A cloud disk file recommendation method and related device based on behavior memory and scene perception

Abstract

The application discloses a cloud disk file recommendation method based on behavior memory and scene perception and related equipment, and the method comprises the following steps: collecting operation logs and context information in response to terminal operation of a target object; using a preset deep learning model to model the operation logs and corresponding time information of historical records to obtain a long-term behavior preference vector of the target object; performing file recommendation matching according to the long-term behavior preference vector to obtain a candidate file set; performing scene perception through a preset classification model and a rule engine based on the real-time collected context information to obtain a scene label; reordering the files in the candidate file set based on the scene label to obtain an ordered file set; and generating a recommended file package through structured integration based on the ordered file set, and pushing the recommended file package to the target object. The application can dynamically adapt to changes in user demand, significantly improves user experience, and can be widely applied to the technical field of data processing.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a cloud drive file recommendation method and related equipment based on behavioral memory and scene awareness. Background Technology

[0002] With the rapid development of cloud computing and mobile internet, cloud storage, as a core tool for personal and enterprise users to store, synchronize, and share files, has been widely used in daily office work, study, and life. However, as the number of files stored by users continues to increase, traditional cloud storage systems, which mainly rely on manual searching, directory browsing, or simple sorting mechanisms based on "recently accessed" or "most frequently opened" to display files, can no longer meet users' needs for efficiently obtaining the content they need. Summary of the Invention

[0003] The main objective of this invention is to propose a cloud drive file recommendation method, device, electronic device, storage medium, and program product based on behavioral memory and scene awareness, aiming to solve at least one problem of the prior art.

[0004] To achieve the above objectives, one aspect of this invention proposes a cloud drive file recommendation method based on behavioral memory and scene awareness, the method comprising: In response to terminal operations of the target object, the system collects the target object's operation logs and context information; the context information includes operation time, device type, location information, and calendar events. Using a pre-defined deep learning model, behavioral memory modeling is performed on the historical operation logs and their corresponding time information to obtain the long-term behavioral preference vector of the target object; File recommendation matching is performed based on long-term behavioral preference vectors to obtain a candidate file set; Based on real-time collected contextual information, scene perception is performed through a pre-set classification model and rule engine to obtain scene labels; Based on scene tags, the files in the candidate file set are reordered to obtain a sorted file set; Based on the sorted file set, a recommendation file package is generated through structured integration, and the recommendation file package is pushed to the target object.

[0005] In some embodiments, a preset deep learning model is used to model the behavioral memory of historical operation logs and their corresponding time information to obtain a long-term behavioral preference vector of the target object, including the following steps: A historical access sequence of the target object is constructed based on the operation logs of the historical records and their corresponding time information; each sequence element in the historical access sequence is marked with corresponding time information; A Transformer model based on self-attention mechanism is used to perform deep modeling of historical access sequences. By using a multi-head attention mechanism to capture behavioral patterns at different time scales, a long-term behavioral preference vector is obtained.

[0006] In some embodiments, file recommendation matching is performed based on long-term behavioral preference vectors to obtain a candidate file set, including the following steps: The similarity between the long-term behavioral preference vector and the type label vector of all files in the recommendation engine is evaluated. Candidate files are selected based on the similarity scores from highest to lowest in the similarity assessment.

[0007] In some embodiments, scene labels are obtained by scene perception based on real-time collected context information and through a preset classification model and rule engine, including the following steps: Based on real-time acquired contextual information, the LightGBM classification model is used to classify the scene and obtain the first scene type; the first scene type includes the classification probability of each scene type; the LightGBM classification model is obtained by nonlinear modeling based on the contextual information of the known scene types. Based on the operation time and the time tag corresponding to the calendar event, the rule engine performs rule determination. When the result of the rule determination meets the preset rule, the scenario type corresponding to the calendar event is determined to be the second scenario type. Among them, the preset rules include the operation time being within the preset range of the time tag or the time difference between the operation time and the time tag being less than the preset time threshold; Scene tags are obtained by summarizing the first scene type and the second scene type.

[0008] In some embodiments, files in the candidate file set are labeled with a first weight, which is determined based on the matching between the file and the long-term behavioral preference vector. The initial sequence of files in the candidate file set is obtained by sorting the files based on the size of the first weight. Based on the scene label, the files in the candidate file set are reordered to obtain a sorted file set, including the following steps: The scene tag is used as a dynamic weighting factor. The second weight is assigned to each file based on the similarity between the scene type that each file in the candidate file set is adapted to and the scene type corresponding to the scene tag. Among them, the second weight is positively correlated with similarity; The ranking weight of each file in the candidate file set is determined based on the sum of the first weight and the second weight. The files in the candidate file set are reordered according to their sorting weights from largest to smallest to obtain a sorted file set.

[0009] In some embodiments, files in the sorted file set are labeled with file types. Based on the sorted file set, a recommended file package is generated through structured integration, including the following steps: The file that appears first in the sorted file set for each file type is taken as the target file for that file type. In response to the target file type corresponding to the preset file type set, target type files of the same type are aggregated and integrated to generate a recommended file package.

[0010] In some embodiments, the method further includes the following steps: Collect feedback data from the target object regarding the recommended file package; the feedback data includes triggered data and non-triggered data. By using the triggered data as positive samples and the non-triggered data as negative samples, positive and negative samples are constructed. In response to preset conditions, the deep learning model and classification model are iteratively optimized using positive and negative samples; Among them, the preset conditions include when the preset update cycle is met or the amount of data for positive and negative samples exceeds the preset storage capacity.

[0011] To achieve the above objectives, another aspect of this invention proposes a cloud drive file recommendation device based on behavioral memory and scene awareness, the device comprising: The data acquisition module is used to respond to terminal operations of the target object and collect the target object's operation logs and context information; the context information includes operation time, device type, location information, and calendar events. The behavioral memory modeling module is used to model the behavioral memory of historical operation logs and their corresponding time information using a preset deep learning model, so as to obtain the long-term behavioral preference vector of the target object. The recommendation engine module is used to perform file recommendation matching based on long-term behavioral preference vectors to obtain a candidate file set; The scene perception module is used to perceive the scene and obtain scene labels based on the real-time collected context information through a preset classification model and rule engine; The recommendation engine module is also used to reorder the files in the candidate file set based on scene tags to obtain a sorted file set; The file package generation module is used to generate recommended file packages based on a sorted file set through structured integration, and then push the recommended file packages to the target objects.

[0012] In some embodiments, the device further includes a feedback acquisition and optimization module, configured to perform the following operations: Collect feedback data from the target object regarding the recommended file package; the feedback data includes triggered data and non-triggered data. By using the triggered data as positive samples and the non-triggered data as negative samples, positive and negative samples are constructed. In response to preset conditions, the deep learning model and classification model are iteratively optimized using positive and negative samples; Among them, the preset conditions include when the preset update cycle is met or the amount of data for positive and negative samples exceeds the preset storage capacity.

[0013] To achieve the above objectives, another aspect of the present invention provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned method.

[0014] To achieve the above objectives, another aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0015] To achieve the above objectives, another aspect of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method.

[0016] The embodiments of the present invention include at least the following beneficial effects: The present invention provides a cloud disk file recommendation method, device, electronic device, storage medium, and program product based on behavioral memory and scene awareness. This solution collects the operation logs and context information of the target object in response to the terminal operation of the target object; wherein, the context information includes operation time, device type, location information, and calendar events; using a preset deep learning model, behavioral memory modeling is performed on the historical operation logs and their corresponding time information to obtain the long-term behavioral preference vector of the target object; file recommendation matching is performed based on the long-term behavioral preference vector to obtain a candidate file set; based on the real-time collected context information, scene awareness is performed through a preset classification model and rule engine to obtain scene tags; based on the scene tags, the files in the candidate file set are reordered to obtain a sorted file set; based on the sorted file set, a recommended file package is generated through structured integration, and the recommended file package is pushed to the target object. This invention collects user operation logs and contextual information (such as operation time, device type, location information, and calendar events), and uses a deep learning model to model behavioral memory, obtaining a long-term behavioral preference vector. This enables a deep understanding of users' historical behavior and effectively avoids the one-sidedness of traditional methods that rely solely on "recently accessed" or "most frequently opened" data. Furthermore, this invention reorders the candidate file set through scene awareness (using a classification model and rule engine) and generates a recommended file package. This ensures that the recommendation results are not only based on long-term preferences but also fit the real-time scenario, significantly reducing the time users spend searching for files and improving file retrieval efficiency. Specifically, the overall method of this invention combines long-term behavioral modeling and real-time scene analysis, dynamically adapting to changes in user needs and overcoming the shortcomings of existing cloud storage systems' static sorting mechanisms, thus significantly improving the user experience. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of an implementation environment for a cloud disk file recommendation method based on behavioral memory and scene awareness provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating a cloud drive file recommendation method based on behavioral memory and scene awareness provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the unfolding process of step S200 provided in the embodiment of the present invention; Figure 4 This is a schematic diagram of the unfolding process of step S300 provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the unfolding process of step S400 provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the unfolding process of step S500 provided in the embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of a cloud disk file recommendation system based on behavioral memory and scene awareness provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the business process for cloud disk file recommendation based on behavioral memory and scene awareness provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of a cloud disk file recommendation device based on behavioral memory and scene awareness provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of this invention; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this invention as detailed in the appended claims.

[0019] It is understood that the terms “first,” “second,” etc., used in this invention may be used herein to describe various concepts, but unless specifically stated otherwise, these concepts are not limited by these terms. These terms are used only to distinguish one concept from another. For example, first information may also be referred to as second information without departing from the scope of embodiments of the invention, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to determination” as used herein may be interpreted as “when…” or “when…” or “in response to determination.”

[0020] The terms “at least one,” “multiple,” “each,” “any,” etc., used in this invention, “at least one” includes one, two, or more than two; “multiple” includes two or more than two; “each” refers to each of the corresponding multiple; and “any” refers to any one of the multiple.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0022] To facilitate understanding of the technical solution of this invention, the technical terms that may be applied in the technical solution of this invention will first be explained: Cloud Drive: An online file storage and management service based on cloud computing technology, which supports users to synchronize, access and share files across multiple devices.

[0023] Behavior Memory Model: A long-term behavior modeling mechanism based on users' historical operational behaviors (such as file access, download, search, etc.) to characterize users' usage habits and preferences, and supports dynamic updates.

[0024] Scene Awareness: A technical capability that identifies the user's current usage scenario (such as "before a business trip" or "preparing a weekly report") by integrating contextual information such as time, device, calendar events, and geographical location.

[0025] Multimodal data refers to contextual data from different sources with different structural types, including operation logs, file metadata, timestamps, calendar events, device information, etc.

[0026] Recommendation Candidate Set: In the recommendation process, a set of files that may be recommended is initially selected based on the behavioral memory model, and will be re-ranked according to the scenario later.

[0027] File metadata describes file attributes, including filename, type (e.g., .docx, .pptx), size, creation time, modification time, tags, etc.

[0028] Contextual signals reflect real-time information about the user's current state or environment, such as the current time, device, network environment, and calendar events, and are used to support scene recognition.

[0029] In related technologies, traditional cloud storage systems mainly rely on manual searching, directory browsing, or simple sorting mechanisms based on "recently accessed" or "most frequently opened" to display files, which can hardly meet users' needs for efficiently obtaining the content they need.

[0030] In view of this, this invention provides a cloud drive file recommendation method and related device based on behavioral memory and scene awareness. This solution collects the target object's operation logs and context information in response to the target object's terminal operations. The context information includes operation time, device type, location information, and calendar events. Using a preset deep learning model, behavioral memory modeling is performed on the historical operation logs and their corresponding time information to obtain a long-term behavioral preference vector for the target object. File recommendation matching is performed based on the long-term behavioral preference vector to obtain a candidate file set. Based on the real-time collected context information, scene awareness is performed using a preset classification model and rule engine to obtain scene tags. Based on the scene tags, the files in the candidate file set are reordered to obtain a sorted file set. Based on the sorted file set, a recommended file package is generated through structured integration and pushed to the target object. This invention collects user operation logs and contextual information (such as operation time, device type, location information, and calendar events), and uses a deep learning model to model behavioral memory, obtaining a long-term behavioral preference vector. This enables a deep understanding of users' historical behavior and effectively avoids the one-sidedness of traditional methods that rely solely on "recently accessed" or "most frequently opened" data. Furthermore, this invention reorders the candidate file set through scene awareness (using a classification model and rule engine) and generates a recommended file package. This ensures that the recommendation results are not only based on long-term preferences but also fit the real-time scenario, significantly reducing the time users spend searching for files and improving file retrieval efficiency. Specifically, the overall method of this invention combines long-term behavioral modeling and real-time scene analysis, dynamically adapting to changes in user needs and overcoming the shortcomings of existing cloud storage systems' static sorting mechanisms, thus significantly improving the user experience.

[0031] It is understood that the cloud drive file recommendation method based on behavioral memory and scene awareness provided by this invention can be applied to any computer device with data processing and computing capabilities, and this computer device can be various terminals or servers. When the computer device in the embodiment is a server, the server is an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Optionally, the terminal can be a smartphone, tablet, laptop, or desktop computer, but it is not limited to these.

[0032] like Figure 1 The diagram shown is a schematic representation of an implementation environment provided by an embodiment of the present invention. (Refer to...) Figure 1The implementation environment includes at least one terminal 102 and a server 101. The terminal 102 and the server 101 can be connected via a network, either wirelessly or via a wired connection, to complete data transmission and exchange.

[0033] Server 101 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0034] Additionally, server 101 can also be a node server in a blockchain network. Blockchain is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms.

[0035] Terminal 102 can be a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. Terminal 102 and server 101 can be directly or indirectly connected via wired or wireless communication, and this embodiment of the invention does not impose any limitations.

[0036] For example, based on Figure 1 The implementation environment shown in this embodiment of the invention provides a cloud disk file recommendation method based on behavioral memory and scene awareness. The following description uses the application of this cloud disk file recommendation method based on behavioral memory and scene awareness in server 101 as an example. It can be understood that this cloud disk file recommendation method based on behavioral memory and scene awareness can also be applied in terminal 102.

[0037] Reference Figure 2 , Figure 2 This is an optional flowchart of a cloud disk file recommendation method based on behavioral memory and scene awareness provided in an embodiment of the present invention. The executing entity of the cloud disk file recommendation method based on behavioral memory and scene awareness can be any of the aforementioned computer devices (including servers or terminals). Figure 2 The method may include, but is not limited to, steps S100 to S600.

[0038] Step S100: In response to the terminal operation of the target object, collect the operation log and context information of the target object; The context information includes operation time, device type, location information, and calendar events; For example, in some specific implementations, when a user logs in on the terminal and begins operation, the data acquisition module automatically starts, collecting the user's real-time operation logs, including file opening records, recently accessed directory locations, search keywords, etc. Simultaneously, the terminal also collects contextual information, such as the operation time being Friday afternoon, the current device being a PC, and the calendar event containing a "weekly report submission" entry, and uploads this information to the cloud server via an encrypted transmission protocol.

[0039] The data acquisition module runs on the user terminal side, including PC clients, mobile applications, and web browser plugins. This module is responsible for continuously recording the user's file operations, such as opening, editing, sharing, and downloading files, while extracting file metadata such as filename, type, size, modification time, and tags. In addition, this module also collects multi-dimensional contextual signals, including timestamps of user operations, device type, user's geographical location information, and calendar events, and uploads these to the cloud server via encryption protocols to ensure data integrity and security.

[0040] Step S200: Using a preset deep learning model, perform behavioral memory modeling on the historical operation logs and their corresponding time information to obtain the long-term behavioral preference vector of the target object. It should be noted that in some embodiments, such as Figure 3 As shown, step S200 may include the following steps: S210, constructing a historical access sequence of the target object based on the operation log of the historical record and its corresponding time information; wherein, each sequence element in the historical access sequence is labeled with corresponding time information; S220, using a Transformer model based on a self-attention mechanism to perform deep modeling of the historical access sequence, capturing behavioral patterns at different time scales through a multi-head attention mechanism, and obtaining a long-term behavioral preference vector.

[0041] Specifically, the multi-head attention mechanism of the Transformer model used in this embodiment of the invention can capture behavioral patterns at different time scales (such as short-term, medium-term, and long-term), thereby gaining a more comprehensive understanding of user habits and avoiding long-term dependencies that may be ignored by simple ranking mechanisms. Specifically, the self-attention mechanism allows the model to process historical access sequences in parallel and efficiently learn key behaviors in the sequence. Compared with traditional time series models (such as RNNs), it is more suitable for large-scale user data and can effectively improve the scalability and real-time performance of the recommendation system.

[0042] For example, in some specific implementations, the cloud-based behavioral memory modeling module calls the Transformer model to calculate the user's historical operation sequence. The model identifies a behavioral pattern in which the user opened "Work Summary.docx" and "Project Progress.pptx" on multiple Friday afternoons, thus concluding that the user has a long-term behavioral pattern of "periodicly writing weekly reports." This pattern is stored as a behavioral memory vector in the user profile database and is used as one of the input features in this recommendation request.

[0043] The behavioral memory modeling module, deployed in the cloud, is used to characterize users' long-term operating habits. This module employs a Transformer model based on a self-attention mechanism to deeply model users' historical access sequences and their temporal information. Through a multi-head attention mechanism, the model can capture behavioral patterns at different time scales, such as "writing weekly reports every Friday" or "compiling financial statements at the beginning of each month." The modeling output is a vectorized representation of users' long-term behavioral preferences, stored in a user profile database, providing a stable behavioral foundation for subsequent recommendations.

[0044] Step S300: Perform file recommendation matching based on long-term behavioral preference vectors to obtain a candidate file set; It should be noted that in some embodiments, such as Figure 4 As shown, step S300 may include the following steps: S310, performing a similarity assessment between the long-term behavioral preference vector and the type label vectors of all files in the recommendation engine; S320, selecting a candidate file set based on the similarity assessment results from largest to smallest.

[0045] Specifically, this embodiment of the invention evaluates the similarity between long-term behavioral preference vectors and file type label vectors, and filters them from largest to smallest similarity, thereby achieving accurate content-based matching and effectively avoiding the coarse recommendations of traditional methods based solely on access frequency or time. Specifically, the similarity evaluation mechanism can quickly identify the most relevant candidate set from massive files, reducing unnecessary computational overhead, improving recommendation response speed, and enabling users to quickly obtain the files they need.

[0046] For example, in some specific implementations, based on the output of the behavioral memory model, the recommendation engine first generates an initial candidate file set. This set contains documents related to the user's historical weekly report writing, such as "Work Summary.docx", "Project Progress.pptx", and "Weekly Report Template.xlsx", along with files associated with recent projects.

[0047] Step S400: Based on the real-time collected context information, scene perception is performed through a preset classification model and rule engine to obtain scene labels; It should be noted that in some embodiments, such as Figure 5 As shown, step S400 may include the following steps: S410, based on the real-time collected context information, the LightGBM classification model is used to classify the scene to obtain a first scene type; wherein, the first scene type includes the classification probability of each scene type; the LightGBM classification model is obtained by nonlinear modeling based on the context information of the known scene type; S420, based on the operation time and the time tag corresponding to the calendar event, the rule engine performs rule determination, and when the result of the rule determination meets the preset rule, the scene type corresponding to the calendar event is determined as the second scene type; wherein, the preset rule includes the operation time being within the preset range of the time tag or the time difference between the operation time and the time tag being less than the preset time threshold; S430, the scene tag is obtained by summarizing the first scene type and the second scene type.

[0048] In some optional implementations, the implementation logic of step S430 includes, but is not limited to: 1. Selecting one: when the second scene type is empty, the scene type with the highest probability in the first scene type is used as the scene label; when the second scene type is not empty, the second scene type is directly used as the scene label; 2. Weighted determination: when the second scene type is not empty, the probability of the scene type corresponding to the second scene type in the first scene type is weighted (for example, the probability of the relevant scene type is directly increased by a certain value); when the second scene type is not empty, the weighted assignment is 0 (that is, the probability of the scene type in the first scene type is not adjusted), then the scene type with the highest probability in the first scene type can be used as the scene label.

[0049] Specifically, this embodiment of the invention utilizes the LightGBM classification model based on nonlinear modeling, which can efficiently handle complex relationships of contextual information (such as operation time and location) and accurately output scene type probabilities. Simultaneously, the rule engine performs rule determination based on time tags and calendar events, supplementing the model's recognition capabilities under specific events (such as meetings) and effectively avoiding misclassification that might occur with a single model. Furthermore, by combining the model and rules, this embodiment of the invention can quickly adapt to dynamically changing scenarios (such as a user suddenly participating in an activity), ensuring that recommended files are highly relevant to the current scenario, thus improving the practicality of recommendations and user satisfaction.

[0050] For example, in some specific implementations, the scene awareness module receives contextual information from the data acquisition module and analyzes it in conjunction with the LightGBM classification model and the rule engine. The classification model determines that the user is in "office mode" based on time, device, and operation context, while the rule engine further determines that the user's current scenario is "weekly report writing" based on the explicit condition that "time = Friday afternoon and calendar contains 'weekly report'". The final output scene label is "writing weekly report".

[0051] Step S500: Based on scene tags, reorder the files in the candidate file set to obtain a sorted file set; It should be noted that the files in the candidate file set are labeled with a first weight, which is determined based on the matching degree between the file and the long-term behavioral preference vector. The initial sequence of files in the candidate file set is obtained by sorting the files according to the size of the first weight. In some embodiments, such as... Figure 6 As shown, step S500 may include the following steps: S510, using scene tags as dynamic weighting factors, assigning a second weight to each file based on the similarity between the scene type adapted to each file in the candidate file set and the scene type corresponding to the scene tag; wherein, the second weight is positively correlated with similarity; S520, determining the ranking weight of each file in the candidate file set based on the sum of the first weight and the second weight; S530, re-sorting the files in the candidate file set according to the ranking weight from largest to smallest to obtain a ranked file set.

[0052] Specifically, in this embodiment of the invention, a second weight (positively correlated with scene similarity) is assigned to each file based on scene tags, and the final ranking weight is determined by combining the first weight (based on behavioral preferences). This makes the recommendation results consider both long-term preferences and adapt to real-time scenarios, avoiding the rigidity problem of traditional static ranking. Specifically, the re-ranking mechanism in this embodiment of the invention ensures that in specific scenarios (such as work mode vs. leisure mode), the most relevant files are recommended first, reducing the search burden for users in different scenarios and improving the intelligence level of the recommendation system.

[0053] For example, in some specific implementations, reordering can be achieved through the following process steps: 1. Initial state, generating candidate file set: Based on the user's long-term behavioral preferences (e.g., he often processes project reports and meeting minutes on Monday mornings), the system generates an initial candidate file set through similarity matching and performs preliminary sorting according to the degree of matching with long-term preferences (i.e., the first weight).

[0054] 2. Scene Awareness and Scene Tag Acquisition: On a Monday morning at 9:00 AM, a user logged into the cloud drive on their company computer. The system collects contextual information in real time: Operation Time: Monday morning at 9:00 AM (working hours on a weekday); Device Type: Company desktop computer; Calendar Event: The calendar records a "Project A Quarterly Review Meeting" at 9:30 AM.

[0055] The system analyzes the above information using the LightGBM classification model and rule engine. The rule engine finds that the current time is within half an hour before the calendar event "Project A Quarterly Review Meeting," which meets the preset rules. Therefore, the system determines the current scenario label as "About to attend a project review meeting."

[0056] 3. Dynamic Re-ranking and Summarizing of First and Second Weights: The system uses the scene tag "About to Attend a Project Review Meeting" as a dynamic weighting factor to calculate the second weight for each file in the candidate file set. This weight is determined based on the file's similarity (relevance) to the current scene. The system calculates the final weight for each file according to the ranking weight rules and re-ranks them accordingly.

[0057] 4. Final output, sorted file set: After reordering, the final sorted file set can be obtained.

[0058] Step S600: Based on the sorted file set, a recommendation file package is generated through structured integration, and the recommendation file package is pushed to the target object; It should be noted that the files in the sorted file set are labeled with file types. In some embodiments, the recommended file package is generated by structured integration based on the sorted file set, which may include the following steps: taking the top-ranked file of each file type in the sorted file set as the target type file of the corresponding file type; and in response to the target file type corresponding to the preset file type set, summarizing and integrating the target type files of the same type to generate the recommended file package.

[0059] Specifically, this embodiment of the invention integrates the top-ranked files of each file type into a recommended file package, ensuring the comprehensiveness of the recommended content (such as documents, images, videos, etc.) and meeting diverse user needs. This avoids the problem of limited file types that may arise with traditional methods. Furthermore, the recommended file package is presented in a structured format, facilitating quick browsing and access to a group of related files, reducing the time spent manually searching and browsing directories, and thus improving file management efficiency.

[0060] For example, in some specific implementations, generating the recommendation file package can be achieved as follows: 1. Preliminary Steps: Generating a Sorted File Set: Through behavioral memory and scene awareness, the system has identified that the user is currently in a "business trip preparation" scenario (e.g., detected through the calendar event "Traveling to City A for Business Trip" and the location information "Airport"). Furthermore, the system has generated a sorted file set highly relevant to the business trip, with the files tagged by their file type (e.g., contracts, itineraries, images, etc.).

[0061] Example of a sorted file set: 1. Cooperation Agreement with Client in City A.pdf (File Type: Contract); 2. Business Trip Itinerary in City A.xlsx (File Type: Itinerary); 3. Product Introduction Presentation.pptx (File Type: Presentation); 4. Scanned ID Card.jpg (File Type: Document Image); 5. Company Authorization Letter.pdf (File Type: Contract); 6. Hotel Booking Confirmation.pdf (File Type: Itinerary).

[0062] 2. Structured integration to generate a thematic file package: This package responds to a pre-defined set of file types relevant to the "business trip preparation" scenario. This set is not simply a collection of all types, but rather defines the core file types required under the semantic feature of "business trip." For example, the pre-defined target file types are: contracts, itineraries, and document images. The execution steps are as follows: The file that appears first in the sorted file set for each file type is selected as the target file type. Specifically: Regarding contract types, the one listed first is: Cooperation Agreement with Clients in City A.pdf; For itinerary types, the one listed first is: A City Business Trip Itinerary.xlsx; For document image types, the one listed first is: ID card scan.jpg; Files of the same type are aggregated and integrated to generate a recommended file package.

[0063] Specifically, based on the semantic features of the "business trip" usage scenario, this embodiment of the invention intelligently combines these highly related documents (customer contracts, itineraries, identity documents) into a logical whole.

[0064] 3. Final Output, Recommended File Package: The system generates a recommended file package named "Essential File Package for Business Trips in City A". User Interaction: In the cloud drive interface, this "file package" is presented to the user as an independent unit that can be expanded, downloaded, or shared with one click.

[0065] It should be noted that, in some embodiments, the method may further include the following steps: collecting feedback data from the target object to the recommended file package; wherein the feedback data includes triggered data and non-triggered data; constructing positive and negative samples by using the triggered data as positive samples and the non-triggered data as negative samples; and iteratively optimizing the deep learning model and the classification model using the positive and negative samples in response to preset conditions; wherein the preset conditions include satisfying a preset update cycle or the amount of positive and negative sample data exceeding a preset storage capacity.

[0066] Specifically, this embodiment of the invention collects user feedback data on recommended file packages (triggered data is used as positive samples, and non-triggered data is used as negative samples) and uses iterative optimization of deep learning models and classification models, enabling the system to adapt to changes in user preferences and continuously improve recommendation accuracy. In addition, this embodiment of the invention ensures that the recommendation system maintains high performance in long-term use by regularly updating or triggering optimization based on data volume, overcoming the shortcomings of existing cloud disk systems that lack feedback loops, thereby enhancing the reliability of the system and user stickiness.

[0067] For example, in some implementations, when a user chooses to open a recommended file or ignore it, the feedback collection module records the user's behavior and sends it as positive or negative samples to the cloud. In subsequent training, this feedback will be used to optimize the Transformer model and the LightGBM scene classifier, enabling the recommendation results to better adapt to the user's individual needs.

[0068] To explain in detail the principle of the technical solution of the present invention, the overall process of the present invention will be described below with reference to some specific embodiments. It is easy to understand that the following is an explanation of the technical principle of the present invention and should not be regarded as a limitation of the present invention.

[0069] First, it should be noted that some existing cloud storage products have introduced basic recommendation functions, typically based on collaborative filtering or content similarity algorithms, to push files that users may be interested in. However, this type of method has the following technical drawbacks: (1) Lack of context awareness: Most existing recommendation systems rely solely on historical access behavior and fail to combine contextual information such as the user's current time, device, and calendar events. They cannot identify the user's real usage scenario (such as "preparing a weekly report" or "about to go on a business trip"), resulting in a disconnect between the recommendation results and actual needs. (2) Superficial behavioral modeling: Most systems use static or short-term behavioral statistics (such as access frequency in the last 7 days), lacking "behavioral memory" modeling of users' long-term usage habits, making it difficult to capture periodic and regular operation patterns; (3) The recommendation granularity is coarse and the initiative is weak: the recommendation results are mostly a list of single files, lacking the ability to package in a specific context (such as "business trip file package"), and the triggering mechanism is passive, making it impossible to achieve the service upgrade of "proactive prediction + intelligent reminder".

[0070] Therefore, existing cloud drive recommendation technology remains at the "tool-type" stage and has failed to evolve into an "intelligent assistant-type" system. There is an urgent need for an intelligent recommendation method that can deeply integrate user behavior memory and real-time scene perception to improve file retrieval efficiency, reduce user cognitive load, and activate dormant data resources.

[0071] In view of this, the present invention aims to solve the technical problems of existing cloud storage systems in terms of insufficient intelligence, weak scenario understanding, and low user search efficiency in file recommendation. Most current cloud storage products only provide simple recommendation functions based on "recently accessed" or "most frequently opened," lacking in-depth modeling of users' long-term usage habits and failing to dynamically judge based on contextual information such as the user's current time, device, and calendar events. This results in recommendations that are out of touch with actual needs, and limited improvement in user experience.

[0072] In one specific implementation, taking a user opening the cloud storage system via a corporate cloud storage client on a Friday afternoon to prepare for writing a weekly report as an example, this embodiment of the invention implements the following steps: Step 1: Triggering and Data Collection; When a user logs in and begins operation on the terminal, the data collection module automatically starts, collecting the user's real-time operation logs, including file opening records, recently accessed directory locations, search keywords, etc. Simultaneously, the terminal also collects contextual information, such as the operation time being Friday afternoon, the current device being a PC, and the calendar event containing a "weekly report submission" entry, and uploads this information to the cloud server via an encrypted transmission protocol.

[0073] Step Two: Behavioral Memory Modeling; The cloud-based behavioral memory modeling module calls the Transformer model to calculate the user's historical operation sequence. The model identified a behavioral pattern in which the user opened "Work Summary.docx" and "Project Progress.pptx" on multiple Friday afternoons, thus concluding that the user has a long-term behavioral pattern of "periodicly writing weekly reports." This pattern is stored as a behavioral memory vector in the user profile database and used as one of the input features in this recommendation request.

[0074] Step 3: Candidate Set Generation; Based on the output of the behavioral memory model, the recommendation engine first generates a preliminary candidate file set. This set contains documents related to the user's historical weekly report writing, such as "Work Summary.docx", "Project Progress.pptx", and "Weekly Report Template.xlsx", along with files associated with recent projects.

[0075] Step 4: Scene Awareness and Recognition; The scene awareness module receives contextual information from the data acquisition module and analyzes it using the LightGBM classification model and rule engine. The classification model determines that the user is in "office mode" based on time, device, and operation context, while the rule engine further determines the user's current scenario as "weekly report writing" based on the explicit condition "time = Friday afternoon and calendar contains 'weekly report'". The final output scene label is "writing weekly report".

[0076] Step 5: Re-ranking of Recommendation Results; The recommendation engine introduces scenario labels as weighting factors into the candidate set to re-rank the candidate files. In this scenario, "Work Summary.docx" and "Project Progress.pptx," which are strongly related to weekly report writing, are promoted to the top of the recommendation list, while files with lower relevance are demoted.

[0077] Step Six: Generation and Display of Recommendation Results; The sorted results are structured and integrated by the file package generation module. The system combines the recommended files into a "Weekly Report File Package," which includes "Work Summary.docx," "Project Progress.pptx," and "Weekly Report Template.xlsx." This file package is pushed to the user through the cloud drive client interface, supporting one-click opening or batch download.

[0078] Step 7: User Feedback Collection; When a user chooses to open a recommended file or ignore it, the feedback collection module records the user's behavior and sends it as positive and negative samples to the cloud. In subsequent training, this feedback will be used to optimize the Transformer model and the LightGBM scene classifier, enabling the recommendation results to better adapt to the individual needs of users.

[0079] Implementation Results: Through the above process, the system can proactively push highly relevant file packages when a user starts operating the terminal and enters the weekly report writing stage, significantly reducing the time users spend manually searching for files. Compared to traditional recommendation methods based on "recently accessed" or "frequently used files," this embodiment can more accurately identify user intent, improving file retrieval efficiency and user satisfaction with recommendations.

[0080] In another implementation, taking the example of a user launching the system via a mobile cloud storage application before a business trip to quickly access the necessary travel information, this embodiment of the invention implements the following steps: Step 1: Triggering and Data Collection; When a user opens the cloud drive application on their mobile device, the data collection module immediately runs, collecting data related to the current operation. This includes: device type is a smartphone, geolocation information shows the user is near an airport, system time is 10:00 AM, and a "business trip" record is shown in the calendar events. The above information, along with the operation log, is uploaded to the cloud server via an encrypted communication protocol.

[0081] Step Two: Behavioral Memory Modeling; The cloud-based behavioral memory modeling module calls the Transformer model to process the user's historical file access sequences. The system identifies that the user frequently opened "Contract.pdf", "Client Introduction.pptx", and "ID Card Scan.jpg" during past business trips, and infers that these files constitute a set of frequently used files during business trips. The model's output is stored as long-term behavioral preferences and provides a reference for this recommendation.

[0082] Step 3: Candidate Set Generation; The recommendation engine first generates a candidate set containing files such as "Contract.pdf", "Customer Introduction.pptx", and "ID Card Scan.jpg" based on the output of the behavioral memory model. Additionally, to ensure scenario diversity, the candidate set also includes other documents related to recent business dealings, such as "Business Trip Expense Template.xlsx".

[0083] Step 4: Scene Awareness and Recognition; The scene awareness module receives context data uploaded by the terminal and uses the LightGBM classification model to model the input features, determining that the user is highly likely to be in the "business trip preparation" state. Simultaneously, the rule engine matches the condition "geographical location = near the airport and calendar events include 'business trip'", thus directly confirming the current scene as "business trip preparation". The final output scene label is "business trip scene".

[0084] Step 5: Re-ranking of Recommendation Results; The recommendation engine re-ranks the candidate set based on the scene-aware results. Contract documents, itineraries, and identification documents that are highly relevant to the "business trip scenario" are given higher weight and ranked at the top of the recommendation list; other files such as "reimbursement template.xlsx" are ranked in a secondary position.

[0085] Step Six: File Package Generation and Result Display; The file package generation module integrates relevant files according to semantics and function, automatically generating a "Business Trip File Package". This package contains "Contract.pdf", "Client Introduction.pptx", "ID Card Scan.jpg", and "Business Trip Itinerary.xlsx", and is pushed to the user via the mobile cloud drive interface. Users can choose to download with one click or open the package with one click, greatly improving the efficiency of using multiple files collaboratively.

[0086] Step 7: User Feedback Collection; When a user clicks to download or ignores a recommendation, the feedback collection module records the interaction result. Clicking behavior is marked as a positive sample, and ignoring or closing behavior is marked as a negative sample, which is used as training data input for subsequent model optimization. Through feedback loop, the system can gradually improve the accuracy of recommendations for "business trip scenarios" in future recommendations.

[0087] Implementation Results: Through this embodiment, the present invention can proactively provide structured file packages when users are about to travel for business, solving the pain point of traditional recommendation systems where files are scattered and users still need to manually organize them. This solution not only reduces operating costs but also improves the proactive service capabilities and scenario adaptability of the cloud storage system, making the recommendation results more in line with the user's immediate needs.

[0088] In some specific embodiments, the overall architecture of the present invention adopts a closed-loop model of "terminal collection—cloud modeling—intelligent recommendation—feedback optimization". Based on the technical solutions of the embodiments of the present invention, a cloud disk file recommendation system based on behavioral memory and scene awareness can be realized, such as... Figure 7 and Figure 8As shown, its core components include a data acquisition module, a behavior memory modeling module, a scene perception module, a recommendation engine module, a file package generation module, and a feedback collection and optimization module. These modules collaborate through a cloud service interface to achieve real-time modeling and personalized recommendations of user actions and contextual scenarios. Specifically: (1) The data acquisition module runs on the user terminal side, including PC clients, mobile applications, and web browser plugins. This module is responsible for continuously recording the user's file operation behavior, such as opening, editing, sharing, and downloading files, and extracting file metadata such as filename, type, size, modification time, and tags. In addition, this module also collects multi-dimensional context signals, including the timestamp of the user's operation, the type of device used, the user's geographical location information, and calendar events, and uploads them to the cloud server through an encryption protocol to ensure the integrity and security of the data.

[0089] (2) The behavioral memory modeling module is deployed in the cloud to characterize users' long-term operating habits. This module uses a Transformer model based on a self-attention mechanism to perform in-depth modeling of users' historical access sequences and their time information. Through a multi-head attention mechanism, the model can capture behavioral patterns at different time scales, such as "writing weekly reports every Friday" or "organizing financial statements at the beginning of each month". The output of the modeling is a vectorized representation of users' long-term behavioral preferences, which is stored in the user profile database to provide a stable behavioral basis for subsequent recommendations.

[0090] (3) The scene perception module also runs in the cloud and is mainly responsible for identifying the user's real-time usage scenario. This module integrates multi-source contextual information from the data acquisition module, uses the LightGBM classification model to perform non-linear modeling of complex features, and combines it with a rule engine to handle explicit conditions. For example, if the time is Friday afternoon and the calendar event contains "weekly report", the rule engine can directly identify the current scenario as "weekly report writing". Through the combination of model and rules, the system can ensure both efficiency and accuracy, and output scene labels that reflect the user's current intention.

[0091] (4) The recommendation engine module uses the candidate file set output by the behavioral memory model as a basis and the results of the scene perception module as a dynamic weighting factor to reorder the candidate files based on their relevance. In this process, the recommendation results not only depend on the static preferences of long-term behavioral modeling, but can also be adjusted according to the real-time context to achieve personalized recommendations that vary from person to person, time to time, and scene to scene.

[0092] (5) The file package generation module further structures the recommendation results. Based on the content correlation between files and the semantic features of the usage scenario, this module combines multiple related files into a recommended file package. For example, when a user is about to go on a business trip, the system can automatically generate a "business trip file package" containing contracts, itineraries, and scanned copies of documents, and support users to open or download it with one click. This recommendation method effectively solves the problem of fragmented recommendation results in traditional systems and significantly reduces the user's operational burden.

[0093] (6) Finally, the feedback collection and optimization module is responsible for monitoring the user's interaction with the recommendation results, including actions such as clicking, downloading, ignoring, and closing. The relevant feedback data will be used as positive and negative samples to input the behavior modeling and scene recognition modules for iterative training and online optimization of the model. Through continuous closed-loop feedback, the system can continuously adapt to the dynamic changes in user preferences, ensuring the long-term effectiveness and intelligence level of the recommendation results.

[0094] It should be noted that the core of this invention lies in constructing an intelligent and proactive cloud drive file recommendation mechanism by deeply integrating user behavior memory and real-time scene perception. Compared with existing technologies, this invention innovates in several key technical aspects, significantly improving the accuracy, timeliness, and user experience of recommendations. The technical points of this invention's embodiments include, but are not limited to: User Behavior Memory Mechanism Based on Long-Term Behavior Modeling: This invention proposes a dynamically updated user behavior memory model. It utilizes sequence modeling methods (such as RNNs and Transformers) or graph neural networks (GNNs) to model the user's historical file access sequences, capturing their periodic and regular operational patterns (such as editing weekly reports every Friday or organizing financial data at the beginning of each month). Compared to existing static statistical methods that only sort by short-term access frequency or time, this invention can more accurately characterize users' long-term usage habits, avoiding distortion of recommendation results due to short-term behavioral fluctuations. This technique endows the system with "memory capabilities," significantly improving the personalization and stability of recommendations.

[0095] A scene-aware recognition method integrating multi-source context: This invention integrates multi-dimensional contextual signals such as time, device, network, calendar events, and geographic location, and employs a collaborative approach of rule engine and machine learning classification model (such as LightGBM) to accurately identify the user's current usage scenario (e.g., "about to go on a business trip" or "preparing for a meeting"). Most existing technologies rely solely on single behavioral or time information, lacking the ability to understand complex scenarios. This invention, through multimodal data fusion and scene semantic modeling, achieves a technological leap from "behavioral response" to "intent prediction," enabling the recommendation system to proactively adapt to the user's specific context, significantly improving the relevance and usability of recommendations.

[0096] A scenario-driven recommendation generation and re-ranking mechanism: When generating recommendation results, this invention not only outputs a candidate set based on a behavioral memory model, but also introduces scenario semantics as a dynamic weighting factor to re-rank candidate files. For example, in the "before a business trip" scenario, the system automatically increases the recommendation priority of documents such as contracts, PPTs, and scanned copies of ID cards; in the "reimbursement cycle" scenario, it prioritizes recommending invoice templates and approval process documents. This technology breaks through the limitations of traditional recommendation systems that are "one-size-fits-all" or "sorted only by popularity," achieving refined recommendations that are "differentiated by scenario," effectively improving user click-through rates and satisfaction with the recommendation results.

[0097] This invention supports intelligent packaging and recommendation of structured file packages: It further supports combining multiple related files into "recommended file packages" for overall delivery, such as "business trip file packages," "project report packages," and "monthly settlement packages," and provides one-click download or open functionality. This function is based on scene semantics and file correlation analysis, solving the problems of fragmented recommendation results and the need for manual organization by users in existing technologies. This technology significantly reduces user operating costs, improves the efficiency of multi-file collaborative use, and enhances the system's proactive service capabilities.

[0098] A closed-loop feedback mechanism supports continuous model optimization: This invention designs a user feedback collection module to record user actions such as clicking, ignoring, and closing recommendation results, and uses these as positive and negative samples for periodic model updates or online learning. Compared to existing systems that lack a feedback loop and suffer from long-term static model operation, this invention achieves dynamic evolution of the recommendation strategy, ensuring that the system can continuously optimize as user habits change, possessing good adaptability and long-term availability.

[0099] In summary, through the synergistic effect of the aforementioned technical means, this invention effectively solves key technical problems such as low intelligence, lack of scenario understanding, and insufficient initiative in existing cloud drive recommendation systems. It realizes the transformation of file recommendation from "passive search" to "proactive service," significantly improving user search efficiency, file reuse rate, and overall user experience. It has outstanding substantive features and significant progress.

[0100] Specifically, this invention proposes a cloud drive file recommendation method based on behavioral memory and scene awareness. By constructing a user behavior memory model, the system can learn and remember users' periodic operation patterns, such as writing weekly reports and organizing monthly expense reports, forming a long-term characterization of user habits. Simultaneously, the system collects multi-source contextual signals, including time, device type, network environment, and calendar events, and uses a combination of rules and models to identify the user's current specific usage scenario, such as "preparing for a meeting," "about to go on a business trip," or "end-of-month settlement." When generating recommendation results, the system not only relies on historical behavior but also introduces scene semantics as a weighting adjustment factor to reorder candidate files and supports generating structured recommendation content in the form of "file packages," achieving more accurate and proactive intelligent push notifications.

[0101] The core innovation of this invention lies in the deep integration of "behavioral memory" and "scene awareness," breaking through the limitations of traditional recommendation systems that rely solely on static behavioral statistics, and achieving a technological upgrade from passive response to proactive prediction. This method can significantly reduce the time cost for users to find files, improve file reuse rates, reduce storage waste caused by repeated uploads, and enhance the intelligent service capabilities of cloud storage systems. Ultimately, this invention provides a cloud storage recommendation technology solution that better meets users' actual needs and possesses contextual understanding capabilities, demonstrating good practicality, scalability, and industrial application prospects.

[0102] This invention is primarily applied to the file recommendation function of personal cloud storage and enterprise cloud storage systems. Typical use cases include: Office scenario: Users regularly organize work reports, project progress, financial settlements and other documents in the company's internal cloud storage. The system can proactively recommend relevant documents based on historical behavior and time patterns to avoid users searching repeatedly.

[0103] Business travel scenario: Before a business trip, users open the mobile cloud drive. The system can combine calendar events and geolocation information to proactively push a "business trip document package", which includes contracts, itineraries, documents, etc., and supports one-click download.

[0104] Learning scenario: Students organize course materials in their personal cloud storage. The system can automatically push relevant courseware, notes and reference materials before class based on the course schedule and past operating habits.

[0105] By providing automated recommendations in different scenarios, this invention can significantly improve the efficiency of users finding files and reduce the cognitive burden caused by scattered file storage.

[0106] Compared with the prior art, the present invention has at least the following beneficial effects: First, the Transformer-based behavioral memory modeling mechanism proposed in this invention can extract periodic and regular features from users' long-term historical operation sequences. Compared with existing recommendation methods that only rely on short-term access frequency or time ranking, this mechanism has a stronger ability to capture long-term dependencies, enabling the recommendation results to remain stable and personalized, and avoiding distortion due to short-term behavioral fluctuations. Therefore, users no longer need to manually search for repetitive files, significantly reducing operational costs.

[0107] Secondly, this invention integrates a scene perception method with multi-dimensional contextual information, combined with the LightGBM classification model and rule engine. This not only accurately identifies explicit scenes (such as "weekly report writing"), but also infers implicit intentions (such as "business trip preparation") through feature modeling. This method overcomes the limitations of existing recommendation technologies that rely solely on historical behavior and lack real-time adaptability, enabling recommendation systems to possess "intent prediction" capabilities and significantly improving the relevance and immediacy of recommendations.

[0108] Furthermore, the file package generation mechanism proposed in this invention can combine multiple related files and push them to the user based on the semantics of the scenario. Compared with traditional single-file recommendations, this mechanism solves the problems of fragmented recommendation results and the need for users to reorganize them. For example, in a business trip scenario, the system can directly generate a complete file package containing contracts, itineraries, and identification documents, achieving "one-click access" and greatly improving the efficiency of file usage.

[0109] Finally, the feedback collection and optimization mechanism designed in this invention enables the recommendation system to form a complete closed loop. By using user actions such as clicks, downloads, and ignores as feedback samples, the system can continuously learn and dynamically optimize, ensuring that the recommendation model can evolve over time and adapt to changes in user habits. This adaptive capability guarantees the stability and scalability of this invention during long-term use.

[0110] In summary, this invention, through the synergistic effect of modules such as behavioral memory modeling, scene perception, file package recommendation, and feedback optimization, overcomes the bottlenecks of traditional cloud drive recommendation systems, including low intelligence, insufficient scene understanding, and poor recommendation initiative, achieving a technological upgrade from "passive search" to "proactive service." This solution not only significantly improves users' file retrieval efficiency and the system's intelligence level but also possesses promising prospects for industrial application and promotional value, demonstrating outstanding substantive features and significant progress.

[0111] like Figure 9 As shown, this embodiment of the invention also provides a cloud drive file recommendation device 900 based on behavioral memory and scene awareness, which can implement the above-mentioned method. This device may include: The data acquisition module 910 is used to collect the operation logs and context information of the target object in response to the terminal operation of the target object; wherein, the context information includes operation time, device type, location information and calendar events; The behavioral memory modeling module 920 is used to model the behavioral memory of the operation logs of the historical records and their corresponding time information using a preset deep learning model, so as to obtain the long-term behavioral preference vector of the target object. Recommendation engine module 930 is used to perform file recommendation matching based on long-term behavioral preference vectors to obtain a candidate file set; The scene perception module 940 is used to perceive the scene and obtain scene labels based on the real-time collected context information through a preset classification model and rule engine. The recommendation engine module 930 is also used to reorder the files in the candidate file set based on scene tags to obtain a sorted file set; The file package generation module 950 is used to generate recommended file packages based on a sorted file set through structured integration, and then push the recommended file packages to the target objects.

[0112] In some embodiments, the apparatus may further include a feedback acquisition and optimization module for performing the following operations: Collect feedback data from the target object regarding the recommended file package; the feedback data includes triggered data and non-triggered data. By using the triggered data as positive samples and the non-triggered data as negative samples, positive and negative samples are constructed. In response to preset conditions, the deep learning model and classification model are iteratively optimized using positive and negative samples; Among them, the preset conditions include when the preset update cycle is met or the amount of data for positive and negative samples exceeds the preset storage capacity.

[0113] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0114] This invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0115] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0116] like Figure 10 As shown, Figure 10 The hardware structure of an electronic device 1000 according to another embodiment is illustrated. The electronic device 1000 includes: The processor 1001 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (aSIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention. The memory 1002 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RaM). The memory 1002 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1002 and is called and executed by the processor 1001. Input / output interface 1003 is used to implement information input and output; The communication interface 1004 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 1005 transmits information between various components of the device (e.g., processor 1001, memory 1002, input / output interface 1003, and communication interface 1004); The processor 1001, memory 1002, input / output interface 1003 and communication interface 1004 are connected to each other within the device via bus 1005.

[0117] The electronic device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0118] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0119] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0120] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0121] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0122] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0123] The cloud drive file recommendation method, device, electronic device, storage medium, and program product based on behavioral memory and scene awareness provided in this invention embodiment collects the target object's operation logs and context information in response to the target object's terminal operations. The context information includes operation time, device type, location information, and calendar events. Using a preset deep learning model, behavioral memory modeling is performed on the historical operation logs and their corresponding time information to obtain a long-term behavioral preference vector for the target object. File recommendation matching is performed based on the long-term behavioral preference vector to obtain a candidate file set. Based on the real-time collected context information, scene awareness is performed using a preset classification model and rule engine to obtain scene tags. Based on the scene tags, the files in the candidate file set are reordered to obtain a sorted file set. Based on the sorted file set, a recommended file package is generated through structured integration and pushed to the target object. This invention collects user operation logs and contextual information (such as operation time, device type, location information, and calendar events), and uses a deep learning model to model behavioral memory, obtaining a long-term behavioral preference vector. This enables a deep understanding of users' historical behavior and effectively avoids the one-sidedness of traditional methods that rely solely on "recently accessed" or "most frequently opened" data. Furthermore, this invention reorders the candidate file set through scene awareness (using a classification model and rule engine) and generates a recommended file package. This ensures that the recommendation results are not only based on long-term preferences but also fit the real-time scenario, significantly reducing the time users spend searching for files and improving file retrieval efficiency. Specifically, the overall method of this invention combines long-term behavioral modeling and real-time scene analysis, dynamically adapting to changes in user needs and overcoming the shortcomings of existing cloud storage systems' static sorting mechanisms, thus significantly improving the user experience.

[0124] The embodiments described in this invention are for the purpose of more clearly illustrating the technical solutions of the embodiments of this invention, and do not constitute a limitation on the technical solutions provided by the embodiments of this invention. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this invention are also applicable to similar technical problems.

[0125] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present invention, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0126] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0127] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0128] The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of the present invention should be within the scope of the claims of the present invention.

Claims

1. A cloud drive file recommendation method based on behavioral memory and scene awareness, characterized in that, The method includes the following steps: In response to terminal operations of the target object, the operation log and context information of the target object are collected; wherein, the context information includes operation time, device type, location information and calendar events; Using a pre-defined deep learning model, behavioral memory modeling is performed on the historical operation logs and their corresponding time information to obtain the long-term behavioral preference vector of the target object; Based on the long-term behavioral preference vector, file recommendation matching is performed to obtain a candidate file set; Based on the context information collected in real time, scene perception is performed through a preset classification model and rule engine to obtain scene labels; Based on the scene tags, the files in the candidate file set are reordered to obtain a sorted file set; Based on the sorted file set, a recommendation file package is generated through structured integration, and the recommendation file package is pushed to the target object; The process of obtaining scene labels based on the real-time collected context information through a preset classification model and rule engine includes the following steps: Based on the real-time collected context information, the LightGBM classification model is used to classify the scene and obtain a first scene type; wherein, the first scene type includes the classification probability of each scene type; the LightGBM classification model is obtained by nonlinear modeling based on the context information of the known scene types; Based on the operation time and the time tag corresponding to the calendar event, the rule engine performs rule determination. When the result of the rule determination meets the preset rule, the scenario type corresponding to the calendar event is determined to be the second scenario type. The preset rules include the operation time being within a preset range of the time tag or the time difference between the operation time and the time tag being less than a preset time threshold. The scene tags are obtained by summarizing the first scene type and the second scene type; The candidate file set is labeled with a first weight, which is determined based on the matching between the file and the long-term behavioral preference vector. The initial sequence of the files in the candidate file set is obtained by sorting them according to the size of the first weight. The step of re-sorting the files in the candidate file set based on the scene label to obtain a sorted file set includes the following steps: The scene tag is used as a dynamic weighting factor, and a second weight is assigned to each file based on the similarity between the scene type adapted to each file in the candidate file set and the scene type corresponding to the scene tag. The second weight is positively correlated with the similarity. The ranking weight of each file in the candidate file set is determined based on the sum of the first weight and the second weight. The files in the candidate file set are reordered in descending order of their sorting weights to obtain the sorted file set.

2. The method according to claim 1, characterized in that, The step of using a preset deep learning model to model the behavioral memory of the historical operation logs and their corresponding time information to obtain the long-term behavioral preference vector of the target object includes the following steps: The historical access sequence of the target object is constructed based on the operation logs of the historical records and their corresponding time information; wherein, each sequence element in the historical access sequence is marked with the corresponding time information; The historical access sequence is modeled in depth using a Transformer model based on a self-attention mechanism. The behavior patterns at different time scales are captured by a multi-head attention mechanism to obtain the long-term behavior preference vector.

3. The method according to claim 1, characterized in that, The step of performing file recommendation matching based on the long-term behavioral preference vector to obtain a candidate file set includes the following steps: The long-term behavioral preference vector is compared with the type label vectors of all files in the recommendation engine to evaluate their similarity. The candidate file set is obtained by filtering the similarity results from largest to smallest in the similarity assessment.

4. The method according to claim 1, characterized in that, The files in the sorted file set are tagged with file types. The process of generating a recommended file package based on the sorted file set through structured integration includes the following steps: The file that appears first in the sorted file set for each of the file types is taken as the target type file for that file type. In response to the target file type corresponding to the preset file type set, the target type files of the same type are aggregated and integrated as the generated recommended file package.

5. The method according to any one of claims 1 to 4, characterized in that, The method further includes the following steps: Collect feedback data from the target object regarding the recommended file package; wherein the feedback data includes triggered data and non-triggered data; The triggered data is used as a positive sample, and the untriggered data is used as a negative sample to construct a positive and negative sample. In response to preset conditions, the deep learning model and the classification model are iteratively optimized using the positive and negative samples. The preset conditions include meeting a preset update cycle or the amount of data for the positive and negative samples exceeding a preset storage capacity.

6. A cloud drive file recommendation device based on behavioral memory and scene awareness, characterized in that, The device includes: The data acquisition module is used to collect the operation logs and context information of the target object in response to the terminal operation of the target object; wherein, the context information includes operation time, device type, location information and calendar events; The behavioral memory modeling module is used to perform behavioral memory modeling on the historical operation logs and their corresponding time information using a preset deep learning model, so as to obtain the long-term behavioral preference vector of the target object. The recommendation engine module is used to perform file recommendation matching based on the long-term behavioral preference vector to obtain a candidate file set; The scene perception module is used to perceive the scene and obtain scene labels based on the context information collected in real time, through a preset classification model and rule engine; The recommendation engine module is also used to reorder the files in the candidate file set based on the scene tags to obtain a sorted file set; The file package generation module is used to generate a recommended file package based on the sorted file set through structured integration, and push the recommended file package to the target object; The process of obtaining scene labels based on the real-time collected context information through a preset classification model and rule engine includes the following steps: Based on the real-time collected context information, the LightGBM classification model is used to classify the scene and obtain a first scene type; wherein, the first scene type includes the classification probability of each scene type; the LightGBM classification model is obtained by nonlinear modeling based on the context information of the known scene types; Based on the operation time and the time tag corresponding to the calendar event, the rule engine performs rule determination. When the result of the rule determination meets the preset rule, the scenario type corresponding to the calendar event is determined to be the second scenario type. The preset rules include the operation time being within a preset range of the time tag or the time difference between the operation time and the time tag being less than a preset time threshold. The scene tags are obtained by summarizing the first scene type and the second scene type; The candidate file set is labeled with a first weight, which is determined based on the matching between the file and the long-term behavioral preference vector. The initial sequence of the files in the candidate file set is obtained by sorting them according to the size of the first weight. The step of re-sorting the files in the candidate file set based on the scene label to obtain a sorted file set includes the following steps: The scene tag is used as a dynamic weighting factor, and a second weight is assigned to each file based on the similarity between the scene type adapted to each file in the candidate file set and the scene type corresponding to the scene tag. The second weight is positively correlated with the similarity. The ranking weight of each file in the candidate file set is determined based on the sum of the first weight and the second weight. The files in the candidate file set are reordered in descending order of their sorting weights to obtain the sorted file set.

7. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 5.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

Images