Intelligent recommendation method and system based on deep learning

By calculating gray correlation vectors and regularization processing, data quality is improved, combined with principal component analysis and fine pruning technology, the problems of too many data dimensions and too many model neurons are solved, and the accuracy and stability of the intelligent recommendation model are improved.

CN118227877BActive Publication Date: 2024-11-08GUANGZHOU NIRVANA TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410305942.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-18
Publication Date
2024-11-08
Estimated Expiration
2044-03-18

Smart Images

  • Figure CN118227877B_ABST
    Figure CN118227877B_ABST
Patent Text Reader

Abstract

The present invention discloses an intelligent recommendation method and system based on deep learning, and the method includes data collection, data preprocessing, data set division, establishment of intelligent recommendation model and intelligent recommendation. The present invention belongs to the field of data processing technology, and specifically refers to an intelligent recommendation method and system based on deep learning. The scheme reduces data redundancy by calculating grey correlation vectors, improves data quality by regularization processing, extracts key information by weight, reduces complexity by principal component analysis, and improves model training efficiency; based on the design of fine pruning and evolutionary components, effectively reduces the connections in the model network, improves model operation efficiency and generalization ability; based on the design of mutation and crossover strategies, improves model diversity, and introduces anomaly detection components to improve model robustness.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and specifically to an intelligent recommendation method and system based on deep learning. Background Art

[0002] The intelligent recommendation method based on deep learning is a method that combines deep learning technology and recommendation systems. It aims to model user behavior and interests through deep learning models and use these models to provide personalized recommendation services. However, general recommendation methods have problems such as poor quality of original data, inability to extract key information, and too many data dimensions, which leads to inaccurate recommendation models. General recommendation methods have problems such as too many model neurons, resulting in weak model generalization ability, and neurons that have a negative impact on model performance, which affects model stability. Summary of the invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides an intelligent recommendation method and system based on deep learning. In view of the problems that general recommendation methods have poor original data quality, cannot extract key information, and have too many data dimensions, resulting in inaccurate recommendation models, this solution reduces data redundancy by calculating the grey correlation vector, improves data quality by regularization processing, extracts key information by weight, reduces data dimensionality processing by principal component analysis, and improves model training efficiency; in view of the problems that general recommendation methods have too many model neurons, resulting in weak model generalization ability, and there are neurons that have a negative impact on model performance, affecting model stability, this solution is based on the design of fine pruning and evolutionary components to effectively reduce the connections in the model network and improve the model operation efficiency and generalization ability; based on the design of mutation and crossover strategies to improve model diversity, anomaly detection components are introduced to improve model robustness.

[0004] The technical solution adopted by the present invention is as follows: The intelligent recommendation method based on deep learning provided by the present invention comprises the following steps:

[0005] Step S1: data collection;

[0006] Step S2: data preprocessing;

[0007] Step S3: data set division;

[0008] Step S4: Establishing an intelligent recommendation model;

[0009] Step S5: Intelligent recommendation.

[0010] Furthermore, in step S1, the data collection is to collect user data, item data, interaction data and user behavior feedback level; the user data includes basic information of the user and user preference data; the item data includes basic information of the item and feature data of the item; the interaction data is the user's click, collection and purchase behavior data on the item; the user behavior feedback level is the user's preference level for the recommendation results; and the user behavior feedback level is used as a data label.

[0011] Furthermore, in step S2, the data preprocessing is to perform data cleaning, data conversion and standardization on the collected data; the data cleaning is to perform missing value processing, outlier processing and duplicate value processing on the collected data; the data conversion is to convert the cleaned data into a vector form; the standardization is to perform standardization on the converted data based on Z-Score standardization.

[0012] Furthermore, in step S3, the data set division specifically includes the following steps:

[0013] Step S31: Calculate the grey relational vector, and calculate the grey relational vector R=(r(1), r(2), ..., r(m)) based on the preprocessed data; wherein r(1), r(2) and r(m) are the grey relational degrees of the first variable, the second variable and the mth variable respectively;

[0014] Step S32: Regularization processing: the data sample is subjected to regularization processing, and the formula used is as follows:

[0015]

[0016]

[0017]

[0018] In the formula, x ij is the value of the jth feature of the ith sample; norm(·) is the regularization process; x i1 、x i2 and x in are the values ​​of the first feature, the second feature, and the nth feature of the i-th sample; μ(x i ) and sig(x i ) are the mean and standard deviation of all features of the i-th sample respectively; is the value of the i-th sample after regularization; x * is the matrix composed of all samples after regularization; and are the values ​​of the first sample, the second sample, and the mth sample after regularization processing;

[0019] Step S33: weighting, weighting the regularized data samples and the grey relational vector to obtain a weighted matrix G. The formula used is as follows:

[0020]

[0021] Step S34: Calculate the covariance matrix, calculate the covariance matrix cov(G) of the weighted matrix G, and extract the eigenvalues and the corresponding eigenvector

[0022] Step S35: Sorting, sorting the eigenvalues ​​in descending order to obtain the corresponding sorted eigenvectors Select the first t features to retain and get the transformation matrix

[0023] Step S36: Linear change, using the transformation matrix W * Perform linear transformation on the original data to implement principal component analysis based on grey correlation and obtain the new data set Y * , the formula used is as follows:

[0024] Y * =x * ·W * ;

[0025] Step S37: Data set division: divide the new data set into a training set and a test set. The training set is used to train the model; the test set is used to test the model performance.

[0026] Furthermore, in step S4, the establishment of the intelligent recommendation model specifically includes the following steps:

[0027] Step S41: Model architecture design. The established deep learning model uses the transfer learning paradigm to import the weights of the convolution stage. The model transfers the extracted features through the network. These features correspond to the neurons evolved by the evolutionary components of the model. The network output of the model is a vector containing 2048 features, which serves as the bottom layer input of the neural network. The bottom layer of the neural network consists of an input layer that receives 2048 feature vectors, a hidden layer consisting of 512 neurons, and an output layer that defines the number of label categories.

[0028] Step S42: Pruning design, the model network has fewer connections than the standard fully connected layer, where each neuron is connected to the next group of neurons, called a sparse layer; the genetic algorithm is responsible for finding the best pruning pattern for the sparse layer, and the chromosome representing the pruning pattern is decoded into an adjacency matrix; as a result of the pruning operation, the matrix and the layer itself are sparse;

[0029] Step S43: Evolutionary component design, the network population is evolved using a genetic algorithm, and a binary coding strategy is used to indicate whether a neuron is active; if the gene value of a neuron is 1, it is active; if it is 0, it is inactive; based on this direct coding method, each gene uniquely determines a neuron in the decoding network; the initialization of the chromosome corresponds to random discrete initialization in the range of [0, 1], and the binary tournament selection method is selected to initialize the chromosome; the chromosome is used to represent the connection of the neuron; including:

[0030] Step S431: Initialization, using the binary tournament selection method to initialize the chromosome; the performance of the model containing neurons based on the chromosome representation is used as the chromosome fitness value;

[0031] Step S432: Design the crossover probability, the formula used is as follows:

[0032] Jc=baseJc·(1-τ) tc ;

[0033] Where Jc is the crossover probability; baseJc is the base crossover probability; τ is the decay factor; tc is the current number of iterations;

[0034] Step S433: Design a crossover strategy, the operator defines two new individuals from two parents; given these two parents p and q, the formula used to generate offspring is as follows:

[0035]

[0036]

[0037] Where p' i and q' i are two newly generated offspring individuals; p i and q i are the i-th gene values ​​of parents p and q respectively; r is a random number between 0 and 1;

[0038] Step S434: Design a mutation strategy, the formula used is as follows:

[0039]

[0040] In the formula, δ m is the mutation value; c c is the current chromosome; c to is the optimal chromosome in the population; l is the parameter that controls the mutation range; f c and f to are the fitness values ​​of the current chromosome and the best chromosome in the population, respectively;

[0041] Step S435: Anomaly detection component, with an anomaly threshold pre-set, uses softmax with temperature scaling parameters as the anomaly score of each sample. If the score of the instance is lower than the defined threshold, it will output 1 and perform a pruning operation; otherwise, it will output 0; it is expressed as follows:

[0042]

[0043] In the formula, f ON (x; T) is the abnormal score of the neuron corresponding to the chromosome, x is the input data of the neuron corresponding to the chromosome; T is the temperature parameter that controls the degree of scaling; N is the total number of categories, j is the category index; h i (x) and h j (x) are the unscaled probabilities that the model predicts the input to be of class i and class j, respectively;

[0044] Step S436: Pruning determination, the model uses the cross entropy loss function, and a loss threshold is set in advance. When the loss function value of the model for the test set is lower than the loss threshold, the model training is completed; if the maximum number of iterations is reached, the population is reinitialized for pruning; otherwise, iterative pruning continues.

[0045] Furthermore, in step S5, the intelligent recommendation is based on a trained deep learning model, which collects user data, item data and interaction data in real time, and sets a feedback level threshold in advance. If the model's predicted user behavior feedback level for the recommended item is higher than the feedback level threshold, it is considered that the user is satisfied with the recommended item, and the recommendation is then made.

[0046] The deep learning-based intelligent recommendation system provided by the present invention includes a data acquisition module, a data preprocessing module, a data set division module, an intelligent recommendation model establishment module and an intelligent recommendation module;

[0047] The data collection module collects user data, item data, interaction data and user behavior feedback level, and sends the data to the data preprocessing module;

[0048] The data preprocessing module performs data cleaning, data conversion and standardization on the collected data, and sends the data to the data set division module;

[0049] The data set partitioning module improves data quality by calculating grey correlation vectors and regularization processing, extracts key information by weight, and performs data dimension reduction processing by principal component analysis; and sends the data to the intelligent recommendation model establishment module;

[0050] The intelligent recommendation model building module designs fine pruning and evolution components; designs mutation and crossover strategies, introduces anomaly detection components to implement the establishment of the intelligent recommendation model, and sends data to the intelligent recommendation module;

[0051] The intelligent recommendation module implements intelligent recommendations for users based on the trained deep learning model.

[0052] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0053] (1) In view of the problems of poor quality of original data, inability to extract key information, and excessive data dimensions leading to inaccurate recommendation models in general recommendation methods, this solution reduces data redundancy by calculating the grey correlation vector, improves data quality by regularization, extracts key information by weight, and reduces complexity by principal component analysis to reduce data dimensionality, thereby improving model training efficiency.

[0054] (2) In response to the problems of too many model neurons in general recommendation methods, which lead to weak model generalization ability, the existence of neurons that have a negative impact on model performance, and the impact on model stability, this solution is based on the design of fine pruning and evolutionary components to effectively reduce the connections in the model network and improve the model operation efficiency and generalization ability; based on the design of mutation and crossover strategies to improve model diversity, and introduce anomaly detection components to improve model robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A schematic diagram of the process of the intelligent recommendation method based on deep learning provided by the present invention;

[0056] Figure 2 A schematic diagram of an intelligent recommendation system based on deep learning provided by the present invention;

[0057] Figure 3 is a schematic flow chart of step S3;

[0058] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0060] In the description of the present invention, it is necessary to understand that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred system or element must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as a limitation on the present invention.

[0061] Example 1, see Figure 1 The present invention provides an intelligent recommendation method based on deep learning, which comprises the following steps:

[0062] Step S1: Data collection, collecting user data, item data, interaction data and user behavior feedback level;

[0063] Step S2: Data preprocessing, data cleaning, data conversion and standardization of the collected data;

[0064] Step S3: Data set division, improving data quality by calculating grey correlation vector and regularization, extracting key information by weight, and reducing data dimension by principal component analysis;

[0065] Step S4: Establish an intelligent recommendation model, design fine pruning and evolution components; design mutation and crossover strategies, and introduce anomaly detection components to achieve the establishment of an intelligent recommendation model;

[0066] Step S5: Intelligent recommendation.

[0067] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the user data includes the user's basic information and the user's preference data; the item data includes the item's basic information and the item's feature data; the interaction data is the user's click, collection and purchase behavior data for the item; the user behavior feedback level is the user's preference level for the recommendation result; and the user behavior feedback level is used as a data label.

[0068] Example 3, see Figure 1 This embodiment is based on the above embodiment. In step S2, the data cleaning is to process missing values, outliers and duplicate values ​​of the collected data; the data conversion is to convert the cleaned data into a vector form; and the standardization is to standardize the converted data based on Z-Score standardization.

[0069] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S3, the data set division specifically includes the following steps:

[0070] Step S31: Calculate the grey relational vector, and calculate the grey relational vector R=(r(1), r(2), ..., r(m)) based on the preprocessed data; wherein r(1), r(2) and r(m) are the grey relational degrees of the first variable, the second variable and the mth variable respectively;

[0071] Step S32: Regularization processing: the data sample is subjected to regularization processing, and the formula used is as follows:

[0072]

[0073]

[0074]

[0075] In the formula, x ij is the value of the jth feature of the ith sample; norm(·) is the regularization process; x i1 、x i2 and x in are the values ​​of the first feature, the second feature, and the nth feature of the i-th sample; μ(x i ) and sig(x i ) are the mean and standard deviation of all features of the i-th sample respectively; is the value of the i-th sample after regularization; x * is the matrix composed of all samples after regularization; and are the values ​​of the first sample, the second sample, and the mth sample after regularization processing;

[0076] Step S33: weighting, weighting the regularized data samples and the grey relational vector to obtain a weighted matrix G. The formula used is as follows:

[0077]

[0078] Step S34: Calculate the covariance matrix, calculate the covariance matrix cov(G) of the weighted matrix G, and extract the eigenvalues and the corresponding eigenvector

[0079] Step S35: Sorting, sorting the eigenvalues ​​in descending order to obtain the corresponding sorted eigenvectors Select the first t features to retain and get the transformation matrix

[0080] Step S36: Linear change, using the transformation matrix W *Perform linear transformation on the original data to implement principal component analysis based on grey correlation and obtain the new data set Y * , the formula used is as follows:

[0081] Y * =x * ·W * ;

[0082] Step S37: Data set division: divide the new data set into a training set and a test set. The training set is used to train the model; the test set is used to test the model performance.

[0083] By performing the above operations, the general recommendation methods have the problems of poor original data quality, inability to extract key information, and too many data dimensions leading to inaccurate recommendation models. This solution reduces data redundancy by calculating the grey correlation vector, improves data quality through regularization, extracts key information through weighted processing, reduces complexity through data dimensionality reduction through principal component analysis, and improves model training efficiency.

[0084] Example 5, see Figure 1 Based on the above embodiment, in step S4, establishing an intelligent recommendation model specifically includes the following steps:

[0085] Step S41: Model architecture design. The established deep learning model uses the transfer learning paradigm to import the weights of the convolution stage. The model transfers the extracted features through the network. These features correspond to the neurons evolved by the evolutionary components of the model. The network output of the model is a vector containing 2048 features, which serves as the bottom layer input of the neural network. The bottom layer of the neural network consists of an input layer that receives 2048 feature vectors, a hidden layer consisting of 512 neurons, and an output layer that defines the number of label categories.

[0086] Step S42: Pruning design, the model network has fewer connections than the standard fully connected layer, where each neuron is connected to the next group of neurons, called a sparse layer; the genetic algorithm is responsible for finding the best pruning pattern for the sparse layer, and the chromosome representing the pruning pattern is decoded into an adjacency matrix; as a result of the pruning operation, the matrix and the layer itself are sparse;

[0087] Step S43: Evolutionary component design, the network population is evolved using a genetic algorithm, and a binary coding strategy is used to indicate whether a neuron is active; if the gene value of a neuron is 1, it is active; if it is 0, it is inactive; based on this direct coding method, each gene uniquely determines a neuron in the decoding network; the initialization of the chromosome corresponds to random discrete initialization in the range of [0, 1], and the binary tournament selection method is selected to initialize the chromosome; the chromosome is used to represent the connection of the neuron; including:

[0088] Step S431: Initialization, using the binary tournament selection method to initialize the chromosome; the performance of the model containing the neurons based on the chromosome representation is used as the chromosome fitness value;

[0089] Step S432: Design the crossover probability, the formula used is as follows:

[0090] Jc=baseJc·(1-τ) tc ;

[0091] Where Jc is the crossover probability; baseJc is the base crossover probability; τ is the decay factor; tc is the current iteration number;

[0092] Step S433: Design a crossover strategy, the operator defines two new individuals from two parents; given these two parents p and q, the formula used to generate offspring is as follows:

[0093]

[0094]

[0095] Where p' i and q' i are two newly generated offspring individuals; p i and q i are the i-th gene values ​​of parents p and q respectively; r is a random number between 0 and 1;

[0096] Step S434: Design a mutation strategy, the formula used is as follows:

[0097]

[0098] In the formula, δ m is the mutation value; c c is the current chromosome; c to is the optimal chromosome in the population; l is the parameter that controls the mutation range; f c and f to are the fitness values ​​of the current chromosome and the best chromosome in the population, respectively;

[0099] Step S435: Anomaly detection component, with an anomaly threshold pre-set, uses softmax with temperature scaling parameters as the anomaly score of each sample. If the score of the instance is lower than the defined threshold, it will output 1 and perform a pruning operation; otherwise, it will output 0; it is expressed as follows:

[0100]

[0101] In the formula, f ON(x; T) is the abnormal score of the neuron corresponding to the chromosome, x is the input data of the neuron corresponding to the chromosome; T is the temperature parameter that controls the degree of scaling; N is the total number of categories, j is the category index; h i (x) and h j (x) are the unscaled probabilities that the model predicts the input to be of class i and class j, respectively;

[0102] Step S436: Pruning determination, the model uses the cross entropy loss function, and a loss threshold is set in advance. When the loss function value of the model for the test set is lower than the loss threshold, the model training is completed; if the maximum number of iterations is reached, the population is reinitialized for pruning; otherwise, iterative pruning continues.

[0103] By performing the above operations, this solution can effectively reduce the connections in the model network and improve the model operation efficiency and generalization ability based on the design of fine pruning and evolutionary components, in order to solve the problems of too many model neurons in general recommendation methods, resulting in weak model generalization ability, the existence of neurons that have a negative impact on model performance, and the impact on model stability. It can also improve model diversity based on the design of mutation and crossover strategies, and introduce anomaly detection components to improve model robustness.

[0104] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S5, intelligent recommendation is based on a trained deep learning model, which collects user data, item data and interaction data in real time, and sets a feedback level threshold in advance. If the model's predicted user behavior feedback level for the recommended item is higher than the feedback level threshold, it is considered that the user is satisfied with the recommended item, and then the recommendation is made.

[0105] Embodiment 7, see Figure 2 , this embodiment is based on the above embodiment, and the deep learning-based intelligent recommendation system provided by the present invention includes a data acquisition module, a data preprocessing module, a data set division module, an intelligent recommendation model establishment module and an intelligent recommendation module;

[0106] The data collection module collects user data, item data, interaction data and user behavior feedback level, and sends the data to the data preprocessing module;

[0107] The data preprocessing module performs data cleaning, data conversion and standardization on the collected data, and sends the data to the data set division module;

[0108] The data set partitioning module improves data quality by calculating grey correlation vectors and regularization processing, extracts key information by weight, and performs data dimension reduction processing by principal component analysis; and sends the data to the intelligent recommendation model establishment module;

[0109] The intelligent recommendation model building module designs fine pruning and evolution components; designs mutation and crossover strategies, introduces anomaly detection components to implement the establishment of the intelligent recommendation model, and sends data to the intelligent recommendation module;

[0110] The intelligent recommendation module implements intelligent recommendations for users based on the trained deep learning model.

[0111] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0112] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

[0113] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. Intelligent recommendation method based on deep learning, characterized by: The method comprises the following steps: Step S1: Data collection, collecting user data, item data, interaction data and user behavior feedback level; Step S2: Data preprocessing, data cleaning, data conversion and standardization of the collected data; Step S3: Data set division, improving data quality by calculating grey correlation vector and regularization, extracting key information by weight, and reducing data dimension by principal component analysis; Step S4: Establish an intelligent recommendation model, design fine pruning and evolution components; design mutation and crossover strategies, and introduce anomaly detection components to achieve the establishment of an intelligent recommendation model; Step S5: Intelligent recommendation: In step S4, the establishment of the intelligent recommendation model specifically includes the following steps: Step S41: Model architecture design. The established deep learning model uses the transfer learning paradigm to import the weights of the convolution stage. The model transfers the extracted features through the network. These features correspond to the neurons evolved by the evolutionary components of the model. The network output of the model is a vector containing 2048 features, which serves as the bottom layer input of the neural network. The bottom layer of the neural network consists of an input layer that receives 2048 feature vectors, a hidden layer consisting of 512 neurons, and an output layer that defines the number of label categories. Step S42: Pruning design, the model network has fewer connections than the standard fully connected layer, where each neuron is connected to the next group of neurons, called a sparse layer; the genetic algorithm is responsible for finding the best pruning pattern for the sparse layer, and the chromosome representing the pruning pattern is decoded into an adjacency matrix; as a result of the pruning operation, the matrix and the layer itself are sparse; Step S43: Evolutionary component design, the network population is evolved using a genetic algorithm, and a binary coding strategy is used to indicate whether a neuron is active; if the gene value of a neuron is 1, it is active; if it is 0, it is inactive; based on this direct coding method, each gene uniquely determines a neuron in the decoding network; the initialization of the chromosome corresponds to random discrete initialization in the range of [0, 1], and the binary tournament selection method is selected to initialize the chromosome; the chromosome is used to represent the connection of the neuron; including: Step S431: Initialization, using the binary tournament selection method to initialize the chromosome; the performance of the model containing neurons based on the chromosome representation is used as the chromosome fitness value; Step S432: Design the crossover probability, the formula used is as follows: Jc=baseJc·(1-τ) tc ; Where Jc is the crossover probability; baseJc is the base crossover probability; τ is the decay factor; tc is the current number of iterations; Step S433: Design a crossover strategy, the operator defines two new individuals from two parents; given these two parents p and q, the formula used to generate offspring is as follows: Where p' i and q' i are two newly generated offspring individuals; p i and q i are the i-th gene values ​​of parents p and q respectively; r is a random number between 0 and 1; Step S434: Design a mutation strategy, the formula used is as follows: In the formula, δ m is the mutation value; c c is the current chromosome; c to is the optimal chromosome in the population; l is the parameter that controls the mutation range; f c and f to are the fitness values ​​of the current chromosome and the best chromosome in the population, respectively; Step S435: Anomaly detection component, with an anomaly threshold pre-set, uses softmax with temperature scaling parameters as the anomaly score of each sample. If the score of the instance is lower than the defined threshold, it will output 1 and perform a pruning operation; otherwise, it will output 0; it is expressed as follows: In the formula, f ON x; T is the abnormal score of the neuron corresponding to the chromosome, x is the input data of the neuron corresponding to the chromosome; T is the temperature parameter that controls the degree of scaling; N is the total number of categories, j is the category index; h i (x) and h j (x) are the unscaled probabilities that the model predicts the input to be of class i and class j, respectively; Step S436: Pruning determination, the model uses the cross entropy loss function, and a loss threshold is pre-set. When the loss function value of the model for the test set is lower than the loss threshold, the model training is completed; if the maximum number of iterations is reached, the population is reinitialized for pruning; otherwise, iterative pruning continues; in step S3, the data set division specifically includes the following steps: Step S31: Calculate the grey relational vector, and calculate the grey relational vector R=(r1, r2, ..., rm) based on the preprocessed data; wherein r1, r2 and rm are the grey relational degrees of the first variable, the second variable and the mth variable respectively; Step S32: Regularization processing: the data sample is subjected to regularization processing, and the formula used is as follows: In the formula, x ij is the value of the jth feature of the ith sample; norm(·) is the regularization process; x i1 、x i2 and x in are the values ​​of the first feature, the second feature, and the nth feature of the i-th sample; μ(x i ) and sig(x i ) are the mean and standard deviation of all features of the i-th sample respectively; is the value of the i-th sample after regularization; x * is the matrix composed of all samples after regularization; and are the values ​​of the first sample, the second sample, and the mth sample after regularization processing; Step S33: weighting, weighting the regularized data samples and the grey relational vector to obtain a weighted matrix G. The formula used is as follows: Step S34: Calculate the covariance matrix, calculate the covariance matrix cov(G) of the weighted matrix G, and extract the eigenvalues and the corresponding eigenvector Step S35: Sorting, sorting the eigenvalues ​​in descending order to obtain the corresponding sorted eigenvectors Select the first t features to retain and get the transformation matrix Step S36: Linear change, using the transformation matrix W * Perform linear transformation on the original data to implement principal component analysis based on grey correlation and obtain the new data set Y * , the formula used is as follows: Y * =x * ·W * ; Step S37: Data set division: divide the new data set into a training set and a test set. The training set is used to train the model; the test set is used to test the model performance.

2. The deep learning-based intelligent recommendation method according to claim 1, characterized in that: In step S1, the data collection is to collect user data, item data, interaction data and user behavior feedback level; the user data includes basic information of the user and user preference data; the item data includes basic information of the item and feature data of the item; the interaction data is the user's click, collection and purchase behavior data on the item; the user behavior feedback level is the user's preference level for the recommendation results; the user behavior feedback level is used as a data label.

3. The deep learning-based intelligent recommendation method according to claim 1, characterized in that: In step S2, the data preprocessing is to perform data cleaning, data conversion and standardization on the collected data; the data cleaning is to perform missing value processing, outlier processing and duplicate value processing on the collected data; the data conversion is to convert the cleaned data into a vector form; the standardization is to perform standardization on the converted data based on Z-Score standardization.

4. The deep learning-based intelligent recommendation method according to claim 1, characterized in that: In step S5, the intelligent recommendation is based on a trained deep learning model, which collects user data, item data and interaction data in real time and sets a feedback level threshold in advance. If the model's predicted user behavior feedback level for the recommended item is higher than the feedback level threshold, it is considered that the user is satisfied with the recommended item and the recommendation is made.

5. An intelligent recommendation system based on deep learning, used to implement the intelligent recommendation method based on deep learning as described in any one of claims 1 to 4, characterized in that: It includes data collection module, data preprocessing module, data set division module, intelligent recommendation model building module and intelligent recommendation module; The data collection module collects user data, item data, interaction data and user behavior feedback level, and sends the data to the data preprocessing module; The data preprocessing module performs data cleaning, data conversion and standardization on the collected data, and sends the data to the data set division module; The data set partitioning module improves data quality by calculating grey correlation vectors and regularization processing, extracts key information by weight, and performs data dimension reduction processing by principal component analysis; and sends the data to the intelligent recommendation model establishment module; The intelligent recommendation model building module designs fine pruning and evolution components; designs mutation and crossover strategies, introduces anomaly detection components to implement the establishment of the intelligent recommendation model, and sends data to the intelligent recommendation module; The intelligent recommendation module implements intelligent recommendations for users based on the trained deep learning model.