An e-commerce user screening method and device based on big data, and a medium

Abstract

The application discloses an e-commerce user screening method and device based on big data, and a medium, relates to the technical field of big data analysis, and comprises the following steps: extracting the space-time features of video frame sequence data through a three-dimensional convolutional neural network, encoding user behavior sequence through a gated recurrent, obtaining a behavior vector, fusing the behavior vector and the space-time features of the video frame sequence data by using a multi-head attention layer, obtaining an encrypted video feature vector, simultaneously extracting user portrait labels, industry classification labels and content vertical classification labels of a business order account, and obtaining account attribute labels; receiving the encrypted video feature vector and the account attribute labels, performing safe aggregation by using a Paillier homomorphic encryption algorithm, and generating a business order portrait vector; calculating a user causal effect value according to historical user behavior data; and introducing a causal inference double-tower model to perform multi-dimensional feature weighted sorting on a preliminary recommendation list, so that the accuracy of e-commerce user screening and the intelligent level of the system are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data analysis, and particularly relates to an e-commerce user screening method and device based on big data and a medium. BACKGROUND

[0002] With the rapid development of Internet technology and the wide popularity of mobile terminal devices, the content marketing mode in the e-commerce platform gradually evolves from the traditional graphic text promotion to the video and interaction direction. The deep integration of short video platforms and e-commerce platforms gives rise to a large number of content seeding and goods carrying scenarios based on commercial video, so that the demand of users for personalized recommendation is increasing. How to accurately identify potential high conversion intention users and realize efficient matching has become a research hotspot in the field of e-commerce intelligent recommendation.

[0003] Traditional methods usually process visual content features and user behavior data independently, and lack explicit modeling of high-order correlations between the two. At the same time, in the cross-modal feature fusion process, existing methods mostly use simple splicing or linear weighting methods, which are difficult to capture the complex nonlinear relationship between user intention and video content. Although homomorphic encryption technology has been introduced, its application in multi-modal data aggregation and causal inference matching has not been fully explored, resulting in difficulties in achieving accurate user screening while taking into account data security and expressiveness. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an e-commerce user screening method based on big data to solve the problem of difficult to balance cross-modal feature deep fusion and secure aggregation under the condition of accurate user matching.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides an e-commerce user screening method based on big data, which comprises the following steps: calling a business order video metadata through a video storage server API, and generating video frame sequence data by using an FFmpeg tool; capturing user interaction behavior with the business order video by using an SDK to form a user behavior sequence; extracting the spatiotemporal features of the video frame sequence data by using a three-dimensional convolutional neural network, and encoding the user behavior sequence by using a gated recurrent to obtain a behavior vector; fusing the behavior vector and the spatiotemporal features of the video frame sequence data by using a multi-head attention layer to obtain an encrypted video feature vector; extracting a user portrait label, an industry classification label and a content vertical classification label of a business order account to obtain an account attribute label; receiving the encrypted video feature vector and the account attribute label, and performing security aggregation by using a Paillier homomorphic encryption algorithm to generate a business order portrait vector; calculating a user causal effect value according to historical user behavior data; inputting the user causal effect value and the business order portrait vector into a causal inference double-tower matching model, nonlinearly mapping the user causal effect value into a user causal representation vector by using a three-layer fully connected network, globally normalizing the business order portrait vector by using a two-layer fully connected network to output a normalized business order representation vector, and calculating a user-business order matching degree score by using a cosine similarity function; comparing the user-business order matching degree score with a matching degree threshold to screen a target business order account to form a preliminary video ID list; extracting time features and content quality features from the business order video metadata and the user behavior sequence, and performing weighted fusion to form multidimensional features to arrange the preliminary video ID list to obtain a business order video recommendation sequence.

[0008] As a preferred scheme of the e-commerce user screening method based on big data, the user behavior sequence is formed by the following steps,

[0009] The API of the video storage server is used to request a metadata interface with the business order video ID as a parameter to obtain a JSON response of video duration, resolution and key frame timestamp, and the FFmpeg tool is called to perform intelligent frame extraction to output video frame sequence data.

[0010] The video player event bus is listened to by using the SDK to capture event types, operation timestamps and video progress values, and original behavior data streams are formed by using data cleaning and format standardization.

[0011] The original behavior data streams are time-sequentially matched with corresponding video frame sequence data, and the interaction behavior of the user with the business order video is recorded to output a user behavior sequence.

[0012] As a preferred scheme of the e-commerce user screening method based on big data, the encrypted video feature vector is obtained by the following steps,

[0013] The video frame sequence data is input into a three-dimensional convolutional neural network, and the video frame sequence data is convolved and pooled in the spatial and temporal dimensions, visual appearance features and time sequence motion features are extracted, and the spatial and temporal features are output through a full connection layer;

[0014] The user behavior sequence is mapped into a dense vector through an embedding layer, and is input into a gated loop, the long-term dependence between the user behavior sequences is learned by using a reset gate and an update gate mechanism, and a behavior vector is obtained;

[0015] The spatial and temporal features and the behavior vector are linearly transformed respectively to generate a query matrix, a key matrix and a value matrix, the attention weight of the spatial and temporal features to the user behavior sequence is calculated through an attention mechanism, a fusion feature vector is output, and the Paillier homomorphic encryption algorithm is used for encryption processing, and an encrypted video feature vector is output.

[0016] As a preferred scheme of the e-commerce user screening method based on big data, the user portrait label, the industry classification label and the content vertical classification label of the merchant single account are extracted to obtain the account attribute label, and the specific steps are as follows,

[0017] The user basic data of the merchant single account is called through a database interface, and the user portrait label is generated through aggregation calculation and entropy value analysis; meanwhile, the commodity category published by the merchant single account is mapped and matched to generate the industry classification label;

[0018] The video content is analyzed through an NLP model, the content vertical classification label is generated by fusing multi-modal recognition results, and the user portrait label, the industry classification label and the content vertical classification label are structured and packaged as the account attribute label.

[0019] As a preferred scheme of the e-commerce user screening method based on big data, the user causal effect value is calculated according to historical user behavior data, and the specific steps are as follows,

[0020] The encrypted video feature vector and the account attribute label are received, the encrypted video feature vector is weighted and summed in a ciphertext state to obtain an encrypted state aggregation vector; the encrypted state aggregation vector is decrypted, and is fused with the account attribute label to generate a merchant single portrait vector;

[0021] The historical user features, the processing variables and the result variables are collected, the historical user features are input, the processing variables are labels, XGBoost is trained to obtain an intervention prediction value, the historical user features are input, and the result variables are labels, DNN is trained to obtain a result prediction value;

[0022] The difference between the treatment variable and the intervention predicted value result variable and the difference between the result variable and the result predicted value are taken as the intervention residual and the result residual, and linear regression is performed using the least squares method to obtain the causal effect estimate.

[0023] As a preferred scheme of the e-commerce user screening method based on big data, the user-business order matching degree score is calculated by a cosine similarity function, and the specific steps are as follows,

[0024] Based on the YouTube DNN double-tower architecture, the input of the user tower is replaced from the traditional user behavior features to the causal effect value, the input of the business order tower is replaced from a single feature to the global business order portrait vector, and the causal effect estimation and unbiased calibration function are introduced in the output layer to form a causal inference double-tower matching model.

[0025] The user causal effect value and the business order portrait vector are input into the causal inference double-tower matching model, the user causal effect value is input into a three-layer fully connected network by the user tower, and after nonlinear transformation by the PReLU and Tanh activation functions, a user causal representation vector is output.

[0026] The global business order portrait vector is input into a two-layer fully connected network by the business order tower, and after processing by the Swish activation function, an L2 normalization is performed to output a standardized business order representation vector.

[0027] The dot product of the standardized business order representation vector and the user causal representation vector is calculated by a cosine similarity function, and a user-business order matching degree score is obtained by calibration through a Sigmoid function.

[0028] As a preferred scheme of the e-commerce user screening method based on big data, the preliminary video ID list is formed, and the specific steps are as follows,

[0029] The real-time commercial target and peak period traffic of e-commerce are collected, and the matching degree threshold is calculated by a weighted formula;

[0030] The user-business order matching degree score is compared with the matching degree threshold to obtain the qualified business order account ID, the qualified business order account ID is aggregated and filtered to exclude invalid accounts, and a target business order account set is output.

[0031] According to the target business order account set, the video IDs that meet the conditions are queried from the business order video metadata, and a de-duplication process is performed to output a preliminary video ID list.

[0032] As a preferred scheme of the e-commerce user screening method based on big data, the specific steps of obtaining the business order video recommendation sequence are as follows,

[0033] Temporal features are extracted from the metadata of commercial video, content quality features are calculated through user behavior sequences, and weighted fusion is performed to form multi-dimensional features;

[0034] Multi-dimensional features are input into a deep reinforcement learning ranking model, and the video ranking position action is output through a policy network. The model is then optimized according to a multi-objective reward function to obtain a commercial video recommendation sequence.

[0035] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the big data-based e-commerce user screening method described in the first aspect of the present invention.

[0036] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the big data-based e-commerce user screening method described in the first aspect of the present invention.

[0037] The beneficial effects of this invention are as follows: by introducing a causal inference dual-tower model, user-side modeling is driven by causal effect values, which enhances the interpretability and accuracy of matching. Furthermore, deep reinforcement learning is used to perform multi-dimensional feature weighted sorting on the initial recommendation list, taking into account both short-term click and long-term conversion goals. Under the premise of ensuring privacy and security, the entire chain of optimization from content understanding and user intent recognition to secure matching and intelligent sorting is achieved, which significantly improves the accuracy of e-commerce user screening and the level of system intelligence. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Fig. 1 This is a flowchart illustrating a big data-based method for screening e-commerce users.

[0040] Fig. 2 A flowchart for generating user causal effect values.

[0041] Fig. 3 A flowchart for forming a sequence of user behaviors.

[0042] Fig. 4 A flowchart for obtaining feature vectors of encrypted video. Detailed Implementation

[0043] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0044] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0045] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0046] Reference Figs. 1-4 This is one embodiment of the present invention, which provides a method for e-commerce user screening based on big data, including the following steps:

[0047] S1. Retrieve the video metadata of the commercial order through the video storage server API, generate video frame sequence data using FFmpeg, and capture the user's interaction behavior with the commercial order video through the SDK to form a user behavior sequence.

[0048] S1.1. Through the API of the video storage server, request the metadata interface with the commercial video ID as the parameter to obtain the JSON response of video duration, resolution and keyframe timestamps, call the FFmpeg tool to perform intelligent frame extraction, and output video frame sequence data.

[0049] Specifically, the video storage server's application programming interface requests the metadata interface with the merchant order video ID as a parameter to obtain JSON format response data containing video duration, resolution, and keyframe timestamps; the FFmpeg command-line tool is called to intelligently extract frames based on the frame timestamp information at a base frequency of 1 frame per second, and automatically extracts up to 5 frames per second for the detected product display area, generating video frame sequence data sorted by timestamp.

[0050] S1.2. By listening to the video player's event bus through the SDK, the event type, operation timestamp, and video progress value are captured, and the raw behavioral data stream is formed through data cleaning and format standardization.

[0051] Specifically, the software development kit integrated into the client application listens to the video player's event bus, capturing event types such as play, pause, seek, and skip. It also records the operation timestamp (accurate to milliseconds) and the corresponding video progress value (in seconds) for each event. The captured raw data is cleaned, removing abnormal records such as invalid event types, out-of-order timestamps, or video progress values ​​exceeding the actual video duration. The timestamps are uniformly converted to UTC time format, and the video progress values ​​are uniformly converted to percentage format, then arranged in ascending order by operation timestamp to form the raw behavioral data stream.

[0052] S1.3. Perform time-series matching between the original behavior data stream and the corresponding video frame sequence data, record the user's interaction behavior with the commercial video, and output the user behavior sequence.

[0053] Specifically, based on the timestamp of each frame in the video frame sequence data, a time tolerance window is set (500 milliseconds in the example). Each event in the original behavior data stream is traversed. For each event, within the positive and negative tolerance range of the operation timestamp, the corresponding video frame at the time point in the video frame sequence data is searched. When the difference between the operation timestamp of the event and the timestamp of a certain video frame is within the tolerance range, it is associated with the matched video frame. The event type, operation timestamp, video progress value, and the identifier of the matched video frame are recorded. After the traversal is completed, all successfully matched records are sorted in ascending order of operation timestamp to generate a user behavior sequence.

[0054] It should be noted that the lower limit of the tolerance range is based on the tolerance window being significantly larger than the video frame interval (33ms), for example, at least 100ms. The upper limit depends on the tolerance window not being too large, otherwise user actions at different points in time will be incorrectly associated with unrelated video frames, such as incorrectly matching a user's comment on the beginning of the video to the end of the video.

[0055] It should also be noted that the above content has been agreed to by the user and is used for legitimate purposes.

[0056] S2. Extract the spatiotemporal features of video frame sequence data through a three-dimensional convolutional neural network, and obtain the behavior vector by gated recurrent encoding of user behavior sequence. Use a multi-head attention layer to fuse the behavior vector and the spatiotemporal features of video frame sequence data to obtain encrypted video feature vector. At the same time, extract the user profile tags, industry classification tags and content vertical classification tags of the business account to obtain account attribute tags.

[0057] S2.1. Input the video frame sequence data into a three-dimensional convolutional neural network, perform convolution and pooling on the video frame sequence data in the spatial and temporal dimensions, extract visual appearance features and temporal motion features, and output spatiotemporal features through a fully connected layer.

[0058] Specifically, video frame sequence data is input into a 3D convolutional neural network. The 3D convolutional neural network uses 3D convolutional kernels to perform convolution operations simultaneously on the spatial dimension (height, width) and temporal dimension (sequential frame order) of the video frame sequence data, so as to capture both static visual elements in the image and dynamic change information between frames, and obtain a 3D feature map. The 3D feature map is then reduced in dimensionality by a 3D pooling layer (e.g., using max pooling, with an example value of 2x2x2) to reduce computation and enhance the spatial-temporal invariance of the features. After multiple alternating 3D convolution and 3D pooling operations, an abstract feature map containing visual appearance and temporal motion information is extracted. The abstract feature map is then input into a fully connected layer. The fully connected layer integrates the distributed feature representations and maps them to the feature space, outputting spatiotemporal features.

[0059] It should be noted that the training of the 3D convolutional neural network involves collecting video frame sequence data and corresponding video content labels; initializing the weight parameters of the 3D convolutional neural network using a He normal distribution; defining the cross-entropy loss function as the optimization objective; configuring the Adam optimizer with a learning rate of 0.001 (example value); inputting the video frame sequence data into the 3D convolutional neural network using a batch training method; performing forward propagation to obtain the output predicted labels; obtaining the cross-entropy loss value between the predicted labels and the true labels; obtaining the gradients of the parameters of each layer of the 3D convolutional neural network using the backpropagation algorithm; applying gradient clipping to limit the gradient norm to within 1.0 to prevent gradient explosion; updating the weight parameters of the 3D convolutional neural network using the Adam optimizer; and obtaining the trained 3D convolutional neural network.

[0060] S2.2. The user behavior sequence is mapped into a dense vector through the embedding layer and input into the gated loop. The long-term dependency relationship between the user behavior sequence is learned by the reset gate and update gate mechanism to obtain the behavior vector.

[0061] Specifically, discrete event types (such as play and pause) in the user behavior sequence are transformed into dense vector representations of fixed dimensions (64 dimensions in the example). The dense vectors are then sequentially input into a gating loop. At each time step, the gating loop obtains candidate hidden states by resetting the gate to control the degree of forgetting historical information; it obtains the current hidden state by updating the gate to control the fusion ratio between the current input information and the historical hidden states. The gating loop updates the hidden states step by step through the loop, learns the long-term dependencies and patterns between user behavior sequences across time steps, obtains a context vector representing the entire user behavior sequence, and defines the context vector as the behavior vector.

[0062] It should be noted that long-term dependencies between user behavior sequences refer to the fact that the type, intensity, or pattern of events occurring earlier in the sequence can have a significant impact on events that occur several time steps later, or that there is a predictable correlation between them. For example, in a user's viewing sequence of a commercial video, there is a positive long-term dependency between the early event "liking the video at the 5th second" and the later event "sharing the video at the 90th second." That is, the early positive interaction behavior (liking) is very likely to predict that a deeper conversion behavior (sharing) will occur later.

[0063] S2.3. Perform linear transformations on the spatiotemporal features and behavior vectors respectively to generate query matrix, key matrix and value matrix. Calculate the attention weight of spatiotemporal features on user behavior sequence through an attention mechanism, output fused feature vector, and encrypt it using the Paillier homomorphic encryption algorithm to output encrypted video feature vector.

[0064] Specifically, the spatiotemporal features and behavior vectors are linearly transformed using independent trainable weight matrices to generate corresponding query, key, and value matrices. The attention weight distribution is obtained by calculating the dot product of the query and key matrices and normalizing them using the Softmax function. The attention weight distribution represents the degree of attention paid by the spatiotemporal features to different elements in the user behavior sequence. The value matrix is ​​weighted and summed using the attention weights to generate a fused feature vector that integrates the spatiotemporal features and the contextual information of the behavior sequence. Each numerical element in the fused feature vector is encrypted using the public key of the Paillier homomorphic encryption algorithm, converting the plaintext values ​​into ciphertext and outputting the encrypted video feature vector.

[0065] S2.4. Retrieve basic user data of merchant order accounts through database interface, and generate user profile tags through aggregation calculation and entropy value analysis; at the same time, map and match the product categories published by merchant order accounts to generate industry classification tags.

[0066] Specifically, the system retrieves basic user data from merchant accounts via a database interface, including user demographic attributes (such as age, gender, and region) and behavioral data (such as activity level and purchasing power). It then aggregates this basic user data using structured query language, calculating the distribution ratio of each attribute dimension (e.g., calculating the percentage of users in different age groups). The system applies entropy analysis to calculate the information entropy value of each attribute dimension's distribution ratio, assessing the dispersion of the attribute distribution. A lower entropy value indicates a more concentrated distribution and more significant characteristics. Based on this, significant features are selected to form user profile tags. Finally, the system retrieves product category data published by merchant accounts and maps it to industry classification standards (e.g., mapping the product category "smartphones" to the "consumer electronics" industry), generating industry classification tags.

[0067] S2.5. Analyze video content using an NLP model, integrate multimodal recognition results to generate content vertical classification tags, and structure user profile tags, industry classification tags, and content vertical classification tags into account attribute tags.

[0068] Specifically, a natural language processing (NLP) model is used to analyze text information (such as video title, description, and subtitle text) extracted from the metadata of commercial video. The text content category is output through a word embedding layer and a classification layer. Visual features are extracted through a convolutional neural network and the scene recognition result (such as "indoor live broadcast", "outdoor sports" and "food preparation") is output through a classification layer. The text content category output by the NLP model and the scene recognition result are fused in a multimodal manner. A weighted average (e.g., 0.6 weight is assigned to the visual result and 0.4 weight is assigned to the text result) is used to generate vertical content classification tags. User profile tags, industry classification tags and vertical content classification tags are encapsulated into structured account attribute tags in a predetermined format (e.g., JSON format).

[0069] It should be noted that the natural language processing model is trained by extracting text information (video title, description, subtitle text) and corresponding manually labeled content category tags from video metadata to obtain a training dataset. Based on the training dataset, the text is converted into vectors through a word embedding layer using the cross-entropy loss function. After feature extraction, a 32-dimensional classification probability is output through a fully connected layer. The Adam optimizer is used to minimize the cross-entropy loss, and a mixed precision and distributed strategy is adopted to accelerate the process. After continuous iteration, the trained natural language processing model is obtained.

[0070] S3. Receive encrypted video feature vectors and account attribute tags, use Paillier homomorphic encryption algorithm for secure aggregation, generate business order profile vectors; calculate user causal effect values ​​based on historical user behavior data.

[0071] Receive encrypted video feature vectors and account attribute tags. In ciphertext state, perform weighted summation on the encrypted video feature vectors to obtain an encrypted state aggregate vector. Decrypt the encrypted state aggregate vector and merge it with the account attribute tags to generate a business portrait vector.

[0072] Specifically, the system receives encrypted video feature vectors and account attribute tags. In ciphertext mode, it assigns a corresponding weight value to each encrypted video feature vector. Based on the Paillier homomorphic encryption algorithm, it performs weighted processing on the product of ciphertext and plaintext for each encrypted video feature vector. Then, through ciphertext addition operations supported by the Paillier homomorphic encryption algorithm, it sums all the weighted encrypted video feature vectors to generate an encrypted state aggregate vector. The encrypted state aggregate vector is transmitted to a trusted computing node holding the Paillier private key. The Paillier private key is used for decryption to obtain the plaintext aggregate feature vector. The plaintext aggregate feature vector is then concatenated with the account attribute tags to generate a business order profile vector.

[0073] Historical user features, processing variables, and outcome variables are collected. Using historical user features as input and processing variables as labels, XGBoost is trained to obtain intervention prediction values. Using historical user features as input and outcome variables as labels, DNN is trained to obtain outcome prediction values.

[0074] Specifically, historical user characteristics, processing variables, and outcome variables are collected. Historical user characteristics include user profile attributes and behavioral statistical features. Processing variables represent whether a user accepts a specific intervention, and outcome variables represent business target indicators. Using historical user characteristics as input features and processing variables as prediction labels, XGBoost (Extreme Gradient Boosting Tree Model) is used for training. XGBoost learns the mapping relationship between features and processing variables through the gradient boosting tree algorithm and outputs the predicted probability value of the processing variables as the intervention prediction value.

[0075] Simultaneously, historical user characteristics are used as input features and outcome variables are used as prediction labels to train a deep neural network model. The deep neural network model learns the complex relationship between features and outcome variables through multi-layer nonlinear transformations (e.g., the interaction effect between users' historical click behavior and current video features, the nonlinear influence of users' demographic attributes on conversion rates, and the deep association between behavioral sequence patterns and purchase intentions), and outputs the predicted value of the outcome variable as the result prediction value.

[0076] It should be noted that the outcome variables include user click behavior metrics (e.g., click / non-click), user conversion behavior metrics, user interaction behavior metrics, and user viewing depth metrics.

[0077] A processing variable is a binary variable (0 or 1) used to indicate whether a user has been subjected to a specific intervention or exposure.

[0078] The difference between the treatment variable and the intervention prediction value, and the difference between the outcome variable and the outcome prediction value, are respectively taken as the intervention residual and the outcome residual. Linear regression is then performed using the least squares method to obtain the causal effect estimate.

[0079] Specifically, the difference between the treatment variable and the intervention prediction value is taken as the intervention residual, and the difference between the outcome variable and the outcome prediction value is taken as the outcome residual. The intervention residual is taken as the independent variable, and the outcome residual is taken as the dependent variable. The least squares method is used to perform linear regression analysis to obtain the causal effect estimate.

[0080] It should be noted that the expression for calculating the causal effect estimate is as follows:

[0081] ;

[0082] in, This represents the estimated causal effect. Indicates the first The residuals of the intervention variable for each sample, Indicates the first The residuals of the outcome variables for each sample Represents the total sample size. This represents the sample index variable.

[0083] S4. Input the user causal effect value and the order profile vector into the causal inference dual-tower matching model. The user causal effect value is nonlinearly mapped to the user causal representation vector through a three-layer fully connected network. The global order profile vector is processed by a two-layer fully connected network and normalization to output a standardized order representation vector. The user-order matching score is calculated by using the cosine similarity function.

[0084] S4.1. Based on the YouTube DNN dual-tower architecture, the input of the user tower is replaced from traditional user behavior features to causal effect values, and the input of the business order tower is replaced from a single feature to a global business order profile vector. Causal effect estimation and unbiased calibration functions are introduced in the output layer to form a causal inference dual-tower matching model.

[0085] Specifically, based on the YouTube DNN dual-tower architecture, the input to the user tower is replaced from traditional user behavior features to user causal effect values ​​calculated through dual machine learning; the input to the order tower is replaced from a single feature to a global order profile vector securely aggregated through federated learning. The federated learning secure aggregation includes each platform encoding local order features into vectors, encrypting the vectors using the Paillier homomorphic encryption algorithm, the central server weighting and aggregating the encrypted vectors according to platform weights, and decrypting to obtain the global order profile vector; in the output layer, an unbiased calibration function such as the Sigmoid function based on causal effect estimation is added after the cosine similarity calculation to form a causal inference dual-tower matching model.

[0086] It should be noted that each platform refers to an independent data holder participating in federated learning data collaboration, such as brand owner's own data platform, third-party merchant data platform, advertising service provider data platform, and logistics service provider data platform in e-commerce platforms; the platform weight is derived by comprehensively evaluating the data quality and data volume of each platform.

[0087] It should be noted that, to train the causal inference dual-tower matching model, historical user causal effect values, historical global order profile vectors, and real interaction labels are prepared. The parameters of the causal inference dual-tower matching model are initialized, including the weights, biases, and calibration function parameters of the fully connected layers of the user tower and the order tower. In each training iteration, the historical user causal effect values ​​are input into the user tower, and the user causal representation vector is calculated through a three-layer fully connected network (using PReLU and Tanh activation functions). The historical global order profile vector is input into the order tower, processed through a two-layer fully connected network (using the Swish activation function), and then L2 normalized to obtain the standardized order representation vector. The cosine similarity between the two representation vectors is calculated, and the similarity score is processed through an unbiased calibration function (Sigmoid function) to output the predicted matching probability value.

[0088] Obtain the binary cross-entropy loss between the predicted matching probability and the actual interaction label. Using the Adam optimizer, obtain the gradient of the loss function with respect to all parameters of the causal inference dual-tower matching model through the backpropagation algorithm. Update all parameters of the causal inference dual-tower matching model and repeat the iteration multiple times to obtain the trained causal inference dual-tower matching model.

[0089] S4.2. Input the user causal effect value and the business profile vector into the causal inference dual-tower matching model. The user tower inputs the user causal effect value into a three-layer fully connected network. After nonlinear transformation by PReLU and Tanh activation functions, it outputs the user causal representation vector.

[0090] Specifically, the user causal effect value is input into the user tower in the causal inference dual-tower matching model. The first layer of the user tower maps the user causal effect value to a 256-dimensional vector and performs a nonlinear transformation using the parameterized modified linear unit (PReLU) activation function. The second fully connected layer converts the 256-dimensional vector into a 192-dimensional vector and processes it again using the PReLU activation function to further learn complex feature interactions. The third fully connected layer compresses the 192-dimensional vector to 128 dimensions and processes it using the hyperbolic tangent (Tanh) activation function to smoothly compress the output value to the range of [-1, 1], outputting the user causal representation vector.

[0091] S4.3. The Merchant Order Tower inputs the global merchant order profile vector into a two-layer fully connected network, processes it through the Swish activation function, performs L2 normalization, and outputs a standardized merchant order representation vector.

[0092] Specifically, the first fully connected network performs a linear transformation on the global order profile vector using a weight matrix and a bias vector, outputting an intermediate feature representation. This intermediate feature representation undergoes a non-linear transformation using the Swish activation function before being input into the second fully connected network. The second fully connected network then performs a linear transformation on the output of the Swish activation function using a weight matrix and a bias vector, generating the original order representation vector. Finally, the original order representation vector is L2 normalized to output a standardized order representation vector.

[0093] It should be noted that the weight matrix and bias vector are trainable parameters of the quotient single tower, which are learned by minimizing the binary cross-entropy loss function through the backpropagation algorithm and the Adam optimizer during the overall training process of the causal inference dual-tower matching model.

[0094] S4.4. Calculate the dot product of the standardized business order representation vector and the user causal representation vector using the cosine similarity function, and calibrate using the sigmoid function to obtain the user-business order matching score.

[0095] Specifically, the dot product of the standardized business order representation vector and the user causal representation vector is calculated using the cosine similarity function. Since the standardized business order representation vector has been L2 normalized to a magnitude of 1, the dot product result is directly equal to the cosine similarity value. The cosine similarity value is then input into the Sigmoid function for calibration. After calibration, the user-business order matching score is output.

[0096] It should be noted that the expression for calculating the dot product of the standardized commercial order representation vector and the user causal representation vector is as follows:

[0097] ;

[0098] in, This represents the dot product of the standardized order representation vector and the user causal representation vector. Represents the user's causal representation vector. Represents the standardized commercial order representation vector. Indicates the first The component values ​​of the user causal representation vector in each dimension. Indicates the first The component values ​​of the standardized commercial order representation vector in each dimension. This indicates a dimension index.

[0099] It should be noted that 128 refers to the dimension of the user causal representation vector and the standardized commercial order representation vector.

[0100] S5. Compare the user-merchant order matching score with the matching threshold to filter target merchant order accounts and form a preliminary video ID list.

[0101] S5.1. Collect real-time e-commerce business objectives and peak-hour traffic, and calculate the matching degree threshold using a weighted formula.

[0102] Specifically, collect real-time GMV targets and benchmark click-through rates for e-commerce, and combine them with current traffic conditions, such as peak traffic exceeding 100,000 clicks per minute; calculate the matching degree threshold using a weighted formula.

[0103] It should be noted that the expression for calculating the matching degree threshold is:

[0104] ;

[0105] in, Indicates the matching threshold. Indicates the target weight for real-time GMV. This indicates the real-time GMV target (such as the total merchandise transaction volume of RMB 1 million to be completed on the day, the sales volume of RMB 150,000 to be achieved in the current hour, or the average daily GMV target of RMB 500,000 for the remaining period of the month). Indicates the baseline click-through rate weight. This represents the baseline click-through rate.

[0106] S5.2. Compare the user-merchant order matching score with the matching threshold to obtain the qualified merchant order account ID. Aggregate the qualified merchant order account IDs, filter out invalid accounts, and output the target merchant order account set.

[0107] Specifically, the user-merchant order matching score is compared with the matching threshold, and merchant order account IDs with user-merchant order matching scores greater than or equal to the matching threshold are selected as qualified merchant order account IDs; based on the ban flag and last activity time field in the account status database, merchant order account IDs with a status of ban or long-term inactivity are excluded. For example, accounts with an activity time earlier than 30 days are considered invalid; the target merchant order account set is output.

[0108] S5.3. Based on the target merchant account set, query the video IDs that meet the conditions from the merchant video metadata, perform deduplication, and output a preliminary video ID list.

[0109] Specifically, based on the account identifiers in the target merchant account set, all video records belonging to the target merchant account set are retrieved from the merchant video metadata, and the corresponding video IDs are extracted. For the video content corresponding to the obtained video IDs, keyframes are extracted from the merchant video (e.g., using FFmpeg to extract keyframes at fixed time intervals or scene changes). Each keyframe is grayscaled and uniformly scaled to a fixed size. The scaled image is then obtained, and a discrete cosine transform is performed on the scaled image, retaining low-frequency coefficients. The mean of the low-frequency coefficients is obtained, and the low-frequency coefficients are compared with the mean of the low-frequency coefficients. Values ​​greater than the mean of the low-frequency coefficients are recorded as 1, and values ​​less than the mean of the low-frequency coefficients are recorded as 0, generating a binary sequence. The binary sequences of all keyframes are weighted and averaged to generate a fixed-length digital fingerprint. The similarity of the video content is determined by the Hamming distance between the fixed-length digital fingerprints. Duplicate video IDs are removed based on the similarity of the video content to obtain a preliminary list of video IDs.

[0110] S6. Extract time features and content quality features from the video metadata and user behavior sequence, and perform weighted fusion to form multi-dimensional features. Arrange the preliminary video ID list to obtain the video recommendation sequence.

[0111] Temporal features are extracted from the metadata of commercial video, content quality features are calculated through user behavior sequences, and weighted fusion is performed to form multi-dimensional features.

[0112] Specifically, the video release timestamp and video duration data are extracted from the metadata of the commercial video. The release timestamp is converted into specific release time features, and the video duration is converted into duration classification features (for example, the hour portion of the release timestamp is extracted and converted into a 24-hour hour code, such as the release timestamp "2023-10-05 08:30:00" being converted into an hour feature value of 8; the day of the week information in the release timestamp is extracted and converted into a day of the week code, such as Thursday being converted into code 4). The user viewing completion rate, the number of user interaction behaviors, and the number of repeated views are statistically analyzed through user behavior sequences. Based on the statistical values, content quality features are obtained. The time features and content quality features are concatenated into vectors to form the original feature vector. The features in each dimension of the original feature vector are subjected to Min-Max normalization. The fusion weights are assigned according to the importance of the features, and a weighted sum is performed to generate a multi-dimensional feature vector.

[0113] It should be noted that the expression for calculating content quality characteristics is as follows:

[0114] ;

[0115] in, Indicates the weight of viewing completeness. Indicates the weighting of interactive behavior ratios. Indicates the weight of repeated viewing times. Indicates the viewing duration. Indicates the number of interactions. This indicates the number of times the same user has repeatedly played the same video. Indicates content quality characteristics.

[0116] It should be noted that each parameter in the expression is normalized using the corresponding benchmark values ​​(total video duration, maximum historical interaction count, and maximum historical repeated viewing count) before being used in the calculation to eliminate differences in dimensions.

[0117] Multi-dimensional features are input into a deep reinforcement learning ranking model, and the video ranking position action is output through a policy network. The model is then optimized according to a multi-objective reward function to obtain a commercial video recommendation sequence.

[0118] Specifically, multi-dimensional features are input into the deep reinforcement learning ranking model. The policy network in the deep reinforcement learning ranking model receives multi-dimensional features as state input, calculates the action probability distribution of each video at different ranking positions in the recommendation sequence through neural network forward propagation, samples and outputs specific video ranking position actions based on the probability distribution, and optimizes according to the multi-objective reward function to output a commercial video recommendation sequence.

[0119] It should be noted that the deep reinforcement learning ranking model is trained by collecting historical state-action-reward data from the experience replay buffer. The state is a multi-dimensional feature, the action is the historical ranking position, and the reward is calculated by a multi-objective reward function (the immediate value of each recommendation action is quantified by assigning weights to different business objectives and performing a weighted sum, based on core indicators such as click behavior and conversion behavior (e.g., click is assigned a weight of 0.5, conversion is assigned a weight of 0.3), while user experience indicators such as viewing completion rate are also included, and the reward is obtained through the multi-objective reward function). The policy network receives the current state, outputs the probability distribution of each possible ranking action through forward propagation of the neural network, and obtains the policy loss using a proximal policy optimization algorithm. The value network obtains the state value loss, minimizes the overall loss function through the Adam optimizer, and uses the backpropagation algorithm to synchronously update the parameters of the policy network and the value network. This is repeated for multiple rounds of iterative training. When the ranking action output by the policy network can stably obtain a higher cumulative reward, the trained deep reinforcement learning ranking model is obtained.

[0120] This embodiment also provides a computer device applicable to the e-commerce user screening method based on big data, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the e-commerce user screening method based on big data as proposed in the above embodiment.

[0121] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0122] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the big data-based e-commerce user screening method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0123] In summary, this invention enhances the interpretability and accuracy of matching by introducing a causal inference dual-tower model to drive user-side modeling with causal effect values. Furthermore, it utilizes deep reinforcement learning to perform multi-dimensional feature weighted sorting on the initial recommendation list, taking into account both short-term click and long-term conversion goals. Under the premise of ensuring privacy and security, it achieves end-to-end optimization from content understanding and user intent recognition to secure matching and intelligent sorting, significantly improving the accuracy of e-commerce user screening and the level of system intelligence.

[0124] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for screening e-commerce users based on big data, characterized in that: include, The video storage server API is used to retrieve the metadata of the commercial video, and FFmpeg is used to generate video frame sequence data. The SDK is used to capture the user's interaction with the commercial video to form a user behavior sequence. Spatiotemporal features of video frame sequence data are extracted by a 3D convolutional neural network, and user behavior sequences are encoded by gated recurrent encoding to obtain behavior vectors. The behavior vectors and spatiotemporal features of video frame sequence data are fused by a multi-head attention layer to obtain encrypted video feature vectors. At the same time, user profile tags, industry classification tags and content vertical classification tags of commercial accounts are extracted to obtain account attribute tags. The system receives encrypted video feature vectors and account attribute tags, uses the Paillier homomorphic encryption algorithm for secure aggregation, and generates a business order profile vector; it also calculates the user causal effect value based on historical user behavior data. The specific steps for calculating the user causal effect value based on historical user behavior data are as follows. Receive encrypted video feature vectors and account attribute tags. In the encrypted state, perform a weighted summation of the encrypted video feature vectors to obtain an encrypted state aggregate vector. Decrypt the encrypted state aggregate vector and merge it with the account attribute tags to generate a business portrait vector. Historical user features, processing variables, and outcome variables are collected. Using historical user features as input and processing variables as labels, XGBoost is trained to obtain intervention prediction values. Using historical user features as input and outcome variables as labels, DNN is trained to obtain outcome prediction values. The difference between the treatment variable and the intervention prediction value outcome variable and the difference between the outcome variable and the outcome prediction value are respectively used as the intervention residual and the outcome residual. Linear regression is then performed using the least squares method to obtain the causal effect estimate. The user causal effect value and the business order profile vector are input into the causal inference dual-tower matching model. The user causal effect value is nonlinearly mapped into the user causal representation vector through a three-layer fully connected network. The global business order profile vector is processed by a two-layer fully connected network and normalization, and a standardized business order representation vector is output. The user-business order matching score is calculated by the cosine similarity function. The user-merchant order matching score is compared with the matching threshold to filter target merchant order accounts and form a preliminary video ID list; Temporal and content quality features are extracted from the metadata of the commercial video and the user behavior sequence, and weighted and fused to form multi-dimensional features. The preliminary video ID list is then arranged to obtain the commercial video recommendation sequence.

2. The e-commerce user screening method based on big data as described in claim 1, characterized in that: The specific steps for forming a user behavior sequence are as follows: By using the API of the video storage server, a request is made to the metadata interface with the commercial video ID as a parameter to obtain a JSON response containing the video duration, resolution, and keyframe timestamps. The FFmpeg tool is then called to perform intelligent frame extraction and output video frame sequence data. By listening to the video player's event bus through the SDK, the event type, operation timestamp, and video progress value are captured, and the raw behavioral data stream is formed through data cleaning and format standardization. The original behavior data stream is matched temporally with the corresponding video frame sequence data, and the user's interaction behavior with the commercial video is recorded to output the user behavior sequence.

3. The e-commerce user screening method based on big data as described in claim 1, characterized in that: The specific steps for obtaining the feature vector of the encrypted video are as follows: Video frame sequence data is input into a three-dimensional convolutional neural network. Convolution and pooling are performed on the video frame sequence data in the spatial and temporal dimensions to extract visual appearance features and temporal motion features. Spatiotemporal features are output through a fully connected layer. User behavior sequences are mapped into dense vectors through an embedding layer and input into a gated loop. The long-term dependencies between user behavior sequences are learned using reset and update gate mechanisms to obtain behavior vectors. Linear transformations are performed on the spatiotemporal features and behavior vectors to generate query matrix, key matrix, and value matrix. The attention weight of spatiotemporal features to user behavior sequence is calculated through an attention mechanism, and a fused feature vector is output. The Paillier homomorphic encryption algorithm is used for encryption processing to output an encrypted video feature vector.

4. The e-commerce user screening method based on big data as described in claim 1, characterized in that: The process of extracting user profile tags, industry category tags, and content vertical category tags from merchant accounts to obtain account attribute tags involves the following steps: The user base data of the merchant order account is retrieved through the database interface, and user profile tags are generated through aggregation calculation and entropy analysis. At the same time, the product categories published by the merchant order account are mapped and matched to generate industry category tags. By analyzing video content using NLP models and integrating multimodal recognition results, vertical content classification tags are generated. User profile tags, industry classification tags, and vertical content classification tags are then structured and encapsulated into account attribute tags.

5. The e-commerce user screening method based on big data as described in claim 1, characterized in that: The specific steps for calculating the user-merchant matching score using the cosine similarity function are as follows. Based on the YouTube DNN dual-tower architecture, the input of the user tower is replaced from traditional user behavior features to causal effect values, and the input of the business order tower is replaced from a single feature to a global business order profile vector. Causal effect estimation and unbiased calibration functions are introduced in the output layer to form a causal inference dual-tower matching model. The user causal effect value and the business order profile vector are input into the causal inference dual-tower matching model. The user tower inputs the user causal effect value into a three-layer fully connected network. After nonlinear transformation by PReLU and Tanh activation functions, the user causal representation vector is output. The Merchant Order Tower inputs the global merchant order profile vector into a two-layer fully connected network, processes it through the Swish activation function, performs L2 normalization, and outputs a standardized merchant order representation vector. The dot product of the standardized business order representation vector and the user causal representation vector is calculated using the cosine similarity function, and then calibrated using the sigmoid function to obtain the user-business order matching score.

6. The e-commerce user screening method based on big data as described in claim 1, characterized in that: The specific steps for forming a preliminary video ID list are as follows. Collect real-time e-commerce business objectives and peak-hour traffic, and calculate the matching degree threshold using a weighted formula; The user-merchant order matching score is compared with the matching threshold to obtain the qualified merchant order account ID. The qualified merchant order account IDs are aggregated, and invalid accounts are filtered out to output the target merchant order account set. Based on the target merchant account set, query the video IDs that meet the criteria from the merchant video metadata, perform deduplication, and output a preliminary list of video IDs.

7. The e-commerce user screening method based on big data as described in claim 1, characterized in that: The specific steps for obtaining the commercial video recommendation sequence are as follows: Temporal features are extracted from the metadata of commercial video, content quality features are calculated through user behavior sequences, and weighted fusion is performed to form multi-dimensional features; Multi-dimensional features are input into a deep reinforcement learning ranking model, and the video ranking position action is output through a policy network. The model is then optimized according to a multi-objective reward function to obtain a commercial video recommendation sequence.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the big data-based e-commerce user screening method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the big data-based e-commerce user screening method according to any one of claims 1 to 7.

Images