the present specification.

[0021] FIG10 is a flowchart of a product feature data prediction model training method according to an embodiment of this specification.

[0022] FIG11 is a flow chart of a method for training a product feature data prediction model according to an embodiment of this specification.

[0023] FIG12 is a flowchart of a product feature data prediction model training method according to an embodiment of this specification.

[0024] FIG13 is a flowchart of a product feature data prediction model training method according to an embodiment of this specification.

[0025] FIG14 is a flow chart of a method for predicting product characteristic data according to an embodiment of this specification.

[0026] FIG15 is a schematic diagram of the structure of a product feature data prediction model training device according to an embodiment of this specification.

[0027] FIG16 is a schematic diagram of the structure of a product feature data prediction model training device according to an embodiment of this specification.

[0028] FIG17 is a schematic diagram of the structure of a product feature data prediction device according to an embodiment of this specification.

[0029] FIG18 is a schematic structural diagram of an electronic device according to an embodiment of this specification. DETAILED DESCRIPTION

[0030] The following is an explanation and description of the technical solutions of the various embodiments of this specification in conjunction with the drawings of the various embodiments of this specification. However, the following embodiments are only preferred embodiments of the various embodiments of this specification and are not exhaustive. Based on the various embodiments in the implementation manner, other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the various embodiments of this specification.

[0031] The terms "first," "second," "third," etc. in the description and claims of this specification and the accompanying drawings are used to distinguish different objects rather than to describe a specific order. In addition, the terms "including," "having," and any variations thereof are intended to cover non-exclusive inclusions.

[0032] Figure 1 shows a schematic diagram of an application scenario in which the product feature data prediction model outputs predicted values. The product feature data prediction model in the figure can be loaded onto a product feature data prediction platform or APP application on a terminal device (such as a smart phone, tablet computer, desktop, laptop, notebook computer), ultra-mobile personal computer (UMPC), handheld computer, PC device, personal digital assistant (PDA), virtual reality device, etc., for use by organizations or individuals who design, organize and launch products.

[0033] The organization or individual who designs, organizes, and launches a product, based on their needs, inputs the classification product feature data from the sample data of the product to be predicted into the device interface shown in FIG2. The product feature data prediction platform or APP application generates a predicted value of the product feature data of the product to be predicted based on the received input information. FIG2 takes the sales forecast of a high-end financial product as an example. Before releasing a high-end financial product, the issuer will conduct market research and risk assessment to determine the classification product features of the product to be predicted, such as the closed period and rate of return. When the issuer enters the classification product feature data of the product to be predicted, such as "rate of return is 3%, rate of return ranking is 5, closed period is 45 days" into the input interface and clicks the "Generate sales forecast value of the product to be predicted button", the product feature data prediction platform or APP application generates a picture of the "sales forecast tree model of the product to be predicted" and a text message of "sales forecast value of the product to be predicted is 16" based on the received input information. FIG2 is intended to illustrate the interactive process in an application scenario and is not limited to picture information and text information. That is, the generated content can be one or both of text and image data (i.e., the predicted value of the product feature data of the product to be predicted in the embodiments of this specification).

[0034] Referring to Figure 3, a schematic diagram of a scenario for deploying a product feature data prediction model is shown. The scenario includes a server 100 and a terminal device 200. The server 100 is in communication with the terminal device 200. There may be multiple terminal devices 200 connected to the server 100, and the server may store the product feature data prediction model.

[0035] On the one hand, server 100 can respond to instructions from terminal device 200 and load the model into the terminal device. For example, the need to obtain the model from the server may be determined based on the actual operations of an organization or individual involved in designing, organizing, and launching a product. For example, upon receiving product feature data for classification in sample data of a product to be predicted, it may be determined that the model needs to be obtained from server 100 and loaded into terminal device 200.

[0036] Server 100, on the other hand, can be used for model training. For example, a training platform is built into server 100. This training platform can train an initial tree model based on sample data from historical product sales information to obtain a product feature data prediction tree model. The trained product feature data prediction tree model can be stored on server 100 or locally via a software development kit.

[0037] Before this specification elaborates on the product feature data prediction model training method in combination with one or more embodiments, it first introduces the various terms involved in the product feature data prediction model training method.

[0038] Product prediction features: refers to the features that we hope to predict or infer through the product feature data prediction model, such as house prices, sales amounts, stock returns, product categories, customer satisfaction, user purchase decisions, etc.

[0039] Product features for classification: refers to features used to classify or categorize products in addition to the product feature data of the product to be predicted. These features can be numerical, discrete, or categorical to represent certain attributes or characteristics of the product. Moreover, these product features for classification are fixed and can be known before the product is entered into the product feature data prediction model.

[0040] Product feature data prediction value: refers to the result or output value that you want to predict or infer through the model. It can be a continuous variable, such as house price, sales amount, stock return, etc., or a discrete variable, such as product category, customer satisfaction, user purchase decision, etc.

[0041] For example, in the financial field, when the product feature data prediction tree model trained by the training method provided in multiple embodiments of this specification is applied to the sales prediction of high-end financial products, the "product prediction feature" is the sales volume of high-end financial products, the "product feature data prediction value" is the sales prediction value of high-end financial products, and the "product feature for classification" is the important features of high-end financial products obtained by the issuer through market research and risk assessment before releasing the high-end financial products, such as the rate of return, the ranking of returns among similar products, the closed period, and the credit rating of the issuing institution.

[0042] Tree models make decisions based on a tree structure, building a tree by splitting the features of the input data. A tree model consists of nodes and branches, with each node representing a feature or attribute, and a branch representing a decision based on the feature's value. In classification problems, tree models gradually split the data into different categories; in regression problems, tree models gradually divide the data into different numerical ranges.

[0043] Root Node: In a tree model, the root node is the starting point of the tree and has no parent node. The root node represents the initial state of the entire dataset or subset. The root node analyzes the characteristics of the dataset and selects optimal splitting features and splitting points to divide the data into different subsets.

[0044] Leaf Nodes: In a tree model, a leaf node is a terminal node with no child nodes. A leaf node is the endpoint of the tree model's decision-making process, representing the final prediction or classification result for the input data. Furthermore, the number of leaf nodes determines the complexity and granularity of the tree model.

[0045] Dynamic collection: A dynamic collection is a collection in which the nodes in the collection change dynamically.

[0046] Please refer to FIG4 , which shows a flowchart of a product feature data prediction model training method provided in one embodiment of this specification.

[0047] As shown in FIG4 , the product feature data prediction model training method may include at least the following steps.

[0048] Step 402 : Acquire a plurality of sample data including product historical feature data, and select product prediction features and a plurality of product features for classification.

[0049] It is understandable that each sample data contains historical sales information of a product, and multiple sample data contain historical sales information of multiple products.

[0050] Optionally, in this embodiment, methods for selecting several product features for classification include but are not limited to the following.

[0051] 1. Based on domain expertise and experience, and by understanding the product's sales background and market demand, manually select several classification product features that have a high correlation with the sample data classification scheme.

[0052] 2. Calculate the correlation or correlation coefficient between the product features and the sample data classification scheme, and select several product features that have a high correlation with the sample data classification scheme. Correlation analysis can be performed using statistical methods (such as the Pearson correlation coefficient) or machine learning algorithms (such as random forests).

[0053] 3. Use feature selection algorithms to automatically select several product features that are relevant to the sample data classification scheme. These algorithms can rank and select product features based on their importance or relevance, such as the Chi-square test, information gain, and L1 regularization.

[0054] Step 404: preset a root node of the tree model and store all sample data in the root node.

[0055] Step 406, by performing branching operations on the tree model based on the classification product feature data and product prediction feature data corresponding to the sample data in the node, a product feature data prediction tree model is obtained. The product feature data prediction value corresponding to each node in the product feature data prediction tree model is calculated using the product prediction feature data corresponding to the sample data in each node.

[0056] It can be understood that the product feature data prediction value corresponding to each node in the product feature data prediction tree model can be calculated through the product prediction feature data corresponding to the sample data in each node after obtaining the product feature data prediction tree model, or it can be calculated through the product prediction feature data corresponding to the sample data in each node in the process of obtaining the product feature data prediction tree model. This embodiment does not limit this.

[0057] Optionally, the method for obtaining the product feature data prediction value corresponding to each node in the product feature data prediction tree model by calculating the product feature data corresponding to the sample data in each node can be to calculate the average value of the product feature data corresponding to all sample data in each node, and use the average value as the product feature data prediction value corresponding to each node, or it can be to perform a weighted average operation on the product prediction feature data corresponding to all sample data in each node, where the weight can be set according to the importance of the product corresponding to the sample data, etc., and this embodiment does not limit this.

[0058] The technical concept of this embodiment is as follows: First, a tree model structure is used as the basic structure of the product feature data prediction model. Initially, there is only one preset root node in the tree model without any branches. All sample data containing historical feature data of the product is stored in the root node, that is, all products are temporarily divided into the same category of products. Then, by performing branch operations on the tree model based on the classification product feature data and product prediction feature data corresponding to the sample data in the node, it can be understood that multiple branch operations can be performed on the tree model based on the classification product feature data and product prediction feature data corresponding to the sample data in the node, so as to make the classification of products more detailed, thereby making the product feature data prediction tree model achieve better prediction results. Moreover, the product feature data prediction value corresponding to each node in the product feature data prediction tree model is calculated by the product prediction feature data corresponding to the sample data in each node.

[0059] This embodiment obtains a product feature data prediction tree model by performing branching operations on the tree model based on several product feature data for classification and product prediction feature data. It can determine all possible branching schemes by considering the correlation between several product features for classification, and can improve the accuracy of the product feature data prediction model to a certain extent. At the same time, a multi-branch tree model structure is adopted to select multiple branches according to the actual situation of the product prediction feature data of the sample data and the product features for classification, so that the trained product feature data prediction tree model has greater flexibility and adaptability, and can clearly and intuitively represent the branch structure and decision path of the product feature data prediction model, thereby increasing the interpretability of the model.

[0060] Please refer to FIG5 , which shows a flowchart of a product feature data prediction model training method provided in one embodiment of this specification.

[0061] As shown in FIG5 , the product feature data prediction model training method may include at least the following steps.

[0062] Step 502 : Acquire a plurality of sample data including product historical feature data, and select product prediction features and a plurality of product features for classification.

[0063] Step 504: preset a root node of the tree model and store all sample data in the root node.

[0064] The above steps 502 to 504 may refer to steps 402 to 404 and will not be repeated here.

[0065] Step 506: construct a dynamic set and store the root node in the dynamic set.

[0066] Therefore, there is at least one node in the initial dynamic set, namely the root node, and all sample data are stored in the root node, that is, all sample data are classified into one category at the beginning of model training, and then the tree model branching operation is performed on the node, that is, the sample data is classified.

[0067] Step 508: Determine the branching scheme of the remaining nodes based on the classification product feature data and product prediction feature data corresponding to the sample data in the remaining nodes in the dynamic set, and determine whether there is a newly generated node in the branching scheme. If so, store the newly generated node in the dynamic set; if not, treat the node as a leaf node and delete it from the dynamic set.

[0068] It can be understood that the product features for classification are used as the classification basis for the sample data, and the tree model branching operation is performed on the sample data in the remaining nodes in the dynamic set to determine the branching scheme of the remaining nodes. The branching scheme for the remaining nodes determined here can be based on the combination of the value ranges of multiple product features for classification to classify the sample data, or it can be based on the value range of one product feature for classification to classify the sample data.

[0069] It can be understood that the branching scheme of the remaining nodes determined based on the classification product feature data and product prediction feature data corresponding to the sample data in the remaining nodes in the dynamic set may not generate new nodes, that is, there are no newly generated nodes, or it may generate new nodes.

[0070] Step 510, repeat the previous step 508 until there is no node in the dynamic set, so as to obtain a product feature data prediction tree model, and the product feature data prediction value corresponding to each node in the product feature data prediction tree model is calculated by the product prediction feature data corresponding to the sample data in each node.

[0071] The technical concept of this embodiment is as follows: First, the structure of a tree model is used as the basic structure of the product feature data prediction model. At the beginning, there is only one preset root node in the tree model without any branches. All sample data containing historical feature data of the product is stored in the root node, that is, all products are temporarily divided into the same category of products. Next, a dynamic set is constructed and the root node is stored in the dynamic set. Then, based on the classification product feature data and product prediction feature data corresponding to the sample data in the remaining nodes in the dynamic set, a branching scheme for the remaining nodes is determined. At this time, the remaining node must be the root node. If the branching scheme does not generate a new node, that is, no new node is added to the dynamic set, then the remaining node is treated as a leaf node and deleted from the dynamic set. Since the dynamic set is empty at this time, it means that there are no remaining nodes to continue the tree model branching operation, that is, there is no plan to further classify the products in the sample data. The training process is terminated to obtain the trained product feature data prediction tree model. If the branching scheme is to branch the root node as the remaining node into n new nodes, the root node as the remaining node is deleted from the dynamic set, and the newly generated n nodes are stored in the dynamic set. At this time, there are nodes in the dynamic set again, indicating that there are still remaining nodes to continue the tree model branching operation. The operation of "determining the branching scheme for the remaining nodes based on the classification product feature data and product prediction feature data corresponding to the sample data in the remaining nodes in the dynamic set" is repeated until there are no nodes in the dynamic set, thereby obtaining the trained product feature data prediction tree model.

[0072] This embodiment can dynamically adjust the node structure of the tree model according to actual branching conditions through the construction of a dynamic set and the management of the remaining nodes in the dynamic set, making the construction process of the product feature data prediction tree model more flexible and efficient.

[0073] Please refer to FIG6 , which shows a flowchart of a product feature data prediction model training method provided in yet another embodiment of this specification.

[0074] Compared with the example of the flow chart of the product feature data prediction model training method shown in Figure 5, the example shown in Figure 6 adds the process of determining the branching schemes of the remaining nodes based on the classification product feature data and the product prediction feature data corresponding to the sample data in the remaining nodes in the dynamic set, including: in the process of determining the branching schemes of the remaining nodes based on the classification product feature data and the product prediction feature data corresponding to the sample data in the remaining nodes in the dynamic set, the classification product features that have been used on the branch path where the remaining nodes are located are not used, and only one classification product feature is used in the process of determining the branching schemes of each remaining node.

[0075] It can be understood that the branch path where the remaining nodes are located refers to a path from the root node to the current remaining node in the tree model, and is independent of other nodes at the same depth as the current remaining node. Each node on the branch path represents a value or condition of a product feature for classification. The branch path defines the classification flow and decision-making process of sample data in the tree model.

[0076] The technical concept of this embodiment is: based on the classification product feature data and product prediction feature data corresponding to the sample data in the remaining nodes in the dynamic set, the branching scheme of the remaining nodes is determined to classify the sample data according to the value range of a type of classification product feature, and the classification product features that have been used on the branch path where the remaining nodes are located are no longer used.

[0077] For example, please refer to Figure 7. Figure 7 takes the sales forecast of high-end financial products as an example to introduce the specific implementation steps of the product feature data prediction model training method of this embodiment: the product's yield rate, the ranking of the yield among similar products, and the closed period are selected as three types of product features for classification of high-end financial product sample data. It is assumed that the root node determines the branching scheme of the root node based on these three types of product features for classification and the product prediction feature data of the sample data, which is to divide the value of the yield feature at the root node into three intervals of (-∞, 0), [0, 5%), and [5%, +∞), and generate three new nodes according to the three divided intervals, namely node a, node b, and node c, that is, delete the root node from the dynamic set, and the sample data in the root node is also divided into the corresponding three nodes according to the value interval of the yield feature, and the three nodes are stored in the dynamic set. At this time, the dynamic set has three nodes: node a, node b, and node c. Then, taking the remaining node c in the dynamic set as an example, in the process of performing tree model branching operations on the sample data in node c using the dynamic programming algorithm based on the three classification product features, first determine that the branch path where node c is located and the classification product feature used is the yield feature, and then perform tree model branching operations on the sample data in node c based on each of the remaining two classification product features (yield ranking feature, closed cycle feature). For example, based on the two value intervals of the yield ranking feature, it is divided into two new nodes as a tree model branching operation plan, based on the three value intervals of the yield ranking feature, it is divided into three new nodes as another tree model branching operation plan, based on the closed cycle The three value intervals of the closed cycle feature are divided into three new nodes as another branch operation plan, and then the branch plan of node c is determined based on the two categories of classification using product features and the product prediction feature data of the sample data in node c, which is to divide the value of the closed cycle feature at node c into three intervals of [0, 30), [30, 90) and [90, +∞), and generate three new nodes based on the two divided intervals, namely node d, node e and node f. The sample data in node c is also divided into the corresponding three nodes according to the value interval of the closed cycle feature, and the three new nodes are stored in the dynamic set. At this time, the dynamic set has five nodes: node a, node b, node d, node e and node f.Next, taking the remaining node e in the dynamic set as an example, we first determine that the classification product features already used on the branch path where node e is located are the yield feature and the closed cycle feature. Then, we perform a tree model branching operation on the sample data in node e based on the remaining one-category classification product feature (the yield ranking feature). Based on the product prediction feature data of the sample data in node e, we determine the branching scheme for node e to be: dividing the value of the yield ranking feature at node e into three intervals: [1, 3), [3, 10), and [10, +∞). Based on the three divided intervals, we determine to generate three new nodes. Since the branch paths where these three newly generated nodes are located have already used the three selected classification product features, the tree model branching operation cannot be performed on these three newly generated nodes at this time. That is, the final branching scheme for these three newly generated nodes is not to generate new nodes. As can be seen, since the three classification product features of yield, yield ranking, and closed cycle are selected for high-end financial products, the depth of the high-end financial product sales prediction tree model obtained after training is four layers.

[0078] Based on the above, it can be seen that the beneficial effects brought about by the technical solution of this embodiment include at least: by limiting the use of classification product features that have been used on the branch path where the remaining nodes are located, repeated judgment of the same features on the same branch path can be avoided, thereby reducing unnecessary calculations and redundant judgments. At the same time, by limiting the use of only one type of classification product features in each tree model branch operation, the calculation time complexity of determining the branch scheme is greatly reduced, and the execution efficiency of the model is greatly improved; by limiting the classification product features to not being reused on a branch path and using one type of classification product features on a branch node for branch operations, the structural depth of the model is simplified, the complexity of the model is reduced, the branch path can be made clearer and more intuitive, and it helps to enhance the interpretability of the model.

[0079] Please refer to FIG8 , which shows a flowchart of a product feature data prediction model training method provided in yet another embodiment of this specification.

[0080] Compared with the example of the flow chart of the product feature data prediction model training method shown in Figure 6, the example shown in Figure 8 adds a branching scheme for determining the remaining nodes based on the product feature data for classification and the product prediction feature data corresponding to the sample data in the remaining nodes in the dynamic set. For convenience, the figure only shows the following added steps. The branching scheme for determining the remaining nodes based on the product feature data for classification and the product prediction feature data corresponding to the sample data in the remaining nodes in the dynamic set can include at least the following steps.

[0081] Step 802 : Calculate the remaining node prediction errors corresponding to the remaining nodes based on the product prediction feature data corresponding to the sample data in the remaining nodes.

[0082] It can be understood that the remaining node prediction errors corresponding to the remaining nodes refer to the prediction errors of the respective branch solutions used to evaluate the remaining nodes, which are related to the predicted values ​​of the product feature data corresponding to the remaining nodes.

[0083] Step 804 : performing branching operations on the remaining nodes based on the classification product feature data corresponding to the sample data in the remaining nodes to obtain branching solutions for the remaining nodes.

[0084] Preferably, in this embodiment, a dynamic programming algorithm is used to perform branch operations on the remaining nodes based on the product feature data for classification corresponding to the sample data in the remaining nodes. The dynamic programming algorithm is a well-known algorithm that is usually used to solve optimization problems, including optimal classification problems. The dynamic programming algorithm can be used to solve some sub-problems related to better classification, such as feature selection, feature weight calculation, and determination of better classification decision solutions. Specifically, for a product feature for classification, the computational time complexity of the exhaustive traversal method is O(2 n ), while the dynamic programming algorithm is O(n 2 ), where n represents the number of values ​​of the product feature used for classification.

[0085] Step 806 , based on the product prediction feature data corresponding to the sample data in the branch nodes in each branch solution, calculate the branch solution prediction error corresponding to each branch solution.

[0086] It can be understood that the branch scheme prediction error corresponding to the branch scheme refers to the prediction error of the branch scheme for evaluating all newly generated branch nodes of the remaining nodes after the branch operation is performed on the remaining nodes, which is independent of the predicted value of the product feature data corresponding to the remaining nodes.

[0087] Optionally, in this embodiment, methods for calculating the remaining node prediction errors corresponding to the remaining nodes or calculating the branch solution prediction errors corresponding to each branch solution include but are not limited to the following methods.

[0088] 1. Use the mean squared error (MSE) as a measure of the prediction error of the remaining nodes or the branching solution prediction error of each branching solution. The mean squared error is the average of the squares of the differences between the predicted values ​​and the true values.

[0089] 2. Mean Absolute Error (MAE): The mean absolute error is the average of the absolute values ​​of the differences between the predicted values ​​and the true values. Compared to the mean squared error, the mean absolute error is more robust to outliers.

[0090] 3. Log Loss: For binary or multi-classification problems, log loss can be used as a measure of the prediction error of the remaining nodes or the prediction error of the branch solutions. Log loss is the average of the negative logarithm of the difference between the predicted probability and the true label.

[0091] 4. Cross-entropy Loss: Similar to logarithmic loss, for multi-classification problems, cross-entropy loss can also be used to measure the prediction error of the remaining nodes or the prediction error of each branch solution. Cross-entropy loss is the average of the negative logarithm of the difference between the predicted probability distribution and the true probability distribution.

[0092] Step 808 : determining a final branching solution for the remaining nodes based on the remaining node prediction errors corresponding to the remaining nodes and the branching solution prediction errors corresponding to each branching solution.

[0093] Based on the above, it can be seen that the product feature data prediction model training method provided in multiple embodiments of this specification, in the process of determining the branch schemes of the remaining nodes based on the classification product feature data corresponding to the sample data in the remaining nodes in the dynamic set and the product prediction feature data, comprehensively considers the remaining node prediction errors corresponding to the remaining nodes and the branch scheme prediction errors corresponding to each branch scheme, and determines the final branch scheme of the remaining nodes according to specific rules or algorithms. For example, the branch scheme with the smaller branch scheme prediction error corresponding to each branch scheme can be selected as the final branch scheme of the remaining node, or the branch scheme prediction errors corresponding to multiple branch schemes can be comprehensively considered by weighted averaging and other methods to determine the final branch scheme of the remaining node, etc., so as to improve the accuracy and reliability of the product feature data prediction model training.

[0094] Please refer to FIG9 , which shows a flowchart of a product feature data prediction model training method provided in yet another embodiment of this specification.

[0095] Compared with the example of the flow chart of the product feature data prediction model training method shown in Figure 8, the example shown in Figure 9 adds the calculation of the remaining node prediction errors corresponding to the remaining nodes. For convenience, only the following added steps are shown in the figure. At least the following steps can be included in the calculation of the remaining node prediction errors corresponding to the remaining nodes.

[0096] Step 902: Divide the sample data in the remaining nodes into training sample data and verification sample data.

[0097] In this embodiment, according to the training purpose, the sample data in the remaining nodes are divided into training sample data and verification sample data in a certain ratio, such as 7:3 or 8:2.

[0098] Step 904 : Based on the product prediction feature data corresponding to the training sample data in the remaining nodes, the remaining node data prediction values ​​corresponding to the remaining nodes are calculated.

[0099] Optionally, based on the product prediction feature data corresponding to the training sample data in the remaining nodes, the method for calculating the remaining node data prediction value corresponding to the remaining nodes can be to calculate the average value of the product prediction feature data in the training sample data corresponding to the remaining nodes, and use the average value as the remaining node data prediction value corresponding to the remaining nodes, or it can be to perform a weighted average operation on the product prediction feature data in the training sample data corresponding to the remaining nodes, where the weight can be set according to the importance of the product corresponding to the training sample data, etc. This embodiment does not limit this.

[0100] Step 906 , based on the remaining node data prediction values ​​corresponding to the remaining nodes and the product prediction feature data corresponding to the verification sample data in the remaining nodes, calculate the remaining node prediction errors corresponding to the remaining nodes.

[0101] In this embodiment, the training sample data is used to calculate the remaining node data prediction value corresponding to the remaining node, and combined with the product prediction feature data corresponding to the verification sample data in the remaining node, the remaining node prediction error corresponding to the remaining node is determined according to specific rules or algorithms, such as mean square error, mean absolute error or logarithmic loss, etc., thereby improving the accuracy and reliability of the product feature data prediction model training.

[0102] Please refer to FIG10 , which shows a flowchart of a product feature data prediction model training method provided in yet another embodiment of this specification.

[0103] Compared with the example of the flow chart of the product feature data prediction model training method shown in Figure 8, the example shown in Figure 10 adds the calculation of the remaining node prediction errors corresponding to the remaining nodes. For convenience, only the following additional steps are shown in the figure. At least the following steps can be included in the calculation of the remaining node prediction errors corresponding to the remaining nodes.

[0104] Step 1002: Divide the sample data in the remaining nodes into training sample data and verification sample data.

[0105] Step 1004 , calculating the mean of the product prediction feature data corresponding to the training sample data in the remaining nodes to serve as the remaining node data prediction values ​​corresponding to the remaining nodes.

[0106] Step 1006 , calculating the variance between the product prediction feature data corresponding to each verification sample data in the remaining nodes and the remaining node data prediction value corresponding to the remaining nodes, to obtain the remaining node prediction error corresponding to the remaining nodes.

[0107] For example, taking the sales prediction of high-end financial products as an example, the specific implementation steps of determining the remaining node prediction error corresponding to one of the remaining nodes in this embodiment are introduced: assuming that there are 8 sample data in the remaining node, these 8 sample data are divided into training sample data and verification sample data in a ratio of 1:1, the sales values ​​of high-end financial products in the training sample data (that is, the product prediction feature data mentioned in multiple embodiments of this specification) are 3, 5, 6, and 8 respectively, and the sales values ​​of high-end financial products in the verification sample data are 3, 6, 7, and 10 respectively. The mean of the product prediction feature data values ​​of each training sample data of the remaining nodes is calculated, that is, (3+5+6+8) / 4=5.5, then the sales prediction value of the high-end financial product corresponding to the remaining node (that is, the remaining node prediction value corresponding to the remaining node mentioned in multiple embodiments of this specification) is 5.5, and the variance of the product prediction feature data value of each verification sample data of the remaining node and the remaining node prediction value corresponding to the remaining node is calculated, that is, [(3-5.5) 2 +(6-5.5) 2 +(7-5.5) 2 +(10-5.5) 2 ]÷4=7.25, then the sales forecast error of the high-end financial products corresponding to the remaining nodes (that is, the remaining node forecast error corresponding to the remaining nodes mentioned in multiple embodiments of this specification) is 7.25.

[0108] The product feature data prediction model training method provided in this embodiment is a scheme in which the predicted value of the remaining node data corresponding to each remaining node is represented by the mean of the product feature data values ​​of each training sample data of the node. Therefore, even if the amount of sample data is small, this training method can be used to obtain an effective estimate. At the same time, by calculating the variance of the product prediction feature data corresponding to each verification sample data in the remaining node and the predicted value of the remaining node data corresponding to the remaining node, as the remaining node data prediction error corresponding to the remaining node, a relatively accurate basis is provided for subsequently determining the final branching scheme of the remaining node based on the remaining node prediction error corresponding to the remaining node and the branching scheme prediction error corresponding to each branching scheme.

[0109] Please refer to FIG11 , which shows a flowchart of a product feature data prediction model training method provided in yet another embodiment of this specification.

[0110] Compared with the example of the flow chart of the product feature data prediction model training method shown in Figure 8, the example shown in Figure 11 adds the calculation of the branch scheme prediction error corresponding to each branch scheme. For convenience, only the following additional steps are shown in the figure. At least the following steps can be included in the calculation of the branch scheme prediction error corresponding to each branch scheme.

[0111] Step 1102 : Calculate the branch node prediction error corresponding to each branch node in the branch solution based on the product prediction feature data corresponding to the sample data in each branch node in the branch solution.

[0112] Step 1104 : Calculate the sum of the branch node prediction errors corresponding to the branch nodes in the branch solution to serve as the branch solution prediction error corresponding to the branch solution.

[0113] It can be understood that this embodiment only shows the calculation method of the branch scheme prediction error corresponding to one branch scheme of the remaining nodes, and the calculation method of the branch scheme prediction error corresponding to other branch schemes of other remaining nodes can refer to the calculation method shown in this embodiment.

[0114] Optionally, based on the product prediction feature data corresponding to the sample data in each branch node in the branch scheme, methods for calculating the branch node prediction error corresponding to each branch node in the branch scheme include but are not limited to the following.

[0115] 1. Use the mean squared error (MSE) as a measure of the error in the prediction of the product feature data corresponding to each node generated by the tree model branching operation. The mean squared error is the average of the squares of the differences between the predicted value and the true value.

[0116] 2. Mean Absolute Error (MAE): The mean absolute error is the average of the absolute values ​​of the differences between the predicted values ​​and the true values. Compared to the mean squared error, the mean absolute error is more robust to outliers.

[0117] 3. Log Loss: For binary or multi-classification problems, log loss can be used as a measure of the error in predicting the product feature data for each node generated by the tree model branching operation. Log loss is the average of the negative logarithm of the difference between the predicted probability and the true label.

[0118] 4. Cross-entropy Loss: Similar to logarithmic loss, for multi-classification problems, cross-entropy loss can also be used to measure the prediction error of product feature data corresponding to each node generated by the tree model branch operation. Cross-entropy loss is the average of the negative logarithm of the difference between the predicted probability distribution and the true probability distribution.

[0119] Please refer to FIG12 , which shows a flowchart of a product feature data prediction model training method provided in yet another embodiment of this specification.

[0120] Compared with the example of the flow chart of the product feature data prediction model training method shown in Figure 8, the example shown in Figure 12 adds product prediction feature data corresponding to the sample data in each branch node in the branch scheme, and calculates the branch node prediction error corresponding to each branch node in the branch scheme, which can at least include the following steps.

[0121] Step 1202: Divide the sample data in the branch node into training sample data and verification sample data.

[0122] Step 1204 : Based on the product prediction feature data corresponding to the training sample data in the branch node, a branch node data prediction value corresponding to the branch node is calculated.

[0123] Optionally, based on the product prediction feature data corresponding to the training sample data in the branch node, the method for calculating the branch node data prediction value corresponding to the branch node can be to calculate the average value of the product prediction feature data in the training sample data corresponding to the branch node, and use the average value as the branch node data prediction value corresponding to the branch node, or it can be to perform a weighted average operation on the product prediction feature data in the training sample data corresponding to the branch node, where the weight can be set according to the importance of the product corresponding to the training sample data in the branch node, etc. This embodiment does not limit this.

[0124] Step 1206 , based on the branch node data prediction value corresponding to the branch node and the product prediction feature data corresponding to the verification sample data in the branch node, calculate the branch node prediction error corresponding to the branch node.

[0125] In this embodiment, the training sample data is used to calculate the branch node data prediction value corresponding to the branch node, and combined with the product prediction feature data corresponding to the verification sample data in the branch node, the branch node prediction error corresponding to the branch node is determined according to specific rules or algorithms, such as the mean square error, mean absolute error or logarithmic loss, etc., thereby improving the accuracy and reliability of the product feature data prediction model training.

[0126] It can be understood that the above steps 1202 to 1206 only illustrate the calculation method of the branch node prediction error corresponding to one branch node of the branch scheme. The calculation method of the branch node prediction error corresponding to other branch nodes of the branch scheme can refer to the calculation method shown in this embodiment.

[0127] Step 1208 : Calculate the sum of the branch node prediction errors corresponding to the branch nodes in the branch solution to serve as the branch solution prediction error corresponding to the branch solution.

[0128] Please refer to FIG13 , which shows a flowchart of a product feature data prediction model training method provided in yet another embodiment of this specification.

[0129] Compared with the example of the flow chart of the product feature data prediction model training method shown in Figure 8, the example shown in Figure 13 adds product prediction feature data corresponding to the sample data in each branch node in the branch scheme, and calculates the branch node prediction error corresponding to each branch node in the branch scheme, which can at least include the following steps.

[0130] Step 1302: Divide the sample data in the branch node into training sample data and verification sample data.

[0131] Step 1304 , calculating the mean of the product prediction feature data corresponding to the training sample data in the branch node to serve as the branch node data prediction value corresponding to the branch node.

[0132] Step 1306 , calculating the variance between the product prediction feature data corresponding to each verification sample data in the branch node and the predicted value of the branch node data corresponding to the branch node, to obtain the branch node prediction error corresponding to the branch node.

[0133] It can be understood that the above steps 1302 to 1306 only illustrate the calculation method of the branch node prediction error corresponding to one branch node of the branch scheme. The calculation method of the branch node prediction error corresponding to other branch nodes of the branch scheme can refer to the calculation method shown in this embodiment.

[0134] Step 1308 : Calculate the sum of the branch node prediction errors corresponding to the branch nodes in the branch solution to serve as the branch solution prediction error corresponding to the branch solution.

[0135] For example, the sales prediction of high-end financial products is used as an example to introduce the specific implementation steps of determining the branch node prediction error corresponding to each branch node in a branch scheme of the remaining nodes in this embodiment: Assume that there are 3 branch nodes generated by a branch scheme of the remaining nodes, namely node x, node y and node z, and the sample data in each branch node is divided into training sample data and verification sample data in a ratio of 1:1. The sales values ​​of high-end financial products in the training sample data of node x (that is, the product prediction feature data values ​​mentioned in multiple embodiments of this specification) are 3 and 5 respectively, the sales values ​​of high-end financial products in the verification sample data are 3 and 6 respectively, and the sales value of high-end financial products in the training sample data of node y is 6. The sales value of high-end financial products in the verification sample data is 7, the sales value of high-end financial products in the training sample data of node z is 8, and the sales value of high-end financial products in the verification sample data is 10. The average sales value of high-end financial products in each training sample data of branch node x generated by a branching scheme of the remaining nodes is calculated, that is, (3+5)÷2=4, then the sales prediction value of the high-end financial products corresponding to branch node x (that is, the branch node data prediction value corresponding to the branch node mentioned in multiple embodiments of this specification) is 4, and the sales prediction error of the high-end financial products corresponding to branch node x (that is, the branch node data prediction error corresponding to the branch node mentioned in multiple embodiments of this specification) is calculated as [(3-4) 2 +(6-4) 2 ]÷2=2.5. Since there is only one training sample data in each of node y and node z, the sales value of the high-end financial product of each training sample data of node y and node z is the sales prediction value of the high-end financial product corresponding to node y and node z. The sales prediction error of the high-end financial product corresponding to branch node y is calculated as (7-6) 2 =1, the sales forecast error of the high-end financial product corresponding to branch node z is (10-8) 2 =4, then the branch solution prediction error corresponding to one branch solution of the remaining nodes is 2.5+1+4=7.5.

[0136] In multiple embodiments of the present specification, the final branching scheme of the remaining nodes is determined based on the remaining node prediction errors corresponding to the remaining nodes and the branching scheme prediction errors corresponding to each branching scheme, including: selecting a branching scheme from all branching schemes as a pending branching scheme based on the branching scheme prediction errors corresponding to each branching scheme of the remaining nodes; judging whether the branching scheme prediction error corresponding to the pending branching scheme is greater than the remaining node prediction error corresponding to the remaining nodes; if so, the final branching scheme is not to generate a new node; if not, the pending branching scheme is used as the final branching scheme.

[0137] Optionally, methods for selecting a branch scheme from all branch schemes as a pending branch scheme include but are not limited to the following.

[0138] 1. Compare the prediction errors of the corresponding branching solutions for the remaining nodes and select the branching solution with the smaller prediction error as the pending branching solution. This method believes that the smaller the prediction error, the more accurate the branching operation results of the branching solution for the remaining nodes, and can provide more reliable predictions.

[0139] 2. Combine multiple evaluation indicators, such as prediction error, the number of nodes generated by the branching scheme, etc., use a comprehensive evaluation method, such as weighted sum or decision tree evaluation indicators (such as information gain, Gini index, etc.), comprehensively consider the importance of each indicator, and select the better branching scheme as the pending branching scheme.

[0140] Please refer to FIG14 , which shows a flow chart of a product feature data prediction method provided in one embodiment of this specification.

[0141] As shown in FIG14 , the product characteristic data prediction method includes at least the following steps.

[0142] Step 1402: Obtain sample data of products to be predicted, where the sample data of products to be predicted includes product feature data for classification.

[0143] Step 1404 : Input the sample data of the product to be predicted into the product feature data prediction tree model trained by the product feature data prediction model training method provided in multiple embodiments of this specification to obtain the product feature data prediction value of the product to be predicted.

[0144] The product feature data prediction method of the embodiment of this specification can use the trained product feature data prediction tree model to output the product feature data prediction value of the product to be predicted and / or the product feature data prediction tree model structure diagram of the product to be predicted based on the classification product feature data in the input product sample data.

[0145] For example, the product feature data prediction method of the embodiment of this specification can be executed on the terminal device shown in Figure 3. The terminal device obtains the trained product feature data prediction model information, as shown in Figure 2. The organization or individual who designs, organizes and launches the product inputs the classification product feature data in the sample data of the product to be predicted on the interface of the terminal device according to the needs, triggering the product feature data prediction model in the terminal device, and then outputs the product feature data prediction value of the product to be predicted and / or the product feature data prediction tree model structure diagram of the product to be predicted on the interface of the terminal device. The trained product feature data prediction model can be stored locally on the terminal device, so that the terminal device can quickly and efficiently load the running model from the hard disk to the memory / video memory, so that the terminal device can load and infer the model to generate the product feature data prediction value of the sample data of the product to be predicted and / or the product feature data prediction tree model structure diagram of the product to be predicted.

[0146] Please refer to Figure 15, which is a schematic diagram of the structure of a product feature data prediction model training device provided in one embodiment of this specification. The technical concept adopted in this embodiment is similar to the technical concept of the training method of the product feature data prediction model described above, and will not be repeated in this embodiment.

[0147] As shown in FIG15 , the product feature data prediction model training device may include at least a data acquisition module 1501, a feature selection module 1502, an initialization storage module 1503, and a training module 1504. Specifically, the data acquisition module 1501 is configured to acquire a plurality of sample data containing historical product feature data; the feature selection module 1502 is configured to select product prediction features and a plurality of classification product features from the sample data; the initialization storage module 1503 is configured to preset a root node of a tree model and store all sample data in the root node; and the training module 1504 is configured to obtain a product feature data prediction tree model by performing branching operations on the tree model based on the classification product feature data and product prediction feature data corresponding to the sample data in the node. The product feature data prediction value corresponding to each node in the product feature data prediction tree model is calculated using the product prediction feature data corresponding to the sample data in each node.

[0148] In another embodiment of the present specification, the training module 1504 obtains a product feature data prediction tree model by performing branching operations on the tree model based on the classification product feature data and product prediction feature data corresponding to the sample data in the node, including: constructing a dynamic set and storing the root node in the dynamic set; determining the branching scheme of the remaining nodes based on the classification product feature data and product prediction feature data corresponding to the sample data in the remaining nodes in the dynamic set, and judging whether there is a newly generated node in the branching scheme; if so, storing the newly generated node in the dynamic set; if not, treating the node as a leaf node and deleting it from the dynamic set; repeating the previous step until there is no node in the dynamic set to obtain the product feature data prediction tree model.

[0149] In another embodiment of the present specification, in the process of determining the branching schemes of the remaining nodes in the training module 1504 based on the classification product feature data and product prediction feature data corresponding to the sample data in the remaining nodes in the dynamic set, the classification product features that have been used on the branch paths where the remaining nodes are located are not used, and only one classification product feature is used in the process of determining the branching schemes for each remaining node.

[0150] Please refer to FIG16 , which is a schematic diagram of the structure of a product feature data prediction model training device provided in one embodiment of this specification.

[0151] As shown in FIG16 , the product feature data prediction model training device may include at least a data acquisition module 1501, a feature selection module 1502, an initialization storage module 1503, and a training module 1504. The training module 1504 may include at least: a node error generation unit 1601 for calculating the remaining node prediction errors corresponding to the remaining nodes based on the product prediction feature data corresponding to the sample data in the remaining nodes; a node branch generation unit 1602 for performing branch operations on the remaining nodes based on the classification product feature data corresponding to the sample data in the remaining nodes to obtain branch solutions for the remaining nodes; a branch error generation unit 1603 for calculating the branch solution prediction errors corresponding to each branch solution based on the product prediction feature data corresponding to the sample data in the branch nodes in each branch solution; and a branch solution determination unit 1604 for determining the final branch solution for the remaining nodes based on the remaining node prediction errors corresponding to the remaining nodes and the branch solution prediction errors corresponding to each branch solution.

[0152] In another embodiment of the present specification, the node error generation unit 1601 calculates the remaining node prediction errors corresponding to the remaining nodes, including: dividing the sample data in the remaining nodes into training sample data and verification sample data; calculating the remaining node data prediction values ​​corresponding to the remaining nodes based on the product prediction feature data corresponding to the training sample data in the remaining nodes; and calculating the remaining node prediction errors corresponding to the remaining nodes based on the remaining node data prediction values ​​corresponding to the remaining nodes and the product prediction feature data corresponding to the verification sample data in the remaining nodes.

[0153] In another embodiment of the present specification, the node error generation unit 1601 calculates the remaining node data prediction values ​​corresponding to the remaining nodes based on the product prediction feature data corresponding to the training sample data in the remaining nodes, including: calculating the mean of the product prediction feature data corresponding to the training sample data in the remaining nodes as the remaining node data prediction values ​​corresponding to the remaining nodes; and calculating the remaining node prediction errors corresponding to the remaining nodes based on the remaining node data prediction values ​​corresponding to the remaining nodes and the product prediction feature data corresponding to the verification sample data in the remaining nodes, including: calculating the variance of the product prediction feature data corresponding to each verification sample data in the remaining nodes and the remaining node data prediction values ​​corresponding to the remaining nodes, to obtain the remaining node prediction errors corresponding to the remaining nodes.

[0154] In another embodiment of the present specification, the branch error generation unit 1603 calculates the branch scheme prediction error corresponding to each branch scheme, including: calculating the branch node prediction error corresponding to each branch node in the branch scheme based on the product prediction feature data corresponding to the sample data in each branch node in the branch scheme; and calculating the sum of the branch node prediction errors corresponding to each branch node in the branch scheme as the branch scheme prediction error corresponding to the branch scheme.

[0155] In another embodiment of the present specification, the branch error generation unit 1603 calculates the branch node prediction error corresponding to each branch node in the branch scheme based on the product prediction feature data corresponding to the sample data in each branch node in the branch scheme, including: dividing the sample data in the branch node into training sample data and verification sample data; calculating the branch node data prediction value corresponding to the branch node based on the product prediction feature data corresponding to the training sample data in the branch node; and calculating the branch node prediction error corresponding to the branch node based on the branch node data prediction value corresponding to the branch node and the product prediction feature data corresponding to the verification sample data in the branch node.

[0156] In another embodiment of the present specification, the branch error generating unit 1603 calculates a branch node data prediction value corresponding to the branch node based on the product prediction feature data corresponding to the training sample data in the branch node, including: calculating the mean of the product prediction feature data corresponding to the training sample data in the branch node as the branch node data prediction value corresponding to the branch node; and calculating a branch node prediction error corresponding to the branch node based on the branch node data prediction value corresponding to the branch node and the product prediction feature data corresponding to the verification sample data in the branch node, including: calculating the variance of the product prediction feature data corresponding to each verification sample data in the branch node and the branch node data prediction value corresponding to the branch node, to obtain the branch node prediction error corresponding to the branch node.

[0157] In another embodiment of the present specification, the branch solution determining unit 1604 determines the final branch solution of the remaining nodes based on the remaining node prediction errors corresponding to the remaining nodes and the branch solution prediction errors corresponding to each branch solution, including: selecting a branch solution from all branch solutions as a pending branch solution based on the branch solution prediction errors corresponding to each branch solution of the remaining nodes; and determining whether the branch solution prediction error corresponding to the pending branch solution is greater than the remaining node prediction error corresponding to the remaining nodes. If so, the final branch solution is not to generate a new node; if not, the pending branch solution is used as the final branch solution.

[0158] Please refer to FIG17 , which is a schematic diagram of the structure of a product feature data prediction device provided in one embodiment of this specification.

[0159] As shown in FIG17 , the product feature data prediction apparatus may include at least an acquisition module 1701 and a prediction module 1702. The acquisition module 1701 is configured to acquire sample data of the product to be predicted, including product feature data for classification; and the prediction module 1702 is configured to input the sample data of the product to be predicted into the product feature data prediction tree model trained using the aforementioned training method to obtain a predicted value of the product feature data of the product to be predicted.

[0160] In this embodiment, the acquisition module 1701 and the prediction module 1702 can both be device terminals, or the acquisition module 1701 can be a device terminal and the prediction module 1702 can be a server. The specific technical concept is similar to the aforementioned product feature data prediction method embodiment, and this embodiment will not be repeated here.

[0161] Please refer to FIG18 , which is a schematic diagram of the structure of an electronic device provided in one embodiment of this specification.

[0162] As shown in FIG. 18 , the electronic device 1800 may include: at least one processor 1801 , at least one network interface 1804 , a user interface 1803 , a memory 1805 , and at least one communication bus 1802 .

[0163] Among them, the communication bus 1802 can be used to realize the connection and communication of the above-mentioned components; the user interface 1803 can include buttons, and the optional user interface can also include a standard wired interface and a wireless interface; the network interface 1804 can include but is not limited to a Bluetooth module, an NFC module, a Wi-Fi module, etc.

[0164] Among them, the processor 1801 may include one or more processing cores. The processor 1801 uses various interfaces and lines to connect the various parts of the entire electronic device 1800, and executes various functions of the electronic device 1800 and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory 1805, and calling data stored in the memory 1805. Optionally, the processor 1801 can be implemented in at least one hardware form of DSP, FPGA, and PLA. The processor 1801 can integrate one or a combination of CPU, GPU, and modem. Among them, the CPU mainly processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing the content to be displayed on the display screen; and the modem is used to handle wireless communications. It is understandable that the above-mentioned modem may not be integrated into the processor 1801, but may be implemented separately through a chip.

[0165] Among them, the memory 1805 may include RAM or ROM. Optionally, the memory 1805 includes a non-transitory computer-readable medium. The memory 1805 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 1805 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing the operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 1805 may also be at least one storage device located away from the aforementioned processor 1801. The memory 1805 as a computer storage medium may include an operating system, a network communication module, a user interface module and an application. The processor 1801 can be used to call the application stored in the memory 1805 and execute the method in one or more of the above-mentioned embodiments.

[0166] The embodiments of this specification also provide a computer-readable storage medium having instructions stored therein that, when executed on a computer or processor, cause the computer or processor to perform one or more steps in the embodiments shown in Figures 4 to 14 above. If the various component modules of the electronic device described above are implemented as software functional units and sold or used as independent products, they may be stored in a computer-readable storage medium.

[0167] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When executed, the program can include the processes of the above-described method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks. The technical features of this embodiment and the implementation scheme can be combined in any manner unless they conflict.

[0168] The above embodiments are merely preferred embodiments of this specification and are not intended to limit the scope of this specification. Without departing from the design spirit of this specification, various modifications and improvements made to the technical solutions of this specification by ordinary technicians in this field should fall within the scope of protection determined by the claims of this specification.

Claims

1. A product feature data prediction model training method includes the following steps: Obtain multiple sample data containing historical product feature data, and select product prediction features and several product features for classification; Preset a root node of the tree model and store all sample data in the root node; A product feature data prediction tree model is obtained by performing branching operations on the tree model based on the classification product feature data and product prediction feature data corresponding to the sample data in the node. The product feature data prediction value corresponding to each node in the product feature data prediction tree model is calculated using the product prediction feature data corresponding to the sample data in each node.

2. The product feature data prediction model training method according to claim 1, wherein the method of performing a branching operation on the tree model based on the classification product feature data corresponding to the sample data in the node and the product prediction feature data to obtain the product feature data prediction tree model comprises: Constructing a dynamic set and storing the root node in the dynamic set; Determine the branching scheme of the remaining nodes based on the classification product feature data and the product prediction feature data corresponding to the sample data in the remaining nodes in the dynamic set, and determine whether there is a newly generated node in the branching scheme. If so, store the newly generated node in the dynamic set; if not, delete the node as a leaf node from the dynamic set; Repeat the previous step until there are no nodes in the dynamic set to obtain a product feature data prediction tree model.

3. The product feature data prediction model training method as described in claim 2, in the process of determining the branching scheme of the remaining nodes based on the classification product feature data corresponding to the sample data in the remaining nodes in the dynamic set and the product prediction feature data, the classification product features that have been used on the branch path where the remaining nodes are located are not used, and only one classification product feature is used in the process of determining the branching scheme of each remaining node.

4. The product feature data prediction model training method according to claim 3, wherein determining the branching scheme of the remaining nodes based on the classification product feature data corresponding to the sample data in the remaining nodes in the dynamic set and the product prediction feature data comprises: Based on the product prediction feature data corresponding to the sample data in the remaining nodes, the remaining node prediction errors corresponding to the remaining nodes are calculated; Perform branching operations on the remaining nodes based on the classification product feature data corresponding to the sample data in the remaining nodes to obtain branching solutions for the remaining nodes; Based on the product prediction feature data corresponding to the sample data in the branch nodes in each branch solution, the branch solution prediction error corresponding to each branch solution is calculated; The final branching scheme of the remaining nodes is determined based on the remaining node prediction errors corresponding to the remaining nodes and the branching scheme prediction errors corresponding to each branching scheme.

5. The product feature data prediction model training method according to claim 4, wherein the step of calculating the remaining node prediction errors corresponding to the remaining nodes comprises: Divide the sample data in the remaining nodes into training sample data and verification sample data; Based on the product prediction feature data corresponding to the training sample data in the remaining nodes, the remaining node data prediction values ​​corresponding to the remaining nodes are calculated; Based on the remaining node data prediction values ​​corresponding to the remaining nodes and the product prediction feature data corresponding to the verification sample data in the remaining nodes, the remaining node prediction errors corresponding to the remaining nodes are calculated.

6. The product feature data prediction model training method according to claim 5, wherein the calculating of the remaining node data prediction values ​​corresponding to the remaining nodes based on the product prediction feature data corresponding to the training sample data in the remaining nodes comprises: Calculate the mean of the product prediction feature data corresponding to the training sample data in the remaining nodes as the remaining node data prediction value corresponding to the remaining nodes; The calculating of the remaining node prediction errors corresponding to the remaining nodes based on the remaining node data prediction values ​​corresponding to the remaining nodes and the product prediction feature data corresponding to the verification sample data in the remaining nodes includes: The variance of the product prediction feature data corresponding to each verification sample data in the remaining nodes and the remaining node data prediction value corresponding to the remaining nodes is calculated to obtain the remaining node prediction error corresponding to the remaining nodes.

7. The product feature data prediction model training method according to claim 4, wherein the step of calculating the branch solution prediction error corresponding to each branch solution comprises: Based on the product prediction feature data corresponding to the sample data in each branch node in the branch scheme, the branch node prediction error corresponding to each branch node in the branch scheme is calculated; The sum of the branch node prediction errors corresponding to each branch node in the branch scheme is calculated as the branch scheme prediction error corresponding to the branch scheme.

8. The product feature data prediction model training method of claim 7, wherein the step of calculating the branch node prediction error corresponding to each branch node in the branch scheme based on the product prediction feature data corresponding to the sample data in each branch node in the branch scheme comprises: Divide the sample data in the branch node into training sample data and verification sample data; Based on the product prediction feature data corresponding to the training sample data in the branch node, the branch node data prediction value corresponding to the branch node is calculated; Based on the branch node data prediction value corresponding to the branch node and the product prediction feature data corresponding to the verification sample data in the branch node, the branch node prediction error corresponding to the branch node is calculated.

9. The product feature data prediction model training method according to claim 8, wherein the step of calculating the branch node data prediction value corresponding to the branch node based on the product prediction feature data corresponding to the training sample data in the branch node comprises: Calculate the mean of the product prediction feature data corresponding to the training sample data in the branch node as the branch node data prediction value corresponding to the branch node; The step of calculating the branch node prediction error corresponding to the branch node based on the branch node data prediction value corresponding to the branch node and the product prediction feature data corresponding to the verification sample data in the branch node includes: The variance of the product prediction feature data corresponding to each verification sample data in the branch node and the branch node data prediction value corresponding to the branch node is calculated to obtain the branch node prediction error corresponding to the branch node.

10. The product feature data prediction model training method according to any one of claims 4 to 9, wherein determining the final branching solution for the remaining nodes based on the remaining node prediction errors corresponding to the remaining nodes and the branching solution prediction errors corresponding to the branching solutions comprises: Based on the branch solution prediction errors corresponding to the branch solutions of the remaining nodes, a branch solution is selected from all the branch solutions as the pending branch solution; Determine whether the branch scheme prediction error corresponding to the pending branch scheme is greater than the remaining node prediction error corresponding to the remaining nodes. If it is greater, the final branch scheme is not to generate new nodes. If it is not greater, the pending branch scheme is used as the final branch scheme.

11. A method for predicting product characteristic data, comprising the following steps: Acquiring sample data of products to be predicted, wherein the sample data of products to be predicted includes product feature data for classification; The sample data of the product to be predicted is input into the product feature data prediction tree model trained by the training method according to any one of claims 1 to 10 to obtain the product feature data prediction value of the product to be predicted.

12. Product feature data prediction model training device, including: A data acquisition module is used to acquire multiple sample data containing historical feature data of products; Feature selection module, used to select product prediction features and several classification product features from sample data; Initialize the storage module, which is used to preset the root node of the tree model and store all sample data into the root node; The training module is used to obtain a product feature data prediction tree model by performing branch operations on the tree model based on the classification product feature data and product prediction feature data corresponding to the sample data in the node. The product feature data prediction value corresponding to each node in the product feature data prediction tree model is calculated using the product prediction feature data corresponding to the sample data in each node.

13. The product feature data prediction model training device according to claim 12, wherein the training module performs a branching operation on the tree model based on the classification product feature data and the product prediction feature data corresponding to the sample data in the node to obtain the product feature data prediction tree model, including: Constructing a dynamic set and storing the root node in the dynamic set; Determine the branching scheme of the remaining nodes based on the classification product feature data and the product prediction feature data corresponding to the sample data in the remaining nodes in the dynamic set, and determine whether there is a newly generated node in the branching scheme. If so, store the newly generated node in the dynamic set; if not, delete the node as a leaf node from the dynamic set; Repeat the previous step until there are no nodes in the dynamic set to obtain a product feature data prediction tree model.

14. The product feature data prediction model training device as described in claim 13, in the process of determining the branching scheme of the remaining nodes based on the classification product feature data corresponding to the sample data in the remaining nodes in the dynamic set and the product prediction feature data in the training module, the classification product features that have been used on the branch path where the remaining nodes are located are not used, and only one classification product feature is used in the process of determining the branching scheme of each remaining node.

15. The product feature data prediction model training device according to claim 14, wherein the training module comprises: a node error generating unit, configured to calculate the remaining node prediction errors corresponding to the remaining nodes based on the product prediction feature data corresponding to the sample data in the remaining nodes; A node branch generation unit is used to perform branch operations on the remaining nodes based on the classification product feature data corresponding to the sample data in the remaining nodes to obtain various branch solutions for the remaining nodes; A branch error generating unit, configured to calculate a branch solution prediction error corresponding to each branch solution based on product prediction feature data corresponding to sample data in a branch node in each branch solution; The branching scheme determining unit is configured to determine a final branching scheme for the remaining nodes based on the remaining node prediction errors corresponding to the remaining nodes and the branching scheme prediction errors corresponding to the respective branching schemes.

16. The product feature data prediction model training device according to claim 15, wherein the node error generation unit calculates the remaining node prediction errors corresponding to the remaining nodes, comprising: Divide the sample data in the remaining nodes into training sample data and verification sample data; Based on the product prediction feature data corresponding to the training sample data in the remaining nodes, the remaining node data prediction values ​​corresponding to the remaining nodes are calculated; Based on the remaining node data prediction values ​​corresponding to the remaining nodes and the product prediction feature data corresponding to the verification sample data in the remaining nodes, the remaining node prediction errors corresponding to the remaining nodes are calculated.

17. The product feature data prediction model training device according to claim 16, wherein the node error generating unit calculates the remaining node data prediction values ​​corresponding to the remaining nodes based on the product prediction feature data corresponding to the training sample data in the remaining nodes, comprising: Calculate the mean of the product prediction feature data corresponding to the training sample data in the remaining nodes as the remaining node data prediction value corresponding to the remaining nodes; The calculating of the remaining node prediction errors corresponding to the remaining nodes based on the remaining node data prediction values ​​corresponding to the remaining nodes and the product prediction feature data corresponding to the verification sample data in the remaining nodes includes: The variance of the product prediction feature data corresponding to each verification sample data in the remaining nodes and the remaining node data prediction value corresponding to the remaining nodes is calculated to obtain the remaining node prediction error corresponding to the remaining nodes.

18. The product feature data prediction model training device according to claim 15, wherein the step of calculating the branch solution prediction error corresponding to each branch solution in the branch error generation unit comprises: Based on the product prediction feature data corresponding to the sample data in each branch node in the branch scheme, the branch node prediction error corresponding to each branch node in the branch scheme is calculated; The sum of the branch node prediction errors corresponding to each branch node in the branch scheme is calculated as the branch scheme prediction error corresponding to the branch scheme.

19. The product feature data prediction model training device according to claim 18, wherein the branch error generating unit calculates the branch node prediction error corresponding to each branch node in the branch scheme based on the product prediction feature data corresponding to the sample data in each branch node in the branch scheme, comprising: Divide the sample data in the branch node into training sample data and verification sample data; Based on the product prediction feature data corresponding to the training sample data in the branch node, the branch node data prediction value corresponding to the branch node is calculated; Based on the branch node data prediction value corresponding to the branch node and the product prediction feature data corresponding to the verification sample data in the branch node, the branch node prediction error corresponding to the branch node is calculated.

20. The product feature data prediction model training device according to claim 19, wherein the branch error generating unit calculates the branch node data prediction value corresponding to the branch node based on the product prediction feature data corresponding to the training sample data in the branch node, including: Calculate the mean of the product prediction feature data corresponding to the training sample data in the branch node as the branch node data prediction value corresponding to the branch node; The step of calculating the branch node prediction error corresponding to the branch node based on the branch node data prediction value corresponding to the branch node and the product prediction feature data corresponding to the verification sample data in the branch node includes: The variance of the product prediction feature data corresponding to each verification sample data in the branch node and the branch node data prediction value corresponding to the branch node is calculated to obtain the branch node prediction error corresponding to the branch node.

21. The product feature data prediction model training device according to any one of claims 15 to 20, wherein the determining unit determines the final branching solution for the remaining nodes based on the remaining node prediction errors corresponding to the remaining nodes and the branching solution prediction errors corresponding to the respective branching solutions, comprising: Based on the branch solution prediction errors corresponding to the branch solutions of the remaining nodes, a branch solution is selected from all the branch solutions as the pending branch solution; Determine whether the branch scheme prediction error corresponding to the pending branch scheme is greater than the remaining node prediction error corresponding to the remaining nodes. If it is greater, the final branch scheme is not to generate new nodes. If it is not greater, the pending branch scheme is used as the final branch scheme.

22. A product characteristic data prediction device comprising: An acquisition module, configured to acquire sample data of products to be predicted, wherein the sample data of products to be predicted includes product feature data for classification; A prediction module is used to input sample data of the product to be predicted into a product feature data prediction tree model trained by the training method according to any one of claims 1 to 10 to obtain a predicted value of the product feature data of the product to be predicted.

23. An electronic device comprising a processor and a memory; The processor is connected to the memory; The memory is used to store executable program code; The processor reads the executable program code stored in the memory to run a program corresponding to the executable program code, so as to execute the method according to any one of claims 1 to 11.

24. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method according to any one of claims 1 to 11 is implemented.

Images