A product marketing recommendation method and system based on user intent

Abstract

The application discloses a product marketing recommendation method and system based on user intention, and relates to the technical field of data analysis; effective historical interaction data of a target user is divided to obtain a front benchmark intention sequence and a rear benchmark intention sequence; the rear benchmark intention sequence is substituted into a similarity calculation network to obtain an intention similarity list; the intention similarity list and the rear benchmark intention sequence are substituted into a teacher model to obtain intention semantic knowledge; the front benchmark intention sequence is substituted into a student model to obtain an initial student model; and the intention semantic knowledge is transferred to the initial student model to train and determine a target commodity recommendation list. Through accurate capture of the evolution trend and mutation characteristics of user intention, the accuracy and personalization level of the recommendation result are significantly improved, and the demand of the platform and the user for high-quality marketing recommendation is better met.

Description

Technical Field

[0001] This invention belongs to the field of data analysis technology, specifically relating to a product marketing recommendation method and system based on user intent. Background Technology

[0002] With the rapid development of the digital economy, including e-commerce and content services, precise product marketing recommendations have become a core means to improve user experience and increase platform conversion efficiency. Users' historical interactions on the platform, such as product clicks, favorites, purchases, and browsing, contain rich information about their potential intentions. Building recommendation models based on this data to achieve personalized recommendations has become a mainstream technological direction in the industry.

[0003] Currently, marketing recommendation methods based on user intent can be mainly divided into two categories: one is time-series recommendation models, which rely on users' historical interaction data to predict products that users may be interested in later by mining the time-series patterns behind the behavior, focusing on fitting the user's past behavior trends; the other is semantically enhanced recommendation models, which introduce external semantic information to calculate the similarity between products and between users and products, supplementing the semantic dimension of intent, thereby improving the accuracy of recommendation results, focusing on mining the semantic features of intent.

[0004] Publication No. CN117522510A discloses a product recommendation method and apparatus, the core of which revolves around marketing gain modeling to achieve recommendation. The method first constructs a full customer sample set based on user data, then splits the full customer sample set into a first recommendation user group and a second recommendation user group based on the marketing situation of each user. Subsequently, based on the characteristic attributes of each user in the two user groups and the behavioral data generated by users in response to marketing situations, a first marketing gain model and a second marketing gain model are constructed respectively. In the actual recommendation stage, the recommendation quantity of the products to be recommended and the user data to be recommended are obtained. Based on the marketing situation and corresponding behavioral data of the users to be recommended, the gain scores of the users are calculated using the above two models respectively. Finally, the target recommendation audience is determined by combining the gain scores and the recommendation quantity, and the push is completed.

[0005] However, neither the two mainstream methods mentioned above nor existing solutions, including the CN117522510A patent, have designed adaptation mechanisms for the abrupt changes in user intent. They are unable to accurately capture the evolution trend of intent, nor can they effectively transfer knowledge from historical intent to future intent. Furthermore, they lack deep integration of semantic similarity and temporal dynamic features, ultimately resulting in recommendation results that lack foresight and accuracy, failing to meet the platform's and users' needs for high-quality personalized recommendations. Summary of the Invention

[0006] The purpose of this invention is to solve the problem of failing to meet the demand of platforms and users for high-quality personalized recommendations, and to propose a product marketing recommendation method and system based on user intent.

[0007] In a first aspect of this invention, a product marketing recommendation method based on user intent is first proposed, the method comprising: The system acquires valid historical interaction data of the target user over a historical time series, and splits the interaction data according to the target time point to obtain a pre-term baseline intent sequence and a post-term baseline intent sequence; the target time point is determined by the similarity between products in the valid historical interaction data; the valid historical interaction data before the target time point is used as the pre-term baseline intent sequence; and the valid historical interaction data after the target time point is used as the post-term baseline intent sequence. The aforementioned baseline intent sequence is substituted into a similarity calculation network to obtain an intent similarity list; A teacher model is constructed, and the intent similarity list and the subsequent baseline intent sequence are substituted into the teacher model for training to obtain intent semantic knowledge; Construct a student model, and train the student model by substituting the aforementioned baseline intent sequence into the student model to obtain an initial student model; The intent semantic knowledge is transferred to the initial student model for training to obtain the target marketing model; The target marketing model is then used to input the target users' valid historical interaction data to obtain a target product recommendation list.

[0008] This invention provides a product marketing recommendation method based on user intent. By accurately identifying the target time point of user intent mutation, it innovatively divides historical interactions into a pre-stage benchmark and a post-stage benchmark sequence. It also utilizes a teacher-student model architecture to achieve effective transfer from post-stage intent semantic knowledge to the pre-stage benchmark model. This not only accurately captures the intent evolution trend and solves the problems of existing technologies being unable to adapt to intent mutations and lacking foresight, but also significantly improves the accuracy and personalization quality of recommendation results by deeply integrating semantic similarity and temporal dynamic features.

[0009] Optionally, determining the target time point includes: The target valid historical interaction data is obtained by data interception of the valid historical interaction data, and the target valid historical interaction data is converted into a time item ID associated sequence; Substitute the time item ID association sequence into the preset item embedding table to obtain the embedding vector corresponding to each item ID, and obtain all embedding vectors as the time-series intent feature sequence. The local difference score is obtained by calculating the average cosine distance between the intent features at time step t and the previous w time steps in the temporal intent feature sequence using a sliding window. Calculate the mean and standard deviation of the local difference scores for all time steps, and convert the local scores into standardized mutation scores based on the mean and standard deviation; The mutation score that retains the local maximum value after merging consecutive time steps with low scores is denoted as the local maximum mutation value; the low score refers to the standardized mutation score that is less than 1. The target time point is determined based on a preset significant mutation threshold and all local maximum mutation values.

[0010] By transforming effective historical interaction data into a temporal intent feature sequence and innovatively employing a sliding window to calculate the average cosine distance between time steps, this method can keenly capture subtle fluctuations and local differences in user interests at a high-dimensional semantic level. Furthermore, through standardized mutation score processing and a merging strategy for consecutive low scores, random noise interference is effectively filtered out, and the local maximum mutation value is accurately located. Finally, by combining a preset threshold to determine the target time point, precise anchoring of the moment of user intent mutation is achieved. This method not only provides solid data support for subsequently dividing historical data into pre- and post-pre-test baseline sequences, but also solves the problem of existing technologies' difficulty in quantifying intent evolution nodes, laying a crucial foundation for the effective transfer of intent knowledge.

[0011] Optionally, substituting the latter-stage baseline intent sequence into a similarity calculation network to obtain an intent similarity list includes: The first convolutional feature is obtained by sequentially performing 7×7 convolution and 3×3 pooling operations on the aforementioned baseline intent sequence. The first convolutional feature is successively substituted into the four residual layers to obtain the second convolutional feature; After performing average pooling on the convolutional features, the future intent attention features are extracted by the target attention module. Substituting the future intent attention features into the fully connected layer yields an intent similarity list; The working principle of the target attention module is as follows: Obtain the input features, and perform a 1×1 convolution operation on the input features to obtain the first attention features; The first attention is divided into N first channel sub-features along the channel dimension. Attention weights are generated for each first channel sub-feature along the Y-axis to obtain the first sub-attention weight. All first channel sub-features are weighted and fused according to all first sub-attention weights to obtain the intermediate attention feature. The intermediate attention feature is further divided into M second channel sub-features along the channel dimension. Attention weights are generated for each second channel sub-feature along the X-axis to obtain second sub-attention weights. All second channel sub-features are weighted and fused according to all second sub-attention weights to obtain future intention attention features.

[0012] By constructing a deep convolutional neural network and an innovative attention mechanism, we achieved refined semantic mining of the subsequent baseline intent sequence. By generating a more discriminative and semantically representative list of intent similarities, we provided high-quality supervision signals for the teacher model, thus laying a solid feature foundation for subsequent knowledge transfer and accurate recommendation.

[0013] Optionally, the intention similarity list and the subsequent baseline intention sequence are substituted into the teacher model for training to obtain intention semantic knowledge, including: The back-end baseline intent sequence is mapped to a vector sequence through an item embedding table, and the position embedding is fused to generate a fused embedding matrix containing temporal information. A Bernoulli distribution masking matrix is ​​generated based on a preset masking probability, and then multiplied with the fusion embedding matrix to obtain the target embedding matrix; The target embedding matrix is ​​fed into the target encoder-decoder to obtain the intent embedding features; Substituting the intent embedding features into a preset dynamic gating network yields an expert weight matrix; Substituting the intent embedding features into K feedforward neural networks respectively, we obtain multiple expert output vectors; The total expert output vector is obtained by weighted summing of all expert output vectors based on the expert weight matrix. The expert's total output vector is substituted into a 1D causal convolutional network to obtain the expert's total causal vector. The expert's total output vector and the expert's total causal vector are then fused to obtain dynamic intent features. The dynamic intent features are then substituted into a fully connected network to obtain a predicted similarity vector; Based on the intent similarity list, it is compared with the predicted similarity vector, and a mean squared error loss function is defined for aligning the predicted similarity vector with the output predicted similarity list in subsequent training. The training is updated until the similarity between the intent similarity list and the output predicted similarity list is greater than a preset threshold, at which point the training is terminated, and the dynamic intent features updated each time are retained as intent semantic knowledge.

[0014] By simulating the sequence prediction task after a sudden change in user intent, the teacher model is forced to learn the deep temporal dependencies and contextual semantic relationships of intent sequences. At the same time, by using a hybrid expert network and a dynamic gating mechanism, specific expert networks can be adaptively activated according to different intent features, realizing refined modeling and parallel processing of complex intent patterns. This not only ensures that the teacher model can mine high-dimensional, structured intent semantic knowledge, but also solidifies these implicit semantic rules through knowledge distillation, providing high-quality and robust supervision signals for the knowledge transfer of subsequent student models, thereby effectively improving the generalization ability and semantic understanding depth of the entire recommendation system.

[0015] Optionally, transferring the intent semantic knowledge to the initial student model for training to obtain the target marketing model includes: The training similarity is obtained by calculating the similarity between the predicted product list of the student model and the predicted product list of the teacher model using cosine similarity. The target similarity is obtained by correcting the training similarity using a decay coefficient; Distillation loss is obtained by distilling the semantic knowledge of intent based on the target similarity. The Adam optimizer is used to iteratively optimize the parameters of the student model, and the optimization loss is calculated through forward propagation. The total loss is obtained by summing the optimization loss and the distillation loss. The model parameters are updated by backpropagation to minimize the total loss until the training termination condition is met, thus obtaining the target marketing model.

[0016] By introducing cosine similarity calculation and combining it with decay coefficient correction, an adaptive knowledge distillation mechanism is constructed. This mechanism can accurately measure the differences in intent prediction between teacher and student models and effectively balance the optimization loss of model fitting data with the distillation loss of transferring teacher knowledge. Then, by using the Adam optimizer for parameter iteration and backpropagation with the goal of minimizing the total loss, efficient knowledge transfer from teacher model to student model is achieved. This solves the problem of the difficulty in effectively transferring historical intent to future intent. Furthermore, the distillation process forces the student model to learn the implicit structured semantic associations of the teacher model, thereby significantly improving the ability to capture users' future intent and the accuracy of recommendations while maintaining a lightweight design.

[0017] In a second aspect of this invention, a product marketing recommendation system based on user intent is proposed, comprising: The target time point determination module is used to acquire valid historical interaction data of the target user in the historical time series, and split the interaction data according to the target time point to obtain a pre-stage baseline intent sequence and a post-stage baseline intent sequence; the target time point is determined by the similarity between products in the valid historical interaction data; the valid historical interaction data before the target time point is used as the pre-stage baseline intent sequence; and the valid historical interaction data after the target time point is used as the post-stage baseline intent sequence. An intent similarity list generation module is used to input the subsequent baseline intent sequence into a similarity calculation network to obtain an intent similarity list; The intent semantic knowledge generation module is used to construct a teacher model. The intent similarity list and the subsequent baseline intent sequence are substituted into the teacher model for training to obtain intent semantic knowledge. The initial student model training module is used to construct a student model by substituting the preceding baseline intent sequence into the student model for training to obtain the initial student model. The target marketing model training module is used to transfer the intent semantic knowledge to the initial student model for training to obtain the target marketing model. The target product recommendation list generation module is used to input the target user's valid historical interaction data into the target marketing model to obtain the target product recommendation list.

[0018] Optionally, the target time point determination module includes: The time item ID association sequence generation module is used to extract target valid historical interaction data from the valid historical interaction data and convert the target valid historical interaction data into a time item ID association sequence. The temporal intent feature sequence determination module is used to substitute the time item ID association sequence into a preset item embedding table to obtain the embedding vector corresponding to each item ID, and obtain all embedding vectors as the temporal intent feature sequence. The local difference score calculation module is used to calculate the average cosine distance between the intent features at time step t and the previous w time steps in the temporal intent feature sequence using a sliding window to obtain the local difference score. The standardized mutation score determination module is used to calculate the mean and standard deviation of the local difference scores corresponding to all time steps, and convert the local scores into standardized mutation scores based on the mean and standard deviation. The local maximum mutation value determination module is used to merge consecutive low-score time steps and retain the mutation score with the local maximum value as the local maximum mutation value; the low score is the standardized mutation score with a standardized mutation score less than 1. The target time point generation module is used to determine the target time point based on a preset significant mutation threshold and all local maximum mutation values.

[0019] Optionally, the intent similarity list generation module includes: The first convolutional feature determination module is used to perform 7×7 convolution and 3×3 pooling operations on the subsequent baseline intent sequence to obtain the first convolutional feature. The second convolutional feature determination module is used to sequentially substitute the first convolutional features into four residual layers to obtain the second convolutional features. The future intent attention feature determination module is used to perform average pooling on the two convolutional features and then input them into the target attention module to extract the future intent attention features; The intent similarity list determination module is used to input the future intent attention features into the fully connected layer to obtain the intent similarity list; The working principle of the target attention module is as follows: Obtain the input features, and perform a 1×1 convolution operation on the input features to obtain the first attention features; The first attention is divided into N first channel sub-features along the channel dimension. Attention weights are generated for each first channel sub-feature along the Y-axis to obtain the first sub-attention weight. All first channel sub-features are weighted and fused according to all first sub-attention weights to obtain the intermediate attention feature. The intermediate attention feature is further divided into M second channel sub-features along the channel dimension. Attention weights are generated for each second channel sub-feature along the X-axis to obtain second sub-attention weights. All second channel sub-features are weighted and fused according to all second sub-attention weights to obtain future intention attention features.

[0020] Optionally, the intent semantic knowledge generation module includes: The fusion embedding matrix generation module is used to map the back-end baseline intent sequence into a vector sequence through an item embedding table, and fuse position embeddings to generate a fusion embedding matrix containing temporal information. The target embedding matrix determination module is used to generate a Bernoulli distribution masking matrix based on a preset masking probability, and then multiply it with the fused embedding matrix to obtain the target embedding matrix; The intent embedding feature generation module is used to input the target embedding matrix into the target encoder-decoder to obtain intent embedding features; The expert weight matrix determination module is used to substitute the intent embedding features into a preset dynamic gating network to obtain the expert weight matrix. An expert output vector determination module is used to substitute the intent embedding features into K feedforward neural networks to obtain multiple expert output vectors; The expert total output vector determination module is used to obtain the expert total output vector by weighted summation of all expert output vectors based on the expert weight matrix. The dynamic intent feature determination module is used to substitute the expert's total output vector into a 1D causal convolutional network to obtain the expert's total causal vector, and to fuse the expert's total output vector and the expert's total causal vector to obtain dynamic intent features; The prediction similarity vector determination module is used to input the dynamic intent features into a fully connected network to obtain the prediction similarity vector; The dynamic intent feature update module is used to compare the intent similarity list with the predicted similarity vector, and define a mean squared error loss function for aligning the predicted similarity vector with the output predicted similarity list in subsequent training. The training is updated until the similarity between the intent similarity list and the output predicted similarity list is greater than a preset threshold, at which point the training is terminated, and the dynamic intent features updated each time are retained as intent semantic knowledge.

[0021] Optionally, the target marketing model training module includes: The training similarity calculation module is used to calculate the training similarity by using cosine similarity to measure the similarity between the predicted product list of the student model and the predicted product list of the teacher model. The target similarity determination module is used to correct the training similarity using a decay coefficient to obtain the target similarity; The distillation loss calculation module is used to distill the semantic knowledge of intent based on the target similarity to obtain the distillation loss. The optimization loss determination module is used to iteratively optimize the parameters of the student model using the Adam optimizer and calculate the optimization loss through forward propagation. The total loss determination module is used to sum the optimization loss and the distillation loss to obtain the total loss; The Minimize Total Loss Training Module is used to update the model parameters by backpropagating with the minimum total loss until the training termination condition is met to obtain the target marketing model.

[0022] The beneficial effects of this invention are: This invention proposes a product marketing recommendation method based on user intent. By identifying the time points of abrupt changes in user intent, historical interaction data is segmented into a pre- and post-preceding baseline intent sequence. The post-preceding sequence is used in conjunction with a similarity calculation network to train a teacher model to extract semantic knowledge of intent, which is then transferred to a student model trained based on the pre-preceding sequence. This effectively achieves knowledge transfer from historical intent to future intent. Simultaneously, this scheme deeply integrates semantic similarity and temporal dynamic features, accurately capturing the evolution trend and abrupt change characteristics of user intent, enhancing the foresight and adaptability of the recommendation model, and ultimately significantly improving the accuracy and personalization of recommendation results, better meeting the needs of platforms and users for high-quality marketing recommendations. Attached Figure Description

[0023] The invention will now be further described with reference to the accompanying drawings.

[0024] Figure 1 A flowchart illustrating a product marketing recommendation method based on user intent, provided as an embodiment of the present invention; Figure 2 A schematic diagram of a similarity calculation network provided in an embodiment of the present invention; Figure 3 This is a framework diagram of a product marketing recommendation system based on user intent, provided for an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0026] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] This invention provides a product marketing recommendation method based on user intent. See also... Figure 1 , Figure 1 A flowchart illustrating a product marketing recommendation method based on user intent, provided as an embodiment of the present invention. The method includes the following steps: S101, Obtain the target user's valid historical interaction data in the historical time series, and split the interaction data according to the target time point to obtain the front-end baseline intent sequence and the back-end baseline intent sequence; S102, Substitute the subsequent baseline intent sequence into the similarity calculation network to obtain an intent similarity list; S103, Construct a teacher model by substituting the list of intent similarities and the subsequent baseline intent sequence into the teacher model for training to obtain intent semantic knowledge; S104, Construct a student model by substituting the previous baseline intent sequence into the student model for training to obtain the initial student model; S105, Transfer the semantic knowledge of intent to the initial student model for training to obtain the target marketing model; S106, Substitute the target user's valid historical interaction data into the target marketing model to obtain the target product recommendation list; The target time point is determined by the similarity between products in the effective historical interaction data; the effective historical interaction data before the target time point is used as the front-end baseline intent sequence; and the effective historical interaction data after the target time point is used as the back-end baseline intent sequence.

[0028] In one implementation, complete interaction data of the target user in the recommendation scenario is collected, including but not limited to actions such as clicking, purchasing, rating, and favoriting. The timestamp of each action and the corresponding interaction item are recorded. The user's valid interaction data is arranged in ascending order by timestamp to form a time-series interaction sequence of the user.

[0029] In one implementation, the teacher model learns local pattern recognition capabilities by processing a large number of local interactions, such as how to quickly respond to changes in user intent and how to identify sudden group behaviors. Through transfer learning, the student model can inherit the generalization ability of the teacher model, thereby achieving the recognition and understanding of users' local intents without being dominated by them. This improves adaptability to novel items or sudden scenarios, enabling the student model to maintain focus on global interests while still responding appropriately to reasonable changes in local behavior, rather than completely ignoring local signals. Existing traditional hybrid training models tend to continue recommending related products even after user interest has faded because historical interaction weights are too high. However, this solution proposes that the student model be trained with continuous interest, so when local interest fades, the model is more inclined to rely on global preferences for recommendations, improving resource utilization efficiency.

[0030] In one implementation, when the local interest pattern changes, only the teacher model can be updated, and then quickly migrated to the student model, without having to retrain the entire system.

[0031] In one implementation, user interaction data often contains a large amount of local, explosive, and non-persistent interests. If these data are mixed for training, the model will be biased by local explosive behaviors, resulting in the recommendation of a large number of irrelevant products after the user's interest has faded, causing waste of resources and user resentment. Therefore, it is necessary to separate local explosive data from historical data.

[0032] In one implementation, the teacher model and the student model share the same architecture, differing only in their input data and training objectives. Both aim to model the user's intent sequence, outputting a feature vector that represents the user's interests, i.e., dynamic intent features, and ultimately generating a prediction list for recommendation through a fully connected layer. The teacher model's input consists of the back-end baseline intent sequence and an intent similarity list; the student model's input consists only of the front-end baseline intent sequence.

[0033] In one implementation, the teacher model comprises the following layers: an embedding layer: fusion of item embeddings and position embeddings, with an embedding dimension of 64; a denoising layer: a Bernoulli distribution masking matrix; an encoding / decoding core: a target encoder / decoder that generates intent embedding features; a dynamic expert hybrid (MoE): a gated network that generates expert weight matrices; an expert layer: K feedforward neural networks (FFNs); a temporal capture layer: a 1D causal convolutional network (capturing sequence dependencies without revealing future information); and an output layer: a fully connected network (mapping to a predicted similarity vector). The back-end baseline intent sequence is mapped to an embedding matrix incorporating temporal information, and some features are randomly masked. High-dimensional dynamic intent features are extracted through the encoder / decoder and MoE modules, combined with causal convolution. The dynamic intent features are input into the fully connected layer to obtain a predicted similarity vector. The mean squared error (MSE) between this predicted vector and the input intent similarity list is calculated. Parameters are updated via backpropagation until the similarity meets a threshold. During training, the updated dynamic intent features are retained and encapsulated as intent semantic knowledge for use by the student model. The embedding layer to the dynamic expert hybrid layer in the student model is completely consistent with that in the teacher model.

[0034] In one embodiment, determining the target time point includes: Data interception is performed on valid historical interaction data to obtain the target valid historical interaction data, and the target valid historical interaction data is converted into a time item ID associated sequence; Substitute the time item ID association sequence into the preset item embedding table to obtain the embedding vector corresponding to each item ID, and obtain all embedding vectors as the time-series intent feature sequence. The local difference score is obtained by calculating the average cosine distance between the intent features at time step t and the previous w time steps in the temporal intent feature sequence using a sliding window. Calculate the mean and standard deviation of the local difference scores for all time steps, and convert the local scores into standardized mutation scores based on the mean and standard deviation; The mutation score that retains the local maximum value after merging consecutive time steps with low scores is called the local maximum mutation value; low scores are standardized mutation scores with a standardized mutation score less than 1. The target time point is determined based on a preset significant mutation threshold and all local maximum mutation values.

[0035] In one implementation, the process of extracting target valid historical interaction data by data truncation involves the following steps: if the length of the valid historical interaction data is less than 40, it indicates insufficient data volume, and the median time point of the sequence is directly taken as the target time point; if the length of the valid historical interaction data is greater than 200, the most recent 200 interaction data are truncated to ensure detection efficiency and accuracy; the preset item embedding table is obtained by substituting the user's valid historical interaction data into the Skip-gram model for lightweight pre-training.

[0036] In one implementation, the sliding window size is determined by the technicians and is set to 5 by default. The target time point is determined based on a preset significant mutation threshold and all local maximum mutation values. Specifically, the standardized mutation score greater than 2 is recorded as the significant mutation threshold. If there are multiple local maximum mutation values ​​in the sequence, the time point with the highest mutation score is selected as the target time point. If all standardized mutation scores are less than 2, the time point with the local maximum mutation value closest to the end of the sequence is selected as the target time point. The final target time point must satisfy the requirement that there are at least 10 interaction data points before and after it.

[0037] In one implementation, the local difference score is 1 minus the mean cosine similarity between the features of the previous w time steps and the current feature; time step t (where t represents any time in the temporal intent feature sequence, and satisfies that at least 5 records exist before that time) and the previous w time steps (default is 5, the value is between 5 and 10, and is determined by the technician); the standardized mutation score is obtained by subtracting the mean from the local difference score and dividing by the standard deviation.

[0038] In one embodiment, substituting the subsequent baseline intent sequence into a similarity calculation network to obtain an intent similarity list includes: The first convolutional feature is obtained by sequentially performing 7×7 convolution and 3×3 pooling operations on the subsequent baseline intent sequence. Substitute the first convolutional feature into the four residual layers in sequence to obtain the second convolutional feature; After performing average pooling on the convolutional features, the future intent attention features are extracted by the target attention module. Substituting the future intent attention features into the fully connected layer yields a list of intent similarities; The working principle of the target attention module is as follows: Obtain the input features, and perform a 1×1 convolution operation on the input features to obtain the first attention features; The first attention is divided into N first channel sub-features along the channel dimension. Attention weights are generated for each first channel sub-feature along the Y-axis to obtain the first sub-attention weight. The intermediate attention features are obtained by weighted fusion of all first channel sub-features based on all first sub-attention weights. The intermediate attention feature is further divided into M second channel sub-features along the channel dimension. Attention weights are generated for each second channel sub-feature along the X-axis to obtain the second sub-attention weights. All second channel sub-features are weighted and fused according to all second sub-attention weights to obtain the future intent attention feature.

[0039] In one implementation, N and M are determined by technical personnel.

[0040] In one implementation, see [link to implementation details]. Figure 2 , Figure 2 This is a schematic diagram of a similarity calculation network provided in an embodiment of the present invention; it includes four residual layers, namely a first residual layer, a second residual layer, a third residual layer and a fourth residual layer, wherein the second residual layer, the third residual layer and the fourth residual layer have the same structure and are all subjected to 1×1 convolution, 3×3 convolution and 1×1 convolution in sequence; the first residual layer is subjected to 3 1×1 convolutions in sequence.

[0041] In one embodiment, the intention similarity list and the subsequent baseline intention sequence are substituted into the teacher model for training to obtain intention semantic knowledge, including: The back-end baseline intent sequence is mapped to a vector sequence through an item embedding table, and the position embedding is fused to generate a fused embedding matrix containing temporal information. A Bernoulli distribution masking matrix is ​​generated based on a preset masking probability, and then multiplied with the fusion embedding matrix to obtain the target embedding matrix; The target embedding matrix is ​​fed into the target encoder-decoder to obtain the intent embedding features; Substitute the intent embedding features into a pre-defined dynamic gating network to obtain the expert weight matrix; Substituting the intent embedding features into K feedforward neural networks respectively, we obtain multiple expert output vectors; The total expert output vector is obtained by weighted summation of all expert output vectors based on the expert weight matrix. Substituting the expert's total output vector into a 1D causal convolutional network yields the expert's total causal vector. The expert's total output vector and the expert's total causal vector are then fused to obtain the dynamic intent feature. The dynamic intent features are substituted into the fully connected network to obtain the predicted similarity vector; Based on the intent similarity list, it is compared with the predicted similarity vector, and a mean squared error loss function is defined for aligning the predicted similarity vector with the output predicted similarity list in subsequent training. The training is updated until the similarity between the intent similarity list and the output predicted similarity list is greater than a preset threshold, at which point the training is terminated, and the dynamic intent features updated each time are retained as intent semantic knowledge.

[0042] In one implementation, the teacher model construction process first defines key hyperparameters: fixed sequence length L=50; future step size W(5-10); item embedding dimension d=64 or 128; masking probability pmask=0.15 for the multimodal dynamic variational autoencoder MDVAE; number of experts K=8 for the conditional hybrid expert model CMoE, causal convolution kernel size kernel=3; weight decay coefficient 0.01, dropout rate 0.5, optimizer is Adam, learning rate 0.0001; First, a tensor structure is received for two types of input data: a baseline intent sequence and an intent similarity list. The baseline intent sequence consists of an item ID, which corresponds to a standardized sequence of future user interactions. The intent similarity list consists of a similarity score in the range [0,1], which corresponds one-to-one with the sequence time. Construct a learnable item embedding table I, initialized with a normal distribution (mean 0, variance 0.01), and iteratively updated during subsequent training; Construct a learnable position embedding matrix Capture sequence time dependencies and use formulas The fusion yields a fusion embedding matrix; where Embedded sequence retrieved for project ID; Define a random masking operation with a masking probability pmask=0.15, generate a Bernoulli distribution masking matrix with the same shape as the fused embedding matrix, and obtain the masking layer by performing partial dimensional masking on the fused embedding matrix using the Bernoulli distribution masking matrix through element-wise multiplication. Constructing the encoder: A two-layer fully connected network is built, with GELU as the activation function; the first layer expands the dimension of the occlusion layer by a factor of 2, and the second layer splits the output of the first layer into its mean. and variance ; through formula The sampling operation is defined to obtain the repetition parameter, where Ɛ is the standard normal distribution sample; Constructing the decoder: Build a single-layer fully connected network, obtain the encoder input Z and then decode it to obtain an output with the same embedding dimension as the input; Add residual connections and layer normalization at both ends of the encoder and decoder to output the denoised intent embedding. ; A dynamic gating network is constructed, consisting of two fully connected layers. The first layer takes the denoised intent embedding as input and obtains output features through ReLU activation. These output features are then substituted into the second layer, and the weight matrix is ​​obtained after softmax normalization, ensuring that the sum of the expert weights at each time step is 1. K independent feedforward neural networks are constructed, each sharing the input intent embedding. Expert output vectors are obtained through a single hidden layer and GELU activation. The expert output vectors are weighted and fused according to their corresponding expert weights to obtain the total expert output vector. A 1D causal convolutional network is built, with left zero padding. The total expert output vector is substituted into the 1D causal convolutional network to obtain the total expert causal vector, which is then calculated using the formula... Dynamic intent features are obtained by fusing the expert's total output vector and the expert's total causal vector. A single-layer fully connected network is constructed. The predicted similarity vector is obtained by inputting dynamic intent features and compared with the input intent similarity list. The mean squared error (MSE) loss function Lsim is defined to align the predicted similarity with the true similarity list in subsequent training. Training continues until the similarity between the aligned predicted similarity and the true similarity list is greater than a preset threshold. This threshold is determined by the technical personnel. The dynamic intent features at each time step are retained as intent semantic knowledge.

[0043] In one embodiment, transferring semantic knowledge of intent to an initial student model for training to obtain a targeted marketing model includes: The training similarity is obtained by calculating the similarity between the predicted product list of the student model and the predicted product list of the teacher model using cosine similarity. The target similarity is obtained by correcting the training similarity using a decay coefficient; Distillation loss is obtained by distilling semantic knowledge of intent based on target similarity. The Adam optimizer is used to iteratively optimize the parameters of the student model, and the optimization loss is calculated through forward propagation. The total loss is obtained by summing the optimization loss and the distillation loss; The model parameters are updated by backpropagation to minimize the total loss until the training termination condition is met, thus obtaining the target marketing model.

[0044] In one implementation, the attenuation coefficient is expressed as follows: Where λ is 0.01, the target similarity is the decay coefficient multiplied by the training similarity; the distillation loss is obtained by distilling the semantic knowledge of intent based on the target similarity, using the formula... ,in α represents the target similarity; α is the regularization coefficient, empirically set to 0.001. All learnable parameters for the student model; This represents the loss during distillation.

[0045] In one implementation, the predicted product list is compared with the actual interactive products in the previous baseline intent sequence, and the optimization loss is calculated using negative log-likelihood loss; the Adam optimizer has a base learning rate of 0.0001, a weight decay of 0.01, and a maximum norm of 1; the total loss is obtained by adding the optimization loss to the distillation loss coefficient and multiplying it by the global distillation loss, with the distillation loss coefficient being 0.03. In one implementation, the training termination condition is that the Top-K recommendation accuracy of the model is evaluated using an independent validation set after each training round. If the Top-10 accuracy of the validation set does not improve after 10 consecutive rounds, or if it decreases, the training is terminated.

[0046] Based on the same inventive concept, embodiments of the present invention also provide a product marketing recommendation system based on user intent. See also Figure 3 , Figure 3 A framework diagram of a product marketing recommendation system based on user intent provided in an embodiment of the present invention includes: The target time point determination module is used to acquire the effective historical interaction data of the target user in the historical time series, and split the interaction data according to the target time point to obtain the front-end baseline intent sequence and the back-end baseline intent sequence; the target time point is determined by the similarity between products in the effective historical interaction data; the effective historical interaction data before the target time point is used as the front-end baseline intent sequence; and the effective historical interaction data after the target time point is used as the back-end baseline intent sequence. The intent similarity list generation module is used to input the subsequent baseline intent sequence into the similarity calculation network to obtain the intent similarity list; The intent semantic knowledge generation module is used to build a teacher model. The intent similarity list and the subsequent baseline intent sequence are substituted into the teacher model for training to obtain intent semantic knowledge. The initial student model training module is used to build a student model. The initial student model is obtained by substituting the baseline intent sequence from the previous stage into the student model for training. The target marketing model training module is used to transfer semantic knowledge of intent to the initial student model for training to obtain the target marketing model. The target product recommendation list generation module is used to input the target users' valid historical interaction data into the target marketing model to obtain the target product recommendation list.

[0047] The product marketing recommendation system based on user intent provided by this invention innovatively divides historical interactions into a pre-benchmark and a post-benchmark sequence by accurately identifying the target time point of user intent mutation. It also utilizes a teacher-student model architecture to achieve effective transfer from post-benchmark semantic knowledge to the pre-benchmark model. This not only accurately captures the intent evolution trend and solves the problems of existing technologies being unable to adapt to intent mutations and lacking foresight, but also significantly improves the accuracy and personalization quality of recommendation results by deeply integrating semantic similarity and temporal dynamic features.

[0048] The foregoing has described one embodiment of the present invention in detail, but this content is merely a preferred embodiment and should not be considered as limiting the scope of the present invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the scope of the claims of this invention.

Claims

1. A product marketing recommendation method based on user intent, characterized in that, The method includes: The system acquires valid historical interaction data of the target user over a historical time series, and splits the interaction data according to the target time point to obtain a pre-term baseline intent sequence and a post-term baseline intent sequence; the target time point is determined by the similarity between products in the valid historical interaction data; the valid historical interaction data before the target time point is used as the pre-term baseline intent sequence; and the valid historical interaction data after the target time point is used as the post-term baseline intent sequence. The aforementioned baseline intent sequence is substituted into a similarity calculation network to obtain an intent similarity list; A teacher model is constructed, and the intent similarity list and the subsequent baseline intent sequence are substituted into the teacher model for training to obtain intent semantic knowledge; Construct a student model, and train the student model by substituting the aforementioned baseline intent sequence into the student model to obtain an initial student model; The intent semantic knowledge is transferred to the initial student model for training to obtain the target marketing model; The target marketing model is then used to input the target users' valid historical interaction data to obtain a target product recommendation list.

2. The product marketing recommendation method based on user intent according to claim 1, characterized in that, Determining the target time point includes: The target valid historical interaction data is obtained by data interception of the valid historical interaction data, and the target valid historical interaction data is converted into a time item ID associated sequence; Substitute the time item ID association sequence into the preset item embedding table to obtain the embedding vector corresponding to each item ID, and obtain all embedding vectors as the time-series intent feature sequence. The local difference score is obtained by calculating the average cosine distance between the intent features at time step t and the previous w time steps in the temporal intent feature sequence using a sliding window. Calculate the mean and standard deviation of the local difference scores for all time steps, and convert the local scores into standardized mutation scores based on the mean and standard deviation; The mutation score that retains the local maximum value after merging consecutive time steps with low scores is denoted as the local maximum mutation value; the low score refers to the standardized mutation score that is less than 1. The target time point is determined based on a preset significant mutation threshold and all local maximum mutation values.

3. The product marketing recommendation method based on user intent according to claim 1, characterized in that, Substituting the aforementioned baseline intent sequence into the similarity calculation network yields an intent similarity list, including: The first convolutional feature is obtained by sequentially performing 7×7 convolution and 3×3 pooling operations on the aforementioned baseline intent sequence. The first convolutional feature is successively substituted into the four residual layers to obtain the second convolutional feature; After performing average pooling on the convolutional features, the future intent attention features are extracted by the target attention module. Substituting the future intent attention features into the fully connected layer yields an intent similarity list; The working principle of the target attention module is as follows: Obtain the input features, and perform a 1×1 convolution operation on the input features to obtain the first attention features; The first attention is divided into N first channel sub-features along the channel dimension. Attention weights are generated for each first channel sub-feature along the Y-axis to obtain the first sub-attention weight. All first channel sub-features are weighted and fused according to all first sub-attention weights to obtain the intermediate attention feature. The intermediate attention feature is further divided into M second channel sub-features along the channel dimension. Attention weights are generated for each second channel sub-feature along the X-axis to obtain second sub-attention weights. All second channel sub-features are weighted and fused according to all second sub-attention weights to obtain future intention attention features.

4. The product marketing recommendation method based on user intent according to claim 1, characterized in that, The intention similarity list and the subsequent baseline intention sequence are substituted into the teacher model for training to obtain intention semantic knowledge, including: The back-end baseline intent sequence is mapped to a vector sequence through an item embedding table, and the position embedding is fused to generate a fused embedding matrix containing temporal information. A Bernoulli distribution masking matrix is ​​generated based on a preset masking probability, and then multiplied with the fusion embedding matrix to obtain the target embedding matrix; The target embedding matrix is ​​fed into the target encoder-decoder to obtain the intent embedding features; Substituting the intent embedding features into a preset dynamic gating network yields an expert weight matrix; Substituting the intent embedding features into K feedforward neural networks respectively, we obtain multiple expert output vectors; The total expert output vector is obtained by weighted summing of all expert output vectors based on the expert weight matrix. The expert's total output vector is substituted into a 1D causal convolutional network to obtain the expert's total causal vector. The expert's total output vector and the expert's total causal vector are then fused to obtain dynamic intent features. The dynamic intent features are then substituted into a fully connected network to obtain a predicted similarity vector; Based on the intent similarity list, it is compared with the predicted similarity vector, and a mean squared error loss function is defined for aligning the predicted similarity vector with the output predicted similarity list in subsequent training. The training is updated until the similarity between the intent similarity list and the output predicted similarity list is greater than a preset threshold, at which point the training is terminated, and the dynamic intent features updated each time are retained as intent semantic knowledge.

5. The product marketing recommendation method based on user intent according to claim 1, characterized in that, The process of transferring the semantic knowledge of intent to the initial student model for training to obtain the target marketing model includes: The training similarity is obtained by calculating the similarity between the predicted product list of the student model and the predicted product list of the teacher model using cosine similarity. The target similarity is obtained by correcting the training similarity using a decay coefficient; Distillation loss is obtained by distilling the semantic knowledge of intent based on the target similarity. The Adam optimizer is used to iteratively optimize the parameters of the student model, and the optimization loss is calculated through forward propagation. The total loss is obtained by summing the optimization loss and the distillation loss. The model parameters are updated by backpropagation to minimize the total loss until the training termination condition is met, thus obtaining the target marketing model.

6. A product marketing recommendation system based on user intent, characterized in that, The system includes: The target time point determination module is used to acquire valid historical interaction data of the target user in the historical time series, and split the interaction data according to the target time point to obtain a pre-stage baseline intent sequence and a post-stage baseline intent sequence; the target time point is determined by the similarity between products in the valid historical interaction data; the valid historical interaction data before the target time point is used as the pre-stage baseline intent sequence; and the valid historical interaction data after the target time point is used as the post-stage baseline intent sequence. An intent similarity list generation module is used to input the subsequent baseline intent sequence into a similarity calculation network to obtain an intent similarity list; The intent semantic knowledge generation module is used to construct a teacher model. The intent similarity list and the subsequent baseline intent sequence are substituted into the teacher model for training to obtain intent semantic knowledge. The initial student model training module is used to construct a student model by substituting the preceding baseline intent sequence into the student model for training to obtain the initial student model. The target marketing model training module is used to transfer the intent semantic knowledge to the initial student model for training to obtain the target marketing model. The target product recommendation list generation module is used to input the target user's valid historical interaction data into the target marketing model to obtain the target product recommendation list.

7. A product marketing recommendation system based on user intent according to claim 6, characterized in that, The target time point determination module includes: The time item ID association sequence generation module is used to extract target valid historical interaction data from the valid historical interaction data and convert the target valid historical interaction data into a time item ID association sequence. The temporal intent feature sequence determination module is used to substitute the time item ID association sequence into a preset item embedding table to obtain the embedding vector corresponding to each item ID, and obtain all embedding vectors as the temporal intent feature sequence. The local difference score calculation module is used to calculate the average cosine distance between the intent features at time step t and the previous w time steps in the temporal intent feature sequence using a sliding window to obtain the local difference score. The standardized mutation score determination module is used to calculate the mean and standard deviation of the local difference scores corresponding to all time steps, and convert the local scores into standardized mutation scores based on the mean and standard deviation. The local maximum mutation value determination module is used to merge consecutive low-score time steps and retain the mutation score with the local maximum value as the local maximum mutation value; the low score is the standardized mutation score with a standardized mutation score less than 1. The target time point generation module is used to determine the target time point based on a preset significant mutation threshold and all local maximum mutation values.

8. A product marketing recommendation system based on user intent according to claim 6, characterized in that, The intent similarity list generation module includes: The first convolutional feature determination module is used to perform 7×7 convolution and 3×3 pooling operations on the subsequent baseline intent sequence to obtain the first convolutional feature. The second convolutional feature determination module is used to sequentially substitute the first convolutional features into four residual layers to obtain the second convolutional features. The future intent attention feature determination module is used to perform average pooling on the two convolutional features and then input them into the target attention module to extract the future intent attention features; The intent similarity list determination module is used to input the future intent attention features into the fully connected layer to obtain the intent similarity list; The working principle of the target attention module is as follows: Obtain the input features, and perform a 1×1 convolution operation on the input features to obtain the first attention features; The first attention is divided into N first channel sub-features along the channel dimension. Attention weights are generated for each first channel sub-feature along the Y-axis to obtain the first sub-attention weight. All first channel sub-features are weighted and fused according to all first sub-attention weights to obtain the intermediate attention feature. The intermediate attention feature is further divided into M second channel sub-features along the channel dimension. Attention weights are generated for each second channel sub-feature along the X-axis to obtain second sub-attention weights. All second channel sub-features are weighted and fused according to all second sub-attention weights to obtain future intention attention features.

9. A product marketing recommendation system based on user intent according to claim 6, characterized in that, The intent semantic knowledge generation module includes: The fusion embedding matrix generation module is used to map the back-end baseline intent sequence into a vector sequence through an item embedding table, and fuse position embeddings to generate a fusion embedding matrix containing temporal information. The target embedding matrix determination module is used to generate a Bernoulli distribution masking matrix based on a preset masking probability, and then multiply it with the fused embedding matrix to obtain the target embedding matrix; The intent embedding feature generation module is used to input the target embedding matrix into the target encoder-decoder to obtain intent embedding features; The expert weight matrix determination module is used to substitute the intent embedding features into a preset dynamic gating network to obtain the expert weight matrix. An expert output vector determination module is used to substitute the intent embedding features into K feedforward neural networks to obtain multiple expert output vectors; The expert total output vector determination module is used to obtain the expert total output vector by weighted summation of all expert output vectors based on the expert weight matrix. The dynamic intent feature determination module is used to substitute the expert's total output vector into a 1D causal convolutional network to obtain the expert's total causal vector, and to fuse the expert's total output vector and the expert's total causal vector to obtain dynamic intent features; The prediction similarity vector determination module is used to input the dynamic intent features into a fully connected network to obtain the prediction similarity vector; The dynamic intent feature update module is used to compare the intent similarity list with the predicted similarity vector, and define a mean squared error loss function for aligning the predicted similarity vector with the output predicted similarity list in subsequent training. The training is updated until the similarity between the intent similarity list and the output predicted similarity list is greater than a preset threshold, at which point the training is terminated, and the dynamic intent features updated each time are retained as intent semantic knowledge.

10. A product marketing recommendation system based on user intent according to claim 6, characterized in that, The target marketing model training module includes: The training similarity calculation module is used to calculate the training similarity by using cosine similarity to measure the similarity between the predicted product list of the student model and the predicted product list of the teacher model. The target similarity determination module is used to correct the training similarity using a decay coefficient to obtain the target similarity; The distillation loss calculation module is used to distill the semantic knowledge of intent based on the target similarity to obtain the distillation loss. The optimization loss determination module is used to iteratively optimize the parameters of the student model using the Adam optimizer and calculate the optimization loss through forward propagation. The total loss determination module is used to sum the optimization loss and the distillation loss to obtain the total loss; The Minimize Total Loss Training Module is used to update the model parameters by backpropagating with the minimum total loss until the training termination condition is met to obtain the target marketing model.

Images