A Big Data Analysis Method for Internet Commodity Transactions

By constructing a multimodal environmental stimulus tensor and causal graph structure, combined with graph convolution embedding and backdoor adjustment mechanism, the problem of existing technologies being unable to distinguish between user needs and environmental impulses is solved. This enables accurate prediction and dynamic intervention of high return risks before users place orders, thereby reducing the return rate and logistics and warehousing costs of e-commerce platforms.

CN123115771APending Publication Date: 2026-11-17HENAN YINGDONG NETWORK TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611536077.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-30
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing big data analysis methods for internet commodity transactions cannot perceive and integrate multimodal environmental information in live streaming scenarios in real time, nor can they distinguish between users' true needs and impulsive intentions in purchasing decisions. This makes it impossible to accurately predict the actual return probability of transactions, and consequently, it is impossible to conduct effective dynamic marketing interventions before placing an order, resulting in high return rates and high logistics and warehousing costs.

Method used

By collecting live video streams, audio streams, and user bullet screen text streams in real time, a multimodal environmental stimulus tensor is constructed. Confusion variable nodes are introduced into the user-product bipartite graph structure, and graph convolutional embedding representation learning is performed. Combined with a backdoor adjustment mechanism, the impulsive intention weights are calculated, and a dynamic return probability prediction network is constructed to trigger a millisecond-level dynamic intervention mechanism.

Benefits of technology

It enables real-time identification of orders with high impulsiveness and high return risk before users place orders, and dynamic intervention to reduce return rates and costs, thereby improving the operational efficiency of e-commerce platforms.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115771A_ABST
    Figure CN123115771A_ABST
Patent Text Reader

Abstract

This invention discloses a big data analysis method for internet commodity transactions, relating to the fields of artificial intelligence and big data analysis. The method involves real-time acquisition and fusion of live video streams, audio streams, and bullet screen text streams to obtain an environmental stimulus tensor. This tensor is used as a confounding variable to construct a causal graph structure using a user-product bipartite graph. Based on a backdoor adjustment mechanism, the method calculates the impulsive intent weight, achieving causal separation between the actual demand intent and the environmental impulsive intent in the purchase decision. Combining user embedding, product embedding, and impulsive intent weights, the method predicts the dynamic return probability. Within a millisecond-level time window before order placement, a dynamic intervention mechanism is triggered based on dual threshold judgments. This invention can identify and proactively intervene in high-impulsive, high-return-risk orders before the transaction occurs, reducing the return rate and reverse logistics costs of e-commerce platforms.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and big data analysis technology, specifically a method for big data analysis of internet commodity transactions. Background Technology

[0002] In recent years, new internet commodity transaction models represented by live-streaming e-commerce and short video sales have occupied a huge market share. Traditional e-commerce big data analysis methods mainly rely on structured log data such as users' historical clicks and browsing time. The core goal of their analysis and optimization is often to improve click-through rate and conversion rate. Such methods usually build feature engineering systems based on user behavior logs and use deep neural networks to perform binary classification prediction of whether users click or purchase. This is a well-known basic technology in the field of e-commerce big data analysis.

[0003] The closest existing technical solution to this invention is an e-commerce conversion rate prediction method based on structured clickstream data. The specific implementation of this method includes: collecting structured behavioral data such as users' historical click logs, browsing dwell time, add-to-cart records, and historical purchase records on e-commerce platforms; performing feature engineering processing on the above structured data, including the construction and discretization encoding of user features, product features, and contextual features; inputting the constructed feature vectors into a deep neural network for training; and outputting a predicted value of the user's click probability or purchase probability for a certain product. The platform sorts, recommends, or places advertisements on products based on the predicted probability values ​​to improve the overall click-through rate and conversion rate indicators. The data input of the above existing solutions is limited to structured user behavior logs, and the optimization goal of the model always focuses on positive conversion indicators such as whether a click or purchase occurs, without involving the causal decomposition of the motivation behind the purchase behavior, or the prediction of the reverse result of whether the order is returned.

[0004] A causal analysis of the existing technologies reveals the following shortcomings: First, because the data collection process in existing solutions only covers structured clickstream logs and fails to collect unstructured real-time data such as audio, visuals, and bullet comments in live streaming scenarios, the models inherently lack the ability to perceive the source of users' instantaneous impulses at the feature input level. This results in their inability to identify whether users' purchasing decisions are influenced by the host's inflammatory tone, visual stimulation, or the herd mentality of bullet comments. Second, because the objective function of existing solutions only uses the correlation statistics of clicks and purchases as the fitting object, a causal inference mechanism is not introduced. Intervention calculations are performed on confounding variables, so the output purchase probability is essentially a correlation result of the mixture of real demand intention and environmental impulse intention. It is impossible to separate the two from the causal level, which makes the model unable to answer the counterfactual question of whether the user would still buy if the environmental stimulus were removed. Third, since the existing solution lacks the ability to decompose the causal motivation of purchase, it is also impossible to build a targeted return probability prediction model on this basis. Furthermore, it is impossible to trigger dynamic marketing intervention measures in the millisecond-level link before the user places an order. It can only passively calculate costs after the return has actually occurred, and cannot reduce the return rate from the source.

[0005] In summary, in live-stream e-commerce scenarios, users' order decisions are often instantly stimulated by the host's highly persuasive tone, the atmosphere of a buying frenzy, and the herd mentality of bullet comments, exhibiting a high degree of impulsiveness. Traditional big data models can only process structured clickstream data and cannot handle unstructured real-time multimodal stream data. This results in a high return rate for e-commerce companies, despite high conversion rates, leading to significant warehousing, logistics, and product damage costs. Therefore, how to dynamically separate users' true needs from their impulsive environmental intentions at the moment of transaction, and predict the actual return probability of the transaction before it is placed, thereby achieving dynamic marketing intervention, is a pressing technical problem in the field of e-commerce big data analysis. To address this, we propose a big data analysis method for internet commodity transactions. Summary of the Invention

[0006] The purpose of this invention is to provide a method for big data analysis of internet commodity transactions.

[0007] The technical problem this invention aims to solve is that existing big data analysis methods for internet commodity transactions cannot perceive and integrate unstructured multimodal environmental information such as the tone of voice of the anchor, visual stimuli, and the emotions of bullet comments in real time during live streaming scenarios. They also cannot distinguish between the user's true needs and intentions in purchasing decisions and the impulsive intentions generated by instantaneous environmental stimuli from a causal perspective. As a result, the platform cannot accurately predict the actual return probability of a transaction in the millisecond-level link before the order is placed, and thus cannot effectively intervene in dynamic marketing for transactions with high impulsiveness and high return risk. Ultimately, this leads to the problem of high return rates and high reverse logistics and warehousing costs on e-commerce platforms.

[0008] To address the aforementioned technical problems, this invention provides a method for big data analysis of internet commodity transactions, comprising the following technical contents:

[0009] The system collects live video streams, live audio streams, and user bullet screen text streams in real time during online commodity transactions. It extracts and fuses features from the collected multimodal data to obtain an environmental stimulus tensor, which is used to characterize the intensity of the external stimulus environment in which the user is located at the time of placing an order.

[0010] A bipartite graph structure between users and products is constructed, and the environmental stimulus tensor is connected to the bipartite graph structure as a confounding variable node to obtain a causal graph structure containing the confounding variable node, wherein the confounding variable node is connected to the purchase decision node and the return result node respectively.

[0011] Graph convolutional embedding representation learning is performed on the user nodes and product nodes in the causal graph structure to obtain user embedding vectors and product embedding vectors.

[0012] Based on the historical interaction relationship between users and products in the bipartite graph structure, the true demand intention variable is calculated. Based on the backdoor adjustment mechanism, the multiple values ​​of the confounding variable node are summed by conditional probability weighting to calculate the causal effect probability of the true demand intention variable on the purchase decision under the condition of eliminating environmental stimuli interference. Combined with the actual probability of the purchase decision occurring under the actual observation condition, the impulsive intention weight is calculated. The impulsive intention weight is used to characterize the proportion of the purchase component driven by environmental stimuli in the transaction.

[0013] By inputting the impulsive intent weight, the user embedding vector, and the product embedding vector into the return risk prediction network, the dynamic return probability of the transaction is obtained.

[0014] Within a preset time window after the user clicks the order instruction but before the order is finally submitted, it is determined whether the weight of the impulsive intention is greater than a preset first threshold and whether the dynamic return probability is greater than a preset second threshold. If both judgment results are yes, a dynamic intervention mechanism is triggered to perform intervention actions on the transaction. If either judgment result is no, the order placement process of the transaction is allowed.

[0015] The present invention also provides an internet commodity transaction big data analysis device for implementing the above method. The device adopts a modular structure design, including a data acquisition and tensor construction module, a causal graph neural network module, a return risk prediction module, and an intervention decision module. The modules are connected in sequence and work together to form a complete technical closed loop from multimodal data acquisition to intervention decision execution.

[0016] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0017] 1. This invention constructs a real-time multimodal environmental stimulus tensor, enabling the model to perceive multi-dimensional impulses in live streaming scenarios, including acoustic, visual, and textual stimuli, thus overcoming the limitation of existing technologies that can only process structured clickstream data in terms of data input.

[0018] 2. This invention introduces environmental stimuli as confounding variable nodes into the user-product bipartite graph and uses a backdoor adjustment mechanism to perform causal intervention calculations, enabling the model to decompose the purchase decision into a true need intention component and an impulsive intention component from a causal perspective. This solves the problem that existing technologies cannot causally decompose purchase motivation and can only obtain confounding correlation results.

[0019] 3. This invention enables the platform to predict the actual return risk of a transaction before the user places an order by constructing a dynamic return probability prediction network based on impulsive intention weights, user embedding vectors, and product embedding vectors. This makes up for the shortcomings of existing technologies that can only passively calculate costs after the return actually occurs.

[0020] 4. This invention triggers a dynamic intervention mechanism based on the dual threshold judgment results of impulsive intent weight and return probability within a millisecond-level time window before order placement. This enables the platform to proactively identify and intervene in high-impulsiveness, high-return-risk orders before the transaction occurs, thereby reducing the return rate of e-commerce platforms and the corresponding reverse logistics and warehousing costs. By dividing each functional link into relatively independent yet interconnected functional modules, each module can independently adjust and replace data collection frequency, graph convolution layer number, or threshold parameters without affecting the normal operation of other modules. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall system processing flow in an embodiment of the present invention;

[0022] Figure 2 This is a schematic diagram of the multimodal feature and tensor construction process in an embodiment of the present invention;

[0023] Figure 3 This is a schematic diagram of the causal inference and intervention decision-making process in an embodiment of the present invention. Detailed Implementation

[0024] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0025] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0026] I. System Overall Architecture:

[0027] The internet commodity transaction big data analysis method provided in this embodiment has a system architecture that can be divided into five functional layers from bottom to top. The first layer is the data acquisition layer, which includes a live video stream acquisition module, a live audio stream acquisition module, a user bullet screen text acquisition module, and a user behavior log acquisition module. Each acquisition module synchronously acquires multiple data streams through a streaming computing engine according to a unified time window. In this embodiment, the duration of the unified time window is set to 500 milliseconds, that is, a batch of raw multimodal data sets to be processed is generated every 500 milliseconds. The second layer is the multimodal feature extraction layer, which includes an acoustic feature extraction submodule, a visual feature extraction submodule, and a natural language sentiment analysis submodule. The three submodules work in parallel. The system then feeds the output feature vectors of each component into the third layer, which is the environmental stimulus tensor construction layer. This layer concatenates the three types of feature vectors and generates a unified environmental stimulus tensor after nonlinear mapping. The fourth layer is the causal graph neural network layer. This layer uses the user-product bipartite graph as its skeleton and connects the environmental stimulus tensor as a confounding variable node into the graph structure. Through graph embedding and backdoor adjustment, it calculates the user's impulsive intention weight and the user embedding vector and product embedding vector. The fifth layer is the decision execution layer, which includes a dynamic return probability prediction module and a millisecond-level intervention decision module. The outputs of these two modules are respectively connected to the platform's existing risk control system and marketing system to execute specific intervention actions such as adjusting the shipping insurance eligibility and displaying a secondary confirmation pop-up.

[0028] II. Construction process of multimodal environmental stimulus tensor:

[0029] In this embodiment, the construction process of the environmental stimulus tensor is as follows: First, the audio stream, video frame sequence and bullet screen text stream of the live broadcast room are collected in real time by the streaming computing engine, and the three types of data are timestamped according to the aforementioned unified time window to obtain the original multimodal data set within the time window.

[0030] Secondly, acoustic features are extracted from the audio stream. A pre-trained acoustic feature extraction model is used to extract speech rate features, volume change features, and pitch change features. The above three types of acoustic sub-features are concatenated to obtain the acoustic feature vector corresponding to the time window. In this embodiment, the dimension of the acoustic feature vector is set to thirty-two dimensions.

[0031] Next, visual features are extracted from the video frame sequence. A lightweight convolutional neural network is used to extract the flicker frequency features and the proportion of close-up product images to obtain the visual feature vector corresponding to the time window. In this embodiment, the dimension of the visual feature vector is set to 32 dimensions.

[0032] Then, natural language processing is performed on the bullet screen text stream, and the sentiment analysis model is used to calculate the emotional polarity and emotional intensity features of the bullet screen group to obtain the text feature vector corresponding to the time window. In this embodiment, the dimension of the text feature vector is set to twenty-four dimensions.

[0033] Finally, the acoustic feature vector, visual feature vector, and text feature vector are concatenated in sequence to obtain a joint feature vector with a dimension of 88. This joint feature vector is then input into a biased nonlinear mapping network for fusion transformation, outputting a 64-dimensional environmental stimulus tensor. The nonlinear mapping network contains a learnable weight matrix and a learnable bias vector, and a nonlinear activation function is used to transform the concatenated joint feature vector, thereby enhancing the environmental stimulus tensor's ability to express multimodal stimulus information. This environmental stimulus tensor numerically reflects the intensity of the anchor's tone of voice, the strength of the visual stimulus, and the emotional fluctuation of the bullet screen group within the time window, and is used to quantitatively describe the intensity of the external stimulus environment in which the user is located at the moment of placing an order.

[0034] III. Dynamic impulse attribution process based on causal inference:

[0035] After obtaining the environmental stimulus tensor, this embodiment further constructs a causal graph structure containing confounding variables, and calculates the user's impulsive intention weights based on this structure. The specific process is as follows.

[0036] First, construct a user-product bipartite graph. The nodes in the graph include a set of user nodes and a set of product nodes. The edges in the graph represent the historical interaction relationships between users and products, specifically including browsing relationships, adding to cart relationships, and historical purchase relationships.

[0037] Secondly, the environmental stimulus tensor obtained above is used as a confounding variable node and connected to the above-mentioned bipartite graph structure. This confounding variable node is connected to the purchase decision node and the return result node respectively, forming a causal graph structure containing the confounding variable. In this causal graph structure, the true demand intention node points to the purchase decision node through a directed edge, which is used to represent the direct causal effect of the true demand intention on the purchase decision. The environmental stimulus confounding variable node points to the purchase decision node and the return result node through two directed edges respectively, which is used to represent that the environmental stimulus has a direct impact on both the purchase decision and the return result. The purchase decision node points to the return result node through a directed edge, which is used to represent that the purchase decision is a prerequisite for the return result. Since the environmental stimulus confounding variable node acts on both the purchase decision (processing variable) and the return result (outcome variable) at the same time, if no causal intervention is performed on it, direct correlation statistical analysis of the purchase decision and the return result will produce estimation bias. This is the key object that the backdoor adjustment mechanism of this invention needs to deal with.

[0038] Next, a graph neural network is used to learn multi-layer graph convolutional embedding representations for user nodes and product nodes, resulting in user embedding vectors and product embedding vectors respectively. In this embodiment, the dimensions of both user embedding vectors and product embedding vectors are set to 128 dimensions.

[0039] The true demand intention variable is obtained by statistically analyzing the historical interaction relationships between user nodes and product nodes in the bipartite graph structure.

[0040] Then, based on the backdoor adjustment mechanism, the conditional probability weighted sum of multiple values ​​of the confounding variable node is performed to calculate the causal effect probability of the true demand intention on the purchase decision under the condition of eliminating environmental stimuli interference. The specific calculation method is as follows:

[0041] ;

[0042] In the above formula, Represents the variables of the purchase decision. Variables representing actual needs and intentions. Operators that represent causal interventions applied to genuine demand intentions, used to indicate interventions that force genuine demand intentions to take a specific value. Indicates environmental stimuli as confounding variables. For a specific value of the variable, This indicates that given the actual need and the specific values ​​of the environmental stimuli... The conditional probability of a purchase decision occurring under certain conditions can be obtained through statistical estimation of historical transaction samples. Indicates the values ​​of environmental stimulus variables. The prior probability can be estimated from historical statistical data. The summation symbol indicates that the conditional probability and prior probability are calculated separately for each possible value of the environmental stimulus variable and then summed, thereby eliminating the bias caused by the environmental stimulus as a confounding variable in the estimation of causal effects.

[0043] In practical implementation, the environmental stimulus tensor can be discretized into several representative values ​​according to the numerical range. For example, it can be divided into three representative values: low intensity stimulus, medium intensity stimulus, and high intensity stimulus. Each representative value corresponds to a prior probability in historical statistical sense and a conditional probability of the purchase decision under that value. By calculating the product of the conditional probability and the prior probability for the above three representative values ​​and summing them, the probability of causal effect after removing the interference of environmental stimuli can be obtained.

[0044] Subsequently, based on the ratio between the calculated causal effect probability and the purchase probability under actual observation conditions, the impulsive intention weight of this transaction is calculated, as follows:

[0045] ;

[0046] In the above formula, This represents the weight of impulsive intent, with a value ranging from zero to one in a closed interval. The larger the value, the higher the proportion of impulsive components driven by environmental stimuli in the transaction. The meaning is the same as described above. This represents the actual probability of a purchase decision occurring under actual observation conditions that include the real impact of environmental stimuli. This actual probability can be directly obtained from the model output within the current time window.

[0047] For example, suppose the probability of a causal effect of a transaction under the condition of eliminating the interference of environmental stimuli is 30%, while the actual probability of the purchase decision occurring under the actual observation condition that includes the real influence of environmental stimuli is 90%. Then the calculated result of the impulsive intention weight of the transaction is about 67%, indicating that about two-thirds of the purchase component in the transaction comes from the instantaneous environmental stimuli rather than the user's own real needs.

[0048] IV. Dynamic Return Probability Prediction Process:

[0049] After obtaining the impulsive intent weight, user embedding vector, and product embedding vector, this embodiment concatenates the above three to obtain a joint feature vector, and inputs the joint feature vector into the return risk prediction network for processing. The return risk prediction network includes a fully connected layer and an output activation layer. The fully connected layer performs a linear transformation on the concatenated joint feature vector, and the output activation layer adopts an activation function that can map the output value to the interval between zero and one. Finally, it outputs the dynamic return probability of the transaction. The magnitude of the dynamic return probability represents the likelihood of the transaction being returned in the future.

[0050] V. Millisecond-level dynamic intervention decision-making process:

[0051] In this embodiment, the specific implementation process of millisecond-level dynamic intervention decision-making is as follows: Based on the statistical analysis results of offline historical transaction data, the impulsive intention weight threshold and the return probability threshold are preset. In this embodiment, the impulsive intention weight threshold is set to 65%, and the return probability threshold is set to 55%. The two serve as the basis for triggering the intervention mechanism. Within the millisecond time window when the user clicks the order button but the order has not yet been finally submitted, the impulsive intention weight and dynamic return probability calculated above are read in real time.

[0052] The system determines whether the impulsive intent weight of the transaction is greater than the set impulsive intent weight threshold and whether the dynamic return probability is greater than the set return probability threshold. If both results are yes, the transaction is determined to be a high-impulsiveness, high-return-risk transaction. Continuing with the previous example, if the impulsive intent weight of the transaction is 67% and the dynamic return probability is 62%, then since both exceed the corresponding preset thresholds, the transaction is determined to be a high-impulsiveness, high-return-risk transaction.

[0053] For orders deemed to be high-impulsive, high-return-risk transactions, a corresponding dynamic intervention mechanism is triggered. This includes at least one of the following intervention actions: canceling the automatic delivery of shipping insurance for the order, inserting an anti-impulsive secondary confirmation pop-up on the order confirmation page, or showing the user a description of the product's real-world usage scenarios. The purpose of these intervention actions is to slow down the user's ordering pace and allow them to reassess their actual purchasing needs with the help of real-world usage scenario information, thereby reducing irrational ordering behavior driven by instantaneous environmental stimuli.

[0054] For transactions not identified as high-impulsiveness and high-return-risk, the order placement process is allowed to proceed normally. Regardless of whether an intervention is triggered, the actual return result of the transaction is obtained after the order completes its lifecycle. This actual return result is then fed back into the offline training dataset for continuous iterative optimization of the fusion network of environmental stimulus tensors, the causal graph neural network, and the return risk prediction network. This forms a complete technical closed loop from data collection, causal attribution, risk prediction to intervention execution and feedback iteration.

[0055] VI. Timing Description of the Online Inference Link:

[0056] In this embodiment, within the complete time window from when the user clicks the order placement instruction to when the order is finally submitted, each processing step is executed sequentially in chronological order. Specifically, this includes reading the static embedding vectors of the user and the product, calculating the environmental stimulus tensor in real time, adjusting the weight of the calculated impulse intention based on graph aggregation and backdoor, performing dynamic return probability prediction, and making intervention decision judgments. All of the above processing steps can be completed within the low latency range allowed by the core e-commerce transaction link, enabling the present invention to achieve real-time risk identification and dynamic intervention before placing an order without affecting the user's normal order placement experience.

[0057] VII. Device Implementation Examples:

[0058] Corresponding to the aforementioned method embodiments, this embodiment also provides an Internet commodity transaction big data analysis device. The device is built with a modular structure and specifically includes a data acquisition and tensor construction module, a causal graph neural network module, a return risk prediction module, and an intervention decision module.

[0059] The data acquisition and tensor construction module is used to acquire live video streams, live audio streams, and user bullet screen text streams in real time. It extracts and fuses features from the acquired multimodal data to obtain environmental stimulus tensors. This module is further divided into acoustic feature extraction sub-modules, visual feature extraction sub-modules, text sentiment analysis sub-modules, and tensor fusion sub-modules. Each sub-module works in parallel and the output results are aggregated into the tensor fusion sub-module for unified processing.

[0060] The causal graph neural network module is used to construct a user-product bipartite graph structure. The environmental stimulus tensor is used as a confounding variable node to be connected to the bipartite graph structure. The user node and product node in the graph structure are learned by graph convolution embedding representation. The impulsive intention weight is calculated based on the backdoor adjustment mechanism. The module is further divided into a graph embedding submodule and a causal intervention calculation submodule. The causal intervention calculation submodule includes a conditional probability estimation submodule and a backdoor adjustment weighted summation submodule. The two work together to complete the calculation of the causal effect of the real demand intention.

[0061] The return risk prediction module is used to concatenate the impulsive intent weights, user embedding vectors, and product embedding vectors and input them into the return risk prediction network to output dynamic return probabilities.

[0062] The intervention decision module is used to determine the result within a preset millisecond time window based on the dual thresholds of impulsive intent weight and dynamic return probability, trigger the corresponding dynamic intervention mechanism, and connect the judgment result to the platform's existing risk control system and marketing system to perform specific intervention actions such as canceling shipping insurance eligibility and displaying a secondary confirmation pop-up.

[0063] VIII. Alternative Implementation Methods:

[0064] In another embodiment of the present invention, the user and product embedding representation learning stage of the causal graph neural network module can be performed without using graph neural networks for embedding learning. Instead, a matrix factorization method combined with a dual robust estimation method is used to estimate the causal effect. That is, the latent vector representations of users and products are obtained through matrix factorization, and then the dual robust estimation method is combined with the propensity score model and the outcome regression model to correct and estimate the causal effect of the real demand intention. In this way, the separation of real demand intention and impulsive intention is also achieved. The computational complexity of this alternative implementation is relatively low, and it is suitable for deployment on small and medium-sized e-commerce platforms with limited computing resources.

[0065] In another embodiment of the present invention, when there are some obfuscated paths in the environmental stimulus variables that are not directly observable, the causal inference step may not use a backdoor adjustment mechanism, but instead use the instrumental variable method to estimate the causal effect. That is, variables that are related to the purchase decision but not directly related to the return result are selected as instrumental variables, such as the anchor's identity and the historical average conversion rate of the live broadcast room. The causal effect of the real demand intention on the purchase decision is indirectly solved by the two-stage least squares estimation method, thereby realizing the causal decomposition of the purchase decision and the calculation of the impulsive intention weight.

[0066] The embodiments described above are merely preferred embodiments of the present invention, used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Other embodiments made by those skilled in the art based on the technical solutions of the present invention without departing from the technical concept of the present invention should all fall within the scope of protection of the present invention.

[0067] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for big data analysis of internet commodity transactions, characterized in that, include: Real-time acquisition of live video streams, live audio streams, and user bullet screen text streams in internet commodity transaction scenarios; feature extraction and fusion of the acquired multimodal data to obtain an environmental stimulus tensor, which is used to characterize the intensity of the external stimulus environment in which the user is located at the time of placing an order. Construct a bipartite graph structure between users and products, and connect the environmental stimulus tensor as a confounding variable node to the bipartite graph structure to obtain a causal graph structure containing the confounding variable node, wherein the confounding variable node is connected to the purchase decision node and the return result node respectively. Graph convolutional embedding representation learning is performed on the user nodes and product nodes in the causal graph structure to obtain user embedding vectors and product embedding vectors. The actual demand intention variable is obtained by statistically calculating the historical interaction relationship between users and products in the bipartite graph structure. Based on the backdoor adjustment mechanism, the conditional probability weighted summation of multiple values ​​of the confounding variable node is performed to calculate the causal effect probability of the real demand intention variable on the purchase decision under the condition of eliminating environmental stimuli interference. Combined with the actual probability of the purchase decision occurring under the actual observation condition, the impulsive intention weight is calculated. The impulsive intention weight is used to characterize the proportion of the purchase component driven by environmental stimuli in the transaction. The impulsive intent weight, the user embedding vector, and the product embedding vector are input into the return risk prediction network to obtain the dynamic return probability of the transaction. Within a preset time window after the user clicks the order instruction but before the order is finally submitted, it is determined whether the weight of the impulsive intention is greater than a preset first threshold and whether the dynamic return probability is greater than a preset second threshold. If both judgment results are yes, the dynamic intervention mechanism is triggered to perform intervention actions on the transaction. If any judgment result is negative, the order placement process for that transaction will be allowed.

2. The method for big data analysis of internet commodity transactions according to claim 1, characterized in that, The step of extracting and fusing features from the collected multimodal data to obtain an environmental stimulus tensor includes: extracting acoustic features from the live audio stream to obtain an acoustic feature vector containing speech rate features, volume change features, and pitch change features; Visual features are extracted from the live video stream to obtain a visual feature vector that includes features of screen flicker frequency and the proportion of close-up product images. Sentiment analysis is performed on the user's bullet screen text stream to obtain a text feature vector containing the emotional polarity and intensity features of the bullet screen group; The acoustic feature vector, the visual feature vector, and the text feature vector are concatenated in sequence to obtain a joint feature vector. The joint feature vector is then input into a biased nonlinear mapping network for fusion transformation to obtain the environmental stimulus tensor.

3. The method for big data analysis of internet commodity transactions according to claim 1, characterized in that, In the bipartite graph structure, the edges represent the historical interaction relationships between users and products, including browsing relationships, adding to cart relationships, and historical purchase relationships.

4. The method for big data analysis of internet commodity transactions according to claim 1, characterized in that, The calculation of the impulsive intention weight includes: calculating the conditional probability of the purchase decision occurring given the true demand intention and the value for each of the multiple values ​​of the confounding variable node, and then weighting and summing the values ​​together with the prior probabilities corresponding to the values ​​to obtain the causal effect probability. Obtain the actual probability of a purchase decision occurring under actual observation conditions that include the real impact of environmental stimuli; The impulsive intention weight is calculated based on the ratio of the causal effect probability to the actual probability, and the impulsive intention weight ranges from 0 to 1.

5. The method for big data analysis of internet commodity transactions according to claim 1, characterized in that, The intervention actions taken on this transaction include at least one of the following: canceling the automatic delivery of shipping insurance for this transaction, inserting an anti-impulse secondary confirmation pop-up on the order confirmation page, or showing the user a description of the product's actual usage scenarios.

6. The method for big data analysis of internet commodity transactions according to claim 1, characterized in that, Also includes: After the transaction is completed, the actual return result of the transaction is obtained and fed back to the offline training dataset for iterative updates of the fusion network corresponding to the environmental stimulus tensor, the causal graph neural network corresponding to the causal graph structure, and the return risk prediction network.

7. The method for big data analysis of internet commodity transactions according to claim 1, characterized in that, The first threshold and the second threshold are preset based on the statistical analysis results of historical transaction data.

8. A big data analysis device for internet commodity transactions, characterized in that, include: The data acquisition and tensor construction module is used to collect live video streams, live audio streams, and user bullet screen text streams in the Internet commodity transaction scenario in real time. It extracts and fuses the features of the collected multimodal data to obtain the environmental stimulus tensor, which is used to characterize the intensity of the external stimulus environment in which the user is at the time of placing an order. The causal graph neural network module is used to construct a bipartite graph structure between users and products. The environmental stimulus tensor is connected to the bipartite graph structure as a confounding variable node to obtain a causal graph structure containing the confounding variable node. Graph convolution embedding representation learning is performed on the user node and product node in the causal graph structure to obtain user embedding vector and product embedding vector. Impulsive intention weights are calculated based on a backdoor adjustment mechanism. The return risk prediction module is used to input the impulsive intention weight, the user embedding vector, and the product embedding vector into the return risk prediction network to obtain the dynamic return probability. The intervention decision module is used to determine, within a preset time window after the user clicks the order instruction but before the order is finally submitted, whether the weight of the impulsive intention is greater than a preset first threshold and whether the dynamic return probability is greater than a preset second threshold. If both judgment results are yes, the dynamic intervention mechanism is triggered to perform intervention actions on the transaction. If any judgment result is negative, the order placement process for that transaction will be allowed.

9. The Internet commodity transaction big data analysis device according to claim 8, characterized in that, The causal graph neural network module includes a graph embedding submodule and a causal intervention calculation submodule. The graph embedding submodule is used to perform multi-layer graph convolution operations on the user node and the product node to obtain the user embedding vector and the product embedding vector. The causal intervention calculation submodule includes a conditional probability estimation submodule and a backdoor adjustment weighted summation submodule, which are used to calculate the impulsive intention weight.

10. The Internet commodity transaction big data analysis device according to claim 8, characterized in that, The intervention decision module is also used to connect the judgment result that triggers the dynamic intervention mechanism to the platform's risk control system and marketing system to execute the intervention action.