Big data mining method and artificial intelligence system for remote order information

Abstract

The embodiment of the present application provides a big data mining method and artificial intelligence system for remote order information. First, by acquiring sample remote order data and its corresponding behavior path and order description portrait, learning data for training neural networks is generated. Among them, a part of the important feature data is hidden, forming hidden training supervision data and hidden neural network learning data. This setting can prompt the basic order description portrait prediction network to more deeply mine and learn the key information hidden in the data. Then, the basic order description portrait prediction network will use these learning data for training, and generate target training error values ​​and reconstruction training error values ​​based on the deviation value between the predicted result and the actual result, so as to serve as a key indicator for evaluating the network performance of the basic order description portrait prediction network, and is also an important basis for optimizing network parameters, thereby significantly improving the understanding and prediction capabilities of user order behavior.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a big data mining method and an artificial intelligence system for remote order information. Background Art

[0002] In online trading platforms such as e-commerce, understanding users' shopping behaviors and accurately predicting their possible future order information is the key to improving service efficiency and business profits. Traditional methods usually rely on statistical analysis or rule-based models, but these methods face difficulties in processing complex and large-scale order data, especially when user behaviors and shopping habits are highly personalized and dynamically changing.

[0003] In addition, neural networks have demonstrated strong predictive capabilities in many fields, including predicting user shopping behavior in e-commerce. However, how to design and train a neural network model that can fully mine and learn deep features in order data remains a challenge. Therefore, a new technical solution is needed to better acquire and process remote order data, design and optimize neural network models, and generate accurate order description portraits to further improve the efficiency and effectiveness of user service decisions. Summary of the invention

[0004] In order to at least overcome the above-mentioned deficiencies in the prior art, the purpose of the embodiments of the present application is to provide a big data mining method and an artificial intelligence system for remote order information.

[0005] Obtain sample remote order data and a sample order behavior path data sequence and a sample order description portrait corresponding to the sample remote order data, and obtain first neural network learning data, hidden training supervision data and hidden neural network learning data corresponding to the sample remote order data; the hidden training supervision data is data that is hidden after the second neural network learning data is subjected to feature hiding processing, the hidden neural network learning data is other data in the second neural network learning data except the hidden training supervision data, and the first neural network learning data and the second neural network learning data are determined from the sample order behavior path data sequence and the sample order description portrait;

[0006] Loading the sample remote order data and the sample order behavior path data sequence into the basic order description portrait prediction network to generate a training order description portrait, and generating a target training error value according to the deviation value between the training order description portrait and the sample order description portrait;

[0007] Loading the hidden neural network learning data and the first neural network learning data into the basic order description portrait prediction network to generate hidden observation data, and generating a reconstruction training error value according to the deviation value between the hidden training supervision data and the hidden observation data;

[0008] Update the network configuration information of the basic order description portrait prediction network according to the target training error value and the reconstructed training error value until the training termination requirements are met, and generate a target order description portrait prediction network;

[0009] Obtain any target remote order data and a target order behavior path sequence corresponding to the target remote order data from the remote order big data, load the target remote order data and the target order behavior path sequence into a target order description portrait prediction network, and generate a target order description portrait corresponding to the target remote order data.

[0010] In a possible implementation of the first aspect, the reconstruction training error value includes an order portrait reconstruction error value, and the obtaining of first neural network learning data, hidden training supervision data, and hidden neural network learning data corresponding to the sample remote order data includes:

[0011] Acquire the sample order behavior path data sequence as first neural network learning data, and acquire the sample order description portrait as second neural network learning data;

[0012] Perform feature hiding processing on the sample order description portrait to obtain order portrait hidden training supervision data and order portrait hidden learning data, obtain the order portrait hidden training supervision data as hidden training supervision data, and obtain the order portrait hidden learning data as hidden neural network learning data;

[0013] The step of loading the hidden neural network learning data and the first neural network learning data into the basic order description portrait prediction network to generate hidden observation data, and generating a reconstruction training error value according to a deviation value between the hidden training supervision data and the hidden observation data, comprises:

[0014] Loading the sample order behavior path data sequence and the order portrait hidden learning data into the basic order description portrait prediction network to generate order portrait hidden observation data;

[0015] An order portrait reconstruction error value is generated based on the deviation value between the order portrait hidden training supervision data and the order portrait hidden observation data.

[0016] In a possible implementation manner of the first aspect, the reconstructed training error value includes a behavior path reconstruction error value, and the obtaining of first neural network learning data, hidden training supervision data, and hidden neural network learning data corresponding to the sample remote order data includes:

[0017] Acquire the sample order description portrait as first neural network learning data, and acquire the sample order behavior path data sequence as second neural network learning data;

[0018] Performing feature hiding processing on the sample order behavior path data sequence to obtain path hiding training supervision data and path hiding learning data, obtaining the path hiding training supervision data as hidden training supervision data, and obtaining the path hiding learning data as hidden neural network learning data;

[0019] The step of loading the hidden neural network learning data and the first neural network learning data into the basic order description portrait prediction network to generate hidden observation data, and generating a reconstruction training error value according to a deviation value between the hidden training supervision data and the hidden observation data, comprises:

[0020] Loading the sample order description portrait and the path hidden learning data into the basic order description portrait prediction network to generate path hidden observation data;

[0021] A behavior path reconstruction error value is generated according to a deviation value between the path hiding training supervision data and the path hiding observation data.

[0022] In a possible implementation of the first aspect, loading the sample remote order data and the sample order behavior path data sequence into a basic order description portrait prediction network to generate a training order description portrait includes:

[0023] Loading the sample remote order data, the sample order behavior path data sequence and the sample order description portrait into a basic order description portrait prediction network;

[0024] Embedding the sample remote order data and the sample order behavior path data sequence respectively to generate a sample remote order embedding representation and a sample order behavior path embedding representation;

[0025] Aggregating the sample remote order embedding representation and the sample order behavior path embedding representation to obtain a target sample embedding vector;

[0026] Based on the sample order description portrait, the order description portrait is predicted for the target sample embedding vector to generate a training order description portrait.

[0027] In a possible implementation of the first aspect, respectively embedding the sample remote order data and the sample order behavior path data sequence to generate a sample remote order embedding representation and a sample order behavior path embedding representation includes:

[0028] Extracting order knowledge vectors from the sample remote order data to generate basic order knowledge vectors, encoding the basic order knowledge vectors to generate embedded representations of the sample remote orders;

[0029] Performing order behavior node mining on the sample order behavior path data in the sample order behavior path data sequence to generate an order behavior directed knowledge graph corresponding to at least one target order behavior node;

[0030] Extract order knowledge vectors for the order behavior labels corresponding to each sample order behavior path data, the node labels corresponding to each target order behavior node, and the path positions corresponding to each target order behavior node, respectively, to generate order behavior label features corresponding to each sample order behavior path data, node label vectors corresponding to each target order behavior node, and path position vectors corresponding to each path position;

[0031] Generate an initial order behavior path sub-vector corresponding to the target order behavior node based on the node label vector, path position vector, order behavior directed knowledge graph, and order behavior label features corresponding to the sample order behavior path data corresponding to the same target order behavior node, and generate a basic order behavior knowledge vector based on the initial order behavior path sub-vectors corresponding to each target order behavior node in each sample order behavior path data;

[0032] The basic order behavior knowledge vector is encoded to generate an embedding representation of the sample order behavior path.

[0033] In a possible implementation of the first aspect, the current basic knowledge vector is the basic order knowledge vector or the basic order behavior knowledge vector, and encoding the current basic knowledge vector to generate a corresponding current sample embedding representation includes:

[0034] Performing weight distribution on the current basic knowledge vector to generate a current weight distribution vector, and converging the current basic knowledge vector and the current weight distribution vector to generate an initial converged vector;

[0035] Performing polynomial expansion on the initial converged vector to generate a current polynomial expansion vector, and converging the current polynomial expansion vector and the initial converged vector to generate a target converged vector;

[0036] The current sample embedding representation is obtained according to the target convergence vector.

[0037] In a possible implementation of the first aspect, the aggregating the sample remote order embedding representation and the sample order behavior path embedding representation to obtain a target sample embedding vector includes:

[0038] Performing weight allocation between feature dimensions on the sample remote order embedding representation and the sample order behavior path embedding representation to generate a converged weight allocation vector;

[0039] Fusion of the sample remote order embedding representation and the converged weight allocation vector to generate a first fused order embedding representation, performing linear discriminant analysis on the first fused order embedding representation according to the first network configuration information to generate a first linear discriminant analysis vector, performing feature mapping on the first linear discriminant analysis vector to generate an order behavior path cleaning vector;

[0040] According to the order behavior path cleaning vector, the convergence weight allocation vector is cleaned to generate a convergence enhancement vector;

[0041] The converged enhancement vector and the sample remote order embedding representation are fused to generate the target sample embedding vector.

[0042] In a possible implementation of the first aspect, the sample order description portrait includes a plurality of sample order portrait overviews distributed based on logic between order portrait relationships;

[0043] The step of predicting the order description portrait of the target sample embedding vector based on the sample order description portrait to generate a training order description portrait includes:

[0044] Determine a target logical node from each portrait relationship logical node corresponding to the sample order description portrait;

[0045] From the sample order description portrait, obtain the sample order portrait overview before the target logical node as the candidate order portrait overview, embed the candidate order portrait overview, and generate the candidate order portrait embedding vector;

[0046] Perform weight assignment on the candidate order portrait embedding vector to generate an initial weight assignment vector, obtain the initial order portrait embedding vector based on the initial weight assignment vector and the candidate order portrait embedding vector, perform blending weight assignment on the initial order portrait embedding vector and the target sample embedding vector to generate a blending weight assignment vector, obtain the pending order portrait embedding vector based on the blending weight assignment vector and the initial order portrait embedding vector, perform polynomial expansion on the pending order portrait embedding vector to generate a target order portrait embedding vector, and obtain the predicted order portrait embedding vector based on the pending order portrait embedding vector and the target order portrait embedding vector;

[0047] Perform order description portrait prediction on the predicted order portrait embedding vector to generate a predicted order portrait overview corresponding to the target logical node;

[0048] Obtain the next portrait relationship logic node as the target logic node, return to the step of obtaining the sample order portrait overview before the target logic node from the sample order description portrait as the candidate order portrait overview, and execute until the traversal processing of all portrait relationship logic nodes is completed, and multiple predicted order portrait overviews are generated;

[0049] The training order description portrait is obtained based on the overview of each predicted order portrait.

[0050] In a possible implementation of the first aspect, the sample learning data sequence corresponding to the basic order description portrait prediction network includes neural network learning data corresponding to a plurality of sample remote order data respectively, the neural network learning data includes sample remote order data and a corresponding sample order behavior path data sequence, a sample order description portrait, a first neural network learning data, hidden training supervision data, hidden neural network learning data, and an order service type, and the sample learning data sequence includes at least one order service type;

[0051] The step of loading the sample remote order data and the sample order behavior path data sequence into a basic order description portrait prediction network to generate a training order description portrait, and generating a target training error value according to a deviation value between the training order description portrait and the sample order description portrait includes:

[0052] Loading the sample remote order data and the corresponding sample order behavior path data sequence and order service type in the sample learning data sequence into the basic order description portrait prediction network to generate a training order description portrait that matches the order service type corresponding to the sample remote order data;

[0053] Generate a target sub-training error value based on the deviation value between the training order description image and the sample order description image corresponding to the same sample remote order data, and obtain a target training error value based on the target sub-training error value corresponding to each sample remote order data;

[0054] The step of loading the hidden neural network learning data and the first neural network learning data into the basic order description portrait prediction network to generate hidden observation data, and generating a reconstruction training error value according to a deviation value between the hidden training supervision data and the hidden observation data, comprises:

[0055] Loading the first neural network learning data, hidden neural network learning data and order service type corresponding to the sample remote order data in the sample learning data sequence into the basic order description portrait prediction network to generate hidden observation data matching the order service type corresponding to the sample remote order data;

[0056] Generate a reconstruction sub-training error value based on the deviation value between the hidden training supervision data and the hidden observation data corresponding to the same sample remote order data, and obtain a reconstruction training error value based on the reconstruction sub-training error value corresponding to each sample remote order data;

[0057] The basic order description portrait prediction network includes a first embedding representation unit, a second embedding representation unit and a fully connected output unit;

[0058] The step of loading the sample remote order data and the sample order behavior path data sequence into a basic order description portrait prediction network to generate a training order description portrait, and loading the hidden neural network learning data and the first neural network learning data into a basic order description portrait prediction network to generate hidden observation data, further includes:

[0059] Loading the sample remote order data into the first embedding representation unit, loading the sample order behavior path data sequence into the second embedding representation unit, generating a converged embedding representation vector according to the embedding representation results of the first embedding representation unit and the second embedding representation unit, loading the converged embedding representation vector and the sample order description portrait into a fully connected output unit, and generating the training order description portrait;

[0060] The hidden neural network learning data and the first neural network learning data are loaded into the second embedding representation unit to generate the hidden observation data.

[0061] According to one aspect of an embodiment of the present application, an artificial intelligence system is provided, which includes a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions, and the machine-executable instructions are loaded and executed by the processor to implement a big data mining method for remote order information in any one of the possible implementations described above.

[0062] According to one aspect of the embodiments of the present application, a computer program product or a computer program is provided, the computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the methods provided in various optional implementations of the above three aspects.

[0063] In the technical solutions provided by some embodiments of the present application, the embodiments of the present application determine the first neural network learning data and the second neural network learning data by obtaining sample remote order data and the corresponding sample order behavior path data sequence and sample order description portrait; load the sample remote order data and the sample order behavior path data sequence into the basic order description portrait prediction network, generate the training order description portrait, and calculate the target training error value; load the hidden neural network learning data and the first neural network learning data into the basic order description portrait prediction network, generate the hidden observation data, and calculate the reconstruction training error value; update the network configuration information of the basic order description portrait prediction network according to the target training error value and the reconstruction training error value, and generate the target order description portrait prediction network; obtain the target remote order data and the target order behavior path sequence, load the target remote order data and the target order behavior path sequence into the target order description portrait prediction network, and generate the target order description portrait. Thus, the prediction accuracy of the order description portrait can be improved.

[0064] That is, the embodiment of the present application first generates learning data for training the neural network by acquiring sample remote order data and its corresponding behavior path and order description portrait. Among them, a part of the important feature data is hidden, forming hidden training supervision data and hidden neural network learning data. This setting can prompt the basic order description portrait prediction network to dig deeper and learn the key information hidden in the data. Then, the basic order description portrait prediction network will use these learning data for training, and generate the target training error value and the reconstruction training error value according to the deviation value between the predicted result and the actual result, so as to serve as the key indicator for evaluating the network performance of the basic order description portrait prediction network, and also an important basis for optimizing network parameters.

[0065] Subsequently, by continuously updating and adjusting the network configuration information, the basic order description portrait prediction network will be gradually optimized during the training process until the preset training termination requirements are met. The target order description portrait prediction network finally generated can effectively extract and learn the key features in the order data, thereby accurately describing and predicting the new order data. Therefore, the embodiments of the present application can significantly improve the understanding and prediction capabilities of user order behaviors, further optimize services such as product recommendations and advertising, and enhance user experience and platform economic benefits. At the same time, the difficulty and complexity of training the basic order description portrait prediction network are increased by means of feature hiding and reconstruction error calculation, thereby effectively improving the generalization ability and prediction accuracy of the basic order description portrait prediction network. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required to be used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be extracted in combination with these drawings without creative work.

[0067] Figure 1 A schematic diagram of a process for a big data mining method for remote order information provided in an embodiment of the present application;

[0068] Figure 2 A schematic block diagram of the structure of an artificial intelligence system for implementing the above-mentioned big data mining method for remote order information provided in an embodiment of the present application. DETAILED DESCRIPTION

[0069] The following description is intended to enable one of ordinary skill in the art to implement and combine the present application, and the description is provided in the context of a specific application scenario and its requirements. It will be apparent to one of ordinary skill in the art that various changes may be made to the disclosed embodiments, and that the general principles defined in the present application may be applied to other embodiments and application scenarios without departing from the principles and scope of the present application. Therefore, the present application is not limited to the described embodiments, but should be given the broadest scope consistent with the claims.

[0070] Figure 1 This is a flow chart of a big data mining method for remote order information provided by an embodiment of the present application. The big data mining method for remote order information is introduced in detail below.

[0071] Step S110, obtaining sample remote order data and a sample order behavior path data sequence and a sample order description portrait corresponding to the sample remote order data, and obtaining first neural network learning data, hidden training supervision data and hidden neural network learning data corresponding to the sample remote order data.

[0072] In this embodiment, the hidden training supervision data is the data that is hidden after the second neural network learning data is processed by feature hiding, the hidden neural network learning data is other data in the second neural network learning data except the hidden training supervision data, and the first neural network learning data and the second neural network learning data are determined from the sample order behavior path data sequence and the sample order description portrait.

[0073] The sample remote order data may refer to a portion of the remote order data extracted from a large amount of remote order data as sample data for training the basic order description profile prediction network. Each sample remote order data includes order number, order item, order time, delivery address and other information.

[0074] The sample order behavior path data sequence may refer to the sequence data corresponding to each sample remote order data, which records the entire process from browsing products to placing orders, and may include the time of browsing products, page jumps, adding products to the shopping cart, placing orders, and paying. This data sequence can be used to train the basic order description profile prediction network to predict the user's order behavior.

[0075] The sample order description portrait may refer to data corresponding to each sample remote order data, describing the user's basic information (such as gender, age, occupation, etc.), shopping habits, preferences, etc. This sample order description portrait can be used to train the basic order description portrait prediction network to predict the user's order description.

[0076] The first neural network learning data refers to data determined from a sample order behavior path data sequence and a sample order description portrait, and is used to train a neural network to generate an order description portrait.

[0077] The second neural network learning data refers to data determined from a sample order behavior path data sequence and a sample order description portrait, and is used to train a neural network to generate hidden observation data.

[0078] The hidden training supervision data refers to the data that is hidden after the second neural network learning data is processed by feature hiding, and is used to train the neural network to reconstruct the sample remote order data.

[0079] The hidden neural network learning data refers to other data in the second neural network learning data except the hidden training supervision data, which is used to train the neural network to generate hidden observation data.

[0080] Specifically, a part of the large amount of remote order big data can be extracted as sample remote order data. These sample remote order data include order number, order item, order time, delivery address and other information. At the same time, it is also necessary to obtain the sample order behavior path data sequence of each sample remote order data. This sample order behavior path data sequence may include the time when the user browses the product, page jumps, adds the product to the shopping cart, places an order, pays and other behaviors. In addition, it is also necessary to obtain the sample order description portrait of each sample remote order data. This sample order description portrait may include the user's basic information (such as gender, age, occupation, etc.), the user's shopping habits, the user's preferences, etc. Then, it is necessary to determine the first neural network learning data and the second neural network learning data from the sample order behavior path data sequence and the sample order description portrait. For example, the first neural network learning data is used to train the neural network to generate the order description portrait, and the second neural network learning data is used to train the neural network to generate hidden observation data.

[0081] Step S120, loading the sample remote order data and the sample order behavior path data sequence into the basic order description portrait prediction network, generating a training order description portrait, and generating a target training error value based on the deviation value between the training order description portrait and the sample order description portrait.

[0082] For example, in this step, the AI ​​system will load the sample remote order data and the sample order behavior path data sequence into a basic order description portrait prediction network. This basic order description portrait prediction network will predict the order description portrait based on the input data, that is, the training order description portrait. The AI ​​system will calculate the deviation between the predicted training order description portrait and the actual sample order description portrait, and use this deviation as the target training error value.

[0083] Step S130, loading the hidden neural network learning data and the first neural network learning data into the basic order description portrait prediction network to generate hidden observation data, and generating a reconstruction training error value based on the deviation value between the hidden training supervision data and the hidden observation data.

[0084] For example, in this step, the artificial intelligence system will load the hidden neural network learning data and the first neural network learning data into the basic order description portrait prediction network to generate hidden observation data. The hidden observation data can be regarded as a reconstruction or prediction of the original sample remote order data. The artificial intelligence system will calculate the deviation between the hidden training supervision data and the hidden observation data, and use this deviation as the reconstruction training error value. This reconstruction training error value can help the artificial intelligence system determine which parts of the neural network learning data are most important for predicting the order description portrait, thereby facilitating more effective training.

[0085] Step S140, updating the network configuration information of the basic order description portrait prediction network according to the target training error value and the reconstructed training error value until the training termination requirements are met, thereby generating a target order description portrait prediction network.

[0086] For example, in this step, the AI ​​system will update the network configuration information of the basic order description portrait prediction network according to the target training error value and the reconstruction training error value, such as the number of neural network layers, the number of neurons, the learning rate and other parameters. Then, through continuous iteration of this process, until the target training error value and the reconstruction training error value both meet the training termination requirements, the AI ​​system will finally obtain the optimized target order description portrait prediction network.

[0087] Step S150, obtain any target remote order data and the target order behavior path sequence corresponding to the target remote order data from the remote order big data, load the target remote order data and the target order behavior path sequence into the target order description portrait prediction network, and generate a target order description portrait corresponding to the target remote order data.

[0088] When the artificial intelligence system has obtained the target order description portrait prediction network through training, the target order description portrait prediction network can be used to predict the order description portrait of a new remote order. The following is a detailed example:

[0089] Suppose an e-commerce platform receives a new remote order with order number #20230101, the purchased product is a laptop, the order time is January 1, 2023, and the delivery address is District B, City A. The e-commerce platform wants to use the artificial intelligence system to predict the order description profile of this order, including the user's basic information, shopping habits, preferences, etc.

[0090] First, the AI ​​system will obtain the data of this new remote order, that is, the order number #20230101, the purchased product is a laptop, the order time is January 1, 2023, the delivery address is District B, City A, and the order is placed through Channel C. This data is called the target remote order data.

[0091] Next, the artificial intelligence system can obtain the target order behavior path sequence corresponding to the target remote order data. For example, suppose the user first searches for "laptop" on the platform, then browses several different laptops, and finally selects one of them to place an order. Then, the target order behavior path sequence may include behavior data such as "searching for laptops", "browsing laptop 1", "browsing laptop 2", and "selecting laptop A".

[0092] Then, the artificial intelligence system can load the target remote order data and the target order behavior path sequence into the trained target order description portrait prediction network. This target order description portrait prediction network can predict the order description portrait of the order based on the input data. For example, the predicted order description portrait may include: the user is a young office worker who likes to buy electronic products with good performance, has high brand loyalty, and may pay attention to information such as product price, performance, and brand.

[0093] Finally, the artificial intelligence system outputs the predicted description portrait of the target order, providing data support and decision-making reference for the e-commerce platform's customer service, marketing promotion and other businesses.

[0094] Based on the above steps, the embodiment of the present application determines the first neural network learning data and the second neural network learning data by obtaining sample remote order data and the corresponding sample order behavior path data sequence and sample order description portrait; loads the sample remote order data and the sample order behavior path data sequence into the basic order description portrait prediction network, generates a training order description portrait, and calculates the target training error value; loads the hidden neural network learning data and the first neural network learning data into the basic order description portrait prediction network, generates hidden observation data, and calculates the reconstruction training error value; updates the network configuration information of the basic order description portrait prediction network according to the target training error value and the reconstruction training error value, and generates a target order description portrait prediction network; obtains the target remote order data and the target order behavior path sequence, loads the target remote order data and the target order behavior path sequence into the target order description portrait prediction network, and generates the target order description portrait. Thus, the prediction accuracy of the order description portrait can be improved.

[0095] That is, the embodiment of the present application first generates learning data for training the neural network by acquiring sample remote order data and its corresponding behavior path and order description portrait. Among them, a part of the important feature data is hidden, forming hidden training supervision data and hidden neural network learning data. This setting can prompt the basic order description portrait prediction network to dig deeper and learn the key information hidden in the data. Then, the basic order description portrait prediction network will use these learning data for training, and generate the target training error value and the reconstruction training error value according to the deviation value between the predicted result and the actual result, so as to serve as the key indicator for evaluating the network performance of the basic order description portrait prediction network, and also an important basis for optimizing network parameters.

[0096] Subsequently, by continuously updating and adjusting the network configuration information, the basic order description portrait prediction network will be gradually optimized during the training process until the preset training termination requirements are met. The target order description portrait prediction network finally generated can effectively extract and learn the key features in the order data, thereby accurately describing and predicting the new order data. Therefore, the embodiments of the present application can significantly improve the understanding and prediction capabilities of user order behaviors, further optimize services such as product recommendations and advertising, and enhance user experience and platform economic benefits. At the same time, the difficulty and complexity of training the basic order description portrait prediction network are increased by means of feature hiding and reconstruction error calculation, thereby effectively improving the generalization ability and prediction accuracy of the basic order description portrait prediction network.

[0097] In a possible implementation, the reconstruction training error value includes an order portrait reconstruction error value, and the step S110 may include:

[0098] Step S111, obtaining the sample order behavior path data sequence as the first neural network learning data, and obtaining the sample order description portrait as the second neural network learning data.

[0099] Step S112, perform feature hiding processing on the sample order description portrait to obtain order portrait hidden training supervision data and order portrait hidden learning data, obtain the order portrait hidden training supervision data as hidden training supervision data, and obtain the order portrait hidden learning data as hidden neural network learning data.

[0100] For example, this process involves two different data inputs and a feature hiding process. 1. Obtain a sample order behavior path data sequence as the first neural network learning data. This part of the data usually contains various behavior records of users on the e-commerce platform, such as browsing, searching, clicking, purchasing, etc. For example, if a user searches for "mobile phone" successively, clicks on several brands of mobile phones, and then buys one, these behaviors and the order in which they occur constitute the order behavior path data. Then, obtain a sample order description portrait as the second neural network learning data: This part of the data is mainly about the description information of the order, which may include the product category, price, purchase time, etc. of the order. For example, an order may be described as: "Purchased an Apple mobile phone for 1,000 yuan."

[0101] The sample order description portrait is subjected to feature hiding to obtain order portrait hidden training supervision data and order portrait hidden learning data. This process can be understood as a certain masking or modification operation on the original order description portrait. For example, a part of the features (such as price) can be randomly selected and replaced with a specific tag (such as "UNKNOWN") to generate a hidden order description portrait. The original order description portrait (unmasked) is retained as hidden training supervision data for subsequent error calculation; and the masked order description portrait is used as hidden neural network learning data and input into the model for training. Obtaining the order portrait hidden training supervision data as hidden training supervision data and obtaining the order portrait hidden learning data as hidden neural network learning data is to use the two types of data generated in the previous step for model training. The model needs to try to restore the masked features to learn the intrinsic structure of the data. Therefore, the basic order description portrait prediction network can be trained by feature hiding and reconstruction, so that the basic order description portrait prediction network can better understand and abstract the intrinsic structure of the data.

[0102] Step S130 may include:

[0103] Step S131, loading the sample order behavior path data sequence and the order portrait hidden learning data into the basic order description portrait prediction network to generate order portrait hidden observation data.

[0104] Step S132, generating an order portrait reconstruction error value based on the deviation value between the order portrait hidden training supervision data and the order portrait hidden observation data.

[0105] For example, in the previous step, we have obtained the order behavior path data and the masked order description portrait (i.e., order portrait hidden learning data). Now, these two types of data are input into the basic order description portrait prediction network. For example, the user's behavior data may include "searching for mobile phones", "clicking on Apple brands", etc., while the hidden order description portrait may be "purchasing an Apple phone worth UNKNOWN yuan".

[0106] Then, the order portrait hidden observation data is output based on the input sample order behavior path data sequence and the order portrait hidden learning data. This order portrait hidden observation data can be regarded as the prediction of the masked features by the basic order description portrait prediction network. For example, the basic order description portrait prediction network may predict "a 1,000 yuan Apple mobile phone was purchased".

[0107] The order portrait hidden training supervision data is the original, unmasked order description portrait, such as "purchased an Apple mobile phone for 1,000 yuan". This data is compared with the prediction results of the basic order description portrait prediction network (order portrait hidden observation data), and the deviation between the two is calculated, that is, the prediction error. Common error calculation methods include mean square error or cross entropy loss. Therefore, by training the basic order description portrait prediction network to predict the masked features, and adjusting the parameters of the basic order description portrait prediction network by calculating the prediction error, the basic order description portrait prediction network's understanding of the data structure and prediction ability are improved.

[0108] In a possible implementation manner, the reconstruction training error value includes a behavior path reconstruction error value, and step S110 may further include:

[0109] Step S113, obtaining the sample order description portrait as the first neural network learning data, and obtaining the sample order behavior path data sequence as the second neural network learning data.

[0110] Step S114, performing feature hiding processing on the sample order behavior path data sequence to obtain path hiding training supervision data and path hiding learning data, obtaining the path hiding training supervision data as hidden training supervision data, and obtaining the path hiding learning data as hidden neural network learning data.

[0111] For obtaining a sample order description portrait as the first neural network learning data, for example, there is an order data of a user purchasing a product, which includes the user's basic information (such as gender, age, occupation, etc.), shopping habits, preferences, etc. These constitute the sample order description portrait, and this part of data will be used for training the basic order description portrait prediction network.

[0112] For example, the sample order behavior path data sequence is obtained as the second neural network learning data, which usually includes various behavior records of users on the e-commerce platform, such as browsing, clicking, purchasing, etc. For example, the user searched for "iPhone 12" successively, then clicked to view the detailed information of a certain iPhone 12 phone, and finally completed the purchase. This series of behaviors constitutes a behavior path data sequence, which will also be used for the training of the basic order description portrait prediction network.

[0113] The sample order behavior path data sequence is processed by feature hiding to obtain path hiding training supervision data and path hiding learning data. The feature hiding processing here can be understood as some kind of masking or modification operation on the original behavior path data so that the model can learn and predict the hidden information from other information. For example, the "iPhone 12" in the above behavior path can be replaced with a specific marker "UNKNOWN", so that the original behavior path is transformed into a new behavior path containing hidden information. Among them, the new behavior path containing "UNKNOWN" is the path hiding training supervision data, and the remaining data (that is, data other than "UNKNOWN") is the path hiding learning data.

[0114] In the process of obtaining the path hidden training supervision data as the hidden training supervision data and obtaining the path hidden learning data as the hidden neural network learning data, the path hidden training supervision data and the path hidden learning data obtained in the previous step can be directly used as the input data of the basic order description portrait prediction network.

[0115] Therefore, by processing the original order data, various types of learning data suitable for basic order description portrait prediction network training are generated. At the same time, the difficulty and complexity of model training are increased through feature hiding methods, thereby improving the model's learning ability and prediction accuracy.

[0116] Step S130 may further include:

[0117] Step S131, loading the sample order description portrait and the path hidden learning data into the basic order description portrait prediction network to generate path hidden observation data.

[0118] Step S132: generating a behavior path reconstruction error value according to a deviation value between the path hiding training supervision data and the path hiding observation data.

[0119] For example, the sample order description profile (e.g., user's basic information, shopping habits, etc.) and path hidden learning data (e.g., behavior path data other than "UNKNOWN") obtained in the previous step can be used as input and loaded into the basic order description profile prediction network. Then, the basic order description profile prediction network will calculate based on these input data and output the predicted behavior path, which is the so-called path hidden observation data.

[0120] There may be a certain deviation between the predicted behavior path (i.e., path-hidden observation data) and the actual behavior path (i.e., path-hidden training supervision data). This deviation reflects the prediction accuracy of the basic order description portrait prediction network. This deviation, that is, the behavior path reconstruction error value, can be calculated by comparing the predicted results and the actual results. For example, if the predicted behavior path is "search UNKNOWN, click UNKNOWN, buy UNKNOWN", and the actual behavior path is "search iPhone 12, click iPhone 12, buy iPhone 12", then the behavior path reconstruction error value is the degree of deviation between the path-hidden training supervision data and the path-hidden observation data. This behavior path reconstruction error value will be used for subsequent network optimization and updates.

[0121] In this way, by calculating the error between the predicted result and the actual result of the basic order description portrait prediction network, we can know the prediction accuracy of the basic order description portrait prediction network in the current state, thereby providing a basis for the optimization and update of the basic order description portrait prediction network.

[0122] In a possible implementation, step S120 may include:

[0123] Step S121, loading the sample remote order data, the sample order behavior path data sequence and the sample order description portrait into the basic order description portrait prediction network.

[0124] Step S122, embedding the sample remote order data and the sample order behavior path data sequence respectively to generate a sample remote order embedding representation and a sample order behavior path embedding representation.

[0125] For example, the sample remote order data and the sample order behavior path data sequence are converted into numerical representation, i.e., embedded representation. Embedded representation is a method of compressing high-dimensional data into a low-dimensional space, which can retain the important features of the original data. In this embodiment, methods such as word embedding or vector embedding can be used to convert the sample remote order data and the sample order behavior path data sequence into numerical representation.

[0126] Step S123, aggregating the sample remote order embedding representation and the sample order behavior path embedding representation to obtain a target sample embedding vector.

[0127] For example, this step includes merging the remote order embedding representation and the order behavior path embedding representation to generate a target sample embedding vector. The merging method can be simple concatenation, weighted average, or more complex network structures such as attention mechanism. The target sample embedding vector can be used to represent the overall characteristics of the order in order to predict the order description portrait.

[0128] Step S124: Based on the sample order description portrait, the order description portrait is predicted for the target sample embedding vector to generate a training order description portrait.

[0129] In a possible implementation, step S122 may include:

[0130] Step S1221, extracting order knowledge vectors from the sample remote order data to generate basic order knowledge vectors, encoding the basic order knowledge vectors to generate embedded representations of the sample remote orders.

[0131] For example, this step includes using natural language processing (NLP) techniques, such as Bag of Words, TF-IDF, or Word Embedding, to vectorize the sample remote order data. In this way, keywords and features in the sample remote order data can be extracted and converted into numerical vectors. For example, an encoder can be used to convert the basic order knowledge vector into a vector representation of a fixed length. The encoder can use deep learning models such as recurrent neural networks (RNN), long short-term memory networks (LSTM), or transformers. Through the encoded representation, the sample remote order data can be converted into a numerical vector for subsequent processing.

[0132] Step S1222, performing order behavior node mining on the sample order behavior path data in the sample order behavior path data sequence, and generating an order behavior directed knowledge graph corresponding to at least one target order behavior node.

[0133] For example, graph mining techniques such as frequent pattern mining, association rule mining or social network analysis can be used to mine key nodes and relationships in order behavior path data. These key nodes and relationships can form a directed knowledge graph of order behavior to represent the complex relationships and structures of order behavior paths.

[0134] Step S1223, extract order knowledge vectors for the order behavior labels corresponding to each sample order behavior path data, the node labels corresponding to each target order behavior node, and the path positions corresponding to each target order behavior node, to generate order behavior label features corresponding to each sample order behavior path data, node label vectors corresponding to each target order behavior node, and path position vectors corresponding to each path position.

[0135] For example, natural language processing (NLP) techniques, such as Bag of Words, TF-IDF, or Word Embedding, can be used to vectorize the order behavior labels corresponding to each sample order behavior path data, the node labels corresponding to each target order behavior node, and the path positions corresponding to each target order behavior node. In this way, important features and relationships in the sample order behavior path data can be extracted and converted into numerical vectors.

[0136] Step S1224, based on the node label vector, path position vector, order behavior directed knowledge graph corresponding to the same target order behavior node, and the order behavior label features corresponding to the sample order behavior path data, generate the initial order behavior path sub-vector corresponding to the target order behavior node, and based on the initial order behavior path sub-vectors corresponding to each target order behavior node in each sample order behavior path data, generate the basic order behavior knowledge vector.

[0137] For example, a model such as a graph neural network (GNN) or a structured neural network (Structured Neural Network) can be used to fuse the node label vector, path position vector, order behavior directed knowledge graph, and order behavior label features corresponding to the sample order behavior path data corresponding to the same target order behavior node to generate an initial order behavior path subvector corresponding to the target order behavior node. This initial order behavior path subvector can represent the characteristics and importance of the target order behavior node in the order behavior path.

[0138] Then, the initial order behavior path sub-vectors corresponding to each target order behavior node can be aggregated to generate a basic order behavior knowledge vector. The aggregation method can be a simple summation, average value, or a more complex network structure such as attention mechanism. The basic order behavior knowledge vector can represent the overall characteristics and importance of the order behavior path.

[0139] Step S1225, encoding the basic order behavior knowledge vector to generate the sample order behavior path embedding representation.

[0140] For example, an encoder can be used to convert the basic order behavior knowledge vector into a fixed-length vector representation. The encoder can use deep learning models such as recurrent neural networks (RNN), long short-term memory networks (LSTM), or transformers. Through encoding representation, the order behavior path data can be converted into a numerical vector for subsequent processing and use.

[0141] Through the above steps, the order remote order data and order behavior path data can be converted into vector representations, namely the sample remote order embedding representation and the sample order behavior path embedding representation. These vectors can be used to train the basic order description profile prediction network to predict the order description profile and behavior path.

[0142] In a possible implementation, the current basic knowledge vector is the basic order knowledge vector or the basic order behavior knowledge vector, and encoding the current basic knowledge vector to generate a corresponding current sample embedding representation includes:

[0143] First, weight allocation is performed on the current basic knowledge vector to generate a current weight allocation vector, and the current basic knowledge vector and the current weight allocation vector are converged to generate an initial converged vector.

[0144] For example, the current basic knowledge vector (such as order behavior type, order behavior time, order behavior frequency, etc.) needs to be weighted. A weight assignment method, such as weighted average (WMA) or weighted maximum (WMM), can be used to assign different weights to each feature. These weights can be set based on the importance or relevance of the feature. Then, the weight assignment vector is converged with the basic knowledge vector to generate an initial converged vector. This convergence process can use simple summation, average value, or more complex network structures, such as attention mechanisms.

[0145] Then, the initial converged vector is polynomially expanded to generate a current polynomial extended vector, and the current polynomial extended vector and the initial converged vector are converged to generate a target converged vector.

[0146] For example, the initial converged vector can be polynomially expanded. This process can use a polynomial expansion method, such as Polynomial Interpolation or Polynomial Feature Transformation, to generate multiple new features for each feature. Then, the polynomial expansion vector is converged with the initial converged vector to generate a target converged vector. This convergence process can use simple summation, average value or more complex network structures, such as attention mechanism.

[0147] Finally, the current sample embedding representation is obtained based on the target convergence vector. For example, these target convergence vectors can be used as input to train a basic order description profile prediction network (such as a multi-layer perceptron (MLP) or a convolutional neural network (CNN)) to predict the relationship between the order description profile and the order behavior path. Finally, a sample embedding representation can be obtained, which can represent the relationship between the order description profile and the order behavior path. This sample embedding representation can be used for subsequent classification, clustering or other tasks.

[0148] In a possible implementation, step S123 may include:

[0149] Step S1231, performing weight allocation between feature dimensions on the sample remote order embedding representation and the sample order behavior path embedding representation to generate a converged weight allocation vector.

[0150] For example, suppose that the sample remote order embedding representation and the sample order behavior path embedding representation are both 100-dimensional vectors. The sample remote order embedding representation may contain information such as the user's shopping history and preferences, while the sample order behavior path embedding representation may contain the user's behavior information from browsing products, adding to the shopping cart to placing an order.

[0151] In the step of assigning weights between feature dimensions, we first need to determine the importance of each dimension of the feature. This is usually based on experience or obtained through data statistics. For example, if it is found that the user's shopping history has a greater impact on predicting their future purchasing behavior, then the dimension representing the shopping history in the sample remote order embedding representation can be given a higher weight. Similarly, if the user's product browsing behavior before purchase has a greater impact on the final purchase decision, then the dimension representing the browsing behavior in the sample order behavior path embedding representation can also be given a higher weight.

[0152] Specifically, suppose the weights of the 1st, 2nd, and 3rd dimensional features (representing shopping history) in the sample remote order embedding representation are set to 0.7, and the weights of the remaining features are set to 0.3; and the weights of the 1st, 2nd, and 3rd dimensional features (representing browsing behavior) in the sample order behavior path embedding representation are set to 0.7, and the weights of the remaining features are set to 0.3. Then a converged weight distribution vector is obtained.

[0153] In this way, each dimension of features can be given corresponding weights according to its importance, so that these important features can be emphasized more in subsequent processing and the prediction accuracy of the model can be improved.

[0154] For another example, suppose the sample remote order embedding representation contains 10 feature dimensions, each of which corresponds to a different aspect of the order, such as product type, purchase quantity, user history, etc. Similarly, suppose the sample order behavior path embedding representation also contains 10 feature dimensions, representing different behaviors of users in the process of completing an order, such as browsing products, adding to shopping carts, and choosing payment methods.

[0155] In order to distribute weights between feature dimensions, a machine learning-based method can also be used, such as using models such as a gradient boosting decision tree or a neural network to learn the importance of each feature.

[0156] First, you need to prepare a set of training data, which contains sample remote order embedding representations and sample order behavior path embedding representations as input features, as well as corresponding target variables (such as the label of the order description portrait or the actual result of the behavior path).

[0157] Then, the training data is trained using the selected machine learning model (such as gradient boosted decision tree). During the training process, the model learns the predictive power of each feature for the target variable. After the training is completed, the feature importance evaluation results provided by the model can be used to determine the weights of each feature dimension. These weights reflect the importance of each feature in predicting the target variable.

[0158] Based on the feature importance evaluation results, a converged weight allocation vector can be created. Each element of the converged weight allocation vector corresponds to a feature dimension and is assigned a corresponding weight value. The weight value can be a normalized value to ensure that the sum of all weights is 1.

[0159] For example, if the importance of a feature dimension is evaluated as 0.2, and the average importance of other feature dimensions is 0.1, the weight of this dimension can be set to a relatively high value, such as 0.3. The weights of other dimensions are adjusted accordingly based on their importance.

[0160] Once the convergence weight assignment vector is obtained, it can be applied to subsequent convergence steps. For example, when calculating the target sample embedding vector, these weights can be used to adjust the way the remote order embedding representation and the order behavior path embedding representation are merged. By weighted merging of different feature dimensions by weight, it can be ensured that the importance and contribution of each feature dimension are taken into account when generating the target sample embedding vector. Such a process can be adjusted and optimized according to specific application scenarios and data sets. At the same time, different machine learning models and evaluation indicators can be tried to further improve the accuracy and effectiveness of weight assignment.

[0161] Step S1232: Fuse the sample remote order embedding representation and the converged weight allocation vector to generate a first fused order embedding representation; perform linear discriminant analysis on the first fused order embedding representation according to the first network configuration information to generate a first linear discriminant analysis vector; perform feature mapping on the first linear discriminant analysis vector to generate an order behavior path cleaning vector.

[0162] For example, after obtaining the sample remote order embedding representation and the converged weight allocation vector, the next step is to fuse them together. This fusion process can be a simple weighted average or a more complex operation. For example, assuming that the weighted average is selected, then for each dimension of the feature, its value in the first fused order embedding representation is the weighted average of the sample remote order embedding representation and the converged weight allocation vector in that dimension.

[0163] Next, according to the first network configuration information (such as model parameters, activation function, etc.), a linear discriminant analysis is performed on the first fused order embedding representation. Linear discriminant analysis is a dimensionality reduction technique that aims to map high-dimensional data to a low-dimensional space while maintaining the distinction between different categories as much as possible. In this process, a new vector, the first linear discriminant analysis vector, is obtained.

[0164] Then, it is necessary to perform feature mapping on the first linear discriminant analysis vector to generate an order behavior path cleaning vector. The purpose of feature mapping is to transform the original feature space into a new feature space, in which the structure or distribution characteristics of the data may be more conducive to subsequent processing and analysis. For example, a kernel function can be used to map data to a high-dimensional space, or a deep neural network can be used to learn more complex feature representations.

[0165] Through this series of operations, a new vector that can better express the original data and is suitable for subsequent processing is obtained - the order behavior path cleaning vector. Through the above steps, the remote order embedding representation and the convergence weight allocation vector can be fused, and the cleaning vector related to the order behavior path can be obtained through linear discriminant analysis and feature mapping. This cleaning vector can be used for subsequent order processing and analysis tasks, such as order classification, anomaly detection, etc. It should be noted that the specific implementation details and parameter settings may need to be adjusted and optimized according to the actual data set and application scenario.

[0166] Step S1233, cleaning the convergence weight allocation vector according to the order behavior path cleaning vector to generate a convergence enhancement vector.

[0167] For example, after obtaining the order behavior path cleaning vector, it can be used to further optimize the previously obtained convergence weight distribution vector. This optimization process is usually called cleaning.

[0168] For example, suppose the order behavior path cleaning vector is a 100-dimensional vector, each dimension contains a value between 0 and 1, representing the importance of the feature in predicting user behavior. At the same time, there is already a 100-dimensional convergence weight distribution vector.

[0169] During the cleaning process, each dimension of the order behavior path cleaning vector can be multiplied by the corresponding dimension of the convergence weight allocation vector. For example, if the first dimension value of the cleaning vector is 0.8, and the first dimension value of the convergence weight allocation vector is 0.5, then the first dimension value of the new vector after cleaning (that is, the convergence enhancement vector) is 0.8*0.5=0.4. In this way, the original convergence weight allocation vector can be adjusted to better reflect the importance of each feature in predicting user behavior. Through the above steps, the convergence enhancement vector is obtained. This vector retains the original feature information while adding important information learned from the user behavior path, so that it can more accurately describe and predict user behavior.

[0170] Step S1234, fusing the converged enhancement vector and the sample remote order embedding representation to generate the target sample embedding vector.

[0171] For example, next, the converged enhancement vector and the sample remote order embedding representation need to be fused to generate the target sample embedding vector. This process can be done in a variety of ways, including but not limited to simple vector addition, vector concatenation, or more complex methods such as multiplication or weighted average.

[0172] In a possible implementation, the sample order description portrait includes a plurality of sample order portrait overviews distributed based on logic between order portrait relationships. For example, the sample order portrait overview may include information such as the user's shopping habits, product preferences, and consumption level.

[0173] Then, based on this sample order description portrait, it is necessary to predict the order description portrait of the target sample embedding vector (the one mentioned above) and generate a training order description portrait.

[0174] The step S124 may include:

[0175] Step S1241, determining the target logical node from each portrait relationship logical node corresponding to the sample order description portrait.

[0176] For example, the portrait relationship logic node refers to the various features or attributes that constitute the user order description portrait, and there may be a certain logical relationship between them. For example, a user's shopping habits, consumption level, product preferences, etc. can all be regarded as portrait relationship logic nodes.

[0177] When determining the target logical node from these logical nodes, it is usually selected based on the needs of the prediction task and the degree of influence of each node on the prediction result. For example, if the prediction of the user's next purchase behavior is being conducted, the logical node related to shopping habits and product preferences may be selected as the target logical node.

[0178] For example, suppose there are the following logical nodes: "user age", "user gender", "shopping frequency", "shopping time" and "shopping category". When predicting the types of products that users may buy in the future, "shopping frequency", "shopping time" and "shopping category" may be more important, so these three may be selected as target logical nodes.

[0179] After the target logical nodes are determined, further analysis and processing can be performed on these target logical nodes, such as extracting corresponding features, embedding representations, and then making predictions based on these representations.

[0180] Step S1242, from the sample order description portrait, obtain the sample order portrait overview before the target logical node as the candidate order portrait overview, embed the candidate order portrait overview, and generate a candidate order portrait embedding vector.

[0181] For example, in the sample order description portrait, "the sample order portrait overview before the target logical node" may refer to the part of the order portrait information before the target logical node in the time sequence or logical sequence. This information is used as the candidate order portrait overview.

[0182] For example, if the target logical node is "shopping category", then the sample order profile overview before this node may include "user age", "user gender", "shopping frequency" and "shopping time". This information can help better understand the user's shopping habits and provide a basis for predicting the types of goods they may buy.

[0183] Embedding these candidate order portrait overviews means converting them into a numerical form for computer processing. There are many specific embedding methods, such as using word embedding to process text information and using one-hot encoding to process classification information. The generated candidate order portrait embedding vector is the embedding representation of the candidate order portrait overview. This vector contains a lot of information, but the form has changed from the original text or classification to a numerical value, which makes it possible to perform various mathematical operations on it, such as weighting, fusion, etc.

[0184] Step S1243, weight the candidate order portrait embedding vectors to generate an initial weight allocation vector, obtain the initial order portrait embedding vector based on the initial weight allocation vector and the candidate order portrait embedding vector, perform fusion weight allocation on the initial order portrait embedding vector and the target sample embedding vector to generate a fusion weight allocation vector, obtain the pending order portrait embedding vector based on the fusion weight allocation vector and the initial order portrait embedding vector, perform polynomial expansion on the pending order portrait embedding vector to generate a target order portrait embedding vector, and obtain the predicted order portrait embedding vector based on the pending order portrait embedding vector and the target order portrait embedding vector.

[0185] For example, let's say there is a three-dimensional candidate order portrait embedding vector [0.2, 0.3, 0.5]. Through some method (such as training, experience, etc.), the weight of each dimension is obtained, assuming it is [0.5, 0.3, 0.2]. Then, this weight vector is the initial weight distribution vector. Then, multiply each dimension of the candidate order portrait embedding vector by its corresponding weight, and then sum it to get a new value to form the initial order portrait embedding vector. Next, similar to the above operation, except that this time the weight distribution is performed between the initial order portrait embedding vector and the target sample embedding vector. The pending order portrait embedding vector is obtained based on the blended weight distribution vector and the initial order portrait embedding vector. The operation of this step is also weighted summation. The pending order portrait embedding vector is polynomially expanded to generate the target order portrait embedding vector. This is to increase the interaction between features and capture more complex patterns. For example, a two-dimensional vector [0.2, 0.3] can be polynomially expanded to [0.2, 0.3, 0.2^2, 0.3^2, 0.2*0.3]. The predicted order portrait embedding vector is obtained according to the pending order portrait embedding vector and the target order portrait embedding vector. This may involve some models based on machine learning or deep learning, which are used to predict the results according to the input feature vector.

[0186] In general, this process is to weight, fuse and expand various features to generate a feature vector that can be used for prediction.

[0187] Step S1244, perform order description portrait prediction on the predicted order portrait embedding vector to generate a predicted order portrait overview corresponding to the target logical node.

[0188] Step S1245, obtain the next portrait relationship logical node as the target logical node, return to the step of obtaining the sample order portrait overview before the target logical node from the sample order description portrait as the candidate order portrait overview, and execute until the traversal processing of all portrait relationship logical nodes is completed to generate multiple predicted order portrait overviews.

[0189] Step S1246, obtaining the training order description portrait based on the overview of each predicted order portrait.

[0190] Next, it is necessary to obtain the next portrait relationship logic node as the target logic node, and return to step S1242 to continue processing the new target logic node. This process will be repeated until all portrait relationship logic nodes have been processed. Finally, the training order description portrait is obtained based on the overview of each predicted order portrait. This description portrait is the final output of the network, which will be used for subsequent tasks such as user behavior prediction, product recommendation, etc.

[0191] In one possible implementation, the sample learning data sequence corresponding to the basic order description portrait prediction network includes neural network learning data corresponding to multiple sample remote order data respectively, and the neural network learning data includes sample remote order data and corresponding sample order behavior path data sequence, sample order description portrait, first neural network learning data, hidden training supervision data, hidden neural network learning data and order service type, and the sample learning data sequence includes at least one order service type.

[0192] The step S120 may include:

[0193] Step A110, load the sample remote order data and the corresponding sample order behavior path data sequence and order service type in the sample learning data sequence into the basic order description portrait prediction network to generate a training order description portrait that matches the order service type corresponding to the sample remote order data.

[0194] Step A120, generate a target sub-training error value based on the deviation value between the training order description portrait and the sample order description portrait corresponding to the same sample remote order data, and obtain a target training error value based on the target sub-training error values ​​corresponding to each sample remote order data.

[0195] For example, this process is a neural network training process, which aims to train the basic order description portrait prediction network through sample learning data sequences and generate training order description portraits. The following is a specific scenario description of each step:

[0196] First, there is a basic order description portrait prediction network and a corresponding sample learning data sequence. This data sequence includes multiple sample remote order data and corresponding neural network learning data. Among them, the neural network learning data includes sample remote order data, corresponding sample order behavior path data sequence, sample order description portrait, first neural network learning data, hidden training supervision data, hidden neural network learning data and order service type.

[0197] Next, you need to load the sample remote order data, sample order behavior path data sequence, and order service type into the basic order description portrait prediction network. This is a standard neural network training process. The network will learn based on the input data and generate a training order description portrait that matches the order service type corresponding to the sample remote order data.

[0198] Then, it is necessary to calculate the deviation between the training order description portrait and the sample order description portrait corresponding to the same sample remote order data to generate the target sub-training error value. This deviation value can be understood as the difference between the predicted result (training order description portrait) and the actual result (sample order description portrait). Common calculation methods include mean square error, cross entropy, etc.

[0199] Finally, the target training error value needs to be obtained based on the target sub-training error values ​​corresponding to each sample remote order data. This error value represents the overall prediction effect and is usually used to monitor the learning process of the network and to update the network parameters in back propagation.

[0200] In general, this process is a process of using neural networks to make predictions and adjusting model parameters based on the prediction results, with the aim of enabling the model to better predict order description portraits.

[0201] On this basis, step S130 may include: loading the first neural network learning data, hidden neural network learning data and order service type corresponding to the sample remote order data in the sample learning data sequence into the basic order description portrait prediction network, and generating hidden observation data matching the order service type corresponding to the sample remote order data. Then, a reconstruction sub-training error value is generated based on the deviation value between the hidden training supervision data and the hidden observation data corresponding to the same sample remote order data, and a reconstruction training error value is obtained based on the reconstruction sub-training error values ​​corresponding to each sample remote order data.

[0202] This process is part of neural network training and is mainly used to generate hidden observation data and calculate the reconstruction training error value.

[0203] First, the hidden neural network learning data and the first neural network learning data need to be loaded into the basic order description portrait prediction network. These data may include features of sample remote orders, such as user information, product information, etc., as well as hidden data related to them, such as hidden layer representations obtained through autoencoders or other masking methods.

[0204] The basic order description portrait prediction network learns based on the input data and generates hidden observation data. These data can be regarded as an internal representation of the input data by the basic order description portrait prediction network, which reflects the basic order description portrait prediction network's understanding and abstraction of the input data.

[0205] Next, we need to calculate the deviation between the hidden training supervision data and the hidden observation data corresponding to the same sample remote order data to generate a reconstructed sub-training error value. This reconstructed sub-training error value can be understood as the difference between the network prediction result (hidden observation data) and the true result (hidden training supervision data). Usually, a certain loss function is used to calculate this reconstructed sub-training error value, such as the mean square error or cross entropy loss function.

[0206] Finally, the reconstruction training error value needs to be obtained according to the reconstruction sub-training error values ​​corresponding to each sample remote order data. This reconstruction training error value reflects the overall prediction effect and is usually used to monitor the learning process of the basic order description portrait prediction network and to update the parameters of the basic order description portrait prediction network in back propagation.

[0207] The basic order description portrait prediction network includes a first embedding representation unit, a second embedding representation unit and a fully connected output unit.

[0208] The method of loading the sample remote order data and the sample order behavior path data sequence into the basic order description portrait prediction network to generate a training order description portrait, and loading the hidden neural network learning data and the first neural network learning data into the basic order description portrait prediction network to generate hidden observation data, further includes: loading the sample remote order data into the first embedding representation unit, loading the sample order behavior path data sequence into the second embedding representation unit, generating a converged embedding representation vector based on the embedding representation results of the first embedding representation unit and the second embedding representation unit, loading the converged embedding representation vector and the sample order description portrait into the fully connected output unit to generate the training order description portrait. Then, the hidden neural network learning data and the first neural network learning data are loaded into the second embedding representation unit to generate the hidden observation data.

[0209] For example, first, the sample remote order data is loaded into the first embedding representation unit. For example, if the remote order data contains information such as the user's age and gender, then this information will be sent to the first embedding representation unit for processing. At the same time, the sample order behavior path data sequence is loaded into the second embedding representation unit. These data may include user behavior records on the platform, such as browsing, clicking, purchasing and other events and their timestamps. The first embedding representation unit and the second embedding representation unit will embed the input data respectively, and then merge the embedding representation results of the two parts to generate a converged embedding representation vector. This converged embedding representation vector integrates the information of the order data and the behavior path data.

[0210] Next, the converged embedding representation vector and the sample order description image are loaded into the fully connected output unit to generate the training order description image. The fully connected output unit predicts the order description image based on the input feature vector. On the other hand, the hidden neural network learning data and the first neural network learning data are loaded into the second embedding representation unit to generate hidden observation data. This process is similar to the previous steps, except that the hidden data is processed this time.

[0211] In general, this process involves operations such as neural network embedding, feature fusion, and prediction. The goal is to generate training order description portraits and hidden observation data based on the input sample data.

[0212] Figure 2 The hardware structure of the artificial intelligence system 100 provided in the embodiment of the present application for implementing the above-mentioned big data mining method for remote order information is shown as follows: Figure 2 As shown, the artificial intelligence system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0213] In an alternative embodiment, the artificial intelligence system 100 may be a single server or a server group. The server group may be centralized or distributed (for example, the artificial intelligence system 100 may be a distributed system). In an alternative embodiment, the artificial intelligence system 100 may be local or remote. For example, the artificial intelligence system 100 may access information and / or data stored in the machine-readable storage medium 120 via a network. For another example, the artificial intelligence system 100 may be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In an alternative embodiment, the artificial intelligence system 100 may be implemented on a cloud platform. By way of example only, the cloud platform may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, or the like, or any combination thereof.

[0214] The machine-readable storage medium 120 may store data and / or instructions. In an alternative embodiment, the machine-readable storage medium 120 may store data obtained from an external terminal. In an alternative embodiment, the machine-readable storage medium 120 may store data and / or instructions used by the artificial intelligence system 100 to execute or use to complete the exemplary methods described in this application. In an alternative embodiment, the machine-readable storage medium 120 may include a mass storage, a removable storage, a volatile read-write memory, a read-only memory, or the like, or any combination thereof. Exemplary mass storage may include a magnetic disk, an optical disk, a solid-state disk, or the like. Exemplary removable storage may include a flash drive, a floppy disk, an optical disk, a memory card, a compressed disk, a magnetic tape, or the like.

[0215] During the specific implementation process, multiple processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the big data mining method for remote order information in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0216] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned artificial intelligence system 100. The implementation principles and technical effects are similar and will not be repeated in this embodiment.

[0217] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the above-mentioned big data mining method for remote order information is implemented.

[0218] Similarly, it should be noted that in order to simplify the description disclosed in this application, thereby helping to understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof. Similarly, it should be noted that in order to simplify the description disclosed in this application, thereby helping to understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof.

Claims

1. A big data mining method for remote order information, characterized in that: The method comprises: Obtain sample remote order data and a sample order behavior path data sequence and a sample order description portrait corresponding to the sample remote order data, and obtain first neural network learning data, hidden training supervision data and hidden neural network learning data corresponding to the sample remote order data; the hidden training supervision data is data hidden after the second neural network learning data is processed by feature hiding, and the hidden neural network learning data is other data in the second neural network learning data except the hidden training supervision data, the first neural network learning data and the second neural network learning data are determined from the sample order behavior path data sequence and the sample order description portrait, and the sample order behavior path data sequence refers to sequence data corresponding to each sample remote order data, recording the entire process of the user from browsing products to placing an order; Loading the sample remote order data and the sample order behavior path data sequence into the basic order description portrait prediction network to generate a training order description portrait, and generating a target training error value according to the deviation value between the training order description portrait and the sample order description portrait; Loading the hidden neural network learning data and the first neural network learning data into the basic order description portrait prediction network to generate hidden observation data, and generating a reconstruction training error value according to the deviation value between the hidden training supervision data and the hidden observation data; Update the network configuration information of the basic order description portrait prediction network according to the target training error value and the reconstructed training error value until the training termination requirements are met, and generate a target order description portrait prediction network; Obtain any target remote order data and a target order behavior path sequence corresponding to the target remote order data from the remote order big data, load the target remote order data and the target order behavior path sequence into a target order description portrait prediction network, and generate a target order description portrait corresponding to the target remote order data.

2. The big data mining method for remote order information according to claim 1 is characterized in that: The reconstructed training error value includes an order portrait reconstruction error value, and the obtaining of first neural network learning data, hidden training supervision data, and hidden neural network learning data corresponding to the sample remote order data includes: Acquire the sample order behavior path data sequence as first neural network learning data, and acquire the sample order description portrait as second neural network learning data; Perform feature hiding processing on the sample order description portrait to obtain order portrait hidden training supervision data and order portrait hidden learning data, obtain the order portrait hidden training supervision data as hidden training supervision data, and obtain the order portrait hidden learning data as hidden neural network learning data; The step of loading the hidden neural network learning data and the first neural network learning data into the basic order description portrait prediction network to generate hidden observation data, and generating a reconstruction training error value according to a deviation value between the hidden training supervision data and the hidden observation data, comprises: Loading the sample order behavior path data sequence and the order portrait hidden learning data into the basic order description portrait prediction network to generate order portrait hidden observation data; An order portrait reconstruction error value is generated based on the deviation value between the order portrait hidden training supervision data and the order portrait hidden observation data.

3. The big data mining method for remote order information according to claim 1 is characterized in that: The reconstructed training error value includes a behavior path reconstruction error value, and the obtaining of the first neural network learning data, hidden training supervision data, and hidden neural network learning data corresponding to the sample remote order data includes: Acquire the sample order description portrait as first neural network learning data, and acquire the sample order behavior path data sequence as second neural network learning data; Performing feature hiding processing on the sample order behavior path data sequence to obtain path hiding training supervision data and path hiding learning data, obtaining the path hiding training supervision data as hidden training supervision data, and obtaining the path hiding learning data as hidden neural network learning data; The step of loading the hidden neural network learning data and the first neural network learning data into the basic order description portrait prediction network to generate hidden observation data, and generating a reconstruction training error value according to a deviation value between the hidden training supervision data and the hidden observation data, comprises: Loading the sample order description portrait and the path hidden learning data into the basic order description portrait prediction network to generate path hidden observation data; A behavior path reconstruction error value is generated according to a deviation value between the path hiding training supervision data and the path hiding observation data.

4. The big data mining method for remote order information according to claim 1 is characterized in that: The step of loading the sample remote order data and the sample order behavior path data sequence into a basic order description portrait prediction network to generate a training order description portrait includes: Loading the sample remote order data, the sample order behavior path data sequence and the sample order description portrait into a basic order description portrait prediction network; Embedding the sample remote order data and the sample order behavior path data sequence respectively to generate a sample remote order embedding representation and a sample order behavior path embedding representation; Aggregating the sample remote order embedding representation and the sample order behavior path embedding representation to obtain a target sample embedding vector; Based on the sample order description portrait, the order description portrait is predicted for the target sample embedding vector to generate a training order description portrait.

5. The big data mining method for remote order information according to claim 4 is characterized in that: The embedding representation of the sample remote order data and the sample order behavior path data sequence respectively to generate a sample remote order embedding representation and a sample order behavior path embedding representation comprises: Extracting order knowledge vectors from the sample remote order data to generate basic order knowledge vectors, encoding the basic order knowledge vectors to generate embedded representations of the sample remote orders; Performing order behavior node mining on the sample order behavior path data in the sample order behavior path data sequence to generate an order behavior directed knowledge graph corresponding to at least one target order behavior node; Extract order knowledge vectors for the order behavior labels corresponding to each sample order behavior path data, the node labels corresponding to each target order behavior node, and the path positions corresponding to each target order behavior node, respectively, to generate order behavior label features corresponding to each sample order behavior path data, node label vectors corresponding to each target order behavior node, and path position vectors corresponding to each path position; Generate an initial order behavior path sub-vector corresponding to the target order behavior node based on the node label vector, path position vector, order behavior directed knowledge graph, and order behavior label features corresponding to the sample order behavior path data corresponding to the same target order behavior node, and generate a basic order behavior knowledge vector based on the initial order behavior path sub-vectors corresponding to each target order behavior node in each sample order behavior path data; The basic order behavior knowledge vector is encoded to generate an embedding representation of the sample order behavior path.

6. The big data mining method for remote order information according to claim 5 is characterized in that: The current basic knowledge vector is the basic order knowledge vector or the basic order behavior knowledge vector. The current basic knowledge vector is encoded to generate a corresponding current sample embedding representation, including: Performing weight distribution on the current basic knowledge vector to generate a current weight distribution vector, and converging the current basic knowledge vector and the current weight distribution vector to generate an initial converged vector; Performing polynomial expansion on the initial converged vector to generate a current polynomial expansion vector, and converging the current polynomial expansion vector and the initial converged vector to generate a target converged vector; The current sample embedding representation is obtained according to the target convergence vector.

7. The big data mining method for remote order information according to claim 4 is characterized in that: The aggregating the sample remote order embedding representation and the sample order behavior path embedding representation to obtain a target sample embedding vector includes: Performing weight allocation between feature dimensions on the sample remote order embedding representation and the sample order behavior path embedding representation to generate a converged weight allocation vector; Fusion of the sample remote order embedding representation and the converged weight allocation vector to generate a first fused order embedding representation, performing linear discriminant analysis on the first fused order embedding representation according to the first network configuration information to generate a first linear discriminant analysis vector, performing feature mapping on the first linear discriminant analysis vector to generate an order behavior path cleaning vector; According to the order behavior path cleaning vector, the convergence weight allocation vector is cleaned to generate a convergence enhancement vector; The converged enhancement vector and the sample remote order embedding representation are fused to generate the target sample embedding vector.

8. The big data mining method for remote order information according to claim 4 is characterized in that: The sample order description portrait includes a plurality of sample order portrait overviews distributed based on the logic between order portrait relationships, and the sample order portrait overview includes the user's shopping habits, product preferences, and consumption level information; The step of predicting the order description portrait of the target sample embedding vector based on the sample order description portrait to generate a training order description portrait includes: Determine a target logical node from each portrait relationship logical node corresponding to the sample order description portrait, wherein the portrait relationship logical node refers to each feature or attribute constituting the user order description portrait; From the sample order description portrait, obtain the sample order portrait overview before the target logical node as the candidate order portrait overview, embed the candidate order portrait overview, and generate the candidate order portrait embedding vector; Perform weight assignment on the candidate order portrait embedding vector to generate an initial weight assignment vector, obtain the initial order portrait embedding vector based on the initial weight assignment vector and the candidate order portrait embedding vector, perform blending weight assignment on the initial order portrait embedding vector and the target sample embedding vector to generate a blending weight assignment vector, obtain the pending order portrait embedding vector based on the blending weight assignment vector and the initial order portrait embedding vector, perform polynomial expansion on the pending order portrait embedding vector to generate a target order portrait embedding vector, and obtain the predicted order portrait embedding vector based on the pending order portrait embedding vector and the target order portrait embedding vector; Perform order description portrait prediction on the predicted order portrait embedding vector to generate a predicted order portrait overview corresponding to the target logical node; Obtain the next portrait relationship logic node as the target logic node, return to execute the step of obtaining the sample order portrait overview before the target logic node from the sample order description portrait as the candidate order portrait overview, until the traversal processing of all portrait relationship logic nodes is completed, and multiple predicted order portrait overviews are generated; The training order description portrait is obtained based on the overview of each predicted order portrait.

9. The big data mining method for remote order information according to claim 1, characterized in that: The sample learning data sequence corresponding to the basic order description portrait prediction network includes neural network learning data corresponding to a plurality of sample remote order data respectively, the neural network learning data includes sample remote order data and corresponding sample order behavior path data sequence, sample order description portrait, first neural network learning data, hidden training supervision data, hidden neural network learning data and order service type, and the sample learning data sequence includes at least one order service type; The step of loading the sample remote order data and the sample order behavior path data sequence into a basic order description portrait prediction network to generate a training order description portrait, and generating a target training error value according to a deviation value between the training order description portrait and the sample order description portrait includes: Loading the sample remote order data and the corresponding sample order behavior path data sequence and order service type in the sample learning data sequence into the basic order description portrait prediction network to generate a training order description portrait that matches the order service type corresponding to the sample remote order data; Generate a target sub-training error value based on the deviation value between the training order description image and the sample order description image corresponding to the same sample remote order data, and obtain a target training error value based on the target sub-training error value corresponding to each sample remote order data; The step of loading the hidden neural network learning data and the first neural network learning data into the basic order description portrait prediction network to generate hidden observation data, and generating a reconstruction training error value according to a deviation value between the hidden training supervision data and the hidden observation data, comprises: Loading the first neural network learning data, hidden neural network learning data and order service type corresponding to the sample remote order data in the sample learning data sequence into the basic order description portrait prediction network to generate hidden observation data matching the order service type corresponding to the sample remote order data; Generate a reconstruction sub-training error value based on the deviation value between the hidden training supervision data and the hidden observation data corresponding to the same sample remote order data, and obtain a reconstruction training error value based on the reconstruction sub-training error value corresponding to each sample remote order data; The basic order description portrait prediction network includes a first embedding representation unit, a second embedding representation unit and a fully connected output unit; The step of loading the sample remote order data and the sample order behavior path data sequence into a basic order description portrait prediction network to generate a training order description portrait, and loading the hidden neural network learning data and the first neural network learning data into a basic order description portrait prediction network to generate hidden observation data, further includes: Loading the sample remote order data into the first embedding representation unit, loading the sample order behavior path data sequence into the second embedding representation unit, generating a converged embedding representation vector according to the embedding representation results of the first embedding representation unit and the second embedding representation unit, loading the converged embedding representation vector and the sample order description portrait into a fully connected output unit, and generating the training order description portrait; The hidden neural network learning data and the first neural network learning data are loaded into the second embedding representation unit to generate the hidden observation data.

10. An artificial intelligence system, characterized in that: The artificial intelligence system includes a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions, and the machine-executable instructions are loaded and executed by the processor to implement the big data mining method for remote order information according to any one of claims 1 to 9.

Images