User online intention analysis method based on artificial intelligence and big data e-commerce platform

Abstract

The invention discloses a user online intention analysis method based on artificial intelligence and a big data e-commerce platform, and the method comprises the steps: collecting user behavior data from different e-commerce platforms, recording related context information, and converting the user behavior data and the context information into feature vectors; building an intention recognition model based on a transformer model Transform model, training the Transform model by using the collected data, optimizing the model by using a cross entropy loss function in the training process, and evaluating the performance of the model on a test set; and deploying the trained Transform model into a recommendation system, collecting behavior data of the user on different platforms in real time, and inputting the behavior data into the Transform model to obtain the current online intention of the user. According to the method, the online shopping intention of the user can be effectively identified, the accuracy of a recommendation system is improved, more personalized commodity recommendation is provided for the user, and thus the user experience is improved.

Description

Technical Field

[0001] This invention belongs to the field of big data e-commerce technology, and in particular relates to an artificial intelligence method for analyzing user online intent and a big data e-commerce platform. Background Technology

[0002] With the rapid development of the internet, the e-commerce industry has grown rapidly, and more and more users are shopping online through diverse platforms. However, this diversified shopping environment presents e-commerce platforms with a challenge: how to accurately identify users' shopping intentions across multiple platforms in order to provide personalized product recommendations. Traditional user intent analysis methods mainly rely on rule engines or collaborative filtering-based recommendation algorithms. These methods may fail to accurately capture users' true shopping intentions in certain scenarios, leading to a mismatch between recommendation results and user needs. Furthermore, traditional methods may have limitations when processing multi-platform data, making it difficult to accurately capture users' behavioral characteristics across different platforms.

[0003] First, effectively collecting and processing user behavior data from different platforms, and transforming this raw data into feature vectors suitable for deep learning models to perform intent recognition, presents a significant challenge. Furthermore, while ensuring model performance, it's also necessary to consider how to integrate the trained model with a real-world recommendation system to achieve real-time user intent recognition and recommendation.

[0004] Existing user intent analysis methods have limitations in multi-platform environments, making it difficult to accurately identify users' shopping intentions. Therefore, an artificial intelligence-based approach is needed to improve the accuracy and practicality of online intent analysis across multiple platforms. Summary of the Invention

[0005] To address the shortcomings of the existing technology, this invention provides a user online intent analysis method based on artificial intelligence, the method comprising:

[0006] User behavior data is collected from different e-commerce platforms, and relevant contextual information is recorded. The user behavior data and the contextual information are then converted into feature vectors.

[0007] Construct an intent recognition model based on the Transformer model, train the Transformer model using collected data, optimize the model using the cross-entropy loss function during training, and evaluate the model's performance on the test set.

[0008] The trained Transformer model is deployed to the recommendation system. By collecting user behavior data on different platforms in real time and inputting it into the Transformer model, the user's current online intent can be obtained.

[0009] One-hot encoding is used to represent user, platform, product, and user behavior. Continuous features are used to represent time interval, price, and age. The continuous features need to be normalized. All the above features are combined into a feature vector, which is used as the input of the Transformer model.

[0010] In this process, after updating the one-hot encoding representation using a graph neural network (GNN), the updated node representation is combined with the normalized continuous features, and the combined feature vector is input into the Transformer model.

[0011] The method of updating the one-hot encoding representation using a graph neural network (GNN) includes:

[0012] Based on user behavior data on different e-commerce platforms, an interaction network is constructed as a user behavior graph. In the user behavior graph, nodes represent users, products and platforms, and the edges between users and products represent user behaviors. Weights are assigned to the edges between users and products, with different user behaviors corresponding to different weights.

[0013] The graph neural network (GNN) is used to encode the user behavior graph. The GNN updates the representation of each node by aggregating information from neighboring nodes.

[0014] Among them, the edge weight between users and platforms is set to the edge weight determined by the number of times they go online within a preset time period, and the edge weight between products and platforms is set to the edge weight determined by the sales share of products on different platforms.

[0015] Alternatively, the edge weights between users and the platform, and between products and the platform, can be set to fixed weights.

[0016] Here, a vector representation is initialized for each node in the user behavior graph, and a weight matrix is ​​defined for each layer of the graph neural network; the node representation is updated using the adjacency matrix and the weight matrix for each layer.

[0017] The node representation is updated iteratively, and in each iteration, the node representation will capture information about neighboring nodes that are further away.

[0018] After completing the iterative update, the final node representation matrix will be obtained, which contains representations of user, product, and platform nodes.

[0019] The Transformer encoder layer is constructed, which includes a multi-head self-attention sub-layer and a feedforward neural network sub-layer. Each sub-layer also includes residual connections and layer normalization. The number of heads is set to 8, and the hidden layer dimension of the feedforward neural network is 256.

[0020] In the multi-head self-attention sublayer of the encoder, the input embedding is divided into 8 heads, each with its own attention weights;

[0021] In the feedforward neural network sub-layer of the encoder, a feedforward neural network with a hidden layer dimension of 256 is used to perform a non-linear transformation on the data processed by the multi-head self-attention mechanism.

[0022] The encoder output is passed to a fully connected layer to extract information about the user intent. The output dimension of the fully connected layer should be m, where m is the number of categories of the user intent.

[0023] In this process, the Softmax activation function is applied after the fully connected layer to convert the output into a probability distribution.

[0024] For each user behavior sequence in the input data, the output of the Transformer model is a probability distribution representing the predicted probability of each intent category. The intent label with the highest predicted probability from the Transformer model is determined as the intent of the current user behavior sequence.

[0025] The user online intent analysis method may further include providing product recommendations that match the user's intent.

[0026] This invention also discloses an artificial intelligence-based big data e-commerce platform for analyzing user online intent, the big data e-commerce platform comprising:

[0027] The data collection module is used to collect user behavior data and contextual information on different platforms in real time;

[0028] The feature processing module is used to transform user behavior data and contextual information into feature vectors;

[0029] The intent recognition model module is responsible for building and training an intent recognition model based on the Transformer model. The Transformer model is used to capture the user's behavioral sequences on different platforms to identify the user's shopping intent.

[0030] The model training and optimization module is used to train the Transformer model using the collected data and to optimize the Transformer model using the cross-entropy loss function during the training process.

[0031] The performance evaluation module is used to evaluate the performance of the Transformer model on the test set;

[0032] The recommendation system integration module is used to deploy the trained Transformer model into the recommendation system. By collecting user behavior data on different platforms in real time and inputting it into the Transformer model, the module obtains the user's online intent and makes recommendations according to the recommendation strategy corresponding to the online intent.

[0033] The feature processing module is further configured to update the representation of one-hot encoding using a graph neural network (GNN), combine the updated node representation with the normalized continuous features, and input the combined feature vector into the Transformer model.

[0034] The user online intent analysis method and big data e-commerce platform based on the Transformer model of this invention can significantly improve the accuracy of intent recognition. The Transformer model can capture long-distance dependencies in the input sequence, thereby more accurately identifying the user's online intent. By collecting user behavior data on different platforms in real time, the current online intent of the user can be obtained in real time, enabling the recommendation system to respond to user needs promptly. In addition, it can also help to realize personalized recommendations. After accurately identifying the user's online intent, the recommendation system can provide more personalized product recommendations based on the user's online intent, thereby improving user experience and purchase conversion rate. Attached Figure Description

[0035] The above and other objects, features, and advantages of exemplary embodiments of the present disclosure will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the present disclosure are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein:

[0036] Figure 1 This is a flowchart illustrating a user online intent analysis method based on artificial intelligence according to an embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram illustrating a big data e-commerce platform for analyzing online user intent based on artificial intelligence, according to an embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0039] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0040] It should be understood that although the terms first, second, third, etc., may be used to describe... in the embodiments of the present invention, these... should not be limited to these terms. These terms are only used to distinguish... For example, first... may also be referred to as second... without departing from the scope of the embodiments of the present invention, and similarly, second... may also be referred to as first...

[0041] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0042] Depending on the context, the words “if” or “suppose” as used here can be interpreted as “when” or “in response to determination” or “in response to detection.” Similarly, depending on the context, the phrases “if determination” or “if detection (of the stated condition or event)” can be interpreted as “when determination” or “in response to determination” or “when detection (of the stated condition or event)” or “in response to detection (of the stated condition or event).”

[0043] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0044] An AI-based method for analyzing online user intent and a big data e-commerce platform

[0045] like Figure 1 As shown, this invention discloses a method for analyzing user online intent based on the Transformer model, the method comprising the following steps:

[0046] Step 1: Collect user behavior data from different e-commerce platforms, including browsing and purchasing, and determine relevant contextual information, including time intervals.

[0047] Step 2: Transform user behavior data and contextual information into feature vectors. Use one-hot encoding and continuous features to represent the corresponding features, and combine these features into a single feature vector.

[0048] Step 3: Construct an intent recognition model based on the Transformer model. The Transformer model consists of multiple layers of self-attention layers and feedforward neural network layers, capable of capturing long-range dependencies in the input sequence. In this task, the model needs to capture user behavior sequences on different platforms to identify the user's online intent.

[0049] Step 4: Train the Transformer model using the collected data. During training, the cross-entropy loss function can be used to optimize the model.

[0050] Step 5: Evaluate the performance of the Transformer model on the test set, using metrics such as accuracy, recall, and F1 score to measure the model's accuracy in recognizing users' online intent.

[0051] Step 6: Deploy the trained model to the recommendation system. By collecting user behavior data in real time on different platforms and inputting it into the model, the user's current online intent can be obtained.

[0052] By following the steps above, an accurate understanding of user online intent can be obtained using a Transformer-based model.

[0053] Among them, the Transformer model is a deep learning model based on the self-attention mechanism, mainly used for sequence-to-sequence tasks.

[0054] One-hot encoding is a method that converts categorical variables into numerical or numerical combinations, with one encoding per category. It is a non-continuous encoding method.

[0055] The process involves converting user behavior data and contextual information into feature vectors. These vectors are then optimized using a GNN model using one-hot encoded features and combined with continuous features to form a new feature vector. This vector is then formatted to fit the Transformer model before being input into the Transformer model.

[0056] Among them, the GNN model (Graph Neural Network) is a type of deep learning model used to process graph-structured data, which can capture the relationships between nodes and their neighboring nodes.

[0057] The user online intent analysis method may further include providing product recommendations that match the user's online intent.

[0058] The user online intent analysis method may further include collecting user feedback on the recommendation results and adjusting and optimizing the model in real time. Through continuous learning and adjustment, the accuracy and satisfaction of recommendations on different platforms can be improved.

[0059] In one embodiment, one-hot encoding is used to represent users, different platforms, products, and user behavior; continuous features are used to represent time intervals, prices, and ages; and normalization processing is required for these continuous features. All of these features are combined into a feature vector, which serves as the input to the model. This feature vector includes, but is not limited to:

[0060] User behavior characteristics: browsing 1 or purchasing 0 (one-hot encoding is 0 or 1); Product characteristics: using one-hot encoding, including major category characteristics such as women's clothing = 2, men's clothing = 3, electronic products = 4, etc., encoded as integers, and minor category characteristics such as dresses = 8, coats = 9, shirts = 10, etc., in the women's clothing category, encoded as integers; Product price (normalized); Platform characteristics: Platform positioning (quality emphasis = 1, cost-effectiveness emphasis = 2, integer encoding); Contextual characteristics: Time interval (normalized); User profile characteristics: Age (normalized).

[0061] To normalize commodity prices, the Min-Max normalization method can be used. Assume the original price data is P, the minimum price is Pmin, and the maximum price is Pmax. The formula for the normalized price Pnorm is:

[0062] Pnorm = (P - Pmin) / (Pmax - Pmin)

[0063] The normalized commodity price will be limited to the range of [0,1].

[0064] The same Min-Max normalization method can be used for age data. Assume the original age data is A, the minimum age is Amin, and the maximum age is Amax. The formula for the normalized age Anorm is:

[0065] Anorm = (A - Amin) / (Amax - Amin)

[0066] The normalized age range will be limited to [0,1].

[0067] For time interval data, the Min-Max normalization method can also be used. Assume the original time interval data is T, the minimum time interval is Tmin, and the maximum time interval is Tmax. The formula for the normalized time interval Tnorm is:

[0068] Tnorm = (T - Tmin) / (Tmax - Tmin)

[0069] The normalized time interval will be limited to the range of [0,1].

[0070] In one embodiment, suppose that for a certain behavior of a user, its feature vector may be as follows:

[0071]

[0072] In this example, categorical features are represented using integer encoding. Continuous features are normalized.

[0073] When using integer encoding (users, user behavior, products, platforms), it's crucial to ensure the model can correctly process this encoding. Otherwise, the model might misinterpret these numbers as having a magnitude relationship. This invention employs a GNN model to address this issue. Before inputting the integer-encoded categorical features into the model, they can be converted into low-dimensional vectors. This allows the model to capture the correlations between categorical features, rather than treating them as numerical values ​​with a magnitude relationship.

[0074] Positional encoding based on time intervals is added as an additional feature to the input feature vector. The time intervals need to be normalized.

[0075] Assuming the shortest time interval is 1 minute and the longest time interval is 1 day (1440 minutes), if the original time interval is 720 minutes (12 hours), then the normalized time interval is:

[0076] Normalized time interval = (720-1) / (1440-1)≈0.5.

[0077] Next, the normalized time interval can be added to the input feature vector. Suppose we have the following example data: Browsing: 1 (indicating a browsing action occurred); Category: 2 (e.g., women's clothing); Subcategory: 8 (e.g., dresses); Price: 0.4 (product price); Platform Positioning: 1 (indicating a focus on quality); Time Interval: 0.5 (normalized time interval); Age: 0.3 (user age).

[0078] These features can then be combined into a feature vector [ID, 1, 2-8, 0.4, 1, 0.5, 0.3]. This feature vector contains various information about user behavior, such as product category, price, platform, etc., as well as the temporal sequence information of user behavior. These feature vectors can then be combined to form sequence data.

[0079] In one embodiment, user behavior on different e-commerce platforms forms an interaction network. One-hot encoding is a discrete encoding method. When given discrete variable inputs, the Transformer model may experience reduced accuracy in intent recognition because it cannot accurately interpret the numerical relationships between different encodings. Optionally, this invention combines a Graph Neural Network (GNN) with the Transformer model to better capture the behavioral dependencies of users on different platforms.

[0080] The process of combining Graph Neural Networks (GNNs) with Transformer models can be divided into the following steps:

[0081] Step S1: Construct a user behavior graph. Based on user behavior data across different e-commerce platforms, build an interaction network. In this network, nodes can represent users, products, and platforms. The edges between users and products represent user behaviors (such as browsing, purchasing, etc.), and weights are assigned to these edges (different behaviors correspond to different weights). For edges between users and platforms, and between products and platforms, these edges can represent user activity on different platforms and product sales on different platforms. For example, the weight of the edge between users and platforms can be based on activity ratings (e.g., dividing the time period into different intervals based on the number of times online within a week, with different intervals corresponding to different edge weights). Similarly, the weight of the edge between products and platforms can be based on the sales percentage of products on different platforms (divided into different intervals, with different intervals corresponding to different edge weights). Alternatively, when the data volume is too large to update and calculate easily, fixed values ​​can be set based on experience, such as setting all edges between users and platforms, and between products and platforms, to 1.

[0082] Step S2: The graph neural network (GNN) is applied to encode the user behavior graph. The GNN updates the representation of each node by aggregating information from neighboring nodes.

[0083] Step S3: Combine the node representations of the GNN with other features: Combine the node representations (user, product, platform) obtained from the GNN with other features (such as price, time interval, age) to form a new feature vector.

[0084] Step S4: Input the combined feature vector into the Transformer model to perform online user intent recognition.

[0085] By following the steps above, the feature vector is updated using GNN and then input into the Transformer model, enabling better capture of user behavior dependencies on different platforms, thereby improving the accuracy of online intent recognition and recommendation.

[0086] In one embodiment, a graph neural network (GNN) is applied to encode the user behavior graph. The GNN updates the representation of each node by aggregating information from neighboring nodes. This process includes the following steps:

[0087] Step r1: Initialize a vector representation for each node in the graph. Assuming each node is assigned a d-dimensional vector representation, the initialized node representation matrix H^(0) will be a 6xd matrix (1 user node + 3 product nodes + 2 platform nodes).

[0088] Step r2: For each layer of the graph neural network, define a weight matrix W^(l), which maps the node representations from layer l to layer l+1. The dimension of the weight matrix W^(l) is (d_in x d_out), where d_in is the dimension of the node representations in the input layer and d_out is the dimension of the node representations in the output layer.

[0089] Step r3, Aggregate neighbor node information: For each layer, the node representation will be updated using the adjacency matrix A and the weight matrix W^(l). Specifically, the following operations will be performed:

[0090] H^(l+1)=σ(A*H^(l)*W^(l))

[0091] Here, A represents the adjacency matrix of the graph, H^(l) represents the node representation matrix of the l-th layer, W^(l) represents the weight matrix of the l-th layer, and σ is the activation function (ReLU). By iteratively updating the node representations, information in the network structure can be captured. A*H^(l) aggregates the neighbor node information and then multiplies it with the weight matrix W^(l) to achieve the mapping of node representations.

[0092] Step r4: Repeat step L times, where L is the number of layers in the graph neural network. In each iteration, the node representation will capture information from neighboring nodes at greater distances.

[0093] After completing the iterative updates, the final node representation matrix H^(L) is obtained. This matrix contains representations of user, product, and platform nodes, capturing information from the network structure. These representations can be used as features and fed into other models (such as Transformer) for online user intent analysis.

[0094] In one embodiment, a vector representation is initialized for each node in the graph. This can be random initialization or using a pre-trained embedding. Assigning a d-dimensional vector representation to each node embeds the node's information into a continuous vector space, thus better capturing the similarities and relationships between nodes; a larger d-value can capture more information. Optionally, the node representation is initialized randomly, generating a d-dimensional vector representation randomly for each node. Alternatively, if pre-trained embeddings are available (e.g., node attribute-based embeddings), they can be directly used as the initial representation of the nodes. This can accelerate the training process and potentially improve model performance.

[0095] In one embodiment, the user behavior graph has 1 user node (U1), 3 product nodes (P1, P2, P3), and 2 platform nodes (T1, T2), for a total of 6 nodes. It is assumed that each node is assigned a 2-dimensional vector representation (i.e., d = 2).

[0096] First, a 2D vector representation can be initialized for each node. Let's take random initialization as an example:

[0097] H^(0) = [[0.1, 0.3], # Initial representation of user U1

[0098] [0.2,0.4],# Initial representation of product P1

[0099] [0.3,0.1], # Initial representation of product P2

[0100] [0.4,0.2],# Initial representation of product P3

[0101] [0.5, 0.3], # Initial representation of platform T1

[0102] [0.6, 0.4]# Initial representation of platform T2]

[0103] Next, we consider the connections between users, products, and the platform. We've already added edges between users and products based on user behavior data (1 for browsing, 4 for purchasing). Now we need to add edges between users and the platform, and between products and the platform. These edges can represent user activity on different platforms and product sales on different platforms. For example, we can use activity ratings (e.g., dividing different score ranges based on a threshold for online times within a week) as the weight of the edge between users and the platform, or the sales percentage of products on different platforms (divided into thresholds for rating) as the weight of the edge between products and the platform. In this embodiment, to simplify the problem, we assign a fixed weight, for example, 1, to these edges between users and the platform, and between products and the platform. The adjacency matrix A will then be as follows:

[0104] A = [[0,1,4,1,1,1],#U1

[0105] [1,0,0,0,1,1],#P1

[0106] [4,0,0,0,1,0],#P2

[0107] [1,0,0,0,1,1],#P3

[0108] [1,1,1,1,0,0],#T1

[0109] [1,1,0,1,0,0]#T2]

[0110] Then, a graph neural network (such as GCN or GAT) can be applied to update the node representations. In each iteration, the graph neural network updates the node representations based on the adjacency matrix A, the current node representation matrix H^(l), and the weight matrix W^(l).

[0111] For GCN, the update formula is ^(l+1)=σ(A*H^(l)*W^(l)). By iteratively updating the node representations, the graph neural network can capture the relationships between users, products, and the platform. Finally, the node representations learned by the graph neural network can be used as input features and fed into the Transformer model along with other features. In this way, the model can simultaneously consider network structure information and sequence behavior information.

[0112] The weight matrix W^(l) consists of parameters learned during the training of a graph neural network. At the start of training, the weight matrix W^(l) is typically initialized randomly. Then, through gradient descent optimization during training, these weight matrices are progressively adjusted to capture the relationships between nodes in the graph. The activation function σ introduces non-linearity into the neural network. The ReLU (Rectified Linear Unit) activation function is a commonly used activation function, defined as: σ(x) = max(0,x). The ReLU activation function sets negative inputs to 0, while it does not modify inputs greater than 0. Therefore, for inputs greater than 1, the ReLU activation function does not cause information loss. In graph neural networks, the activation function introduces non-linearity to capture complex node relationships. By adjusting the weight matrix W^(l) and iteratively updating the node representations, graph neural networks can learn complex relationships between nodes in the graph.

[0113] The dimension of the weight matrix W^(l) depends on the input and output feature dimensions of each layer in the graph neural network. In layer l, the size of the weight matrix W^(l) is (d_in x d_out), where d_in is the input feature dimension of layer l, and d_out is the output feature dimension of layer l. If, in a certain layer, the input feature dimension is d1 and the output feature dimension is d2, then the size of the weight matrix W^(l) is (d1 x d2). This matrix is ​​continuously updated during training to learn the relationships between nodes. In the example above, there are 6 nodes (1 user node U1, 3 product nodes P1, P2, P3, and 2 platform nodes T1, T2), with the same input and output feature dimensions of 2. Therefore, the size of the node representation matrix H^(l) will be (6 x d), and the size of the weight matrix W^(l) will be (d x d), where d is the dimension of the node feature vector, and d = 2.

[0114] In practice, the adjacency matrix A is normalized to accelerate convergence. Meanwhile, in each layer, the weight matrix W^(l) is updated using gradient descent during training to minimize the loss function.

[0115] When implementing graph neural networks, ready-made graph neural network layers provided by the specialized tool library PyTorch Geometric, such as GCN and GraphSAGE, can be used to easily build and train graph neural network models.

[0116] To implement a Graph Neural Network (GNN) using PyTorch Geometric (PyG), you first need to install the PyG library and then build the graph data structure. In PyG, the `Data` class can be used to represent graph data, requiring the node feature matrix `x`, the edge index matrix `edge_index`, and other optional information (edge ​​features, labels). Next, define the graph neural network model, using predefined graph convolutional layers provided by the PyG library, and define the model's forward propagation process. Then, initialize the model instance, define the loss function and optimizer. Finally, train the model using the training dataset and evaluate it on the validation dataset.

[0117] In this invention, the purpose of updating H^(l+1) is to update the representation of each node (user, product, platform, etc.) in each layer of the graph neural network by aggregating the information of neighboring nodes. After several iterations, a d-dimensional vector representation after graph neural network encoding correction is obtained, capturing the information in the network structure.

[0118] Next, the node representations (users or products) obtained from the GNN can be combined with other features (such as price, platform location, etc.) to form new feature vectors. This allows for the simultaneous consideration of information from the graph structure and other original features in subsequent intent analysis tasks. These combined feature vectors will then be input into other models such as the Transformer to further capture user behavior sequence information across different platforms and perform online intent analysis.

[0119] In one embodiment, it is assumed that a 2D vector representation has been generated for users, products, and e-commerce platform nodes using a GNN, with the time interval encoded as a 1D vector and the behavior type encoded as a 2D one-hot vector (assuming only "browsing" and "purchasing" behaviors). Therefore, the feature vector for each record will be 9-dimensional.

[0120] For example, the GNN of U1 is represented as [0.5, 0.2], the GNN of P1 is represented as [0.3, 0.1], and the GNN of T1 is represented as [0.1, 0.2]. The browsing behavior occurred at 10:00:00 on 2022-04-12. The time interval can be calculated as the normalized time difference (e.g., the time difference with the earliest behavior divided by the maximum time difference), resulting in a time interval encoding of 0.1. The price is normalized to 0.4, and the age is normalized to 0.3.

[0121] Therefore, the combined feature vector of this record is: [0.5,0.2,0.3,0.1,0.1,0.2,0.1,0.4,0.3].

[0122] For other records, a similar method can be used to calculate the combined feature vectors. Finally, these feature vectors can be combined into a matrix, which serves as the input to the Transformer model.

[0123] In one embodiment, the training process of the Transformer model of the present invention includes the following steps:

[0124] In step t1, ensure that the feature vector of each record is 9-dimensional, i.e., GNN representation (6-dimensional, user 2-dimensional + product 2-dimensional + platform 2-dimensional), time interval (1-dimensional), price (1-dimensional) and age (1-dimensional).

[0125] Step t2 uses a linear layer to convert the 9-dimensional feature vector into the input embedding dimension (128 dimensions) required by the model. Setting the input embedding dimension to 128 dimensions indicates that the user behavior sequence comprises a combination of 128 feature vectors. With an input dimension of 9 and an output dimension of 128, the linear layer can be represented as: Linear(9,128).

[0126] Step t3: Construct a Transformer encoder layer, consisting of a multi-head self-attention sublayer and a feedforward neural network sublayer. Each sublayer also includes residual connections and layer normalization. Set the number of heads to 8, and the hidden layer dimension of the feedforward neural network to 256.

[0127] In step t4, in the multi-head self-attention sublayer of the encoder, the input embedding is divided into 8 heads, each with its own attention weights. This allows the model to capture the dependencies between elements in the input sequence in different representation subspaces.

[0128] Step t5: In the feedforward neural network sub-layer of the encoder, a feedforward neural network with a hidden layer dimension of 256 is used to perform a nonlinear transformation on the data processed by the multi-head self-attention mechanism.

[0129] Step t6 involves passing the encoder output to a fully connected layer to extract information about the user's online intent. Assuming there are m intent categories, the output dimension of the fully connected layer should be m. A linear layer can be represented as Linear(128, m). For example, m = 9.

[0130] Step t7: Apply the Softmax activation function after the fully connected layer to convert the output into a probability distribution.

[0131] Step t8: Define an appropriate loss function (cross-entropy loss) to measure the difference between the model's predictions and the true labels. Choose an optimizer (such as Adam, SGD, etc.) to minimize the loss function, thereby optimizing the model parameters.

[0132] Step t9: Train the model using the training dataset and evaluate it on the validation dataset. Evaluation metrics such as accuracy and F1 score can be used to measure model performance. Adjust model parameters and architecture based on the evaluation results to optimize performance.

[0133] The output of the Transformer model can be used to obtain predictions about a user's online intent.

[0134] In one embodiment, there are m intent categories. Assuming the true intent category is c, the true label probability distribution y is a vector of length m, where the c-th element is 1 and the remaining elements are 0. (Model prediction probability distribution) It is a vector of length m, representing the probability predicted by the Transformer model for each class.

[0135] The cross-entropy loss function can be determined as follows:

[0136]

[0137] in,

[0138] L: Cross-entropy loss function value;

[0139] yi: The i-th element in the true label probability distribution. When i = c, yi = 1, otherwise yi = 0;

[0140] The i-th element in the model prediction probability distribution represents the probability that the model predicts the i-th type of intention;

[0141] m: The number of intent categories.

[0142] By minimizing the loss function L, the model's predictions are made closer to the true intent. During training, an optimizer (such as Adam or SGD) is selected to minimize the loss function. The optimizer can automatically perform optimization calculations to obtain optimized model parameters.

[0143] In one embodiment, this invention uses the Hugging Face Transformers library to train a Transformer model, which first requires installing the Hugging Face Transformers library. Then, data preparation is performed, processing the data into a format supported by the Transformers library. These operations are accomplished using the tokenization provided by the library. The Transformer model is then fine-tuned using PyTorch's training and evaluation process. A loss function and optimizer are defined, and the model is trained on the training dataset.

[0144] In one embodiment, to combine features into a single feature vector and format it suitable for input into the Transformer model, all features are concatenated into a single feature vector, and user behavior data across different platforms is arranged chronologically. For each user, a data matrix containing their behavior sequences across different platforms can be generated, with each row representing a feature vector at a given time point. To meet the input requirements of the Transformer model, all user behavior sequences need to be standardized to the same length. A suitable sequence length can be chosen, padding insufficient sequences (e.g., using zero vectors) and truncating excessive sequences. The formatted dataset is then divided into training, validation, and test sets. This facilitates model tuning during training and performance evaluation during testing.

[0145] In one embodiment, based on the input of a GNN and user-featured feature vectors, the following are the types of outputs from the Transformer model and the corresponding recommendation strategies:

[0146] (1) Finding cost-effective products: When users browse similar products on cost-effective platforms, they may be looking for products with more competitive prices. Recommendation strategies can focus on recommending cost-effective products.

[0147] (2) Finding high-quality products: Users browsing products on platforms that emphasize quality are likely more concerned about the quality of the products. Recommendation strategies can focus on recommending higher-quality products.

[0148] (3) Interest in a certain type of product: Users may become interested in similar products when browsing different platforms. Recommendation strategies can focus on recommending similar products that users may be interested in.

[0149] (4) Quick Shopping: Users can quickly find and purchase the products they need on different platforms, spending very little time browsing other products. Recommendation strategies can focus on recommending frequently purchased items from the user's historical purchase records, making it easier for users to quickly find the products they need.

[0150] (5) Understand industry trends and popular products: Users follow and browse the latest technology products, fashion trends, or other popular items on different platforms. Recommendation strategies can focus on recommending recently popular and attention-grabbing products.

[0151] (6) Purchasing high-end and luxury goods: Users browsing products on high-end and luxury goods platforms may be interested in purchasing such items. Recommendation strategies can focus on recommending high-end and luxury goods, such as designer clothing, jewelry, and watches.

[0152] (7) Finding environmentally friendly and sustainable products: Users browse and search for environmentally friendly, organic, or sustainable products on different platforms. Recommendation strategies can focus on recommending environmentally friendly, organic, and sustainable products, such as organic food and eco-friendly household goods.

[0153] (8) Finding customized products and personalized services: Users browse customized products and seek personalized services on different platforms. Recommendation strategies can focus on recommending products that offer customization options and personalized services, such as customized furniture and personalized gifts.

[0154] (9) Interest in products with specific functions or features: Users follow and browse products with specific functions or features on different platforms, such as smart home devices and outdoor sports equipment. Recommendation strategies can focus on recommending products with functions or features that users are interested in.

[0155] Based on the user's online intents mentioned above, an integer label can be assigned to each intent. For example, for the intents (1) to (9) mentioned above, integer labels 1-9 can be assigned accordingly.

[0156] Optionally, some of the user online intents of the above 1-9 categories can be selected for use in the training model, or other types of user online intents can be added, depending on the characteristics of data annotation and business needs.

[0157] In one embodiment, for each sequence of user actions in the input data, the output of the Transformer model is a probability distribution representing the predicted probability for each intent category. For example, the model's output is as follows:

[0158]

[0159]

[0160] In this example, the model predicts the highest probability (0.6) for intent label 2, therefore it can be assumed that the intent corresponding to user behavior sequence 1 is 2. In this way, the model can predict the intent of different user behavior sequences and provide corresponding recommendation strategies for each sequence.

[0161] In one embodiment, such as Figure 2 As shown, the big data e-commerce platform is used for user online intent analysis based on artificial intelligence, and the big data e-commerce platform includes the following modules:

[0162] The data collection module is used to collect user behavior data (including browsing and purchasing) on ​​different platforms in real time and determine relevant contextual information (including time intervals) to provide input data for training and recommendation.

[0163] The feature processing module is used to transform user behavior data and context information into feature vectors, representing the corresponding features using one-hot encoding and continuous features, respectively. The feature processing module is also used to update the one-hot encoding representation using a graph neural network (GNN), combine the updated node representation with the normalized continuous features, and input the combined feature vector into the Transformer model.

[0164] The intent recognition model module is responsible for building and training an intent recognition model based on the Transformer model, which is used to capture user behavior sequences on different platforms to identify the user's online intent.

[0165] The model training and optimization module is used to train the Transformer model using the collected data and to optimize the Transformer model using the cross-entropy loss function during the training process.

[0166] The performance evaluation module is used to evaluate the model's performance on the test set, using metrics such as accuracy, recall, and F1 score to measure the model's accuracy in recognizing users' online intent.

[0167] The recommendation system integration module is used to deploy the trained Transformer model into the recommendation system. By collecting user behavior data on different platforms in real time and inputting it into the Transformer model, the module obtains the user's online intent and makes recommendations according to the recommendation strategy corresponding to the online intent.

[0168] The recommendation system integration module is used to identify users' online intentions in real time. It collects user behavior data on different platforms in real time and inputs it into the trained Transformer model to obtain the user's current online intention. The trained Transformer model is built and trained by the intention recognition model module, which uses a Transformer-based intention recognition model to capture user behavior sequences on different platforms to identify the user's online intention.

[0169] The user online intent analysis method and big data e-commerce platform based on the Transformer model of this invention can significantly improve the accuracy of intent recognition. The Transformer model can capture long-distance dependencies in the input sequence, thereby more accurately identifying the user's online intent. By collecting user behavior data on different platforms in real time, the current online intent of the user can be obtained in real time, enabling the recommendation system to respond to user needs promptly. In addition, it can also help to realize personalized recommendations. After accurately identifying the user's online intent, the recommendation system can provide more personalized product recommendations based on the user's online intent, thereby improving user experience and purchase conversion rate.

[0170] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0171] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0172] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0173] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0174] The units described in the embodiments of this disclosure can be implemented in software or in hardware. The names of the units are not, in some cases, intended to limit the specific unit.

[0175] The preferred embodiments of the present invention have been described above to make the spirit of the present invention clearer and easier to understand, and are not intended to limit the present invention. All modifications, substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope summarized by the appended claims.

Claims

1. A method for analyzing online user intent based on artificial intelligence, the method comprising: User behavior data is collected from different e-commerce platforms, and relevant contextual information is recorded. The user behavior data and the contextual information are then converted into feature vectors. Construct an intent recognition model based on the Transformer model, train the Transformer model using collected data, optimize the model using the cross-entropy loss function during training, and evaluate the model's performance on the test set. The trained Transformer model is deployed to the recommendation system. By collecting user behavior data on different platforms in real time and inputting it into the Transformer model, the user's current online intent can be obtained.

2. The method for analyzing online user intent based on artificial intelligence as described in claim 1, characterized in that, One-hot encoding is used to represent users, platforms, products, and user behavior. Continuous features are used to represent time intervals, prices, and ages. These continuous features need to be normalized. All of the above features are combined into a feature vector, which is used as the input to the Transformer model.

3. The method for analyzing online user intent based on artificial intelligence as described in claim 2, characterized in that, After updating the one-hot encoding representation using a graph neural network (GNN), the updated node representation is combined with the normalized continuous features, and the combined feature vector is input into the Transformer model.

4. The method for analyzing online user intent based on artificial intelligence as described in claim 3, characterized in that, Updating the one-hot encoding representation using a graph neural network (GNN) includes: Based on user behavior data on different e-commerce platforms, an interaction network is constructed as a user behavior graph. In the user behavior graph, nodes represent users, products and platforms, and the edges between users and products represent user behaviors. Weights are assigned to the edges between users and products, with different user behaviors corresponding to different weights. The graph neural network (GNN) is used to encode the user behavior graph. The GNN updates the representation of each node by aggregating information from neighboring nodes.

5. The method for analyzing online user intent based on artificial intelligence as described in claim 4, characterized in that, The edge weight between users and platforms is set to a weight determined by the number of times they go online within a preset time period, and the edge weight between products and platforms is set to a weight determined by the sales share of products on different platforms. Alternatively, the edge weights between users and the platform, and between products and the platform, can be set to fixed weights.

6. The method for analyzing online user intent based on artificial intelligence as described in claim 3, characterized in that, Initialize a vector representation for each node in the user behavior graph. Define a weight matrix for each layer of the graph neural network. Update the node representation using the adjacency matrix and the weight matrix for each layer. The node representation is updated iteratively, and in each iteration, the node representation will capture information about neighboring nodes that are further away. After completing the iterative update, the final node representation matrix will be obtained, which contains representations of user, product, and platform nodes.

7. The method for analyzing online user intent based on artificial intelligence as described in claim 1, characterized in that, Construct a Transformer encoder layer, which includes a multi-head self-attention sub-layer and a feedforward neural network sub-layer. Each sub-layer also includes residual connections and layer normalization. Set the number of heads to 8 and the hidden layer dimension of the feedforward neural network to 256. In the multi-head self-attention sublayer of the encoder, the input embedding is divided into 8 heads, each with its own attention weights; In the feedforward neural network sub-layer of the encoder, a feedforward neural network with a hidden layer dimension of 256 is used to perform a non-linear transformation on the data processed by the multi-head self-attention mechanism. The encoder output is passed to a fully connected layer to extract information about the user intent. The output dimension of the fully connected layer should be m, where m is the number of categories of the user intent.

8. The method for analyzing online user intent based on artificial intelligence as described in claim 6, characterized in that, Apply the Softmax activation function after the fully connected layer to convert the output into a probability distribution; For each user behavior sequence in the input data, the output of the Transformer model is a probability distribution representing the predicted probability of each intent category. The intent label with the highest predicted probability from the Transformer model is determined as the intent of the current user behavior sequence.

9. The method for analyzing online user intent based on artificial intelligence as described in claim 6, characterized in that, The user online intent analysis method may further include providing product recommendations that match the user's intent.

10. A big data e-commerce platform for analyzing online user intent based on artificial intelligence, the big data e-commerce platform comprising: The data collection module is used to collect user behavior data and contextual information on different platforms in real time; The feature processing module is used to transform user behavior data and contextual information into feature vectors; The intent recognition model module is responsible for building and training an intent recognition model based on the Transformer model. The Transformer model is used to capture the user's behavioral sequences on different platforms to identify the user's shopping intent. The model training and optimization module is used to train the Transformer model using the collected data and to optimize the Transformer model using the cross-entropy loss function during the training process. The performance evaluation module is used to evaluate the performance of the Transformer model on the test set; The recommendation system integration module is used to deploy the trained Transformer model into the recommendation system. By collecting user behavior data on different platforms in real time and inputting it into the Transformer model, the module obtains the user's online intent and makes recommendations according to the recommendation strategy corresponding to the online intent.

11. The big data e-commerce platform for user online intent analysis based on artificial intelligence as described in claim 9, characterized in that, The feature processing module is also used to update the representation of one-hot encoding using a graph neural network (GNN), combine the updated node representation with the normalized continuous features, and input the combined feature vector into the Transformer model.

Images