A convolutional network recommendation method, system, device and medium

Abstract

The present application relates to the technical field of machine learning, in particular to a convolution network recommendation method, system, device and medium; the method first converts original sample data into an original sample feature embedding matrix, secondly introduces a random mask to obtain local feature interaction and global feature interaction of the original sample feature embedding matrix, then combines the local feature interaction and the global feature interaction with the original sample feature embedding matrix to obtain a high-order feature interaction matrix, and finally constructs a recommendation model according to the high-order feature interaction matrix to obtain a prediction result, removes meaningless local feature interaction and global noise, reduces the influence of noise feature interaction, suppresses the generation of overfitting, and improves recommendation accuracy.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and more specifically, to a convolutional network recommendation method, system, device, and medium. Background Technology

[0002] Click-through rate (CTR) is a core task in computational advertising and recommendation systems. A typical recommendation system architecture is divided into two phases: recall and ranking. CTR prediction is used in the fine-grained ranking phase of the recommendation system, estimating the probability of a user clicking on an item. The key challenge in CTR prediction currently lies in how to acquire and select meaningful high-order interactions based on data features.

[0003] General linear models, such as LR and FTRL, although they perform well in practical use, lack the ability to learn feature interactions. To overcome this limitation, models based on Factorization Machine (FM) have been proposed. They vectorize features and represent the inner product of feature vectors as interactions between features, achieving quite good results. With the success of deep learning in computer vision and natural language processing, more and more deep learning models have begun to be combined with CTR prediction tasks, giving rise to many deep models, such as DeepFM, xDeepFM, DIN[5], etc. These models input the original features into deep neural networks (DNN) to learn explicit or implicit feature interactions.

[0004] The local connection mechanism of CNNs typically captures the local feature interactions between features of data samples. This mechanism assumes that there is spatial correlation between features. However, in CTR prediction, the order between feature vectors has no real meaning. For example, the feature sequence (A,B,C,D) and (C,D,A,B) correspond to the same semantics of the data sample, which is very different from the sequence of images and sentences. Furthermore, due to the local connection mechanism, CNNs capture many meaningless local feature interactions. Summary of the Invention

[0005] This invention addresses the problem that existing user click-through rate prediction methods lack the ability to learn feature interactions and capture many meaningless local feature interactions. It proposes a convolutional network recommendation method, system, device, and medium. By introducing a random mask, the local and global feature interactions of the original sample feature embedding matrix are obtained and combined to obtain a high-order feature interaction matrix. This removes meaningless local feature interactions and global noise, reduces the impact of noisy feature interactions, suppresses overfitting, and improves recommendation accuracy.

[0006] The specific implementation details of this invention are as follows:

[0007] A convolutional network recommendation method first converts the original sample data into an original sample feature embedding matrix. Then, a random mask is introduced to obtain the local and global feature interactions of the original sample feature embedding matrix. Next, the local and global feature interactions are combined with the original sample feature embedding matrix to obtain a higher-order feature interaction matrix. Finally, a recommendation model is constructed based on the higher-order feature interaction matrix to obtain the prediction result.

[0008] To better realize the present invention, the following steps are further included:

[0009] Step S1: Preprocess the original sample data and convert the preprocessed original sample data into an original sample feature embedding matrix;

[0010] Step S2: Construct a network block including convolutional layers, pooling layers and recombination layers, build a random mask layer in the network block, obtain the local feature interaction vector and global feature interaction vector of the original sample feature embedding matrix from the network block, and reorganize the dimensions of the local feature interaction vector and the global feature interaction vector to obtain a high-order feature interaction matrix;

[0011] Step S3: Combine the higher-order feature interaction matrix with the original sample feature embedding matrix to obtain a combined feature matrix, and use the combined feature matrix as input samples to construct a recommendation model;

[0012] Step S4: Input the sample data to be predicted into the recommendation model to obtain the prediction result, and recommend sample data according to the prediction result. The prediction result is the predicted probability value of whether the sample data to be predicted is clicked.

[0013] To better realize the present invention, step S1 further includes the following steps:

[0014] Step S11: Obtain the original sample data and calculate the feature values ​​of the original sample data;

[0015] Step S12: Filter out the original sample data whose feature values ​​are less than a set threshold, and divide the filtered original sample data into continuous feature data and category feature data;

[0016] Step S13: Call the discretization function to discretize the continuous feature data into categorical feature data and convert it into one-hot format.

[0017] Step S14: Call the embedding layer to compress the category feature data into low-dimensional vectors of the same length and convert them into feature embedding matrices.

[0018] To better realize the present invention, step S2 further includes the following steps:

[0019] Step S21: Construct a network block including convolutional layers, pooling layers and recombination layers, and build a random mask layer in the network block to obtain the objective function of the network block;

[0020] Step S22: Based on the spatial positional relationship of the original sample feature embedding matrix, capture the local interaction relationship of the original sample feature embedding matrix, and obtain the local feature interaction vector of the original sample feature embedding matrix from the convolutional layer and the pooling layer.

[0021] Furthermore, the specific operation of capturing the local interaction relationship of the original sample feature embedding matrix in step S22 is as follows: construct multiple convolution kernel matrices with the same kernel size to capture the local information of the original sample feature embedding matrix, and obtain the local interaction relationship based on the local information.

[0022] Step S23: Obtain the global feature interaction vector of the original sample feature embedding matrix from the recombination layer, and recombine the local feature interaction vector and the global feature interaction vector in terms of dimensions to obtain a higher-order feature interaction matrix.

[0023] To better implement the present invention, step S21 further includes the following steps:

[0024] Step S211: Construct a network block including convolutional layers, pooling layers, and recombination layers;

[0025] Step S212: Decompose the convolutional layer weights of the convolutional layer in low rank within a set period;

[0026] Step S213: Construct a random mask layer in the convolutional layer and the recombination layer, randomly select the parameter structure of the convolutional layer and the parameter structure of the recombination layer, and obtain the transformed convolutional layer weight parameters and recombination layer weight parameters.

[0027] Step S214: Obtain the objective function of the network block based on the transformed convolutional layer weight parameters and recombination layer weight parameters.

[0028] To better implement the present invention, the specific operation of step S212 is as follows: within a set period, the weights of the convolutional layer are decomposed according to the truncated singular value decomposition, and then the low rank is optimized according to the kernel norm regularization of the stochastic subgradient algorithm to remove redundant information in the original sample feature embedding matrix.

[0029] Based on the convolutional network recommendation method proposed above, in order to better realize the present invention, a convolutional network recommendation system is further proposed, including a feature embedding module, a feature convolution module, a classifier module, and a recommendation module;

[0030] The feature embedding module is used to preprocess the original sample data and convert the preprocessed original sample data into an original sample feature embedding matrix.

[0031] The feature convolution module is used to construct a network block including a convolutional layer, a pooling layer and a recombination layer, and to build a random mask layer in the network block. It also obtains the local feature interaction vector and the global feature interaction vector of the original sample feature embedding matrix from the network block, and recombines the dimensions of the local feature interaction vector and the global feature interaction vector to obtain a high-order feature interaction matrix.

[0032] The classifier module is used to combine the high-order feature interaction matrix with the original sample feature embedding matrix to obtain a combined feature matrix.

[0033] The recommendation module is used to construct a recommendation model by taking the combined feature matrix as input samples; inputting the sample data to be predicted into the recommendation model to obtain the prediction result; and recommending sample data based on the prediction result, wherein the prediction result is the predicted probability value of whether the sample data to be predicted is clicked.

[0034] Based on the convolutional network recommendation method proposed above, in order to better realize the present invention, an electronic device is further proposed, including a memory and a processor; the memory is used to store a computer program; when the processor executes the computer program, it implements the above-mentioned convolutional network recommendation method.

[0035] Based on the convolutional network recommendation method proposed above, in order to better realize the present invention, a computer-readable storage medium is further proposed, wherein computer instructions are stored on the computer-readable storage medium; when the computer instructions are executed on the above-mentioned electronic device, the above-mentioned convolutional network recommendation method is implemented.

[0036] The present invention has the following beneficial effects:

[0037] (1) The convolutional network recommendation method proposed in this invention accurately eliminates meaningless feature interactions generated by the feature generation module, reduces the impact of noisy feature interactions, and suppresses overfitting.

[0038] (2) This invention enables the autonomous selection of valuable feature interactions by introducing a random binary mask.

[0039] (3) The present invention implements gradient estimation of discrete parameters based on ARM algorithm, which ensures normal training and prediction of system model. By utilizing different advanced network models and combining with multiple advanced models, the prediction accuracy is significantly improved. Attached Figure Description

[0040] Figure 1 This is a flowchart of the collaborative process for the convolutional network recommendation method provided by the present invention.

[0041] Figure 2 This is a schematic diagram of the feature embedding module.

[0042] Figure 3 This is a schematic diagram of the feature convolution module.

[0043] Figure 4 This is a schematic diagram of the low-rank decomposition module of convolution.

[0044] Figure 5 This is a schematic diagram of the classifier module. Detailed Implementation

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments, and therefore should not be regarded as a limitation on the scope of protection. 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.

[0046] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set up," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0047] Example 1:

[0048] This embodiment proposes a convolutional network recommendation method, which specifically includes the following steps.

[0049] Step S1: Preprocess the original sample data and convert the preprocessed original sample data into an original sample feature embedding matrix.

[0050] Furthermore, step S1 specifically includes the following steps:

[0051] Step S11: Obtain the original sample data and calculate the feature values ​​of the original sample data;

[0052] Step S12: Filter out the original sample data whose feature values ​​are less than a set threshold, and divide the filtered original sample data into continuous feature data and category feature data;

[0053] Step S13: Call the discretization function to discretize the continuous feature data into categorical feature data and convert it into one-hot format.

[0054] Step S14: Call the embedding layer to compress the category feature data into low-dimensional vectors of the same length and convert them into feature embedding matrices.

[0055] Step S2: Construct a network block including convolutional layers, pooling layers, and recombination layers, and build a random mask layer in the network block. Obtain the local feature interaction vector and global feature interaction vector of the original sample feature embedding matrix from the network block, and reorganize the dimensions of the local feature interaction vector and the global feature interaction vector to obtain a high-order feature interaction matrix.

[0056] Furthermore, step S2 specifically includes the following steps:

[0057] Step S21: Construct a network block including convolutional layers, pooling layers and recombination layers, and build a random mask layer in the network block to obtain the objective function of the network block.

[0058] Furthermore, step S21 specifically includes the following steps:

[0059] Step S211: Construct a network block including convolutional layers, pooling layers, and recombination layers.

[0060] Step S212: Decompose the convolutional layer weights of the convolutional layer in low rank within a set period.

[0061] Further, the specific operation of step S212 is as follows: within a set period, the weights of the convolutional layer are decomposed according to the truncated singular value decomposition, and then the low rank is optimized according to the nuclear norm regularization of the stochastic subgradient algorithm to remove redundant information in the original sample feature embedding matrix.

[0062] Step S213: Construct a random mask layer in the convolutional layer and the recombination layer, randomly select the parameter structure of the convolutional layer and the parameter structure of the recombination layer, and obtain the transformed convolutional layer weight parameters and recombination layer weight parameters.

[0063] Step S214: Obtain the objective function of the network block based on the transformed convolutional layer weight parameters and recombination layer weight parameters.

[0064] Step S22: Based on the spatial positional relationship of the original sample feature embedding matrix, capture the local interaction relationship of the original sample feature embedding matrix, and obtain the local feature interaction vector of the original sample feature embedding matrix from the convolutional layer and the pooling layer.

[0065] Furthermore, the specific operation of capturing the local interaction relationship of the original sample feature embedding matrix in step S22 is as follows: construct multiple convolution kernel matrices with the same kernel size to capture the local information of the original sample feature embedding matrix, and obtain the local interaction relationship based on the local information;

[0066] Step S23: Obtain the global feature interaction vector of the original sample feature embedding matrix from the recombination layer, and recombine the local feature interaction vector and the global feature interaction vector in terms of dimensions to obtain a higher-order feature interaction matrix.

[0067] Step S3: Combine the higher-order feature interaction matrix with the original sample feature embedding matrix to obtain a combined feature matrix, and use the combined feature matrix as input samples to construct a recommendation model;

[0068] Step S4: Input the sample data to be predicted into the recommendation model to obtain the prediction result, and recommend sample data according to the prediction result. The prediction result is the predicted probability value of whether the sample data to be predicted is clicked.

[0069] Working principle: This embodiment first converts the original sample data into an original sample feature embedding matrix. Then, a random mask is introduced to obtain the local and global feature interactions of the original sample feature embedding matrix. Next, the local and global feature interactions are combined with the original sample feature embedding matrix to obtain a higher-order feature interaction matrix. Finally, a recommendation model is constructed based on the higher-order feature interaction matrix to obtain the prediction result. By introducing a random mask to obtain the local and global feature interactions of the original sample feature embedding matrix and combining them to obtain the higher-order feature interaction matrix, meaningless local feature interactions and global noise are removed, reducing the influence of noisy feature interactions, suppressing overfitting, and improving recommendation accuracy.

[0070] Example 2:

[0071] This embodiment is based on the above embodiment 1, such as... Figure 1 The proposed system is a convolutional network recommendation system, comprising a feature embedding module, a feature convolution module, a classifier module, and a recommendation module; wherein the feature convolution module includes a trainable ARM mask module and a low-rank convolution decomposition module.

[0072] like Figure 1As shown, this embodiment uses MLP to obtain global feature interactions, which are combined with CNN to achieve a complementary effect. Furthermore, a trainable random binary mask is introduced and multiplied element-wise with the convolutional feature interactions to remove meaningless local feature interactions and global noise, thereby obtaining meaningful high-order feature interactions and improving recommendation accuracy.

[0073] The recommendation system proposed in this embodiment includes the following modules:

[0074] like Figure 2 As shown, the feature embedding module is used to extract advertisement or product samples and their corresponding click tags from the database, and filter out feature domains with feature values ​​less than a set threshold. The features of the input samples are divided into continuous features and categorical features. Considering the information redundancy of continuous features, they are discretized into categorical features. The specific discretization function is as follows:

[0075]

[0076] After discretization, the data is uniformly converted to one-hot encoding, and then compressed into a low-dimensional vector of the same length using an embedding layer, i.e., feature embedding. Next, the categorical features are transformed into feature embeddings; specifically, for each sample, each field i (1≤i≤n)... f Each of these will be represented by a vector e. i ∈R 1×k Let n represent, where n f Where k is the number of fields and k is the length of the feature embedding, therefore, each sample can be transformed into a feature embedding matrix. In addition, to prevent inconsistencies in gradient directions when updating parameters, the feature generation module makes a complete and independent copy of the feature embedding matrix from the deep classifier module.

[0077] like Figure 3 As shown, the feature convolution module is used to capture local feature interactions using CNN convolution modules and global feature interactions using MLP. It acquires feature interactions at different levels by constructing a multi-layered network and then combines them with the original features. It combines convolutional layers, pooling layers, and recombination layers into a single network block, flexibly constructing multiple network blocks to capture novel feature interactions at different levels. The convolutional layer captures the local interaction relationships between features based on the spatial positional relationships of the input sample feature matrix. Multiple convolutional kernels of the same size capture different local information of the matrix respectively. The convolutional kernel matrix is ​​composed of... To represent. The calculation formula is:

[0078]

[0079]

[0080] in, Let represent the i-th feature map of the first convolutional layer, and p and q represent the row and column indices of that feature map.

[0081] After the convolutional layers, max pooling layers are used to filter out the more important feature information. The output formula of the pooling layer of the first network block is as follows:

[0082]

[0083] in, This indicates the height of the pooling window of the first pooling layer. The output of this pooling layer then goes to two places: first, it serves as input to the convolutional layer of the next network block to continue capturing feature interactions; second, it serves as input to the fully connected layer of the current network block to generate entirely new feature interactions.

[0084] After the pooling layer, S 1 In fact, local feature interactions have already been obtained. Next, a reorganization layer is needed to obtain global feature interactions. The actual operation involves S... 1 The output feature interaction vector is completely flattened through a flattening layer, then passed through a fully connected layer, and finally its dimensions are reorganized to obtain a brand-new feature interaction matrix R. 1 Then all generated feature interactions can be represented as:

[0085]

[0086] After concatenating the original features along their feature dimensions, the input to the deep classifier is obtained:

[0087] E = (E, R)

[0088] A trainable ARM masking module is used to filter network parameters by adding random masking layers to each layer of the network combined with L0 regularization, and uses the ARM gradient estimation algorithm to ensure gradient propagation during the training process of the masking layer parameters. Specifically, it adds a random masking layer Z after the model's convolutional layers and reconstruction layers, thus affecting the parameter structure W of the convolutional and reconstruction layers. c W r Make a random selection and transform it into:

[0089]

[0090]

[0091] Here, ⊙ represents element-wise multiplication, and the matrix Z∈{0,1} is a binary matrix, indicating whether to retain the elements of the corresponding weight matrix. These are the parameters of the transformed convolutional and recombined layers. Clearly, this method generates a different network structure for each iteration. L0 regularization is added to the objective function to restrict non-zero elements in matrix Z. The model's objective function is:

[0092]

[0093] Where W represents the weight parameters that require an additional random mask in the model, and λ controls the regularization effect factor of the model, I [c] `c` is an indicator function that outputs 1 if condition `c` is met, and 0 otherwise. `y` is the label. These are model predictions.

[0094] First, since the mask Z is used to select whether to retain parameter values, we assume that each element z of the mask Z in the model follows a Bernoulli distribution, i.e., z ~ Bernoulli(z; π), where the parameter π ∈ [0, 1].

[0095] Considering that the objective function of this model involves a probability distribution, stochastic variational optimization is used.

[12] Based on relevant knowledge, the following inequalities hold:

[0096]

[0097] Where F is an arbitrary function, and p(z) is the distribution function that the random variable z follows. It can be seen that an upper bound on the minimum value of a function is its expected value. Therefore, we can use this theory to replace the function with its expected value, and rewrite the objective function formula (16) as follows:

[0098]

[0099] It is known that the second term of the objective function is differentiable with respect to the parameter Π, allowing for gradient backpropagation of Π. However, the first term, involving probability distribution sampling, remains non-differentiable. To introduce a novel unbiased gradient estimation algorithm, we need to reparameterize the distribution model parameters in the formula.

[13] Prepare a function to convert parameter Π into a new parameter Φ, as shown in the following formula:

[0100] Π=g(Φ)

[0101]

[0102] By introducing the ARM gradient estimation algorithm and utilizing the reparameterization technique, sampling from a Bernoulli distribution is transformed into sampling from a uniform distribution with fixed probability parameters, thus realizing gradient estimation for bivariate random variables and ensuring correct gradient backpropagation. The theorem is as follows:

[0103] Theorem 1: ARM. For a vector containing V binary random variables, define the expected function with respect to z.

[0104]

[0105] The gradient of the variable phi can be expressed as:

[0106]

[0107] in, To find the gradient, we have... and

[0108] like Figure 4 As shown, the convolutional low-rank decomposition module performs low-rank decomposition on the convolutional layers. Every specified period, it applies truncated singular value decomposition to the convolutional layer weights, performing weight replacement. Simultaneously, it uses kernel norm regularization based on the stochastic subgradient descent algorithm to further enhance the low-rank property. Every specified period m, it performs truncated singular value decomposition on the convolutional graph parameters and uses kernel norm regularization based on the stochastic subgradient descent algorithm to further improve the low-rank property, removing redundant information from the acquired features. The specific formula is as follows:

[0109]

[0110]

[0111] Where D(·) is a low-rank tensor estimator, f(·) refers to the model, α is the learning rate, t represents the number of iterations, m is the iteration period of the low-rank decomposition interval, and W t These are model parameters with period t, T z These are the model parameters after the periodic z-low rank decomposition, σ i U represents the singular values ​​of the model parameter matrix. i With V i It is the singular value vector after decomposition.

[0112] like Figure 5 As shown, the classifier module combines the novel high-order feature interactions generated by the aforementioned modules with the original features. It further enhances performance by incorporating various advanced recommendation models, exhibiting strong scalability and robustness. Specifically, it concatenates the high-order feature interactions generated by the aforementioned modules with the original features, inputs this concatenation into the recommendation model, and predicts whether the user will click on the sample.

[0113] Recommendation module and database: Take the product samples to be predicted from the database, remove the products that the user has already purchased, randomly select a certain number of products, use the model to predict the click-through rate, take the samples with a click-through rate of true, sort them using a ranking algorithm, and recommend the top K products to the user.

[0114] In summary, the main steps and procedures can be briefly described as follows:

[0115] 1. The feature embedding module retrieves raw data from the database, performs preprocessing, and converts it into a feature embedding matrix;

[0116] 2. Since the trainable ARM mask module and the low-rank convolution decomposition module are embedded in the feature convolution module during instance execution, after the feature embedding matrix is ​​input into the feature convolution module, the low-rank convolution decomposition module first performs low-rank decomposition on the model features, and then the ARM mask module is used to perform trainable filtering on the generated convolution map to obtain valuable high-order feature interactions.

[0117] 3. Next, the high-order generated features output by the feature convolution module are combined with the original features and input into the advanced recommendation module for prediction;

[0118] 4. Finally, the prediction results, i.e. the probability of whether the sample is clicked, are output and stored in the database module.

[0119] The other parts of this embodiment are the same as those in Embodiment 1 above, so they will not be described again.

[0120] Example 3:

[0121] Based on any one of Embodiments 1-2 above, this embodiment proposes an electronic device, including a memory and a processor; the memory is used to store a computer program; when the processor executes the computer program, it implements the above-described convolutional network recommendation method.

[0122] This embodiment also proposes a computer-readable storage medium storing computer instructions; when the computer instructions are executed on the aforementioned electronic device, the aforementioned convolutional network recommendation method is implemented.

[0123] The other parts of this embodiment are the same as any one of the above embodiments 1-2, so they will not be described again.

[0124] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A convolutional network recommendation method, characterized in that, Specifically, the following steps are included: Step S1: Preprocess the original sample data and convert the preprocessed original sample data into an original sample feature embedding matrix; the original sample data is an advertisement or product sample extracted from the database; Step S2: Construct a network block including convolutional layers, pooling layers and recombination layers, build a random mask layer in the network block, obtain the local feature interaction vector and global feature interaction vector of the original sample feature embedding matrix from the network block, and reorganize the dimensions of the local feature interaction vector and the global feature interaction vector to obtain a high-order feature interaction matrix; Step S3: Combine the higher-order feature interaction matrix with the original sample feature embedding matrix to obtain a combined feature matrix, and use the combined feature matrix as input samples to construct a recommendation model; Step S4: Input the sample data to be predicted into the recommendation model to obtain the prediction result, which is the predicted probability value of whether the sample data to be predicted is clicked; Step S1 specifically includes the following steps: Step S11: Obtain the original sample data and calculate the feature values ​​of the original sample data; Step S12: Filter out the original sample data whose feature values ​​are less than a set threshold, and divide the filtered original sample data into continuous feature data and category feature data; Step S13: Call the discretization function to discretize the continuous feature data into categorical feature data and convert it into one-hot format; Step S14: Call the embedding layer to compress the category feature data into low-dimensional vectors of the same length, and convert them into feature embedding matrices; Step S2 specifically includes the following steps: Step S21: Construct a network block including convolutional layers, pooling layers and recombination layers, and build a random mask layer in the network block to obtain the objective function of the network block; Step S22: Based on the spatial positional relationship of the original sample feature embedding matrix, capture the local interaction relationship of the original sample feature embedding matrix, and obtain the local feature interaction vector of the original sample feature embedding matrix from the convolutional layer and the pooling layer; Step S23: Obtain the global feature interaction vector of the original sample feature embedding matrix from the recombination layer, and recombine the local feature interaction vector and the global feature interaction vector in terms of dimensions to obtain a higher-order feature interaction matrix.

2. The convolutional network recommendation method according to claim 1, characterized in that, Step S21 specifically includes the following steps: Step S211: Construct a network block including convolutional layers, pooling layers, and recombination layers; Step S212: Within a set period, construct a low-rank decomposition model and decompose the convolutional layer weights of the convolutional layer in a low-rank manner. Step S213: Construct a random mask layer in the convolutional layer and the recombination layer, randomly select the parameter structure of the convolutional layer and the parameter structure of the recombination layer, and obtain the transformed convolutional layer weight parameters and recombination layer weight parameters. Step S214: Obtain the objective function of the network block based on the transformed convolutional layer weight parameters and recombination layer weight parameters.

3. The convolutional network recommendation method according to claim 2, characterized in that, The specific operation of step S212 is as follows: within a set period, a low-rank decomposition model is constructed, the weights of the convolutional layer are decomposed according to the truncated singular value decomposition of the low-rank decomposition model, and then the low-rank property is optimized according to the kernel norm regularization of the stochastic subgradient algorithm to remove redundant information in the original sample feature embedding matrix.

4. The convolutional network recommendation method according to claim 2, characterized in that, The specific operation of capturing the local interaction relationship of the original sample feature embedding matrix in step S22 is as follows: construct multiple convolution kernel matrices with the same kernel size to capture the local information of the original sample feature embedding matrix, and obtain the local interaction relationship based on the local information.

5. A convolutional network recommendation system, characterized in that, It includes a feature embedding module, a feature convolution module, a classifier module, and a recommendation module; The feature embedding module is used to preprocess the original sample data and convert the preprocessed original sample data into an original sample feature embedding matrix; the original sample data is an advertisement or product sample extracted from the database. The feature convolution module is used to construct a network block including a convolutional layer, a pooling layer and a recombination layer, and to build a random mask layer in the network block. It also obtains the local feature interaction vector and the global feature interaction vector of the original sample feature embedding matrix from the network block, and recombines the dimensions of the local feature interaction vector and the global feature interaction vector to obtain a high-order feature interaction matrix. The classifier module is used to combine the high-order feature interaction matrix with the original sample feature embedding matrix to obtain a combined feature matrix. The recommendation module is used to construct a recommendation model by taking the combined feature matrix as input samples. The sample data to be predicted is input into the recommendation model to obtain the prediction result, which is the predicted probability value of whether the sample data to be predicted is clicked. The feature embedding module is specifically used to acquire original sample data and calculate the feature values ​​of the original sample data; filter out the original sample data whose feature values ​​are less than a set threshold, and divide the filtered original sample data into continuous feature data and category feature data; call the discretization function to discretize the continuous feature data into category feature data and convert it into one-hot format; call the embedding layer to compress the category feature data into low-dimensional vectors of the same length and convert it into a feature embedding matrix; The feature convolution module is specifically used to construct a network block including convolutional layers, pooling layers, and recombination layers, and to build a random mask layer in the network block to obtain the objective function of the network block; according to the spatial position relationship of the original sample feature embedding matrix, it captures the local interaction relationship of the original sample feature embedding matrix, and obtains the local feature interaction vector of the original sample feature embedding matrix from the convolutional layer and the pooling layer; it obtains the global feature interaction vector of the original sample feature embedding matrix from the recombination layer, and recombines the local feature interaction vector and the global feature interaction vector in terms of dimensions to obtain a high-order feature interaction matrix.

6. An electronic device, characterized in that, It includes a memory and a processor; the memory is used to store a computer program; when the processor executes the computer program, it implements the convolutional network recommendation method as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions; when the computer instructions are executed on the electronic device as described in claim 6, they implement the convolutional network recommendation method as described in any one of claims 1-4.

Images