An AI-based intelligent management method and system for electronic files

Abstract

The present invention proposes an AI-based intelligent management method and system for electronic files. The method belongs to the field of information technology. The method includes: extracting multi-level features of file content based on convolutional neural networks and recursive neural networks, and training an adaptive classification model based on the extracted features; constructing user portraits based on the user's historical operation records and behavior habits based on deep neural networks, introducing reinforcement learning mechanisms, and dynamically adjusting recommendation strategies based on user portraits and current contexts. Through convolutional neural networks and recursive neural networks, multi-level features can be extracted from multimodal files such as text, images, and audio, fully mining the intrinsic information of the files, and improving the accuracy of file classification.

Description

Technical Field

[0001] The present invention proposes an AI-based intelligent management method and system for electronic files, belonging to the field of information technology. Background Art

[0002] With the advent of the digital age, the management of electronic documents has become the key to improving work efficiency. Traditional document management methods often rely on manual operations by users, which are cumbersome and inefficient. In addition, most existing electronic document management systems lack intelligence and personalization, and cannot perform intelligent recommendations and classifications based on users' actual needs. Therefore, there is an urgent need to develop an intelligent electronic document management method based on artificial intelligence to solve the pain points in existing technologies. Summary of the invention

[0003] The present invention provides an AI-based electronic file intelligent management method and system to solve the problems mentioned in the above background technology:

[0004] The present invention proposes an AI-based intelligent management method for electronic files, the method comprising:

[0005] S1. Perform multi-level feature extraction on the file content based on convolutional neural network and recurrent neural network, and train an adaptive classification model based on the extracted features;

[0006] S2. Based on the deep neural network, we build user profiles based on the user's historical operation records and behavior habits, introduce reinforcement learning mechanisms, and dynamically adjust the recommendation strategy based on the user profile and current context;

[0007] S3. Automatically generate intelligent folders based on the similarity of file content and user usage habits, dynamically adjust the file organization structure, combine natural language processing technology to achieve semantic-based search functions, and use reinforcement learning to optimize search result ranking;

[0008] S4. Model file access patterns through deep neural networks and monitor abnormal access behaviors in real time. If abnormal behaviors are found, the system triggers an early warning mechanism and automatically takes response measures according to the preset security policies.

[0009] Furthermore, the S1 includes:

[0010] S11, preprocessing the file content before feature extraction, the preprocessing includes:

[0011] Text preprocessing: perform word segmentation, stop word removal, and stem extraction on text files.

[0012] Image preprocessing: scaling, grayscale, and edge detection of image files.

[0013] Audio preprocessing: perform spectrum analysis and Mel-frequency cepstral coefficient extraction on audio files;

[0014] S12. Using convolutional neural network and recursive neural network, perform multi-level feature extraction on the preprocessed file content; including:

[0015] The convolutional neural network is used to perform convolution operations on the preprocessed text to extract the local features of the text; the recurrent neural network is then used to model the text sequence to capture the contextual dependencies of the text.

[0016] The pre-trained convolutional neural network is used to extract features from the image and obtain high-level semantic information of the image.

[0017] The MFCC features of the audio are convolved through a convolutional neural network, and the temporal features of the audio are captured in combination with a recurrent neural network.

[0018] S13. Based on the extracted multi-level features, an adaptive classification model is constructed, and the classification model is trained through a large amount of labeled file data. The extracted features are input into the trained adaptive classification model, and the classification label of each file is output.

[0019] Furthermore, the S13 includes:

[0020] Based on the cross-modal feature fusion mechanism, the features from three different modalities, text, image, and audio, are fused, and the weights of each modal feature in the fusion process are dynamically adjusted through the attention mechanism or gating mechanism;

[0021] On the basis of feature fusion, the local feature extraction capability from convolutional neural network and the context / sequence modeling capability of recurrent neural network are further integrated;

[0022] Based on the results of feature fusion, hyperparameter tuning and model architecture search are performed through deep neural networks; and the prediction results of multiple base classifiers are integrated through random forests;

[0023] Collect and organize large-scale document datasets containing rich label information, and increase data diversity through data enhancement techniques;

[0024] When labeled data is limited, semi-supervised learning or unsupervised learning methods are introduced to improve model performance using unlabeled data;

[0025] Based on the user feedback mechanism, the model is continuously optimized and iterated through feedback on model classification results in actual applications. Active learning is used to select samples for manual labeling based on user feedback, and the model is retrained to form a closed-loop feedback system.

[0026] Furthermore, the cross-modal feature fusion mechanism is based on fusing features from three different modalities, namely text, image and audio, and dynamically adjusting the weights of each modal feature in the fusion process through an attention mechanism or a gating mechanism; including:

[0027] Unify the dimensions of features extracted from text, images, and audio, and normalize the features through Z-score standardization;

[0028] The text, image, and audio features that have been initially aligned and normalized are concatenated to form a joint feature vector containing all modal information;

[0029] Based on the feature fusion network, the correlation and importance between different modal features are dynamically learned through the attention mechanism or self-attention mechanism;

[0030] Through the gating mechanism, the information flow of different modal features in the fusion process is controlled;

[0031] The quality of the fused features is evaluated through a variety of evaluation indicators; the optimal fusion scheme is selected by comparing the evaluation results under different fusion strategies, and the fused features are further optimized based on the evaluation results.

[0032] Furthermore, the S2 includes:

[0033] S21. Obtain the user's historical operation records from the server background, extract features from the user behavior data through a deep neural network, and build a user portrait;

[0034] S22, taking the user profile and current context as the state space of reinforcement learning, and defining the action space of the recommendation system;

[0035] S23, using a reward function, rewarding or punishing the recommendation system according to user behavior, wherein the user behavior includes whether to click on the recommended file and the stay time;

[0036] S24. The recommendation strategy is trained through reinforcement learning algorithm, so that the system can dynamically adjust the recommendation strategy according to user portrait and context.

[0037] Furthermore, the S22 includes:

[0038] Collect and integrate current multi-dimensional context information, based on the user portrait built in S21, combined with real-time user behavior data, and dynamically update the user portrait through online learning technology;

[0039] Combine the above integrated context information with the dynamically updated user profile to form a high-dimensional state space;

[0040] The action space of the recommendation system is defined, and based on the action space, the recommended actions are dynamically adjusted according to the specific context of the user based on the context-aware strategy.

[0041] Furthermore, the S3 includes:

[0042] S31, using cosine similarity to calculate the similarity of the file contents, and combining the user's historical operation records to analyze the user's usage habits;

[0043] S32, automatically generating intelligent folders based on the similarity of file contents and user usage habits, and dynamically adjusting the file organization structure;

[0044] S33, through the GPT pre-trained language model, the query statement entered by the user is semantically understood, the file content is indexed and constructed, and search based on keywords and natural language is supported;

[0045] S34. Sort the search results by a reinforcement learning algorithm so that relevant files are displayed to the user first.

[0046] Further, the S32 includes:

[0047] Based on the cosine similarity calculation of file content similarity, deep semantic analysis technology is used to understand the semantics of file content. By combining file type and content features, a multi-dimensional file similarity matrix is ​​constructed to capture deeper connections between files.

[0048] Conduct in-depth analysis of users' historical operation records and identify potential patterns of user file usage through cluster analysis and association rule mining technology;

[0049] Based on the similarity of file content and user behavior patterns, the preset smart folder generation algorithm automatically classifies related files into the same folder, and introduces a user feedback mechanism to allow users to name, adjust or delete automatically generated folders;

[0050] Dynamically adjust the file organization structure based on the user's usage habits and workflow, and intelligently recommend or hide related folders based on the user's current work situation based on context-aware technology;

[0051] Smart folders are personalized and optimized based on the user's personal preferences and work style. Based on the reinforcement learning algorithm, the folder generation rules and display strategies are continuously adjusted and optimized according to the user's feedback on the use of smart folders (such as click-through rate, dwell time, etc.).

[0052] Furthermore, the S4 includes:

[0053] S41, extracting features from the user's file access records, modeling the user's file access pattern through a deep neural network, and capturing the user's normal access behavior features;

[0054] S42, monitoring the user's file access behavior in real time, and comparing it with the normal access behavior predicted by the model. If the user's access behavior deviates from the normal pattern, it is determined to be abnormal behavior;

[0055] S43. If abnormal behavior is found, the early warning mechanism is immediately triggered, a warning message is sent to the user or administrator, and response measures are automatically taken according to the preset security policy.

[0056] The present invention proposes an AI-based intelligent management system for electronic files, the system comprising:

[0057] Feature extraction module: Perform multi-level feature extraction on file content based on convolutional neural networks and recursive neural networks, and train an adaptive classification model based on the extracted features;

[0058] Strategy adjustment module: Based on the deep neural network, the user profile is constructed according to the user's historical operation records and behavior habits, and the reinforcement learning mechanism is introduced to dynamically adjust the recommendation strategy according to the user profile and current context;

[0059] Sorting optimization module: automatically generates intelligent folders based on the similarity of file content and user usage habits, dynamically adjusts the file organization structure, combines natural language processing technology to achieve semantic-based search functions, and uses reinforcement learning to optimize the sorting of search results;

[0060] Exception handling module: The file access pattern is modeled through deep neural networks to monitor abnormal access behavior in real time. If abnormal behavior is found, the system triggers an early warning mechanism and automatically takes response measures according to the preset security policies.

[0061] The invention has the following beneficial effects: through convolutional neural networks and recursive neural networks, multi-level features can be extracted from multimodal files such as text, images, and audio, the intrinsic information of the files can be fully mined, and the accuracy of file classification can be improved; feature fusion mechanisms (such as cross-modal feature fusion and attention mechanisms) can effectively integrate data of different modalities and improve the accuracy and robustness of classification models; adaptive classification models can be continuously optimized based on a large amount of annotated data to ensure continuous improvement of classification results; through reinforcement learning technology, the recommendation strategy can be intelligently adjusted according to the user's historical behavior and current context information to achieve personalized recommendations. The combination of user portraits and real-time behavior data can accurately capture user preferences and needs; dynamically updated user portraits and context perception capabilities make recommendations more in line with the user's actual work needs and improve the user experience of the file management system; combined with the similarity analysis of file content and user usage habits, intelligent folders are automatically generated, and the file organization structure is dynamically adjusted according to user behavior. This semantic-based file organization method can effectively reduce the user's search cost and file management difficulty; the user feedback mechanism and personalized optimization make the folder generation more in line with the user's workflow, enhancing the flexibility and practicality of the folder system. Based on natural language processing technology, users can quickly locate the required files through semantic understanding and keyword search. Combined with reinforcement learning, the ranking of search results is optimized to ensure that highly relevant files are displayed first, which improves search efficiency and accuracy; the function of dynamically adjusting search ranking allows the system to continuously optimize according to user behavior and feedback, enhancing the intelligence of the system; deep learning modeling and real-time monitoring of file access patterns can efficiently identify abnormal access behaviors and provide timely early warning responses. The system automatically takes appropriate security measures according to security policies to ensure the security of file data; abnormal behavior detection can not only improve the security of the system, but also avoid potential risks caused by human negligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a step diagram of the method of the present invention;

[0063] Figure 2 This is a system module diagram of the present invention. DETAILED DESCRIPTION

[0064] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0065] One embodiment of the present invention, as Figure 1 As shown, an AI-based intelligent management method for electronic files includes:

[0066] S1. Perform multi-level feature extraction on the file content based on convolutional neural network and recurrent neural network, and train an adaptive classification model based on the extracted features;

[0067] S2. Based on the user's historical operation records and behavior habits, a user profile is constructed based on a deep neural network. A reinforcement learning mechanism is introduced to dynamically adjust the recommendation strategy based on the user profile and current context (such as time, location, device status, etc.);

[0068] S3. Automatically generate intelligent folders based on the similarity of file content and user usage habits, dynamically adjust the file organization structure, combine natural language processing technology to achieve semantic-based search functions, and use reinforcement learning to optimize search result ranking;

[0069] S4. Model file access patterns through deep neural networks and monitor abnormal access behaviors in real time. If abnormal behaviors are found, the system triggers an early warning mechanism and automatically takes response measures according to the preset security policies.

[0070] The working principle of the above technical solution is as follows: use convolutional neural networks (CNN) to process images and audio content, extract low-level features of these multimedia contents (such as edges, textures, audio spectra, etc.) through structures such as convolutional layers and pooling layers, and then combine these low-level features into high-level semantic features through fully connected layers; use recurrent neural networks (RNN) or its variants (such as LSTM, GRU) to process text content, and capture the temporal dependencies in the text (such as sentence structure, paragraph coherence, etc.) through the loop structure, thereby extracting the semantic features of the text; input the extracted semantic features of multimedia content into an adaptive classification model. The adaptive classification model learns through training how to automatically adjust the classification strategy according to the complexity of the file content.For example, for complex files containing multiple media types, the model may adopt a more complex classification strategy, while for files with relatively simple content, a simpler classification strategy is adopted; during the training process, the model will continuously adjust its parameters according to the label data (i.e., files with known classifications) until the preset classification accuracy is achieved; based on the deep neural network, the user's historical operation records and behavioral habits are analyzed to extract user preferences, interests, usage habits and other features; these features are combined into a user portrait to represent the user's personality and needs; a reinforcement learning mechanism is introduced to dynamically adjust the recommendation strategy according to the user portrait and the current context (such as time, location, device status, etc.); the reinforcement learning model will continuously try different recommendation strategies and evaluate the effectiveness of these strategies based on user feedback (such as clicks, browsing time, whether to save, etc.); through continuous iteration and optimization, the model will eventually be able to find the recommendation strategy that best suits the user and provide personalized file recommendations; based on the similarity of file content and user usage habits, smart folders are automatically generated, and similar files are clustered together to form a folder through clustering algorithms (such as K-means, DBSCAN, etc.); folders The name of a folder can be automatically generated based on natural language processing technology to reflect the subject or content of the files in the folder; the smart folder will dynamically adjust its structure based on the user's usage habits and the update of the file content; for example, if the files in a folder are often used or accessed together, then these files may be organized more closely together; combined with natural language processing technology, semantic-based search function is realized; users can search for files by entering natural language queries, and the system will match the semantics of the query with the semantics of the file content to find the most relevant files; reinforcement learning is used to optimize the sorting of search results so that the most relevant files are ranked first; file access patterns are modeled through deep neural networks to learn normal access behavior patterns; this includes features such as access time, access frequency, access device, and access user; file access behavior is monitored in real time and compared with normal access patterns; if abnormal behavior is found (such as unauthorized access, frequent data transmission, etc.), an early warning mechanism is triggered; response measures are automatically taken according to preset security policies; this may include locking files to prevent further access, recording logs for subsequent analysis, notifying administrators for manual intervention, etc.

[0071] The effects of the above technical solution are as follows: by performing multi-level feature extraction on file content through convolutional neural network (CNN) and recurrent neural network (RNN), the semantic features of multimedia content such as text, image, and audio can be accurately captured; the adaptive classification model can automatically adjust the classification strategy according to the complexity of the file content, thereby improving the precision and accuracy of classification; the method can handle various types of file content, including text, image, audio, etc., expanding the application scope of electronic file management; through the processing of deep learning models, a large number of files can be efficiently processed to improve the work efficiency of file management; user portraits can be constructed based on the user's historical operation records and behavioral habits to more accurately understand the user's needs and preferences; the introduction of reinforcement learning mechanism, dynamically adjusting the recommendation strategy according to the user portrait and current context, can provide users with personalized file recommendations and improve user satisfaction and experience; personalized recommendations can reduce the time users spend searching for files and improve work efficiency; at the same time, by recommending files that users may be interested in, the utilization of storage resources can be optimized and unnecessary storage waste can be reduced; according to the similarity and Automatically generate smart folders based on user usage habits, which can simplify the file management process and reduce the workload of users to manually organize files; dynamically adjust the file organization structure to make files more orderly and easy to find; combine natural language processing technology to implement semantic-based search functions, so that users can find files through natural language queries, improving the convenience and efficiency of search; use reinforcement learning to optimize the sorting of search results, put the most relevant files in the front, and further improve the accuracy of search; model file access patterns through deep neural networks, and be able to monitor abnormal access behaviors in real time, such as unauthorized access, frequent data transfer, etc.; once abnormal behavior is detected, the system can immediately trigger the early warning mechanism and automatically take response measures according to the preset security policy, such as locking files, recording logs, notifying administrators, etc., so as to effectively ensure the security and integrity of files; real-time monitoring and early warning mechanisms can detect potential security threats in a timely manner and provide administrators with enough time to respond and deal with them; preset security policies can automatically respond to abnormal behaviors, reduce delays and errors in manual intervention, and improve the efficiency and accuracy of emergency response.

[0072] In one embodiment of the present invention, the S1 includes:

[0073] S11, preprocessing the file content before feature extraction, the preprocessing includes:

[0074] Text preprocessing: perform word segmentation, stop word removal, and stem extraction on text files;

[0075] Image preprocessing: scaling, grayscale, and edge detection of image files;

[0076] Audio preprocessing: Perform spectral analysis and Mel-frequency cepstral coefficients extraction operations on the audio file;

[0077] S12. Use a convolutional neural network and a recurrent neural network to perform multi-level feature extraction on the preprocessed file content, including:

[0078] Perform convolution operations on the preprocessed text through a convolutional neural network to extract local features of the text; then combine with a recurrent neural network to model the text sequence and capture the context dependencies of the text;

[0079] Extract features from the image through a pre-trained convolutional neural network to obtain high-level semantic information of the image;

[0080] Perform convolution operations on the MFCC features of the audio through a convolutional neural network, and combine with a recurrent neural network to capture the temporal features of the audio;

[0081] S13. Based on the extracted multi-level features, construct an adaptive classification model, train the classification model with a large amount of labeled file data, input the extracted features into the trained adaptive classification model, and output the classification label of each file.

[0082] The working principle of the above technical solution is: Preprocess the file content, including:

[0083] Text preprocessing:

[0084] Word segmentation: Split the text file into individual words or phrases for subsequent processing.

[0085] Stop word removal: Remove words that are common in the text but contribute little to the meaning of the text, such as "de", "le", etc.

[0086] Stemming: Restore words to their basic forms, such as restoring "running" to "run", to reduce the impact of lexical diversity on subsequent processing.

[0087] Image preprocessing:

[0088] Scaling: Adjust the image size to meet the requirements of subsequent processing or model input.

[0089] Grayscale conversion: Convert the color image to a grayscale image to simplify the image information and reduce the computational load.

[0090] Edge detection: Detect the edges in the image through an algorithm to extract the structural information of the image.

[0091] Audio preprocessing:

[0092] Spectral analysis: Convert the audio signal into a spectral representation for analyzing the frequency components of the audio.

[0093] Mel-frequency cepstral coefficient extraction: Extracts Mel-frequency cepstral coefficients from audio signals. This is a commonly used feature in audio signal processing and can reflect the speech characteristics of the audio.

[0094] Use convolutional neural networks to perform convolution operations on preprocessed text to extract local features of text, such as the combination pattern of characters or words; combine recursive neural networks to model text sequences and capture the contextual dependencies of text, that is, the semantic connections between words; use pre-trained convolutional neural networks (such as VGG, ResNet, etc.) to extract features from preprocessed images and obtain high-level semantic information of images, such as the shape, color, texture, etc. of objects; perform convolution operations on preprocessed audio MFCC features and use CNN to extract local features of audio; combine RNN to capture the temporal features of audio, that is, the law of change of audio signals over time; build an adaptive classification model based on the extracted multi-level features. Use a large amount of annotated file data to train the classification model, including files with known classifications and their corresponding features; during the training process, the model will continuously adjust its parameters to reduce classification errors and improve classification accuracy; input the extracted features into the trained adaptive classification model, and the model will output the classification label of each file based on the input features; the classification label represents the category or theme of the file and can be used for file organization, retrieval and management.

[0095] The effect of the above technical solution is: by preprocessing the file content, noise can be removed, data can be simplified, and subsequent feature extraction can be more accurate. For example, word segmentation, stop word removal, and stem extraction in text preprocessing can help reduce the diversity of vocabulary and make text features more consistent; scaling, graying, and edge detection in image preprocessing can highlight the key information of the image and reduce the interference of irrelevant information; spectrum analysis and MFCC extraction in audio preprocessing can capture the key features of the audio signal and provide a basis for subsequent analysis; the preprocessing process can simplify the data and reduce the amount of data for subsequent processing, thereby improving the overall processing efficiency. For example, image graying can reduce the amount of calculation and speed up the extraction of image features; using convolutional neural networks (CNN) and recurrent neural networks (RNN) to perform multi-level feature extraction on the preprocessed file content can capture the rich features of different types of files such as text, images, and audio. CNN is good at extracting local features, such as character combinations in text, edges and textures in images, and spectral features in audio; while RNN can capture the contextual dependencies of sequence data, such as semantic coherence in text and temporal features in audio; multi-level feature extraction can extract deep features of file content, which are more abstract and generalized, and help improve the generalization ability of the classification model. That is, the model can better handle unseen file data and improve the accuracy and robustness of classification; the adaptive classification model can automatically adjust the classification strategy according to the complexity of the file content, such as adjusting model parameters and selecting different feature combinations. This adaptive ability enables the model to better adapt to file classification tasks of different types and difficulties; by training the classification model with a large amount of annotated file data, the mapping relationship between file features and classification labels can be learned. The trained model can output accurate classification labels after the extracted features are input, thereby improving the accuracy of file classification; the adaptive classification model can automatically classify files, reducing the workload of manual classification and improving the efficiency of file management. At the same time, accurate classification results help users find the required files faster and improve the convenience of file retrieval.

[0096] In one embodiment of the present invention, the S13 includes:

[0097] S131. Based on the cross-modal feature fusion mechanism, the features from three different modalities, namely text, image and audio, are fused, and the weights of the features of each modality in the fusion process are dynamically adjusted through the attention mechanism or the gating mechanism;

[0098] S132. On the basis of feature fusion, further integrate the local feature extraction capability from the convolutional neural network and the context / sequence modeling capability of the recurrent neural network;

[0099] S133. Based on the result of feature fusion, perform hyperparameter tuning and model architecture search (such as network width and depth adjustment) through deep neural network; and integrate the prediction results of multiple base classifiers through random forest;

[0100] S134. Collect and organize large-scale document datasets containing rich label information, and use data enhancement technology to increase data diversity;

[0101] S135. When labeled data is limited, introduce semi-supervised learning or unsupervised learning methods to improve model performance using unlabeled data;

[0102] S136. Based on the user feedback mechanism, the model is continuously optimized and iterated through feedback on model classification results in actual applications. Active learning is used to select the most informative samples for manual labeling based on user feedback, and the model is retrained to form a closed-loop feedback system.

[0103] The working principle of the above technical solution is as follows: considering the heterogeneity between the features of the three different modalities of text, image, and audio, that is, they have different data structures and representation methods, it is necessary to use a cross-modal feature fusion mechanism to fuse these features; in the fusion process, the weights of each modal feature in the fusion process are dynamically adjusted through an attention mechanism or a gating mechanism to fully utilize the complementarity between different modal features while reducing the impact of noise and redundant information; for text, word embedding (such as Word2Vec or BERT) is combined with the CNN-RNN architecture, where word embedding is used to retain the semantic information of vocabulary, CNN is used to extract local features of text, and RNN is used to capture the structural dependencies and context of sentences. The following relationship; for images, advanced CNN models such as residual networks (ResNet) are introduced to extract deep features. These features contain high-level semantic information of the image, and the global context information of the image is combined to improve the accuracy of classification; for audio, deep CNN is used to process Mel-frequency cepstral coefficients (MFCC) to extract the spectral features of the audio, and combined with long short-term memory networks (LSTM) or gated recurrent units (GRU) to capture the temporal dependencies of the audio; based on the results of feature fusion, hyperparameter tuning and model architecture search are performed through deep neural networks to find the optimal model configuration; random forests are used to integrate the prediction results of multiple base classifiers to improve the accuracy and robustness of classification. Random forests reduce the risk of overfitting of a single model by constructing multiple decision trees and combining their prediction results; collect and organize large-scale file data sets containing rich label information, covering multiple types of files such as text, images, and audio. Increase the diversity of data through data enhancement techniques (such as text synonym replacement, image rotation / cropping, audio speed change, etc.) to improve the generalization ability of the model. Data augmentation can generate more training samples and help the model learn more robust feature representations. When labeled data is limited, semi-supervised learning or unsupervised learning methods are introduced to improve model performance using unlabeled data. Semi-supervised learning improves the generalization ability of the model by using the relationship between labeled and unlabeled data. For example, the self-training method uses the model to predict unlabeled data and uses the prediction results with high confidence as new labeled data to train the model. Unsupervised learning discovers the inherent structure and pattern of the data through methods such as cluster analysis to provide guidance for the labeling process. For example, clustering results can be used to preliminarily classify data, thereby accelerating the labeling process. Based on the user feedback mechanism, the model is continuously optimized and iterated through feedback on the model classification results in actual applications. Active learning is used to select the most informative samples for manual labeling based on user feedback, and the model is retrained to form a closed-loop feedback system.Active learning can efficiently utilize limited annotation resources and improve the performance of the model by selecting the most representative samples for annotation. The combination of user feedback and active learning can continuously adjust and optimize the model to make it more adaptable to the needs and changes of actual applications.

[0104] The effects of the above technical solutions are as follows: through the cross-modal feature fusion mechanism, it is possible to make full use of information from different modalities such as text, images, and audio to improve the comprehensiveness and accuracy of the classification model. At the same time, by using the attention mechanism or the gating mechanism to dynamically adjust the weights of each modal feature, it is possible to cope with the heterogeneity between different modal features and further improve the fusion effect; the model can more accurately understand the diversity and complexity of file content and improve the accuracy and robustness of classification; combining the local feature extraction capability of the convolutional neural network and the context / sequence modeling capability of the recursive neural network, it is possible to more deeply mine the feature information of the file content. For different types of files such as text, images, and audio, targeted processing strategies are adopted respectively to better adapt to their respective characteristics; the model can more finely capture the detailed features and temporal relationships of the file content and improve the precision and accuracy of the classification; through deep neural network hyperparameter tuning and model architecture search, the optimal model configuration can be found to improve the performance of the model. At the same time, using random forest to integrate the prediction results of multiple base classifiers can further reduce the risk of overfitting of the model and improve the stability of classification; the model can better adapt to different data distributions and task requirements, and improve the generalization ability and accuracy of classification; collecting and organizing large-scale file data sets containing rich label information can provide sufficient data support for model training. At the same time, increasing the diversity of data through data enhancement technology can further improve the generalization ability of the model; the model can learn richer feature representations and improve the accuracy and robustness of classification; in the case of limited labeled data, introducing semi-supervised learning or unsupervised learning methods can use unlabeled data to improve model performance, reduce the cost of data annotation, and improve the generalization ability of the model; the model can achieve better classification results under limited labeled data, and can process more unlabeled data to improve overall performance; continuous optimization and iteration of the model based on the user feedback mechanism can ensure that the model always adapts to the needs of actual applications. At the same time, using active learning to select the most informative samples for manual annotation and retraining can further improve the performance of the model; the model can continuously adapt to changes in actual applications and improve classification accuracy and user experience. At the same time, through the combination of active learning and user feedback, limited annotation resources can be efficiently utilized to improve the efficiency and effectiveness of model training.

[0105] In one embodiment of the present invention, the S131 includes:

[0106] Unify the dimensions of features extracted from text, images, and audio, and normalize the features through Z-score standardization;

[0107] The text, image, and audio features that have been initially aligned and normalized are concatenated to form a joint feature vector containing all modal information;

[0108] Based on the feature fusion network, the correlation and importance between different modal features are dynamically learned through the attention mechanism or self-attention mechanism;

[0109] Through the gating mechanism, the information flow of different modal features in the fusion process is controlled;

[0110] The quality of the fused features is evaluated through a variety of evaluation indicators. The optimal fusion scheme is selected by comparing the evaluation results under different fusion strategies, and the fusion features are further optimized based on the evaluation results. The quality of the fused features is evaluated through the following formula:

[0111]

[0112] in, , , , represents the weight coefficient, represents the penalty term for feature consistency score, which is used to calculate the sum of inconsistencies between features of different modalities. represents the mutual information between the i-th and j-th modal features, and N represents the number of modalities; represents the discrimination score, represents the inter-class divergence, It represents the intra-class divergence. The larger the inter-class divergence and the smaller the intra-class divergence, the better the feature discrimination. P represents the improvement score of the feature on the task performance. and Represents the minimum and maximum performance scores, which are used to normalize P to the [0,1] interval.

[0113] The working principle of the above technical solution is as follows: the features extracted from text, images, and audio often have different dimensions and dimensions. And they are dimensionalized; the methods of dimensional unification include feature scaling (such as minimum-maximum scaling, Z-score standardization, etc.), principal component analysis (PCA) or autoencoders, etc. These methods can map features of different dimensions to the same dimensional space while retaining the information of the original features as much as possible; Z-score standardization is a commonly used feature scaling method, which calculates the mean and standard deviation of each feature and converts the feature value into a standard normal distribution with 0 as the mean and 1 as the standard deviation, thereby eliminating the dimensional differences between different features; the text, image, and audio features that have been initially aligned and normalized are spliced ​​together to form a joint feature vector containing all modal information; the splicing process is to expand the feature vectors of different modalities in dimensions so that the joint feature vector can simultaneously contain information from different modalities; the formation of the joint feature vector is a key step in realizing cross-modal feature fusion, which provides a rich source of information for the subsequent fusion process; the feature fusion network is a neural network that can learn the correlation and importance between different modal features. Through the network structure; through the attention mechanism or self-attention mechanism, the feature fusion network can dynamically capture the interaction between different modal features and adjust the weights of the features according to these interactions; the attention mechanism focuses on important features by calculating the correlation scores between different features, thereby achieving effective fusion of features; the gating mechanism (such as the gating structure in GRU or LSTM) is used to control the information flow of different modal features during the fusion process; the gating mechanism learns the importance of each modal feature in a specific task through training, and adjusts the weights of the features according to these importance; during the fusion process, the gating mechanism can dynamically control the flow of information to ensure that important features are fully utilized, while redundant or noisy features are suppressed; the quality of the fused features is evaluated through a variety of evaluation indicators; the selection of evaluation indicators depends on the specific application scenario and task requirements. For example, in classification tasks, classification accuracy and F1 score are commonly used evaluation indicators; according to the evaluation results, the optimal fusion scheme is selected, and the fused features are further optimized as needed.

[0114] The effects of the above technical solution are as follows: through feature scaling, principal component analysis (PCA) or autoencoder methods, the features of different modalities are mapped to the same dimensional space and normalized, which effectively solves the problems of heterogeneity and dimensionality difference between features; the features of different modalities can be compared and fused under the same mathematical framework, providing a good foundation for the subsequent feature fusion process; the text, image, and audio features that have been preliminarily aligned and normalized are spliced ​​to form a joint feature vector containing all modal information; this joint feature vector can fully reflect the diversity and complexity of the file content, and provides a rich source of information for subsequent feature fusion and classification tasks; based on the feature fusion network, the correlation and importance between different modal features are dynamically learned through the attention mechanism or self-attention mechanism; it can achieve effective fusion of features, while highlighting important features, suppressing redundant or noise features, and improving the performance and robustness of the model; through the gating mechanism (such as the gating structure in GRU or LSTM) to control the information flow of different modal features in the fusion process; can dynamically adjust the weight of each modal feature under a specific task to ensure that important features are fully utilized, while redundant or noise features are suppressed, thereby improving the accuracy and generalization ability of the model; the quality of the fused features is evaluated through a variety of evaluation indicators, and the optimal fusion scheme is selected based on the evaluation results; it can ensure the quality of the fused features and further optimize them based on the evaluation results, such as removing redundant features through feature selection methods, or further refining effective information in the fused features through feature transformation methods, thereby improving the performance and efficiency of the model; this technical solution is not only suitable for feature fusion of modalities such as text, image, audio, etc., but can also be extended to feature fusion of other modalities, such as video, sensor data, etc.; it can be flexibly applied to various cross-modal tasks, such as multimedia retrieval, sentiment analysis, event detection, etc., and has broad application prospects and scalability. The above formula can effectively measure and optimize the consistency of feature fusion by calculating the sum of the inconsistencies between features of different modalities. The penalty term helps to reduce the conflict of information between modalities and improve the accuracy and reliability of fused features. The discriminability of features can be evaluated by calculating the inter-class divergence and intra-class divergence. The larger the inter-class divergence and the smaller the intra-class divergence, the better the feature discrimination, which is crucial to improving the accuracy of classification tasks. By evaluating the improvement score of features on task performance, the contribution of features to the final task performance can be quantified. and Used to normalize P to the interval [0,1], which helps to fairly compare the performance of different features or fusion strategies; , , , These weight coefficients allow the model to adjust the importance of different scoring items according to different application scenarios and needs, increasing the flexibility and adaptability of the model. The above formula integrates the evaluation of multiple dimensions such as feature consistency, discrimination and performance improvement, providing a comprehensive perspective to evaluate the quality of feature fusion, which helps to select the optimal fusion scheme and further optimize the fusion features based on the evaluation results. Through the attention mechanism or self-attention mechanism, the model can dynamically learn the correlation and importance between different modal features, which helps to extract more critical and relevant information and improve the performance of the model. The gating mechanism allows the information flow of different modal features in the fusion process to be controlled, which helps the model to more accurately process and fuse information from different modalities and improve the fusion effect.

[0115] In one embodiment of the present invention, the S2 includes:

[0116] S21. Obtain the user's historical operation records from the server background, extract features from the user behavior data through a deep neural network, and build a user profile;

[0117] S22, taking the user profile and current context (such as time, location, device status, etc.) as the state space of reinforcement learning, and defining the action space of the recommendation system, such as recommending files, adjusting the recommendation order, etc.;

[0118] S23, using a reward function, rewarding or punishing the recommendation system according to user behavior, wherein the user behavior includes whether to click on the recommended file and the stay time;

[0119] S24. The recommendation strategy is trained through reinforcement learning algorithm, so that the system can dynamically adjust the recommendation strategy according to user portrait and context.

[0120] The working principle of the above technical solution is as follows: the system first obtains the user's historical operation records from the server background, which include key information such as file access frequency, editing time, and file type preference; then, the deep neural network is used to extract features from these user behavior data. Deep neural networks can automatically learn complex patterns and features in data, thereby effectively extracting key information from user behavior; based on the extracted features, the system builds a user profile. The user profile includes the user's interest preferences, work habits, etc., which is a comprehensive description and summary of the user's behavior pattern; the user profile and the current context (such as time, location, device status, etc.) are used as the state space of reinforcement learning. The state space is the basis for the reinforcement learning algorithm to make decisions, which reflects the current environment and user status; the action space of the recommendation system is defined. The action space includes all possible actions that the recommendation system can take, such as recommending files, adjusting the recommendation order, etc. These actions are the decision results made by the recommendation system based on the current state; the reward function is designed to reward or punish the recommendation system based on user behavior. The reward function is the core component of the reinforcement learning algorithm, which determines the reward value obtained after the recommendation system takes a certain action; user behavior includes key indicators such as whether to click on the recommended file and the stay time. When a user clicks on and carefully reads a file provided by the recommendation system, the system will give a positive reward (+1 point), indicating that the recommendation system has made the right decision. If the user ignores the recommended file, the system will give a slight negative reward (-0.1 point), indicating that the recommendation system needs to improve its decision-making strategy; through the design of the reward function, the recommendation system is encouraged to provide files that better meet user needs, thereby improving user satisfaction and the performance of the recommendation system; the recommendation strategy is trained using a reinforcement learning algorithm. The reinforcement learning algorithm gradually finds the optimal recommendation strategy through continuous trial and error and learning; during the training process, the system dynamically adjusts the recommendation strategy based on user profiles and contexts. Through continuous iteration and optimization, the recommendation system can gradually adapt to the behavior patterns and demand changes of different users; ultimately, the trained recommendation system can intelligently recommend files based on user profiles and current contexts, improving user satisfaction and system performance.

[0121] The effects of the above technical solution are: feature extraction of user historical operation records through deep neural networks to build user portraits, including interest preferences, work habits, etc. This enables the recommendation system to deeply understand the personalized needs of users; the system can provide personalized recommendations based on user portraits to improve the accuracy of recommendations and user satisfaction. Users can more easily find files that they are interested in or need, improving work efficiency and experience; the current context (such as time, location, device status, etc.) is used as one of the state spaces of reinforcement learning. This enables the recommendation system to consider more environmental factors and make more intelligent recommendations; the system can dynamically adjust the recommendation strategy according to the user's current environment, such as recommending work-related files during working hours and recommending entertainment content during leisure time. Context-aware recommendations are more in line with the actual needs of users; the recommendation strategy is trained through reinforcement learning algorithms, so that the system can dynamically adjust the recommendation strategy according to user portraits and contexts; the system can continuously learn and optimize the recommendation strategy to adapt to changes in user behavior and new needs. The ability to dynamically adjust makes the recommendation system more intelligent and flexible; a reward function is designed to reward or punish the recommendation system based on user behavior. This provides a feedback mechanism for users, allowing the recommendation system to know the effectiveness of its recommendations; the reward mechanism can motivate the recommendation system to provide files that better meet user needs, improve the accuracy of recommendations and user satisfaction. At the same time, this also promotes the self-optimization and improvement of the recommendation system.

[0122] In one embodiment of the present invention, the S22 includes:

[0123] Collect and integrate current multi-dimensional context information, based on the user portrait built in S21, combined with real-time user behavior data, and dynamically update the user portrait through online learning technology;

[0124] Combine the above integrated context information with the dynamically updated user profile to form a high-dimensional state space;

[0125] The action space of the recommendation system is defined, and based on the action space, the recommended actions are dynamically adjusted according to the specific context of the user based on the context-aware strategy.

[0126] The working principle of the above technical solution is as follows: the system first collects and integrates the current multi-dimensional context information, including time (such as differences in working habits between weekdays and weekends, mornings and nights), location (office, home or mobile may affect file access preferences), device status (such as screen size, whether to connect peripherals, battery power, etc. may affect the convenience of reading or editing) and network environment (such as high-speed network may prefer to load large files, low-speed network recommends lightweight documents); based on the user portrait constructed in S21, the system combines real-time user behavior data (such as new file access, editing behavior and preference changes) to dynamically update the user portrait through online learning technology. The user portrait can reflect the user's latest interests and preferences in real time; the integrated context information is combined with the dynamically updated user portrait to form a high-dimensional state space. The state space not only contains the basic characteristics of the user (such as interest preferences, work habits), but also incorporates the current specific situational information (such as time, location, device status and network environment). This combination provides a rich contextual background for the recommendation strategy, making the recommendation more accurate and personalized; the action space first includes basic recommended file operations, such as recommending specific files, folders, etc. These operations are the basic functions of the recommendation system and can meet the basic needs of users. In addition to basic operations, the action space also introduces more sophisticated and personalized action options. For example, based on the user's current workflow prediction, a collection of related files is recommended (such as templates and commonly used charts when writing reports); intelligent sorting based on file type (putting frequently edited files in the front row); and even providing content-based summary previews or quick editing entrances. These action options are designed to improve the practicality and user experience of recommendations; the action space should also include an immediate response mechanism for user feedback. The system can adjust subsequent recommendation strategies based on whether the user views the recommended content. The immediate response mechanism enables the recommendation system to continuously learn and optimize its recommendation strategy; based on the action space, the system dynamically adjusts the recommended actions based on the specific context of the user based on the context-aware strategy. For example, in the short time before a meeting, short and important document summaries are recommended first; while when working continuously for a long time, more reference materials and in-depth reading materials may be recommended. This adjustment is designed to improve the relevance and practicality of the recommended content and further enhance the user experience.

[0127] The effects of the above technical solution are: by collecting and integrating multi-dimensional contextual information such as time, location, device status and network environment, the system can have a more comprehensive understanding of the user's current environment and situation, so that the recommendation system can more accurately predict the user's behavior and needs, thereby improving the accuracy and relevance of the recommendation; based on the user portrait constructed in S21, combined with real-time user behavior data, the user portrait is dynamically updated through online learning technology; the real-time update of the user portrait enables the system to keep up with changes in user interests and behaviors, and maintain the timeliness and personalization of recommendations; the integrated contextual information is combined with the dynamically updated user portrait to form a high-dimensional state space; the high-dimensional state space provides a rich contextual background for the recommendation strategy, allowing the system to comprehensively consider the user's basic characteristics and current situation and make More intelligent and accurate recommendations: In addition to basic file recommendation operations, more sophisticated and personalized action options are introduced, such as recommending related file collections based on workflow predictions, and intelligent sorting based on file types, which improves the practicality and user experience of recommendations, allowing users to find the required files more easily and improve work efficiency; the action space includes an immediate response mechanism to user feedback, such as adjusting subsequent recommendation strategies based on whether the user views the recommended content; the immediate feedback mechanism enables the system to continuously learn and optimize recommendation strategies to better adapt to the changing needs of users; based on the context-aware strategy, the recommended actions are dynamically adjusted according to the specific context of the user; this dynamic adjustment improves the relevance and practicality of the recommended content, such as recommending short and important file summaries before a meeting, and recommending more reference materials when working for a long time. This context-aware recommendation strategy significantly enhances the user experience.

[0128] In one embodiment of the present invention, S3 includes:

[0129] S31, using cosine similarity to calculate the similarity of the file contents, combining the user's historical operation records to analyze the user's usage habits, the usage habits including common file combinations and access frequencies; the similarity of the file contents is calculated using the following formula:

[0130]

[0131] in, Represents a file and The similarity score between and Respectively represent files and The eigenvalue at the i-th dimension, n represents the dimension of the eigenvector.

[0132] S32, automatically generating intelligent folders according to the similarity of file contents and the user's usage habits, and dynamically adjusting the file organization structure;

[0133] S33, through the GPT pre-trained language model, the query statement entered by the user is semantically understood, the file content is indexed and constructed, and search based on keywords and natural language is supported;

[0134] S34. Sort the search results through a reinforcement learning algorithm so that the most relevant files are displayed to the user first.

[0135] The working principle of the above technical solution is as follows: In the vector space, the contents of two files can be represented as two vectors. The cosine value of the angle between the two vectors can be used to measure the similarity between them. The closer the cosine value is to 1, the more similar the contents of the two files are; the closer the cosine value is to 0, the greater the difference in the contents of the two files is. By analyzing the user's historical operation records, such as opening, editing, and saving files, the user's commonly used file combinations and access frequencies can be understood. This information helps the system understand the user's preferences and behavior patterns, thereby providing users with more personalized services; based on the similarity of file content and the user's usage habits, the system can automatically generate smart folders to classify similar files together. It can improve the accessibility of files and help users better organize and manage their files; as users use and files change, the system can dynamically adjust the file organization structure to ensure the accuracy and practicality of smart folders. This includes adding new files to appropriate folders, updating the contents of existing folders, etc.; the GPT pre-trained language model is used to semantically understand the query statements entered by the user, index the file content, and support searches based on keywords and natural language; GPT is a powerful pre-trained language model with excellent language understanding and generation capabilities. Through the GPT model, the system can accurately understand the semantics and intent of the query statement entered by the user; based on the understanding ability of the GPT model, the system can index and build the file content to quickly retrieve files related to the user's query. This includes extracting key information of the file, generating summaries, etc.; users can search for files by entering keywords or natural language queries. The system will match the most relevant files based on the semantics of the query statement and the index of the file content, and return them to the user; reinforcement learning is a machine learning method that learns the best behavior strategy through continuous trial and error. In the sorting of search results, the reinforcement learning algorithm can optimize the sorting strategy based on the user's historical search behavior and feedback information; the system will sort the search results according to relevance and give priority to displaying the most relevant files to the user. The reinforcement learning algorithm will continuously adjust the sorting strategy based on feedback information such as user click behavior and dwell time to improve the accuracy of search results and user satisfaction.

[0136] The effects of the above technical solution are as follows: through the cosine similarity algorithm, the similarity between file contents can be accurately calculated, thereby helping users to quickly find similar or related files. It not only takes into account the literal similarity of file contents, but also can capture the semantic similarity of file contents to a certain extent, thereby improving the intelligent level of file management; combined with the user's historical operation records, the user's usage habits are deeply analyzed, including commonly used file combinations and access frequencies; personalized analysis helps the system to better understand user needs and provide users with file management solutions that are more in line with their usage habits; according to the similarity of file contents and user usage habits, smart folders can be automatically generated to classify similar files together; it not only reduces the time for users to manually organize files, but also improves the orderliness and clarity of file organization; it can dynamically adjust the file organization structure to adapt to changes in user needs and file content; through dynamic adjustment, the timeliness and accuracy of file organization are ensured, which is effective. It helps users find the files they need quickly and improves work efficiency. Through the GPT pre-trained language model, it can accurately understand the semantics and intentions of the query statements entered by users. It can more accurately match user queries and file content, and improve the accuracy and relevance of search results. This technical solution not only supports traditional keyword-based search methods, but also supports natural language-based search methods. Through diversified search methods, users can query files in a more natural and flexible way, which improves the ease of use and convenience of the search function. Through reinforcement learning algorithms, the search result sorting strategy can be optimized according to the user's historical search behavior and feedback information. This intelligent sorting method enables the most relevant files to be displayed to users first, improving the accuracy of search results and user satisfaction. It can display search results according to the user's usage habits and preferences, so that the search results are more in line with the user's personalized needs. The personalized display method helps to improve the user's search experience and satisfaction.

[0137] In one embodiment of the present invention, the S32 includes:

[0138] Based on the cosine similarity calculation of file content similarity, deep semantic analysis technology is used to understand the semantics of file content. By combining file type and content features, a multi-dimensional file similarity matrix is ​​constructed to capture deeper connections between files.

[0139] Conduct in-depth analysis of users' historical operation records and identify potential patterns of user file usage through cluster analysis and association rule mining technology;

[0140] Based on the similarity of file content and user behavior patterns, the system automatically groups related files into the same folder according to the preset smart folder generation algorithm. It also introduces a user feedback mechanism to allow users to name, adjust or delete automatically generated folders.

[0141] Dynamically adjust the file organization structure based on the user's usage habits and workflow, and intelligently recommend or hide related folders based on the user's current work situation based on context-aware technology;

[0142] Smart folders are personalized and optimized based on the user's personal preferences and work style. Based on the reinforcement learning algorithm, the folder generation rules and display strategies are continuously adjusted and optimized according to the user's feedback on the use of smart folders (such as click-through rate, dwell time, etc.).

[0143] The working principle of the above technical solution is as follows: introduce deep semantic analysis technology to conduct a more detailed semantic understanding of file content and capture the deep-level associations between files; combine file types (documents, pictures, codes, etc.) and content features (keywords, themes, emotions, etc.) to build a multi-dimensional file similarity matrix; deeply analyze the user's historical operation records, pay attention to common file combinations, access frequencies, and file usage habits in different situations; apply cluster analysis and association rule mining technology to identify potential patterns of user file usage, such as file collections related to specific tasks, frequently accessed file paths, etc.; based on file content similarity and user behavior patterns, according to the preset smart folder generation algorithm, automatically group related files classify automatically generated folders into the same folder; introduce a user feedback mechanism to allow users to name, adjust or delete automatically generated folders; dynamically adjust the file organization structure according to the user's usage habits and workflow, such as adjusting the display order of folders, automatically sorting files, etc.; use context-aware technology to intelligently recommend or hide related folders based on the user's current work situation (such as files being edited, etc.); combine the user's personal preferences and work style to personalize and optimize smart folders, such as adjusting the folder's icon, color and other appearance; apply reinforcement learning algorithms to continuously adjust and optimize the folder generation rules and display strategies based on user feedback on the use of smart folders (such as click-through rate, dwell time, etc.).

[0144] The effects of the above technical solutions are as follows: through deep semantic analysis technologies such as BERT or Transformer, it is possible to understand the file content more carefully and capture deeper connections between files; make file classification and search more accurate, and improve the intelligence level of file management; combine file type and content characteristics to build a multi-dimensional file similarity matrix, which is helpful to comprehensively evaluate the similarity between files; multi-dimensional similarity evaluation makes file classification more reasonable, which is convenient for users to quickly find the required files; deeply analyze the user's historical operation records and explore the user's file usage habits in different situations; through cluster analysis and association rule mining technology, identify potential patterns of user file usage, such as file collections related to specific tasks, frequently accessed file paths, etc.; based on the mining results of user behavior patterns, it can intelligently recommend files or folders that users may need; personalized recommendations reduce the time users spend looking for files and improve work efficiency; according to file content similarity and user behavior patterns, automatically generate intelligent Folders, classify related files into the same folder; automatic classification reduces the time users spend on manually organizing files, making the file organization structure clearer; dynamically adjust the file organization structure according to the user's usage habits and workflow, such as adjusting the display order of folders, automatically sorting files, etc.; based on context-aware technology, intelligently recommend or hide related folders to make file access more convenient; introduce a user feedback mechanism to allow users to name, adjust or delete automatically generated folders; this technical solution can continuously learn and optimize to better meet user needs; combine the user's personal preferences and work style to personalize and optimize smart folders, such as adjusting the folder's icon, color and other appearance; based on the reinforcement learning algorithm, continuously adjust and optimize the folder generation rules and display strategies according to the user's feedback on the use of smart folders; this personalized optimization improves the user's visual experience and satisfaction; use the reinforcement learning algorithm to continuously learn and optimize according to the user's feedback on the use of smart folders.

[0145] In one embodiment of the present invention, the S4 includes:

[0146] S41, extracting features from the user's file access records, modeling the user's file access pattern through a deep neural network, and capturing the user's normal access behavior features;

[0147] S42, monitoring the user's file access behavior in real time, and comparing it with the normal access behavior predicted by the model. If the user's access behavior deviates from the normal pattern, it is determined to be abnormal behavior;

[0148] S43. If abnormal behavior is found, the early warning mechanism is immediately triggered, a warning message is sent to the user or administrator, and response measures are automatically taken according to the preset security policy.

[0149] The working principle of the above technical solution is as follows: the system first collects the user's file access records, including key information such as access time, access frequency and accessed file type; these features are extracted and processed through a deep neural network (DNN) to more accurately capture the user's normal access behavior characteristics; the extracted features are learned and trained using a deep neural network to build a model that can reflect the user's normal file access behavior; the model can identify and predict the user's possible file access behavior in a specific situation; the system monitors the user's file access behavior in real time, including the files accessed, the time of access, the method of access, etc.; the real-time data is continuously input into the previously established model for dynamic comparison and analysis; the user access behavior monitored in real time is used as the basis for the analysis of the user's file access behavior. The system compares the user's access behavior with the normal access behavior predicted by the model; if it is found that the user's access behavior deviates from the normal pattern, such as unauthorized access, frequent data transmission and other abnormal behaviors, the system will immediately identify it; once the system detects abnormal behavior, it will immediately trigger the early warning mechanism; the early warning information will be sent to the user or administrator in time to remind them of potential security risks; according to the preset security policy, the system will automatically take response measures to deal with abnormal behavior; measures include locking files, recording logs, notifying administrators, etc. The system will also record abnormal behavior data, including detailed information such as the type of abnormal behavior, time of occurrence, files involved, etc.; these data can be used for subsequent security analysis and improvements, helping the system to more accurately identify potential security risks and optimize the early warning mechanism.

[0150] The effects of the above technical solution are as follows: by modeling the user's file access pattern through a deep neural network, the user's normal access behavior characteristics can be accurately captured; the user's file access behavior is monitored in real time and compared with the normal behavior predicted by the model. Once access behavior that deviates from the normal pattern is found, such as unauthorized access or frequent data transmission, the system will immediately determine it as abnormal behavior; the real-time monitoring and anomaly detection mechanism can quickly identify potential security threats, stop malicious behavior in time, and significantly improve the security of file access behavior; once abnormal behavior is found, the early warning mechanism will be immediately triggered to send early warning information to the user or administrator; according to the preset security policy, the system will automatically take response measures, such as locking files, recording logs, notifying administrators, etc., to prevent abnormal behavior from causing further damage; the timeliness and accuracy of the early warning and response mechanism can effectively reduce security risks and protect user data security; through automated monitoring and early warning mechanisms, the need for manual intervention is reduced; administrators can focus more on handling early warning information and abnormal situations without spending a lot of time manually monitoring users' file access Behavior; automated management not only improves work efficiency, but also reduces the risk of human error; the system can record abnormal behavior data, including the type of abnormal behavior, time of occurrence, files involved and other detailed information; these data can be used for subsequent security analysis and improvement, helping administrators to have a deeper understanding of users' file access behavior and optimize file management strategies; through timely early warning and response mechanisms, security problems that users may encounter can be quickly resolved; users' trust in system security is enhanced, and user satisfaction and loyalty are improved; personalized security strategies are formulated based on users' access behavior characteristics and security needs; personalized security strategies can better meet user needs and improve user experience; in many industries and organizations, data security and privacy protection are compliance requirements that must be followed; through real-time monitoring and early warning mechanisms, organizations can be helped to promptly discover and deal with potential security risks, and ensure that data security and privacy protection comply with relevant laws and regulations and standards; the system can record and generate detailed audit logs and reports, which are of great significance to the organization's compliance audits and internal investigations.

[0151] One embodiment of the present invention, as Figure 2 As shown, an AI-based intelligent electronic file management system includes:

[0152] Feature extraction module: Perform multi-level feature extraction on file content based on convolutional neural networks and recursive neural networks, and train an adaptive classification model based on the extracted features;

[0153] Strategy adjustment module: Based on the deep neural network, the user profile is constructed according to the user's historical operation records and behavior habits, and the reinforcement learning mechanism is introduced to dynamically adjust the recommendation strategy according to the user profile and current context;

[0154] Sorting optimization module: automatically generates intelligent folders based on the similarity of file content and user usage habits, dynamically adjusts the file organization structure, combines natural language processing technology to achieve semantic-based search functions, and uses reinforcement learning to optimize the sorting of search results;

[0155] Exception handling module: The file access pattern is modeled through deep neural networks to monitor abnormal access behavior in real time. If abnormal behavior is found, the system triggers an early warning mechanism and automatically takes response measures according to the preset security policies.

[0156] The working principle of the above technical solution is as follows: use convolutional neural networks to process image and audio content, extract low-level features of these multimedia contents through structures such as convolutional layers and pooling layers, and then combine these low-level features into high-level semantic features through fully connected layers; use recursive neural networks or their variants to process text content, capture the temporal dependencies in the text through a loop structure, and thus extract the semantic features of the text; input the extracted semantic features of multimedia content into an adaptive classification model. The adaptive classification model learns through training how to automatically adjust the classification strategy according to the complexity of the file content. For example, for complex files containing multiple media types, the model may adopt a more complex classification strategy, while for files with relatively simple content, a simpler classification strategy is adopted; during the training process, the model will continuously adjust its parameters according to the label data until the preset classification accuracy is achieved; based on the deep neural network, the user's historical operation records and behavioral habits are analyzed to extract user preferences, interests, usage habits and other features; these features are combined into a user portrait to represent the user's personality and needs; a reinforcement learning mechanism is introduced to dynamically adjust the recommendation strategy according to the user portrait and the current context; the reinforcement learning model will continuously try different recommendation strategies and evaluate the effectiveness of these strategies based on user feedback; through continuous iteration and optimization, the model will eventually be able to find the recommendation strategy that best suits the user and provide personalized file recommendations; based on the similarity of file content and user usage habits, smart folders are automatically generated, and similar files are clustered together to form a folder through a clustering algorithm; the name of the folder can be automatically generated based on natural language processing technology , to reflect the subject or content of the files in the folder; the smart folder will dynamically adjust its structure according to the user's usage habits and the update of the file content; for example, if the files in a folder are often used or accessed together, then these files may be organized more closely together; combined with natural language processing technology, semantic-based search function is realized; users can search for files by entering natural language queries, and the system will match the semantics of the query with the semantics of the file content to find the most relevant files; reinforcement learning is used to optimize the sorting of search results so that the most relevant files are ranked first; file access patterns are modeled through deep neural networks to learn normal access behavior patterns; this includes features such as access time, access frequency, access device, and access user; file access behavior is monitored in real time and compared with normal access patterns; if abnormal behavior is found, an early warning mechanism is triggered; response measures are automatically taken according to preset security policies; this may include locking files to prevent further access, recording logs for subsequent analysis, notifying administrators for manual intervention, etc.

[0157] The effects of the above technical solution are as follows: by performing multi-level feature extraction on file content through convolutional neural networks and recursive neural networks, the semantic features of multimedia content such as text, images, and audio can be accurately captured; the adaptive classification model can automatically adjust the classification strategy according to the complexity of the file content, thereby improving the precision and accuracy of classification; the method can handle various types of file content, including text, images, audio, etc., expanding the application scope of electronic file management; through the processing of deep learning models, a large number of files can be efficiently processed to improve the work efficiency of file management; user portraits can be constructed based on the user's historical operation records and behavioral habits to more accurately understand the user's needs and preferences; the introduction of reinforcement learning mechanism, dynamically adjusting the recommendation strategy according to the user portrait and the current context, can provide users with personalized file recommendations and improve user satisfaction and experience; personalized recommendations can reduce the time users spend searching for files and improve work efficiency; at the same time, by recommending files that users may be interested in, the utilization of storage resources can be optimized and unnecessary storage waste can be reduced; according to the similarity of file content and user usage Automatically generate smart folders, which can simplify the file management process and reduce the workload of users to manually organize files; dynamically adjust the file organization structure to make files more orderly and easy to find; combine natural language processing technology to implement semantic-based search functions, so that users can find files through natural language queries, improving the convenience and efficiency of search; use reinforcement learning to optimize the sorting of search results, put the most relevant files in the front, and further improve the accuracy of search; model file access patterns through deep neural networks, and be able to monitor abnormal access behaviors in real time, such as unauthorized access, frequent data transfer, etc.; once abnormal behavior is detected, the system can immediately trigger the early warning mechanism and automatically take response measures according to the preset security policy, such as locking files, recording logs, notifying administrators, etc., so as to effectively ensure the security and integrity of files; real-time monitoring and early warning mechanisms can detect potential security threats in a timely manner and provide administrators with sufficient time to respond and deal with them; preset security policies can automatically respond to abnormal behaviors, reduce delays and errors in manual intervention, and improve the efficiency and accuracy of emergency response.

[0158] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. An AI-based intelligent management method for electronic files, characterized in that: The method comprises: S1. Perform multi-level feature extraction on the file content based on convolutional neural network and recurrent neural network, and train an adaptive classification model based on the extracted features; S2, based on the deep neural network, build user profiles according to the user's historical operation records and behavior habits, introduce reinforcement learning mechanism, and dynamically adjust the recommendation strategy according to the user profile and current context; S3. Automatically generate intelligent folders based on the similarity of file content and user usage habits, dynamically adjust the file organization structure, combine natural language processing technology to achieve semantic-based search functions, and use reinforcement learning to optimize search result ranking; S4. Model file access patterns through deep neural networks to monitor abnormal access behaviors in real time. If abnormal behaviors are found, the system triggers an early warning mechanism and automatically takes response measures according to the preset security policies. Said S1 comprises: S11, preprocessing the file content before feature extraction, the preprocessing includes: Text preprocessing: perform word segmentation, stop word removal, and stem extraction on text files. Image preprocessing: scaling, grayscale, and edge detection of image files. Audio preprocessing: perform spectrum analysis and Mel-frequency cepstral coefficient extraction on audio files; S12. Using convolutional neural network and recursive neural network, perform multi-level feature extraction on the preprocessed file content; including: The convolutional neural network is used to perform convolution operations on the preprocessed text to extract the local features of the text; the recurrent neural network is then used to model the text sequence to capture the contextual dependencies of the text. The pre-trained convolutional neural network is used to extract features from the image and obtain high-level semantic information of the image. The MFCC features of the audio are convolved through a convolutional neural network, and the temporal features of the audio are captured in combination with a recurrent neural network. S13, based on the extracted multi-level features, construct an adaptive classification model, train the classification model through a large amount of labeled file data, input the extracted features into the trained adaptive classification model, and output the classification label of each file; The S13 comprises: Based on the cross-modal feature fusion mechanism, the features from three different modalities, text, image, and audio, are fused, and the weights of each modal feature in the fusion process are dynamically adjusted through the attention mechanism or gating mechanism; On the basis of feature fusion, the local feature extraction capability from convolutional neural network and the context / sequence modeling capability of recurrent neural network are further integrated; Based on the results of feature fusion, hyperparameter tuning and model architecture search are performed through deep neural networks; and the prediction results of multiple base classifiers are integrated through random forests; Collect and organize large-scale document datasets containing rich label information, and increase data diversity through data enhancement techniques; When labeled data is limited, semi-supervised learning or unsupervised learning methods are introduced to improve model performance using unlabeled data; Based on the user feedback mechanism, the model is continuously optimized and iterated through feedback on model classification results in actual applications. Active learning is used to select samples for manual labeling based on user feedback, and the model is retrained to form a closed-loop feedback system. The cross-modal feature fusion mechanism is based on fusing features from three different modalities: text, image, and audio. The weights of the features of each modality in the fusion process are dynamically adjusted through an attention mechanism or a gating mechanism, including: Unify the dimensions of features extracted from text, images, and audio, and normalize the features through Z-score standardization; The text, image, and audio features that have been initially aligned and normalized are concatenated to form a joint feature vector containing all modal information; Based on the feature fusion network, dynamic learning is performed through the relevance and importance of different modal features through the attention mechanism or self-attention mechanism; Through the gating mechanism, the information flow of different modal features in the fusion process is controlled; The quality of the fused features is evaluated through a variety of evaluation indicators. The optimal fusion scheme is selected by comparing the evaluation results under different fusion strategies, and the fusion features are further optimized based on the evaluation results. The S3 includes: S31, using cosine similarity to calculate the similarity of the file contents, and combining the user's historical operation records to analyze the user's usage habits; S32, automatically generating intelligent folders according to the similarity of file contents and the user's usage habits, and dynamically adjusting the file organization structure; S33, through the GPT pre-trained language model, the query statement entered by the user is semantically understood, the file content is indexed and constructed, and search based on keywords and natural language is supported; S34, sorting the search results by a reinforcement learning algorithm so that relevant files are displayed to the user first; The S32 includes: Based on the cosine similarity calculation of file content similarity, deep semantic analysis technology is used to understand the semantics of file content. By combining file type and content features, a multi-dimensional file similarity matrix is ​​constructed to capture deeper connections between files. Conduct in-depth analysis of users' historical operation records and identify potential patterns of user file usage through cluster analysis and association rule mining technology; Based on the similarity of file content and user behavior patterns, the preset smart folder generation algorithm automatically classifies related files into the same folder, and introduces a user feedback mechanism to allow users to name, adjust or delete automatically generated folders; Dynamically adjust the file organization structure based on the user's usage habits and workflow, and intelligently recommend or hide related folders based on the user's current work situation based on context-aware technology; Smart folders are personalized and optimized based on the user's personal preferences and work style. Based on the reinforcement learning algorithm and according to the user's feedback on the use of smart folders, the folder generation rules and display strategies are continuously adjusted and optimized.

2. According to claim 1, the AI-based intelligent management method for electronic files is characterized in that: The S2 comprises: S21. Obtain the user's historical operation records from the server background, extract features from the user behavior data through a deep neural network, and build a user portrait; S22, taking user profile and current context as the state space of reinforcement learning, and defining the action space of the recommendation system; S23, using a reward function, rewarding or punishing the recommendation system according to user behavior, wherein the user behavior includes whether to click on the recommended file and the stay time; S24. The recommendation strategy is trained through reinforcement learning algorithm, so that the system can dynamically adjust the recommendation strategy according to user portrait and context.

3. According to claim 2, the AI-based intelligent electronic file management method is characterized in that: The S22 comprises: Collect and integrate current multi-dimensional context information, based on the user portrait built in S21, combined with real-time user behavior data, and dynamically update the user portrait through online learning technology; Combine the above integrated context information with the dynamically updated user profile to form a high-dimensional state space; The action space of the recommendation system is defined, and based on the action space, the recommended actions are dynamically adjusted according to the specific context of the user based on the context-aware strategy.

4. According to the AI-based electronic file intelligent management method of claim 1, it is characterized in that: The S4 comprises: S41, extracting features from the user's file access records, modeling the user's file access pattern through a deep neural network, and capturing the user's normal access behavior features; S42, monitoring the user's file access behavior in real time, and comparing it with the normal access behavior predicted by the model. If the user's access behavior deviates from the normal pattern, it is determined to be abnormal behavior; S43. If abnormal behavior is found, the early warning mechanism is immediately triggered, a warning message is sent to the user or administrator, and response measures are automatically taken according to the preset security policy.

5. A system for implementing the AI-based intelligent electronic file management method as claimed in claim 1, characterized in that: The system comprises: Feature extraction module: Perform multi-level feature extraction on file content based on convolutional neural networks and recursive neural networks, and train an adaptive classification model based on the extracted features; Strategy adjustment module: Based on the deep neural network, the user profile is constructed according to the user's historical operation records and behavior habits, and the reinforcement learning mechanism is introduced to dynamically adjust the recommendation strategy according to the user profile and current context; Sorting optimization module: automatically generates intelligent folders based on the similarity of file content and user usage habits, dynamically adjusts the file organization structure, combines natural language processing technology to achieve semantic-based search functions, and uses reinforcement learning to optimize the sorting of search results; Exception handling module: The file access pattern is modeled through deep neural networks to monitor abnormal access behavior in real time. If abnormal behavior is found, the system triggers an early warning mechanism and automatically takes response measures according to the preset security policies.

Images