Personalized content recommendation method and system based on multi-modal perception

Abstract

The invention relates to the technical field of personalized content recommendation, and discloses a personalized content recommendation method and system based on multi-modal perception, and the method comprises the steps: obtaining multi-modal data, carrying out the preliminary compression of the multi-modal data into low-dimensional potential representation, extracting cross-dimensional correlation features, carrying out the feature fusion, generating a unified feature vector, and obtaining a potential mode. Calculating a correlation coefficient of each dimension and the sensitive attribute, if the correlation coefficient exceeds a threshold value, starting data disturbance processing to obtain a coding representation hiding the sensitive attribute, monitoring a core feature retention rate, if the retention rate is too low, dynamically adjusting disturbance intensity, then performing iterative optimization, obtaining optimized coding features, and performing matching detection with a coding feature library to obtain a coding feature library; and determining a personalized content list to obtain an original recommendation result, and performing privacy protection detection and optimization on the original recommendation result to obtain a final recommendation result. According to the method, the problem of imbalance between privacy protection and recommendation precision in personalized content recommendation is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of personalized content recommendation, in particular to a personalized content recommendation method and system based on multi-modal perception. BACKGROUND

[0002] In the digital age, the importance of personalized content recommendation is increasingly prominent. It is not only closely related to improving user experience, but also has a direct effect on the user appeal and business benefits of the platform. With the analysis of user behavior and preferences, the personalized content recommendation system can push content that matches the user's interests to the user, thereby optimizing the efficiency of information acquisition.

[0003] Under the prior art, the personalized recommendation system only relies on single-dimensional user data, and lacks the ability to integrate different perception dimension information when processing multi-dimensional user data. This scattered information makes it difficult for the system to capture the full picture of user preferences, resulting in a difficult-to-capture user demand in the recommended results. At the same time, the existing method directly exposes sensitive information when processing user data, and lacks an effective encoding mechanism to hide sensitive attributes.

[0004] Therefore, there is a problem of imbalance between privacy protection and recommendation accuracy in the multi-modal data processing of personalized content recommendation, which directly affects the credibility of the personalized content recommendation system and the trust of users on the platform. SUMMARY

[0005] The present application provides a personalized content recommendation method and system based on multi-modal perception, which can solve the problem of imbalance between privacy protection and recommendation accuracy in the multi-modal data processing of personalized content recommendation.

[0006] In a first aspect, to solve the above technical problems, the present application provides a personalized content recommendation method based on multi-modal perception, comprising:

[0007] Obtaining multi-modal data, performing preliminary compression processing on the multi-modal data to obtain a low-dimensional latent representation;

[0008] According to the low-dimensional latent representation, cross-dimension correlation features are obtained and feature fusion processing is performed to obtain a unified feature vector;

[0009] According to the unified feature vector, a latent mode is obtained, and if the sensitive attribute correlation coefficient in the latent mode exceeds a preset sensitive threshold, data perturbation processing is performed to obtain an encoded representation that hides sensitive attributes;

[0010] According to the unified feature vector, a core feature retention rate is calculated, and if the core feature retention rate is lower than a preset retention rate threshold, the encoded representation of the hidden sensitive attribute is adjusted to obtain an adjusted encoded version;

[0011] iteratively optimizing the adjusted encoded version to obtain an optimized encoding feature;

[0012] matching and detecting the optimized encoding feature with a preset encoding feature library to determine a personalized content list and obtain an original recommendation result;

[0013] performing privacy protection detection and optimization on the original recommendation result to obtain a final recommendation result.

[0014] In an optional implementation, the obtaining of the multi-modal data and the preliminary compression processing of the multi-modal data to obtain a low-dimensional latent representation comprises:

[0015] obtaining an interaction record, extracting text, image and audio data from a social platform, a sensor and a log file to obtain an original multi-modal data set; and performing regular sampling

[0016] performing standardization processing on the original multi-modal data to obtain a multi-modal data set with unified format;

[0017] encoding the multi-modal data set by using a preset variational autoencoder to extract modal features and obtain a low-dimensional latent representation.

[0018] In an optional implementation, the obtaining of a cross-dimension correlation feature and the feature fusion processing according to the low-dimensional latent representation to obtain a unified feature vector comprises:

[0019] extracting modal features according to the low-dimensional latent representation to obtain a modal feature set;

[0020] performing dimension reduction processing on the modal feature set by principal component analysis to obtain a cross-dimension feature set;

[0021] performing weight distribution and weighted fusion on the cross-dimension feature set by using an attention mechanism to obtain a unified feature vector.

[0022] In an optional implementation, the obtaining of a latent mode according to the unified feature vector and the data perturbation processing on sensitive attribute-related coefficients exceeding a preset sensitive threshold in the latent mode to obtain an encoded representation hiding the sensitive attribute comprises:

[0023] performing decomposition on the unified feature vector by principal component analysis to obtain a latent mode set;

[0024] calculating a sensitive attribute-related coefficient of any mode in the latent mode set;

[0025] injecting Gaussian noise into mode data with a sensitive attribute-related coefficient exceeding a preset sensitive threshold to obtain perturbed mode data;

[0026] The preset variational autoencoder is used to encode the disturbed pattern data, to obtain an encoded representation of the hidden sensitive attribute.

[0027] In an optional implementation, the core feature retention rate is calculated according to the unified feature vector, and if the core feature retention rate is lower than a preset retention rate threshold, the encoded representation of the hidden sensitive attribute is adjusted to obtain an adjusted encoded version, which comprises:

[0028] The core feature initial representation is determined by principal component analysis on the set of unified feature vectors, and the core feature retention rate is calculated.

[0029] If the core feature retention rate is lower than a preset retention rate threshold, the preset variational autoencoder parameters are adjusted by a gradient descent algorithm to obtain an adjusted parameter configuration.

[0030] The encoded representation of the hidden sensitive attribute is adjusted using the adjusted parameter configuration to obtain a preliminary encoded representation.

[0031] The business data consistency of the preliminary encoded representation is verified, and the part related to the sensitive attribute is again subjected to noise disturbance processing to obtain an adjusted encoded version of the hidden sensitive attribute.

[0032] In an optional implementation, the adjusted encoded version is iteratively optimized to obtain an optimized encoded feature, which comprises:

[0033] The preset variational autoencoder is used to preliminarily encode the multi-modal data according to the adjusted encoded version, to obtain a preliminary optimized feature representation.

[0034] The consistency of the feature distribution of the preliminary optimized feature representation is verified, and the Gaussian distribution sampling is used for regularization processing to obtain a regularized feature representation.

[0035] Based on the regularized feature representation, the preset variational autoencoder parameters are iteratively optimized by a gradient descent algorithm to obtain an optimized parameter configuration.

[0036] The multi-modal data is encoded by the variational autoencoder using the optimized parameter configuration, to obtain the optimized encoded feature.

[0037] In an optional implementation, the optimized encoded feature is matched with a preset encoded feature library to determine a personalized content list, to obtain an original recommendation result, which comprises:

[0038] The optimized encoded feature is matched with a preset encoded feature library to obtain a recommendation candidate set by collaborative filtering.

[0039] rank the contents in the recommendation candidate set by content filtering to obtain a personalized content list;

[0040] encrypt the personalized content list to obtain an original accurate recommendation result.

[0041] In an optional embodiment, the privacy protection detection and optimization on the original recommendation result to obtain a final recommendation result comprises:

[0042] calculate the matching degree of the optimized encoding feature and the pre-stored historical recommendation data, and if the matching degree is lower than a preset sensitive threshold, enhance the cross-dimension integration strength by an attention mechanism to obtain an improved recommendation candidate set;

[0043] perform protection detection on the improved recommendation candidate set, fuse the judgment result of the core feature retention rate, and if the retention rate index meets the privacy sensitive threshold after being associated with the matching degree, perform privacy protection optimization processing, adjust the weight distribution, obtain a privacy-enhanced candidate version, and compare the privacy-enhanced candidate version with the encoding representation of the hidden sensitive attribute, if the retention rate is optimized, obtain a final recommendation set optimized for privacy protection;

[0044] compare the final recommendation set optimized for privacy protection with the current behavior data of the user to obtain a final recommendation result.

[0045] In a second aspect, the present application provides a personalized content recommendation system based on multi-modal perception, comprising:

[0046] a data feature extraction module for obtaining original multi-modal data, performing preliminary compression processing on the original multi-modal data, and generating a low-dimensional latent representation;

[0047] a data feature fusion module for obtaining cross-dimension associated features and performing feature fusion processing to obtain a unified feature vector;

[0048] a data privacy protection module for detecting latent patterns with a sensitive attribute highly correlated coefficient exceeding a preset sensitive threshold, and performing data perturbation processing to obtain an encoding representation of a hidden sensitive attribute;

[0049] a data adjustment module for calculating a core feature retention rate of the encoding representation of the hidden sensitive attribute, and if the core feature retention rate is lower than a retention rate threshold, adjusting the encoding parameters and regenerating an encoding version of the multi-modal data to obtain an adjusted encoding version;

[0050] a data iterative optimization module for performing iterative optimization on the adjusted encoding version to obtain an optimized encoding feature;

[0051] The recommendation generation module is configured to match and detect the optimized encoding features with a preset encoding feature library, determine a personalized content list, and obtain an original recommendation result.

[0052] The privacy protection detection and optimization module is configured to perform privacy protection detection and optimization on the original recommendation result, and obtain a final recommendation result.

[0053] Compared with the prior art, the present application has the following beneficial effects:

[0054] (1) The present application realizes the analysis from a single dimension to multi-dimension by obtaining multi-modal data, preliminarily compressing the multi-modal data into low-dimensional potential representation, extracting cross-dimension correlation features and performing feature fusion to generate a unified feature vector. The present application breaks the information barrier between perception dimensions through feature fusion, and constructs a unified feature expression that can fully reflect the complex interest pattern of a user. This processing method not only significantly improves the integrity and accuracy of feature expression, but also lays a solid data foundation for subsequent accurate recommendation.

[0055] (2) The present application obtains a potential pattern based on the unified feature vector, calculates the correlation coefficient of each dimension and sensitive attribute, and if the correlation coefficient exceeds a preset sensitive threshold, data perturbation processing is started to obtain an encoding representation of the hidden sensitive attribute, while monitoring the core feature retention rate, and if the retention rate is too low, the perturbation strength is dynamically adjusted. The adjusted encoding version is iteratively optimized to obtain an optimized encoding feature. This method only interferes with the data pattern that actually has privacy risks to the minimum necessary extent, so as to effectively confuse the sensitive attribute while maximizing the retention of the core feature for recommendation, so that the system can still maintain high-quality personalized recommendation effect on the premise of strictly protecting the user's sensitive information from being leaked.

[0056] (3) The present application matches and detects the optimized encoding features with a preset encoding feature library to determine a personalized content list, obtains an original recommendation result, and then performs privacy protection detection and optimization on the original recommendation result to obtain a final recommendation result. After generating the preliminary recommendation result, the system does not directly output, but adds a special privacy protection detection and optimization link to perform final security review and calibration on the recommendation result. This design forms an effective secondary protection barrier, which can identify and eliminate the sensitive information association risk that may be left in the previous processing, and ensures that the final output recommendation content is not only accurate and consistent with the user's interest, but also fully meets the privacy protection standard. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is a flowchart of a personalized content recommendation method based on multi-modal perception provided by the first embodiment of the present application;

[0058] Figure 2 is a structure diagram of a personalized content recommendation system based on multi-modal perception provided by the second embodiment of the present application. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0060] With reference to Figure 1 The first embodiment of the present application provides a structure diagram of a personalized content recommendation method based on multi-modal perception, including steps S101 to S107, which are specifically as follows.

[0061] Step S101: Obtain multi-modal data, and perform preliminary compression processing on the multi-modal data to obtain a low-dimensional latent representation.

[0062] Step S102: Obtain cross-dimension correlation features and perform feature fusion processing according to the low-dimensional latent representation to obtain a unified feature vector.

[0063] Step S103: Obtain a latent mode according to the unified feature vector, and if a sensitive attribute correlation coefficient in the latent mode exceeds a preset sensitive threshold, perform data perturbation processing to obtain an encoded representation of a hidden sensitive attribute.

[0064] Step S104: Calculate a core feature retention rate according to the unified feature vector, and if the core feature retention rate is lower than a preset retention rate threshold, adjust the encoded representation of the hidden sensitive attribute to obtain an adjusted encoded version.

[0065] Step S105: Perform iterative optimization on the adjusted encoded version to obtain an optimized encoded feature.

[0066] Step S106: Perform matching detection on the optimized encoded feature and a preset encoded feature library to determine a personalized content list and obtain an original recommendation result.

[0067] Step S107: Perform privacy protection detection and optimization on the original recommendation result to obtain a final recommendation result.

[0068] In step S101, multi-modal data is obtained, and preliminary compression processing is performed on the multi-modal data to obtain a low-dimensional latent representation.

[0069] In an alternative embodiment, the acquiring multi-modal data, and the performing preliminary compression processing on the multi-modal data to obtain a low-dimensional latent representation, comprise:

[0070] The interaction records are acquired, and text, image, and audio data are extracted from social platforms, sensors, and log files to obtain an original multi-modal data set.

[0071] The original multi-modal data is standardized to obtain a multi-modal data set with uniform format.

[0072] The multi-modal data set is encoded using a preset variational autoencoder to extract modal features and obtain a low-dimensional latent representation.

[0073] Wherein, the interaction records obtained from the multi-modal data source can be realized by a pre-designed collection tool. Illustratively, for social platforms, a crawler tool can be used to extract user comments, pictures, and voice messages from Weibo or WeChat; sensor data can be obtained from a smart watch to obtain user heart rate, motion trajectory, etc.; log files can extract user operation records from servers. These data sources collectively constitute the original multi-modal data set.

[0074] It is worth noting that social platform data may contain JSON format text, BMP format images, and WAV format audio, while sensor data may be in CSV format, and log files may be in TXT format. Such format differences will cause subsequent processing difficulties. Data preprocessing can standardize the above data. Illustratively, text data can be uniformly converted to UTF-8 encoding, image data can be adjusted to a uniform resolution such as 256x256 pixels and converted to JPEG format, and audio data can be uniformly converted to 16kHz sampling rate MP3 format. Preferably, text data is standardized by word segmentation and stop word removal; image data is standardized in size and color space by cropping and grayscale processing; audio data is processed by noise reduction and standardization of volume. This standardization processing ensures that different modal data has a consistent input format when encoding.

[0075] It is worth mentioning that the preset variational autoencoder is used to encode the format-consistent multi-modal dataset to extract the latent features. The format-consistent multi-modal dataset is input into the variational autoencoder as a dataset, and in each training iteration, a small batch of samples sampled from the data distribution is input into the encoder. First, a multi-modal encoding strategy is adopted: the image data is processed using a convolutional neural network, the text sequence is processed using a recurrent neural network, and the sound signal is processed using VGGish, so as to respectively map the original data of each modality to a probability distribution space; the encoder outputs a mean vector and a variance vector for each modality sample, and a smooth and continuous latent feature vector is generated by sampling from the distribution through a reparameterization technique. The decoder reconstructs these latent feature vectors into multi-modal data, while being optimized through a double-objective loss function, and the objective function is composed of a reconstruction loss and a KL divergence regularization term: the reconstruction loss measures the negative log-likelihood of the decoder output and the original input, and the KL divergence regularization optimizes the latent representations of the text, image and audio to align in the same latent space, facilitating unified modeling of multi-modal data. The gradient of the objective function with respect to the parameters is calculated through backpropagation, and the network weights are updated using optimization methods such as adaptive moment estimation algorithm.

[0076] In step S102, cross-dimensional association features are obtained according to the low-dimensional latent representation, and feature fusion processing is performed to obtain a unified feature vector.

[0077] According to the low-dimensional latent representation, each modality feature set is extracted.

[0078] The principal component analysis is performed on the each modality feature set to obtain a cross-dimensional feature set.

[0079] The attention mechanism is used to assign weights to the cross-dimensional feature set and perform weighted fusion to obtain a unified feature vector.

[0080] Among them, the interaction record is obtained from the multi-modal data source, and the pre-established collection tool is used to extract visual, auditory and tactile signals from social platforms, sensors and log files. For example, when the interaction record is obtained from the multi-modal data source, efficient collection can be achieved through customized tools. For social platforms, a crawler tool can be designed to extract user-uploaded video clips, bullet screen text and like behavior from short video platforms. Sensor data can collect user step frequency, skin conductance and other tactile signals through a smart bracelet. Log files can extract user click paths and dwell time from application servers. These data represent the respective modality feature sets of visual, auditory and tactile signals.

[0081] It is worth noting that due to the dimensional difference of each modal feature set, direct fusion will lead to unbalanced feature weights. Dimensionality reduction processing can be performed through principal component analysis. Exemplarily, all features are uniformly reduced to 256 dimensions to generate a cross-dimensional feature set. This process can retain more than 90% of the data variance, ensuring minimal information loss. The feature set after dimensionality reduction is uniform in dimension, facilitating cross-modal fusion.

[0082] It is worth noting that the attention mechanism can allocate weights to the cross-dimensional feature set. After the heterogeneous feature vectors extracted from the 256-dimensional visual, auditory and tactile different perception channels enter the fusion layer, the correlation scores between any two feature dimensions are calculated through the self-attention unit to obtain a channel-level attention map; then the original features are weighted and labeled using the scores, and the dimensions highly related to the current user interest are given large weights, so that their information dominates in the subsequent fusion, while the weakly related or redundant dimensions are suppressed to small weights. Finally, the weighted features are interacted through the cross-modal attention module to generate a unified feature vector with interpretable weights. Exemplarily, the actions of the characters in the video may contribute more to emotion analysis, with a weight value of 0.6; the rhythm of the background music has a weight of 0.3; and the tactile signal has a weight of 0.1. The softmax function normalizes these weights to generate a unified weight distribution. Finally, the unified feature vector after fusion can comprehensively reflect the user's interactive behavior.

[0083] In step S103, according to the unified feature vector, the latent mode is obtained, and if the sensitive attribute correlation coefficient in the latent mode exceeds the preset sensitive threshold, data perturbation processing is performed to obtain an encoded representation of the hidden sensitive attribute.

[0084] The unified feature vector is decomposed using principal component analysis to obtain a set of latent modes;

[0085] The sensitive attribute correlation coefficient of any mode in the set of latent modes is calculated;

[0086] Gaussian noise is injected into the mode data whose sensitive attribute correlation coefficient exceeds the preset sensitive threshold to obtain perturbed mode data;

[0087] The perturbed mode data is encoded using a preset variational autoencoder to obtain an encoded representation of the hidden sensitive attribute.

[0088] In the context of multi-modal data fusion, principal component analysis can be used to decompose the unified feature vector to extract a set of latent patterns. Principal component analysis maps high-dimensional features to a lower-dimensional space through a linear transformation, preserving the main data variance. Illustratively, for a 512-dimensional unified feature vector of a short video platform user, principal component analysis can decompose it into several latent patterns, each reflecting a different aspect of user behavior, such as viewing preferences, interaction frequency, or emotional response. Assuming that 10 latent patterns are obtained after decomposition, each represented by a set of feature vectors, 95% of the original data variance is preserved. This decomposition helps reveal the underlying rules of user behavior, facilitating subsequent analysis.

[0089] It is worth noting that to determine whether the sensitive attribute correlation coefficient of any pattern in the set of latent patterns exceeds the preset threshold, the sensitive attribute must first be defined. The sensitive attribute refers to a personal privacy feature dimension that needs to be specially protected and can be inferred from multi-modal behavior data. It can be preset through business rules or learned from labeled data. Specifically, these attributes are not directly stored in the original data fields, but refer to information related to user identity, status, or inherent characteristics. Illustratively, in a short video platform, sensitive attributes include but are not limited to user emotional state or bias tendency, where emotional state can be defined based on positive / negative scores obtained from user comment text sentiment analysis, and bias tendency can be defined by calculating the standard deviation of user click rates on different categories of content. The correlation coefficient can be obtained by calculating the Pearson correlation coefficient between the latent pattern vector and the sensitive attribute vector. Referring to historical sensitive data correlation rules, by analyzing the correlation coefficient between the calculated pattern vector and the sensitive attribute vector in past sensitive events, when the value exceeds 0.7, the success rate of inferring sensitive attributes from the calculated pattern vector will significantly increase. Therefore, the sensitive threshold is set to 0.7. If the coefficient of a certain pattern reaches 0.8, it indicates that the pattern is highly correlated with the sensitive attribute and may leak user privacy. This judgment helps identify data patterns that need to be protected, ensuring the security of subsequent processing.

[0090] It is worth noting that the mode data exceeding the threshold is disturbed by the Gaussian noise injection method, which can effectively hide sensitive information. Among them, the noise mean is usually set to 0 to avoid introducing systematic bias; the variance is dynamically adjusted according to the data sensitivity, and for high-sensitive mode data exceeding the threshold, the standard deviation is 1.5-2 times the data standard deviation, ensuring that the original mode is difficult to be restored after disturbance; for low-sensitive but still need to protect the mode data, a smaller standard deviation is used, which is 0.5-1 times the data standard deviation, which hides sensitive information while retaining the overall distribution characteristics of the data. For example, for a mode with a correlation coefficient of 0.8, Gaussian noise with a mean of 0 and a standard deviation of 0.5-1 times the data standard deviation can be introduced, and the disturbed mode data still retains the main behavior characteristics, but the identifiability of sensitive attributes is significantly reduced. This method can balance data usability and privacy protection in practical applications.

[0091] It is worth noting that the disturbed mode data is encoded by a variational autoencoder to generate an encoded representation that hides sensitive attributes. The variational autoencoder compresses the input data into a low-dimensional representation through an encoder, and then reconstructs the data through a decoder. For example, the disturbed 256-dimensional mode data is input into the variational autoencoder, the encoder generates a 128-dimensional hidden representation, and the decoder attempts to restore the original mode. During training, the hidden representation can effectively retain non-sensitive features, such as user preference for video types, while further weakening sensitive attributes, such as specific emotional intensity. This encoded representation facilitates subsequent tasks such as recommendation system optimization while reducing privacy risks.

[0092] In step S104, according to the unified feature vector, the core feature retention rate is calculated, and if the core feature retention rate is lower than the preset retention rate threshold, the encoded representation of the hidden sensitive attribute is adjusted to obtain an adjusted encoded version.

[0093] For the set of unified feature vectors, the initial representation of the core feature is determined by principal component analysis, and the core feature retention rate is calculated;

[0094] If the core feature retention rate is lower than the preset retention rate threshold, the preset variational autoencoder parameters are adjusted by a gradient descent algorithm to obtain an adjusted parameter configuration;

[0095] The adjusted parameter configuration is used to adjust the encoded representation of the hidden sensitive attribute to obtain a preliminary encoded representation;

[0096] The business data consistency of the preliminary encoded representation is verified, and the part related to the sensitive attribute is again subjected to noise disturbance processing to obtain an adjusted encoded version that hides the sensitive attribute.

[0097] It is worth noting that the core feature evaluation is performed on the unified feature vector set. Through principal component analysis, the principal components with the largest contribution information are extracted from the high-dimensional feature vectors, which constitute the initial representation of the core features. Subsequently, the core feature retention rate is calculated. Exemplarily, for a 1024-dimensional feature vector of a user, including viewing time, like behavior, and comment frequency, etc., principal component analysis can extract the first 20 principal components to generate a unified feature vector set, retaining 90% of the original information. The core feature retention rate can be obtained by calculating the cumulative variance contribution rate of the principal components.

[0098] It is worth noting that the preset retention rate threshold is a critical standard for judging whether the privacy-processed data has damaged the recommendation availability in personalized content recommendation, which is determined through historical data sets. By analyzing the historical unified feature vector set, it is found that when the core feature retention rate is higher than 85%, sufficient user-related information can be obtained through feature extraction to generate accurate and personalized recommended content; therefore, the retention rate threshold is 85%. Those skilled in the art understand that the retention rate threshold of 85% is an empirical value based on the general recommendation scenario, and in actual application, the threshold can be adjusted within the range of 80% to 90% according to the service type or performance requirements. For example, for a high-precision recommendation service, the threshold can be increased to 90%; for a lightweight recommendation scenario, the threshold can be decreased to 80%. When it is detected that the core feature retention rate is lower than the preset retention rate threshold, it indicates that the current privacy protection processing is too aggressive and has damaged the availability of the recommendation. At this time, the adaptive optimization mechanism is started, and the parameters of the preset variational autoencoder are fine-tuned through the gradient descent algorithm. The optimization goal is to adjust the internal weights of the variational encoder. The specific implementation is as follows: first, calculate the loss function based on the features output by the current model and the key recommendation features in the original data, which includes the reconstruction loss and the regularization loss, and solve the gradient of each layer weight of the encoder with respect to the target function through back propagation; then update the parameters along the negative direction of the gradient by a certain proportion of the initial learning rate. For example, enhance the convolution kernel weight of the encoder that retains the video category feature, and weaken the fully connected layer weight that excessively compresses the key information; after each round of fine-tuning, the core feature retention rate is monitored in real time, and if it rises continuously for 3 rounds and stabilizes above the threshold, the iteration is stopped.

[0099] It is worth noting that the adjusted parameters of the variational autoencoder are used for encoding processing of the multi-modal data to generate a preliminary encoding representation. Exemplarily, a 512-dimensional feature vector is input into the variational autoencoder to generate a 256-dimensional preliminary encoding representation.

[0100] After generating the preliminary encoding representation, business data consistency verification must be performed. This verification method can be achieved by comparing the similarity of the encoding representation and the original data, such as calculating the cosine similarity between the preliminary encoding representation and the original data; if the coefficient reaches the preset threshold value 0.7, it is determined that the encoding representation meets the consistency requirement. According to historical data, when the cosine similarity reaches 0.6, the user's core behavior characteristics have been basically retained, and when it is higher than 0.7, the feature direction coincidence degree of the two is more than 70%, which can completely retain the user's core behavior characteristics. At the same time, after adding Gaussian noise to the encoding with a similarity of 0.7, the overall similarity can still be maintained in the interval of 0.6-0.7, which still meets the requirement of retaining the user's core behavior characteristics, and reserves information buffer space for subsequent Gaussian noise disturbance sensitive attributes, taking into account the business availability and privacy protection feasibility. Therefore, the preset threshold value is 0.7. Then, the part related to the sensitive attribute is disturbed by injecting Gaussian noise, so as to realize user privacy protection. Among them, the noise mean is usually set to 0 to avoid introducing systematic bias; the variance is dynamically adjusted according to the data sensitivity, and for high-sensitive mode data exceeding the threshold value, a larger standard deviation of 1.5-2 times the data standard deviation is used to ensure that the original mode is difficult to be restored after disturbance; for low-sensitive but still protected mode data, a smaller variance of 0.5-1 times the data standard deviation is used to hide sensitive information while retaining the overall distribution characteristics of the data. The sensitive attributes include the user's age distribution or emotional tendency.

[0101] Exemplarily, after inputting a 512-dimensional user behavior feature vector into the autoencoder and generating a 256-dimensional preliminary encoding representation, the business data consistency of the preliminary encoding representation with the original data needs to be verified by cosine similarity; if the similarity reaches 0.9, it means that the encoding can reflect the overall behavior trend of the user, such as watching preferences, interaction habits, etc., and meets the consistency requirement. After the encoding meets the consistency, the part related to the sensitive attribute (such as age distribution, emotional tendency) needs to be disturbed by Gaussian noise to protect privacy, for example, identifying a subset vector related to emotional tendency from the 256-dimensional preliminary encoding, and the cosine similarity is 0.75 (exceeding the preset threshold value 0.7), at this time, Gaussian noise with mean value 0 and standard deviation 0.05 is introduced for disturbance, and the cosine similarity of the disturbed encoding version is 0.68 again, so that the disturbed encoding version retains the non-sensitive information of the user's preference for video types while reducing the identifiability of the emotional tendency. This way ensures the balance between data availability and privacy protection. It should be noted that the disturbance and encoding process need to be dynamically adjusted according to business needs, such as updating the noise standard deviation regularly to ensure the balance between privacy protection and data utility.

[0102] In a possible implementation, the adjusted encoding version can be further used for optimization of the recommendation system. The encoding representation retains core behavioral features, such as user preference for funny videos, and weakens sensitive information. This encoding can be directly input into the recommendation model to improve the accuracy of content matching. For example, the recommendation system preferentially pushes short entertaining videos that the user often watches according to the encoding representation, thereby improving the user experience.

[0103] In step S105, the adjusted encoding version is iteratively optimized to obtain an optimized encoding feature.

[0104] According to the adjusted encoding version, the multi-modal data is preliminarily encoded by using a preset variational autoencoder to obtain a preliminary optimized feature representation.

[0105] The consistency of the feature distribution of the preliminary optimized feature representation is verified, and the preliminary optimized feature representation is regularized by Gaussian distribution sampling to obtain a regularized feature representation.

[0106] Based on the regularized feature representation, the parameters of the preset variational autoencoder are iteratively optimized by using a gradient descent algorithm to obtain an optimized parameter configuration.

[0107] The multi-modal data is encoded by using the optimized parameter configuration to obtain the optimized encoding feature.

[0108] It should be noted that the variational autoencoder is used for preliminary encoding of the multi-modal data, aiming to extract more representative feature representations from complex high-dimensional data. For example, in the user behavior analysis scenario of a short video platform, the variational autoencoder compresses the user's multi-modal data such as viewing time, sharing behavior, and search keywords into a low-dimensional latent feature space by introducing probability distribution modeling. Assuming that the original data is a 768-dimensional feature vector, the variational autoencoder can generate a 128-dimensional preliminary optimized feature representation.

[0109] It should be noted that the verification of the consistency of the feature distribution of the preliminary optimized feature representation checks whether it meets the expected business data distribution. Specifically, the consistency can be evaluated by comparing the statistical characteristics such as mean and variance of the preliminary optimized feature representation and the original data. Assuming that the consistency requires the KL divergence of the feature distribution to be less than 0.1, if the calculation result is 0.08, it means that the preliminary optimized feature representation meets the requirements. At this time, the feature representation can be regularized by Gaussian distribution sampling.

[0110] It is worth noting that, based on regularized feature representation, a gradient descent algorithm is used to iteratively optimize the parameters of the variational autoencoder. The optimization objective is to minimize the loss, including reconstruction loss and regularization loss. The iterative optimization loop uses a gradient descent-based optimization algorithm, aiming to adjust the encoder and decoder weights multiple times to make the latent feature space more accurately capture user behavior patterns, thus obtaining the optimized parameter configuration. The core process is as follows: Initialize encoder parameters. and decoder parameters Set the maximum number of iterations N=100, and input randomly sampled user behavior samples. The mean of the latent variables is obtained. and variance Select Latent features are generated by sampling that follows a normal distribution and is obtained through reparameterization techniques. The input is fed into the decoder to obtain the reconstructed data. ', Calculate user behavior samples With reconstructing data 'Comprehensive Loss' The overall loss is the sum of the reconstruction loss and the regularization loss. The reconstruction loss is calculated using the numerical difference method to determine the mean square error, and the regularization loss is calculated using the formula... Realization, i.e., the mean of each dimension of the latent features. ,variance The gradient ∇ of the encoder is calculated by summing according to this formula. and the gradient of the decoder∇ Update the parameters of the input encoder along the negative gradient direction. Decoder parameters Learning rate Set the value to 0.005 and repeat the above steps for iterative optimization until the maximum number of iterations is reached, thus obtaining the optimized encoder and decoder parameter configurations.

[0111] It is worth noting that when encoding multimodal data using the optimized parameter configuration, the original data is re-input into the variational autoencoder to generate the final optimized encoded features. These optimized features retain core behavioral information while reducing redundancy between features through regularization.

[0112] In step S106, the optimized coding features are matched and detected with a preset coding feature library to determine the personalized content list and obtain the original recommendation result.

[0113] The optimized coding features are matched with a preset coding feature library, and a recommendation candidate set is obtained through collaborative filtering;

[0114] Content filtering is used to sort the content in the recommendation candidate set to obtain a personalized content list;

[0115] The personalized content list is encrypted to obtain an original accurate recommendation result.

[0116] It is worth noting that the matching detection operation filters the label with the highest cosine similarity as the matching result by calculating the cosine similarity of the optimized encoding features and the features corresponding to each label in the preset encoding feature library, and the collaborative filtering method is used to filter the preliminary recommendation candidate set. For example, according to the matched "entertainment preference" label, a set of videos that other users with similar behavior patterns like, such as recently popular funny short videos or variety clips, are filtered to form a candidate set. The construction of the preset encoding feature library can be gradually promoted from multiple dimensions. First, the basic interest field framework is determined in combination with the platform core business scenarios (such as short videos, e-commerce, information, etc.), for example, a short video platform can first divide entertainment, education, life, fitness, technology, etc. The first level of interest categories; then, through historical user behavior data (including clicks, viewing time, collections, comment keywords, purchase records, etc.), deep mining is performed to extract key behavior indicators in each interest field, such as "entertainment preference" category can extract "funny video viewing ratio ≥ 60%" "variety content like ratio ≥ 40%" and other core features, and these indicators are converted into standardized feature vectors; then, the industry general interest label system and artificial annotation verification are introduced, the preliminary extracted feature vectors are screened and optimized, and redundant or low correlation features are removed, while new emerging interest directions (such as recently emerging "AI generated content preference" "immersive experience content preference" etc.) corresponding features are supplemented; finally, a dynamic updating mechanism of the feature library is established, and based on the newly added user behavior data and the changes in the platform content ecosystem, the dimension weight of the feature vector is adjusted, new or merged interest categories (such as "fitness" is divided into "home fitness" "outdoor fitness"), and the feature vector is normalized through, for example, cosine similarity, to ensure that the features in the library can accurately match the user behavior encoding and flexibly adapt to the dynamic changes of user interests and platform business.

[0117] In one implementation manner, the collaborative filtering method combines user historical data and real-time interaction data to generate a more time-effective candidate set. Specifically as follows: first, the similarity between the user and other users needs to be calculated through the user's historical data and real-time interaction data. Common similarity calculation methods include cosine similarity and Pearson correlation coefficient. Then, find the n users most similar to the target user, and use the behaviors of these n users to generate recommendations. According to the behaviors of these similar users, extract the content that the target user has not contacted and the click rate of which is higher than 10% within seven days to form a candidate set.

[0118] It is worth mentioning that the content filtering ranking method ranks the preliminary recommended candidate set to generate a personalized content list. For example, the metadata of the videos in the candidate set, such as video length, content tags, and user historical interaction data, are analyzed to prioritize videos that are highly relevant to user preferences. Assuming that a user often watches funny videos that are less than 5 minutes long, such videos are placed at the top of the list. This ranking method ensures that the recommended content is more in line with user interests.

[0119] It is worth mentioning that the personalized content list is encrypted by adding noise. For example, noise is added to the metadata of the recommended list, such as adding random perturbations with a mean of 0 and a standard deviation of 0.01 to the recommendation weight of a video. This processing method protects user privacy while ensuring the accuracy of the recommendation results. The encrypted list can still accurately push funny videos, but user-specific behavior data is not directly exposed.

[0120] In an implementable manner, the noise standard deviation can be dynamically adjusted, such as increasing from 0.01 to 0.02, to balance privacy protection and recommendation effectiveness.

[0121] In step S107, the original recommendation result is detected and optimized for privacy protection to obtain a final recommendation result.

[0122] The matching degree of the optimized encoding features and the pre-stored historical recommendation data is calculated. If the matching degree is lower than a pre-set matching degree threshold, the cross-dimensional integration strength is enhanced through an attention mechanism to obtain an improved recommendation candidate set.

[0123] The improved recommendation candidate set is detected for protection, and the core feature retention rate judgment result is fused. If the retention rate index is associated with the matching degree and is lower than a pre-set privacy sensitivity threshold, privacy protection optimization processing is performed to adjust the weight distribution, obtain a privacy-enhanced candidate version, and compare the privacy-enhanced candidate version with the encoded representation of the hidden sensitive attribute. If the retention rate is optimized, a final recommendation set optimized for privacy protection is obtained.

[0124] The final recommendation set optimized for privacy protection is compared with the user's current behavior data to obtain a final recommendation result.

[0125] It is worth noting that when it is detected that the matching degree of the optimized encoding feature and the pre-stored historical recommendation data is lower than the pre-set matching degree threshold, it indicates that the user interest may have drifted. Among them, the pre-set matching degree threshold refers to analyzing the historical user behavior data, calculating the similarity distribution of the behavior vectors in the interest stable period and the drift period in the user's recent behavior features, and determining a matching degree threshold with the pre-stored historical recommendation data to ensure that it can effectively distinguish between user interest fluctuations and interest migration. To this end, the attention mechanism is activated to dynamically enhance the integration strength of the user multi-dimensional behavior data, and the weight distribution of each behavior dimension is re-evaluated. Illustratively, the optimized encoding feature is matched with the pre-stored historical recommendation data to determine whether it meets the pre-set threshold, such as a cosine similarity of 0.85. If the matching degree is lower than the threshold, it indicates that the user's interest may have changed, and the feature needs to be further optimized. Illustratively, the optimized encoding feature of a certain user has a similarity of only 0.7 with the pre-stored historical recommendation data of the "funny video" label, which is lower than the threshold. At this time, the attention mechanism is introduced to perform weight distribution on the optimized encoding feature. The attention mechanism assigns higher weights to feature dimensions related to food by analyzing the user's recent interaction data, such as frequent pausing of watching food videos, and generates enhanced cross-dimension integrated features, such as a 256-dimensional comprehensive interest vector. This way can more accurately reflect the user's current interest tendency.

[0126] It is worth noting that the privacy protection detection is started on the enhanced candidate set, and the core feature retention rate is analyzed in association with the user matching degree, wherein the association analysis is as follows: in the core feature retention rate, the information entropy of the feature vector is taken as the core feature retention rate index of the enhanced recommendation candidate set, and the result of multiplying the core feature retention rate and the user matching degree is compared with the pre-set privacy threshold 0.8. The pre-set privacy threshold is a critical standard for judging whether the data exists in the personalized content recommendation. A common calculation method is: first, count the benchmark range of the information entropy of the user preference feature vector in the normal scenario, and through a large number of sample analysis, it is found that when the feature vector information entropy of the recommendation candidate set is higher than 0.8, the user preference distribution contained in the vector is more dispersed, and it is difficult to lock the specific user; when the information entropy is lower than 0.8, the preference feature is highly concentrated, and the risk of leaking user identity or privacy demand is significantly increased. Therefore, the privacy threshold 0.8 is calculated to ensure that the threshold can cover the preference feature distribution of most normal users, while accurately identifying high-risk data with excessive concentrated preferences. Those skilled in the art understand that the privacy threshold 0.8 is an empirical calculation value based on the general user group sample, and in actual application, the threshold can be adjusted within the range of 0.75 to 0.85 according to the characteristics of the user group or the sensitivity of the scene. For example, for specific user group recommendation scenarios, the threshold can be raised to 0.85; for general user recommendation scenarios, the threshold can be lowered to 0.75.

[0127] If the retention rate of the core features of the initial recommendation candidate set is lower than a preset privacy threshold after being correlated with the matching degree, privacy protection optimization processing is performed on the candidate set. The weight distribution of the initial recommendation candidate set is adjusted using a logistic regression algorithm to obtain an adjusted weight distribution. Based on the adjusted weight distribution, the initial recommendation candidate set is reordered to generate a privacy-enhanced candidate version. Subsequently, the optimized candidate set is compared and verified with a fully anonymized encoding representation to ensure that, under the premise of optimizing the retention rate, a final recommendation set that balances privacy protection and recommendation utility is ultimately generated. The fully anonymized encoding refers to an encoding that, after being perturbed by injecting Gaussian noise, cannot be directly or indirectly associated with or restored to the original personal information or original sensitive attributes. To achieve full anonymization, corresponding Gaussian noise needs to be injected using a differential privacy algorithm. The core of this method lies in: firstly, determining the global sensitivity of the query function (such as gradient, summation function), i.e., the maximum value of the function output change. ; then through the formula Calculate the standard deviation σ of the noise. This is a privacy parameter, typically set between 0.1 and 10. To mitigate the risk of minor violations of strict differential privacy, it is typically set to 10. -5 Inject a mean of 0 and a standard deviation of into the original data or intermediate results. The output contains Gaussian noise. After this processing, it is impossible to deduce user information from the output.

[0128] For example, the video platform performs privacy protection checks on the improved food recommendation candidate set: First, core features are extracted, and their retention rate is calculated to be 85%; the product of 90% user matching degree is lower than the preset privacy sensitivity threshold of 0.8. Privacy protection optimization is then initiated, adjusting the weight distribution of the candidate set through a logistic regression algorithm, determining the weights of sensitive and non-sensitive features, setting the maximum number of iterations N=100, and the learning rate... Set to 0.1, and based on the current weights, use the Sigmoid activation function to predict the input sample. Compare the predicted value y' with the target value y, and adjust the gradient of each weight using the cross-entropy loss function. Use gradient descent algorithm to update the weights. , For each iteration, new weight values ​​are provided. The weight value to be iterated for each round. Repeat the above steps for iterative optimization until the maximum number of iterations is reached to end the iteration. For example, reduce the weight of the sensitive feature from 0.3 to 0.1, and increase the weight of the non-sensitive feature from 0.2 to 0.4, to generate a privacy-enhanced candidate version. Again, the detection shows that the core feature retention rate has increased to 89%, and the core feature retention rate multiplied by the user matching degree of 90% is 0.801, which has reached the preset privacy sensitive threshold of 0.8. After comprehensive evaluation, it is determined as the final recommended set of privacy protection optimization.

[0129] It is worth noting that before outputting the final result, the final recommended set after privacy optimization is compared with the user's current behavior data to further filter or adjust the recommended content, ensuring that the recommended result meets the privacy protection requirements and accurately matches the user's real-time needs. For example, the platform has completed privacy optimization processing and generated a final recommended candidate set of three categories. The set has been anonymized for sensitive fields, and the core feature retention rate and user matching degree meet the balance standard of privacy protection and recommendation availability. Based on the behavior comparison mechanism, the user's interaction data within the last 1 hour is extracted, and short videos with 3 consecutive clicks and content that is quickly skipped are identified. The categories of the short videos with consecutive clicks and the categories of the content that is quickly skipped are extracted. By matching the optimized recommended set with the above behavior characteristics, the content strongly associated with the category of the short videos with consecutive clicks is selected, and the category of the content that is quickly skipped is simultaneously excluded, and finally the final recommended set after privacy protection optimization that meets the user's current behavior is output.

[0130] Reference Figure 2 The second embodiment of the present application provides a personalized content recommendation system based on multi-modal perception, comprising:

[0131] A data feature extraction module is configured to obtain original multi-modal data, perform preliminary compression processing on the original multi-modal data, and generate low-dimensional latent representation.

[0132] A data feature fusion module is configured to obtain cross-dimension association features and perform feature fusion processing to obtain a unified feature vector.

[0133] A data privacy protection module is configured to detect latent patterns with a sensitive attribute highly correlated coefficient exceeding a preset sensitive threshold, and perform data perturbation processing to obtain an encoded representation hiding the sensitive attribute.

[0134] A data adjustment module is configured to calculate a core feature retention rate for the encoded table hiding the sensitive attribute. If the core feature retention rate is lower than a retention rate threshold, the encoded parameters are adjusted and an encoded version of the multi-modal data is regenerated to obtain an adjusted encoded version.

[0135] a data iteration optimization module configured to iteratively optimize the adjusted encoding version to obtain an optimized encoding feature;

[0136] a recommendation generation module configured to match and detect the optimized encoding feature with a preset encoding feature library, determine a personalized content list, and obtain an original recommendation result;

[0137] a privacy protection detection and optimization module configured to detect and optimize the original recommendation result for privacy protection to obtain a final recommendation result.

[0138] It should be noted that the personalized content recommendation system based on multi-modal perception provided by the embodiments of the present application is used to execute all process steps of the personalized content recommendation method based on multi-modal perception of the above-mentioned embodiments, and the working principles and beneficial effects of the two are one-to-one corresponding, thus no longer being described in detail.

[0139] In summary, the present application discloses a personalized content recommendation method based on multi-modal perception, which comprises obtaining multi-modal data, preliminarily compressing the multi-modal data into low-dimensional latent representation, extracting cross-dimension correlation features and performing feature fusion to generate a unified feature vector. On the basis of the unified feature vector, latent patterns are obtained, the correlation coefficients of each dimension and sensitive attributes are calculated, and if the correlation coefficients exceed a preset sensitive threshold, data perturbation processing is started to obtain an encoding representation of hidden sensitive attributes, while monitoring the core feature retention rate, and if the retention rate is too low, the perturbation strength is dynamically adjusted. Then, the adjusted encoding version is iteratively optimized to obtain an optimized encoding feature, which is matched and detected with the optimized encoding feature to determine a personalized content list and obtain an original recommendation result, and the original recommendation result is detected and optimized for privacy protection to obtain a final recommendation result. The present application realizes the analysis from a single dimension to multi-dimension, constructs a unified feature expression capable of comprehensively reflecting the complex interest patterns of users through feature fusion. After generating a preliminary recommendation result, a special privacy protection detection and optimization link is added, and the recommendation result is finally subjected to safety review and calibration. This design forms an effective secondary protection barrier, which can identify and eliminate the sensitive information correlation risks that may be left in the previous processing, and ensure that the recommended content output is not only accurate and consistent with the user's interest, but also fully meets the privacy protection standards.

[0140] The above-described specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application, and it should be understood that the above-described embodiments are only specific embodiments of the present application and are not used to limit the protection scope of the present application. It is particularly pointed out that any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A personalized content recommendation method based on multimodal perception, characterized in that, include: Acquire multimodal data, perform preliminary compression processing on the multimodal data, and obtain a low-dimensional latent representation; Based on the low-dimensional latent representation, cross-dimensional related features are obtained and feature fusion processing is performed to obtain a unified feature vector. Based on the unified feature vector, a latent pattern is obtained. If the correlation coefficient of the sensitive attribute in the latent pattern exceeds a preset sensitivity threshold, data perturbation processing is performed to obtain the encoded representation of the hidden sensitive attribute. Based on the unified feature vector, the core feature retention rate is calculated. If the core feature retention rate is lower than the preset retention rate threshold, the encoding representation of the hidden sensitive attribute is adjusted to obtain the adjusted encoding version. The adjusted encoding version is iteratively optimized to obtain optimized encoding features; The optimized coding features are matched and detected with a preset coding feature library to determine a personalized content list and obtain the original recommendation result; The original recommendation results are then subjected to privacy protection checks and optimizations to obtain the final recommendation results.

2. The personalized content recommendation method based on multimodal perception according to claim 1, characterized in that, The process of acquiring multimodal data and performing preliminary compression processing on the multimodal data to obtain a low-dimensional latent representation includes: Obtain interaction records and extract text, image, and audio data from social platforms, sensors, and log files to obtain the raw multimodal dataset; The original multimodal data is standardized to obtain a multimodal dataset with a uniform format; The multimodal dataset is encoded using a preset variational autoencoder to extract features from each modality and obtain a low-dimensional latent representation.

3. The personalized content recommendation method based on multimodal perception according to claim 1, characterized in that, The step of obtaining cross-dimensional related features based on the low-dimensional latent representation and performing feature fusion processing to obtain a unified feature vector includes: Based on the low-dimensional latent representation, features of each modality are extracted to obtain feature sets of each modality; Principal component analysis is used to reduce the dimensionality of each modal feature set to obtain a cross-dimensional feature set. An attention mechanism is used to assign weights and perform weighted fusion on the cross-dimensional feature set to obtain a unified feature vector.

4. The personalized content recommendation method based on multimodal perception according to claim 1, characterized in that, The step of obtaining a latent pattern based on the unified feature vector, and if the correlation coefficient of the sensitive attribute in the latent pattern exceeds a preset sensitivity threshold, then performing data perturbation processing to obtain the encoded representation of the hidden sensitive attribute, includes: The unified feature vector is decomposed using principal component analysis to obtain a set of latent patterns; Calculate the correlation coefficient of the sensitive attributes of any pattern in the set of potential patterns; Gaussian noise is injected into the pattern data whose correlation coefficient of the sensitive attribute exceeds a preset sensitivity threshold to obtain perturbed pattern data. The perturbed pattern data is encoded using a preset variational autoencoder to obtain an encoded representation of the hidden sensitive attributes.

5. The personalized content recommendation method based on multimodal perception according to claim 1, characterized in that, The step of calculating the core feature retention rate based on the unified feature vector, and adjusting the encoding representation of the hidden sensitive attributes to obtain the adjusted encoding version if the core feature retention rate is lower than a preset retention rate threshold, includes: For the unified feature vector set, the initial representation of the core features is determined by principal component analysis, and the core feature retention rate is calculated. If the core feature retention rate is lower than the preset retention rate threshold, the preset variational autoencoder parameters are adjusted using the gradient descent algorithm to obtain the adjusted parameter configuration. The modified parameter configuration is used to adjust the encoding representation of the hidden sensitive attributes to obtain a preliminary encoding representation; Verify the consistency of the business data represented by the preliminary encoding, and further process the noise perturbation of the part related to sensitive attributes to obtain an adjusted encoding version that hides the sensitive attributes.

6. The personalized content recommendation method based on multimodal perception according to claim 1, characterized in that, The iterative optimization of the adjusted encoding version to obtain optimized encoding features includes: Based on the adjusted encoding version, a preset variational autoencoder is used to perform preliminary encoding processing on the multimodal data to obtain preliminary optimized feature representations; To verify the consistency of the feature distribution of the preliminary optimized feature representation, regularization processing is performed through Gaussian distribution sampling to obtain a regularized feature representation; Based on the regularized feature representation, the preset variational autoencoder parameters are iteratively optimized using the gradient descent algorithm to obtain the optimized parameter configuration. Using the optimized parameter configuration, the multimodal data is processed by the variational autoencoder to obtain the optimized coding features.

7. The personalized content recommendation method based on multimodal perception according to claim 1, characterized in that, The step of matching and detecting the optimized coding features with a preset coding feature library to determine a personalized content list and obtain the original recommendation result includes: The optimized coding features are matched with a preset coding feature library, and a recommendation candidate set is obtained through collaborative filtering; Content filtering is used to sort the content in the recommendation candidate set to obtain a personalized content list; The personalized content list is encrypted to obtain the original, accurate recommendation results.

8. The personalized content recommendation method based on multimodal perception according to claim 7, characterized in that, The process of performing privacy protection detection and optimization on the original recommendation results to obtain the final recommendation results includes: Calculate the matching degree between the optimized encoding features and the pre-stored historical recommendation data. If the matching degree is lower than the preset matching degree threshold, enhance the cross-dimensional integration strength through an attention mechanism to obtain an improved recommendation candidate set. The enhanced recommendation candidate set is subjected to protection detection, and the core feature retention rate judgment result is integrated. If the retention rate index is lower than the preset privacy sensitivity threshold after being correlated with the matching degree, privacy protection optimization processing is performed, the weight distribution is adjusted, and a privacy-enhanced candidate version is obtained. Then, the privacy-enhanced candidate version is compared with the encoded representation of the hidden sensitive attribute. If the retention rate is optimized, the final recommendation set with privacy protection optimization is obtained. The final recommendation set optimized for privacy protection is compared with the user's current behavior data to obtain the final recommendation result.

9. A personalized content recommendation system based on multimodal perception, characterized in that, include: The data feature extraction module is used to acquire raw multimodal data, perform preliminary compression processing on the raw multimodal data, and generate a low-dimensional latent representation; The data feature fusion module is used to acquire cross-dimensional related features and perform feature fusion processing to obtain a unified feature vector. The data privacy protection module is used to detect potential patterns where the correlation coefficient of sensitive attributes exceeds a preset sensitivity threshold, and to perform data perturbation processing to obtain the encoded representation of the hidden sensitive attributes; The data adjustment module is used to calculate the core feature retention rate of the encoded representation of hidden sensitive attributes. If the core feature retention rate is lower than the retention rate threshold, the encoding parameters are adjusted and the encoded version of the multimodal data is regenerated to obtain the adjusted encoded version. The data iteration and optimization module is used to iteratively optimize the adjusted encoding version and obtain optimized encoding features; The recommendation generation module is used to match and detect the optimized coding features with a preset coding feature library to determine a personalized content list and obtain the original recommendation result; The privacy protection detection and optimization module is used to perform privacy protection detection and optimization on the original recommendation results to obtain the final recommendation results.

Images