A real-time interest tag generation method based on sparse coding

Abstract

The present invention discloses a real-time interest tag generation method based on sparse coding, comprising the following steps: S1, collecting and preprocessing multimodal data of users; S2, extracting features from data, performing sparse coding through a sparse feature selection mechanism of multimodal fusion, and generating a unified interest feature representation; S3, applying a multimodal weighted sparse feature selection algorithm to generate key feature vectors; S4, optimizing the matrix using a holographic evolutionary adaptive optimization algorithm, and selecting a global optimal dictionary; S5, generating a global optimal sparse coding solution through a quantum entanglement adaptive optimization algorithm; S6, optimizing individual interest tags in combination with group behavior patterns, and generating personalized interest tags; S7, generating real-time interest tags in a heterogeneous environment through meta-reinforcement learning and transfer learning; S8, using a fuzzy inference system based on a topological quantum field to adjust rules in real time, and generating a final interest tag. The present invention can dynamically respond to changes in user interests and generate personalized interest tags.

Description

Technical Field

[0001] The present invention relates to the technical field of information processing and user behavior analysis, and in particular to a real-time interest tag generation method based on sparse coding. Background Art

[0002] With the rapid development of big data, artificial intelligence and Internet technology, user personalized recommendation systems play an increasingly important role in various application fields. Personalized recommendation provides customized content and services by analyzing user behaviors, interests and preferences, thereby improving user engagement and satisfaction. However, current personalized recommendation methods still have many challenges and shortcomings in processing multimodal data, responding to user interest changes in real time, and accurately generating interest tags.

[0003] Most existing personalized recommendation systems are based on a single data source or a simple fusion method. These methods mainly rely on static or semi-static data processing models, have a partial understanding of user behavior, and lack in-depth data mining and analysis capabilities. This limitation results in the recommendation system often being unable to respond promptly and accurately to users' complex and changing interests, and the recommendation effect is not ideal. Specifically, the existing technology has obvious defects in the following aspects:

[0004] 1. Singleness and limitations of data processing: Most personalized recommendation systems only process single-modal data, such as text, images, or videos, and fail to fully utilize multimodal information in user behavior. This single-modal data processing method ignores the interrelationships between different modal data, resulting in a lack of comprehensiveness and accuracy in recommendation results. For example, user behavior on social platforms may include text comments, video viewing, and picture sharing, but existing systems are usually unable to effectively integrate these multimodal data, which affects the comprehensive understanding of user interests.

[0005] 2. Lack of real-time response capabilities: Existing recommendation systems usually rely on pre-trained models, which have low model update frequencies and cannot dynamically respond to real-time changes in user interests. This causes the system to lag behind the user's actual interest changes when recommending content, and cannot provide accurate real-time recommendation services. In addition, static models are difficult to adapt to the diverse needs of users at different time points or situations, reducing user experience satisfaction.

[0006] 3. Insufficient feature extraction and optimization capabilities: Current recommendation systems often face problems such as insufficient feature extraction and poor optimization algorithm performance when processing large-scale user data. Traditional feature extraction methods mostly rely on manually designed features and lack deep understanding of data and adaptive optimization capabilities. In addition, existing optimization algorithms usually only focus on a single goal and cannot take into account users' multiple needs and behavior patterns, which makes the generated interest tags unable to effectively reflect users' real interests.

[0007] 4. Insufficient understanding of user behavior: Most existing systems use shallow learning methods or simple statistical models, which cannot deeply understand user behavior patterns and potential interests. This shallow analysis can only capture the user's explicit behavior, but lacks in-depth exploration of the user's implicit interests and potential needs, resulting in low accuracy and personalization of recommendation results.

[0008] 5. Lack of adaptability across platforms and heterogeneous environments: Modern user behaviors often span multiple platforms and different application scenarios, but existing recommendation systems lack effective adaptation and integration mechanisms when processing user data in heterogeneous environments. This limitation makes the recommendation system inconsistent in performance on different platforms or application scenarios, and cannot provide a continuous and consistent user experience.

[0009] Therefore, how to provide a real-time interest tag generation method based on sparse coding is a problem that technical personnel in this field urgently need to solve. Summary of the invention

[0010] One object of the present invention is to propose a real-time interest tag generation method based on sparse coding. The present invention makes full use of multimodal data fusion, quantum entanglement algorithm and fuzzy reasoning system of topological quantum field, and describes in detail the algorithm for dynamically generating and optimizing user personalized interest tags. It has the advantages of real-time response to user interest changes, generation of accurate tags, and improvement of personalized recommendation effect.

[0011] A real-time interest tag generation method based on sparse coding according to an embodiment of the present invention includes the following steps:

[0012] S1. Collect and preprocess the user's multimodal data, including text data, image data, audio data and video data;

[0013] S2. Extract features from the preprocessed multimodal data, perform sparse coding on each modality data through the sparse feature selection mechanism of multimodal fusion, extract cross-modal core features, and generate a unified feature representation of interest;

[0014] S3, applying a multimodal weighted sparse feature selection algorithm to the generated unified interest feature representation, performing weighted fusion, adjusting the weight according to the importance of each modality data, and generating a key feature vector;

[0015] S4, obtaining key feature vectors, adopting holographic evolutionary adaptive optimization algorithm, using holographic storage technology to record and retrieve the historical optimization path of the sparse coding dictionary matrix, optimizing the sparse coding dictionary matrix, and selecting the globally optimal sparse coding dictionary;

[0016] S5. Process the optimized sparse coding dictionary matrix through the quantum entanglement adaptive optimization algorithm, construct a quantum entangled state to parallelly calculate multiple sparse coding solutions, dynamically adjust the entangled state, and determine the global optimal sparse coding solution;

[0017] S6. Based on the global optimal sparse coding solution, a user behavior graph is constructed. The real-time collaborative sparse coding method is used to optimize individual interest tags in combination with the behavior patterns of group users, and individual interests are combined with group trends to generate personalized interest tags.

[0018] S7. In multiple heterogeneous environments, use the heterogeneous environment meta-reinforcement learning algorithm to learn the generated personalized interest tags and build a universal representation of the policy network. Transfer the common features in different environments to the new environment through transfer learning, and generate real-time tags that match the user's current interests by fine-tuning the policy network in real time.

[0019] S8. Based on the generated real-time interest tags, a fuzzy reasoning system based on topological quantum fields is used to map user behavior characteristics to fuzzy sets in high-dimensional topological space, construct fuzzy set associations, and use topological protection states and quantum entangled states to process fuzzy rules in parallel. The rules and membership functions are adjusted in real time according to changes in input data to generate the final interest tags.

[0020] Optionally, S2 includes the following steps:

[0021] S21, extracting features from the preprocessed multimodal data to generate high-dimensional feature representations of each modality;

[0022] S22, inputting the extracted high-dimensional feature representation into a dimensionality reduction processing module, using a dimensionality reduction algorithm combining graph embedding and spectral clustering to reduce the high-dimensional features of each modality to a unified feature dimension, and generating a standardized low-dimensional feature representation;

[0023] S23, the standardized low-dimensional feature representation is fused through the multimodal fusion sparse feature selection mechanism to generate a multimodal joint feature matrix ;

[0024] S24, multi-dimensional nested sparse coding cross-iteration algorithm, multi-modal joint feature matrix Perform sparse coding:

[0025] ;

[0026] in, is the sparse coding weight tensor, is the multimodal joint feature matrix, is the characteristic correlation matrix, which is used to capture the linear and nonlinear relationships between modes. is the feature interaction matrix, which is used to extract high-order interaction features. is the Laplace operator, represents the convolution operation, Represents the concatenation operation of matrices or tensors, which is used to combine different feature representations. is the bias tensor, is a nonlinear activation function, is the feature weighting function, Represents element-wise multiplication of tensors;

[0027] S25. The sparse feature representation obtained by sparse coding optimization is used to generate a unified feature representation of interest.

[0028] Optionally, S4 includes the following steps:

[0029] S41, obtaining a key feature vector generated by a multimodal weighted sparse feature selection algorithm, and using the key feature vector as input for optimizing a sparse coding dictionary matrix;

[0030] S42, using the holographic evolutionary adaptive optimization algorithm to optimize the sparse coding dictionary matrix, and using the holographic storage technology to record and construct the historical optimization path of the sparse coding dictionary matrix to form a historical path matrix ;

[0031] S43, sparse feature representation obtained by combining multi-dimensional nested sparse coding cross-iteration algorithm and the historical path matrix , under the topological quantum field optimization mechanism, the sparse coding dictionary matrix is ​​mapped to the high-dimensional topological space for further optimization:

[0032] ;

[0033] in, is the final optimized sparse coding dictionary matrix, Indicated in Tensor multiplication over dimensions, is the historical path matrix, is the topology optimization field state matrix, express norm, which is used to measure the difference between matrices, is the regularization parameter under the topological quantum field optimization mechanism, which is used to control the difference between the dictionary matrix and the topological optimization field state matrix. Representation calculation and The square of the Euclidean distance between two matrices;

[0034] S44, the final sparse coding dictionary matrix obtained by further optimization , used for subsequent sparse coding calculation and interest tag generation process.

[0035] Optionally, S5 includes the following steps:

[0036] S51. Obtain optimized sparse coding dictionary matrix And key eigenvectors, take them as input to initialize the quantum entangled state matrix and the adaptive weight matrix ;

[0037] S52, based on the initial quantum entangled state matrix Through quantum superposition and quantum tunneling effects, parallel computing is performed to generate multiple sparse coding solutions, and through the dynamic quantum entangled state reorganization mechanism, the correlation of quantum entangled states is dynamically adjusted according to the quality of the solution in each iteration, and the quantum state structure is adaptively adjusted to optimize the sparse coding process;

[0038] S53. During the optimization process, quantum noise suppression technology is applied to reduce quantum noise interference during the optimization process through phase adjustment and quantum state protection mechanism;

[0039] S54, combining the quantum entangled state matrix, adaptive weight matrix, historical path matrix and key eigenvectors, gradually converge the sparse coding solution to the optimal solution in the global scope:

[0040] ;

[0041] in, , , To adjust the importance parameters of different optimization items, is the sparse feature representation obtained by the multi-dimensional nested sparse coding cross-iteration algorithm, is the quantum entangled state matrix, is the adaptive weight matrix, is the topology optimization field state matrix, is the historical path matrix, express norm, which is used to measure the difference between matrices, express norm, represents the mixed norm, Represents the dictionary matrix Some kind of projection operation, Indicates that the sparse features and quantum entangled state matrix The energy expression after joint coding is: Indicated in Tensor multiplication over dimensions;

[0042] S55, through the quantum state measurement process, select the final global optimal sparse coding solution matrix from multiple sparse coding solutions calculated in parallel , and apply it in the subsequent interest tag generation process.

[0043] Optionally, the S6 includes the following steps:

[0044] S61, based on the generated global optimal sparse coding solution matrix , construct a user behavior graph, use the quantum walk mechanism to propagate features on the user behavior graph, and generate an updated behavior feature vector ;

[0045] S62: Update the behavior feature vector And the global optimal sparse coding solution matrix Combined with the above, a multi-level dynamic regularization optimization algorithm is used to perform preliminary optimization of personalized interest tags:

[0046] ;

[0047] in, For the individual interest tags obtained by preliminary optimization, A variable representing the interest tags of individual users, is the key feature vector, represents the projection operation on the global optimal sparse coding solution matrix, Representative User The behavioral feature vector of

[0048] S63, using the real-time collaborative sparse coding method, the individual user interest tags obtained by preliminary optimization are Behavior patterns with group users Combine to generate the final personalized interest tag :

[0049] ;

[0050] in, is the multimodal fusion matrix, capturing the association between different modalities. is the bias matrix, is the activation function, used to introduce nonlinear transformation, Represents the convolution operation, which is used for feature extraction. represents tensor multiplication in the third dimension, represents the Hadamard product, which is used to combine sparse coding solutions and group behavior patterns;

[0051] S64: Generate personalized interest tags As input, it is used in the subsequent interest tag generation and strategy learning process.

[0052] Optionally, the S7 includes the following steps:

[0053] S71. Initialize the heterogeneous environment meta-reinforcement learning algorithm in multiple heterogeneous environments and construct a heterogeneous environment feature set ,in Indicates The feature representation of the environment is based on the environment feature set. and the generated personalized interest tags, and initialize the input layer of the strategy network through adaptive quantum state optimization;

[0054] S72. Generate a universal policy network representation in heterogeneous environments by introducing multimodal embedding and dynamic collaborative policy learning mechanism :

[0055] ;

[0056] in, is the adaptive quantum state optimization function, For the The feature weights in heterogeneous environments are For the The quantum entangled state matrix in the environment, Represents the multimodal embedding function, combined with the user's final label and the modal history matrix , For the Personalized interest tags generated in the environment Dynamic collaborative projection operations for key features , used to further enhance the collaborative optimization capabilities of the general strategy network;

[0057] S73. Based on the multimodal quantum state aggregation optimization method, the general strategy network is represented Apply to new set of environmental features In this paper, adaptive quantum state optimization is used to further adjust the generated real-time policy network representation. , and combined with the multimodal dynamic aggregation feature matrix to generate real-time interest tags :

[0058] ;

[0059] in, is the multimodal dynamic aggregation feature matrix in the new environment, is the key feature matrix in the new environment, is the aggregation operation of quantum state and policy network, is the quantum entangled state matrix in the new environment, is the collaborative projection matrix in the new environment, Represents a tensor product operation;

[0060] S74: Generate real-time interest tags It is used as the final label of the user’s current interests and is used in the subsequent personalized service recommendation process.

[0061] The beneficial effects of the present invention are:

[0062] (1) The present invention significantly improves the accuracy and efficiency of data processing and interest feature extraction by combining multimodal data fusion technology and sparse feature selection mechanism. In particular, in terms of unified representation of multimodal data, the present invention overcomes the limitations of traditional single-modal processing and can quickly integrate multiple modal data such as text, images, audio and video, and generate a high-precision unified interest feature representation.

[0063] (2) By introducing the holographic evolutionary adaptive optimization algorithm and the quantum entanglement adaptive optimization algorithm, the present invention optimizes the selection of sparse coding dictionaries and the generation of the global optimal sparse coding solution. The quantum state is dynamically adjusted to achieve the global optimization of sparse coding, ensuring the high efficiency and global optimality in the process of generating personalized interest tags.

[0064] (3) The present invention improves the real-time response and dynamic adjustment capabilities of personalized interest tags through the fuzzy reasoning system of topological quantum fields. In particular, in terms of real-time adjustment of fuzzy rules and membership functions, the present invention can effectively respond to changes in user interests and generate more accurate personalized interest tags, thereby optimizing the user experience and service quality of the personalized recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0066] Figure 1 A flowchart of a real-time interest tag generation method based on sparse coding proposed by the present invention;

[0067] Figure 2 This is a flow chart of optimizing the sparse coding dictionary matrix based on the holographic evolutionary adaptive optimization algorithm proposed by the present invention;

[0068] Figure 3 A flowchart for generating the final interest tag for the fuzzy reasoning system based on topological quantum field proposed in the present invention. DETAILED DESCRIPTION

[0069] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0070] refer to Figure 1-3 , a real-time interest tag generation method based on sparse coding, comprising the following steps:

[0071] S1. Collect and preprocess the user's multimodal data, including text data, image data, audio data and video data;

[0072] S2. Extract features from the preprocessed multimodal data, perform sparse coding on each modality data through the sparse feature selection mechanism of multimodal fusion, extract cross-modal core features, and generate a unified feature representation of interest;

[0073] S3, applying a multimodal weighted sparse feature selection algorithm to the generated unified interest feature representation, performing weighted fusion, adjusting the weight according to the importance of each modality data, and generating a key feature vector;

[0074] S4, obtaining key feature vectors, adopting holographic evolutionary adaptive optimization algorithm, using holographic storage technology to record and retrieve the historical optimization path of the sparse coding dictionary matrix, optimizing the sparse coding dictionary matrix, and selecting the globally optimal sparse coding dictionary;

[0075] S5. Process the optimized sparse coding dictionary matrix through the quantum entanglement adaptive optimization algorithm, construct a quantum entangled state to parallelly calculate multiple sparse coding solutions, dynamically adjust the entangled state, and determine the global optimal sparse coding solution;

[0076] S6. Based on the global optimal sparse coding solution, a user behavior graph is constructed. The real-time collaborative sparse coding method is used to optimize individual interest tags in combination with the behavior patterns of group users, and individual interests are combined with group trends to generate personalized interest tags.

[0077] S7. In multiple heterogeneous environments, use the heterogeneous environment meta-reinforcement learning algorithm to learn the generated personalized interest tags and build a universal representation of the policy network. Transfer the common features in different environments to the new environment through transfer learning, and generate real-time tags that match the user's current interests by fine-tuning the policy network in real time.

[0078] S8. Based on the generated real-time interest tags, a fuzzy reasoning system based on topological quantum fields is used to map user behavior characteristics to fuzzy sets in high-dimensional topological space, construct fuzzy set associations, and use topological protection states and quantum entangled states to process fuzzy rules in parallel. The rules and membership functions are adjusted in real time according to changes in input data to generate the final interest tags.

[0079] In this implementation, S2 includes the following steps:

[0080] S21, extracting features from the preprocessed multimodal data to generate high-dimensional feature representations of each modality;

[0081] S22, inputting the extracted high-dimensional feature representation into a dimensionality reduction processing module, using a dimensionality reduction algorithm combining graph embedding and spectral clustering to reduce the high-dimensional features of each modality to a unified feature dimension, and generating a standardized low-dimensional feature representation;

[0082] S23, the standardized low-dimensional feature representation is fused through the multimodal fusion sparse feature selection mechanism to generate a multimodal joint feature matrix ;

[0083] S24, multi-dimensional nested sparse coding cross-iteration algorithm, multi-modal joint feature matrix Perform sparse coding:

[0084] ;

[0085] in, is the sparse coding weight tensor, is the multimodal joint feature matrix, is the characteristic correlation matrix, which is used to capture the linear and nonlinear relationships between modes. is the feature interaction matrix, which is used to extract high-order interaction features. is the Laplace operator, represents the convolution operation, Represents the concatenation operation of matrices or tensors, which is used to combine different feature representations. is the bias tensor, is a nonlinear activation function, is the feature weighting function, Represents element-wise multiplication of tensors;

[0086] S25. The sparse feature representation obtained by sparse coding optimization is used to generate a unified feature representation of interest.

[0087] In this implementation manner, S4 includes the following steps:

[0088] S41, obtaining a key feature vector generated by a multimodal weighted sparse feature selection algorithm, and using the key feature vector as input for optimizing a sparse coding dictionary matrix;

[0089] S42, using the holographic evolutionary adaptive optimization algorithm to optimize the sparse coding dictionary matrix, and using the holographic storage technology to record and construct the historical optimization path of the sparse coding dictionary matrix to form a historical path matrix ;

[0090] S43, sparse feature representation obtained by combining multi-dimensional nested sparse coding cross-iteration algorithm and the historical path matrix , under the topological quantum field optimization mechanism, the sparse coding dictionary matrix is ​​mapped to the high-dimensional topological space for further optimization:

[0091] ;

[0092] in, is the final optimized sparse coding dictionary matrix, Indicated in Tensor multiplication over dimensions, is the historical path matrix, is the topology optimization field state matrix, express norm, which is used to measure the difference between matrices, is the regularization parameter under the topological quantum field optimization mechanism, which is used to control the difference between the dictionary matrix and the topological optimization field state matrix. Representation calculation and The square of the Euclidean distance between two matrices;

[0093] S44, the final sparse coding dictionary matrix obtained by further optimization , used for subsequent sparse coding calculation and interest tag generation process.

[0094] In this implementation manner, S5 includes the following steps:

[0095] S51. Obtain optimized sparse coding dictionary matrix And key eigenvectors, take them as input to initialize the quantum entangled state matrix and the adaptive weight matrix ;

[0096] S52, based on the initial quantum entangled state matrix Through quantum superposition and quantum tunneling effects, parallel computing is performed to generate multiple sparse coding solutions, and through the dynamic quantum entangled state reorganization mechanism, the correlation of quantum entangled states is dynamically adjusted according to the quality of the solution in each iteration, and the quantum state structure is adaptively adjusted to optimize the sparse coding process;

[0097] S53. During the optimization process, quantum noise suppression technology is applied to reduce quantum noise interference during the optimization process through phase adjustment and quantum state protection mechanism;

[0098] S54, combining the quantum entangled state matrix, adaptive weight matrix, historical path matrix and key eigenvectors, gradually converge the sparse coding solution to the optimal solution in the global scope:

[0099] ;

[0100] in, , , To adjust the importance parameters of different optimization items, is the sparse feature representation obtained by the multi-dimensional nested sparse coding cross-iteration algorithm, is the quantum entangled state matrix, is the adaptive weight matrix, is the topology optimization field state matrix, is the historical path matrix, express norm, which is used to measure the difference between matrices, express norm, represents the mixed norm, Represents the dictionary matrix Some kind of projection operation, Indicates that the sparse features and quantum entangled state matrix The energy expression after joint coding is: Indicated in Tensor multiplication over dimensions;

[0101] S55, through the quantum state measurement process, select the final global optimal sparse coding solution matrix from multiple sparse coding solutions calculated in parallel , and apply it in the subsequent interest tag generation process.

[0102] In this implementation manner, S6 includes the following steps:

[0103] S61, based on the generated global optimal sparse coding solution matrix , construct a user behavior graph, use the quantum walk mechanism to propagate features on the user behavior graph, and generate an updated behavior feature vector ;

[0104] S62: Update the behavior feature vector And the global optimal sparse coding solution matrix Combined with the above, a multi-level dynamic regularization optimization algorithm is used to perform preliminary optimization of personalized interest tags:

[0105] ;

[0106] in, For the individual interest tags obtained by preliminary optimization, A variable representing the interest tags of individual users, is the key feature vector, represents the projection operation on the global optimal sparse coding solution matrix, Representative User The behavioral feature vector of

[0107] S63, using the real-time collaborative sparse coding method, the individual user interest tags obtained by preliminary optimization are Behavior patterns with group users Combine to generate the final personalized interest tag :

[0108] ;

[0109] in, is the multimodal fusion matrix, capturing the association between different modalities. is the bias matrix, is the activation function, used to introduce nonlinear transformation, Represents the convolution operation, which is used for feature extraction. represents tensor multiplication in the third dimension, represents the Hadamard product, which is used to combine sparse coding solutions and group behavior patterns;

[0110] S64: Generate personalized interest tags As input, it is used in the subsequent interest tag generation and strategy learning process.

[0111] In this implementation manner, the S7 includes the following steps:

[0112] S71. Initialize the heterogeneous environment meta-reinforcement learning algorithm in multiple heterogeneous environments and construct a heterogeneous environment feature set ,in Indicates The feature representation of the environment is based on the environment feature set. and the generated personalized interest tags, and initialize the input layer of the strategy network through adaptive quantum state optimization;

[0113] S72. Generate a universal policy network representation in heterogeneous environments by introducing multimodal embedding and dynamic collaborative policy learning mechanism :

[0114] ;

[0115] in, is the adaptive quantum state optimization function, For the The feature weights in heterogeneous environments are For the The quantum entangled state matrix in the environment, Represents the multimodal embedding function, combined with the user's final label and the modal history matrix , For the Personalized interest tags generated in the environment Dynamic collaborative projection operations for key features , used to further enhance the collaborative optimization capabilities of the general strategy network;

[0116] S73. Based on the multimodal quantum state aggregation optimization method, the general strategy network is represented Apply to new set of environmental features In this paper, adaptive quantum state optimization is used to further adjust the generated real-time policy network representation. , and combined with the multimodal dynamic aggregation feature matrix to generate real-time interest tags :

[0117] ;

[0118] in, is the multimodal dynamic aggregation feature matrix in the new environment, is the key feature matrix in the new environment, is the aggregation operation of quantum state and policy network, is the quantum entangled state matrix in the new environment, is the collaborative projection matrix in the new environment, Represents a tensor product operation;

[0119] S74: Generate real-time interest tags It is used as the final label of the user’s current interests and is used in the subsequent personalized service recommendation process.

[0120] Embodiment 1:

[0121] In order to verify the feasibility of the present invention in implementation, the present invention is applied to the recommendation system of a large e-commerce platform located in New York, USA. The platform receives millions of users every day and sells a wide variety of goods, covering multiple categories such as electronic products, clothing, household goods and food. Due to the diversity and real-time changes of user behavior data, it is difficult for traditional recommendation systems to accurately capture the real-time changes in users' interests, resulting in the recommendation results being unable to meet the personalized needs of users. In this embodiment, we use a real-time interest tag generation method based on sparse coding of the present invention to improve the accuracy of the recommendation system and user satisfaction.

[0122] First, the system collects multimodal behavior data of users in real time, including browsed product pages, searched keywords, watched product videos, etc. These data are quickly preprocessed, key features are extracted, and encoded through the sparse feature selection mechanism of multimodal fusion to generate a unified feature representation of interest. This step provides a high-dimensional sparse feature space for the recommendation system, enabling the system to capture the user's potential points of interest in massive data.

[0123] Next, these extracted feature data are mapped into a high-dimensional topological fuzzy space, and the sparse coding dictionary matrix is ​​optimized in real time through the holographic evolutionary adaptive optimization algorithm. This algorithm not only records and refers to the historical optimization path, but also dynamically adjusts the sparse coding dictionary to ensure that the system can accurately reflect the user's current interest status. The core of this process is that the system can quickly adapt to changes in user interests, thereby optimizing recommendation results in real time.

[0124] To further improve the accuracy of recommendations, we also applied a quantum entanglement adaptive optimization algorithm. This algorithm uses the characteristics of quantum entanglement to optimize multiple sparse coding solutions through parallel computing and select the global optimal solution. Finally, the system generates real-time interest tags based on these optimization results and prioritizes the products that best match the user's current interests.

[0125] After implementing the invention, we tested the system for a week with test data from 100,000 active users covering multiple categories such as electronics, clothing, household items, and food. During the test, the system made personalized recommendations based on users' real-time interest tags and compared them with a control group using a traditional recommendation algorithm. The test results show that the real-time interest tag generation method based on sparse coding significantly improves the recommendation effect.

[0126] Table 1: Improvement of user behavior data

[0127]

[0128] Table 2: Improvement of sales indicators

[0129]

[0130] The data shows that in the electronic product category, after using the method of the present invention, the product click-through rate increased from 7.8% to 9.2%, and the purchase conversion rate increased from 2.4% to 2.8%. In the clothing category, the click-through rate increased from 5.6% to 6.5%, and the conversion rate increased from 1.7% to 1.9%. In the household goods category, the click-through rate increased from 6.2% to 7.1%, and the conversion rate increased from 2.1% to 2.5%. In addition, the average user stay time on the platform increased by 10%, from 5 minutes and 40 seconds before the test to 6 minutes and 15 seconds.

[0131] More significantly, the platform's average daily sales increased by 8% compared to before the test, with sales of electronics and clothing categories increasing by 10% and 7% respectively. The user satisfaction survey results also showed that 80% of users were satisfied or very satisfied with the platform's recommendation results, a significant increase from 67% before the implementation.

[0132] These data fully verify the effectiveness of the present invention in practical applications. Through the real-time interest tag generation method based on sparse coding, the platform can capture the changes in user interests in real time and generate personalized recommendations based on them, thereby significantly improving the accuracy of recommendations and user satisfaction. This not only improves the interaction between users and the platform, but also brings higher sales and user retention rates, providing strong support for the long-term development of the platform.

[0133] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A real-time interest tag generation method based on sparse coding, characterized in that: The steps include: S1. Collect and preprocess the user's multimodal data, including text data, image data, audio data and video data; S2. Extract features from the preprocessed multimodal data, perform sparse coding on each modality data through the sparse feature selection mechanism of multimodal fusion, extract cross-modal core features, and generate a unified feature representation of interest; S3, applying a multimodal weighted sparse feature selection algorithm to the generated unified interest feature representation, performing weighted fusion, adjusting the weight according to the importance of each modality data, and generating a key feature vector; S4, obtaining key feature vectors, adopting holographic evolutionary adaptive optimization algorithm, using holographic storage technology to record and retrieve the historical optimization path of the sparse coding dictionary matrix, optimizing the sparse coding dictionary matrix, and selecting the globally optimal sparse coding dictionary; S5. Process the optimized sparse coding dictionary matrix through the quantum entanglement adaptive optimization algorithm, construct a quantum entangled state to parallelly calculate multiple sparse coding solutions, dynamically adjust the entangled state, and determine the global optimal sparse coding solution; S6. Based on the global optimal sparse coding solution, a user behavior graph is constructed. The real-time collaborative sparse coding method is used to optimize individual interest tags in combination with the behavior patterns of group users, and individual interests are combined with group trends to generate personalized interest tags. S7. In multiple heterogeneous environments, use the heterogeneous environment meta-reinforcement learning algorithm to learn the generated personalized interest tags and build a universal representation of the policy network. Transfer the common features in different environments to the new environment through transfer learning, and generate real-time tags that match the user's current interests by fine-tuning the policy network in real time. S8. Based on the generated real-time interest tags, a fuzzy reasoning system based on topological quantum fields is used to map user behavior characteristics to fuzzy sets in high-dimensional topological space, construct fuzzy set associations, and use topological protection states and quantum entangled states to process fuzzy rules in parallel. The rules and membership functions are adjusted in real time according to changes in input data to generate the final interest tags.

2. According to the method for generating real-time interest tags based on sparse coding in claim 1, it is characterized in that: The S2 specifically includes: S21, extracting features from the preprocessed multimodal data to generate high-dimensional feature representations of each modality; S22, inputting the extracted high-dimensional feature representation into a dimensionality reduction processing module, using a dimensionality reduction algorithm combining graph embedding and spectral clustering to reduce the high-dimensional features of each modality to a unified feature dimension, and generating a standardized low-dimensional feature representation; S23, the standardized low-dimensional feature representation is fused through the multimodal fusion sparse feature selection mechanism to generate a multimodal joint feature matrix ; S24, multi-dimensional nested sparse coding cross-iteration algorithm, multi-modal joint feature matrix Perform sparse coding: ； in, is the sparse coding weight tensor, is the multimodal joint feature matrix, is the characteristic correlation matrix, which is used to capture the linear and nonlinear relationships between modes. is the feature interaction matrix, which is used to extract high-order interaction features. is the Laplace operator, represents the convolution operation, Represents the concatenation operation of matrices or tensors, which is used to combine different feature representations. is the bias tensor, is a nonlinear activation function, is the feature weighting function, Represents element-wise multiplication of tensors; S25. The sparse feature representation obtained by sparse coding optimization is used to generate a unified feature representation of interest.

3. According to the method for generating real-time interest tags based on sparse coding according to claim 1, it is characterized in that: The S4 specifically includes: S41, obtaining a key feature vector generated by a multimodal weighted sparse feature selection algorithm, and using the key feature vector as input for optimizing a sparse coding dictionary matrix; S42, using the holographic evolutionary adaptive optimization algorithm to optimize the sparse coding dictionary matrix, and using the holographic storage technology to record and construct the historical optimization path of the sparse coding dictionary matrix to form a historical path matrix ; S43, sparse feature representation obtained by combining multi-dimensional nested sparse coding cross-iteration algorithm and the historical path matrix , under the topological quantum field optimization mechanism, the sparse coding dictionary matrix is ​​mapped to the high-dimensional topological space for further optimization: ； in, is the final optimized sparse coding dictionary matrix, Indicated in Tensor multiplication over dimensions, is the historical path matrix, is the topology optimization field state matrix, express norm, which is used to measure the difference between matrices, is the regularization parameter under the topological quantum field optimization mechanism, which is used to control the difference between the dictionary matrix and the topological optimization field state matrix. Representation calculation and The square of the Euclidean distance between two matrices; S44, the final sparse coding dictionary matrix obtained by further optimization , used for subsequent sparse coding calculation and interest tag generation process.

4. According to the method for generating real-time interest tags based on sparse coding according to claim 1, it is characterized in that: The S5 specifically includes: S51. Obtain optimized sparse coding dictionary matrix And key eigenvectors, take them as input to initialize the quantum entangled state matrix and the adaptive weight matrix ; S52, based on the initial quantum entangled state matrix Through quantum superposition and quantum tunneling effects, parallel computing is performed to generate multiple sparse coding solutions, and through the dynamic quantum entangled state reorganization mechanism, the correlation of quantum entangled states is dynamically adjusted according to the quality of the solution in each iteration, and the quantum state structure is adaptively adjusted to optimize the sparse coding process; S53. During the optimization process, quantum noise suppression technology is applied to reduce quantum noise interference during the optimization process through phase adjustment and quantum state protection mechanism; S54, combining the quantum entangled state matrix, adaptive weight matrix, historical path matrix and key eigenvectors, gradually converge the sparse coding solution to the optimal solution in the global scope: ； in, , , To adjust the importance parameters of different optimization items, is the sparse feature representation obtained by the multi-dimensional nested sparse coding cross-iteration algorithm, is the quantum entangled state matrix, is the adaptive weight matrix, is the topology optimization field state matrix, is the historical path matrix, express norm, which is used to measure the difference between matrices, express norm, represents the mixed norm, Represents the dictionary matrix Some kind of projection operation, Indicates that the sparse features and quantum entangled state matrix The energy expression after joint coding is: Indicated in Tensor multiplication over dimensions; S55, through the quantum state measurement process, select the final global optimal sparse coding solution matrix from multiple sparse coding solutions calculated in parallel , and apply it in the subsequent interest tag generation process.

5. According to the method for generating real-time interest tags based on sparse coding according to claim 4, it is characterized in that: The S6 specifically includes: S61, based on the generated global optimal sparse coding solution matrix , construct a user behavior graph, use the quantum walk mechanism to propagate features on the user behavior graph, and generate an updated behavior feature vector ; S62: Update the behavior feature vector And the global optimal sparse coding solution matrix Combined with the above, a multi-level dynamic regularization optimization algorithm is used to perform preliminary optimization of personalized interest tags: ； in, For the individual interest tags obtained by preliminary optimization, A variable representing the interest tags of individual users, is the key feature vector, represents the projection operation on the global optimal sparse coding solution matrix, Representative User The behavioral feature vector of S63, using the real-time collaborative sparse coding method, the individual user interest tags obtained by preliminary optimization are Behavior patterns with group users Combine to generate the final personalized interest tag : ； in, is the multimodal fusion matrix, capturing the association between different modalities. is the bias matrix, is the activation function, used to introduce nonlinear transformation, Represents the convolution operation, which is used for feature extraction. represents tensor multiplication in the third dimension, represents the Hadamard product, which is used to combine sparse coding solutions and group behavior patterns; S64: Generate personalized interest tags As input, it is used in the subsequent interest tag generation and strategy learning process.

6. The method for generating real-time interest tags based on sparse coding according to claim 4, characterized in that: The S7 specifically includes: S71. Initialize the heterogeneous environment meta-reinforcement learning algorithm in multiple heterogeneous environments and construct a heterogeneous environment feature set ,in Indicates The feature representation of the environment is based on the environment feature set. and the generated personalized interest tags, and initialize the input layer of the strategy network through adaptive quantum state optimization; S72. Generate a universal policy network representation in heterogeneous environments by introducing multimodal embedding and dynamic collaborative policy learning mechanism : ； in, is the adaptive quantum state optimization function, For the The feature weights in heterogeneous environments are For the The quantum entangled state matrix in the environment, Represents the multimodal embedding function, combined with the user's final label and the modal history matrix , For the Personalized interest tags generated in the environment, Dynamic collaborative projection operations for key features , used to further enhance the collaborative optimization capabilities of the general strategy network; S73. Based on the multimodal quantum state aggregation optimization method, the general strategy network is represented Apply to new set of environmental features In this paper, adaptive quantum state optimization is used to further adjust the generated real-time policy network representation. , and combined with the multimodal dynamic aggregation feature matrix to generate real-time interest tags : ； in, is the multimodal dynamic aggregation feature matrix in the new environment, is the key feature matrix in the new environment, is the aggregation operation of quantum state and policy network, is the quantum entangled state matrix in the new environment, is the collaborative projection matrix in the new environment, Represents a tensor product operation; S74: Generate real-time interest tags It is used as the final label of the user’s current interests and is used in the subsequent personalized service recommendation process.

Images