Remote Digital Service Resource Recommendation Method and System Based on Artificial Intelligence Mining

Abstract

This application provides a method and system for remote digital service resource recommendation based on artificial intelligence mining. First, historical interaction data of the target user in multiple remote service scenarios is obtained, covering interaction records, service resource attribute information, and user behavior sequences, and based on this, a dynamic feature vector representing the user preference distribution and the evolution path of service demand is generated. It is input into a multi-modal recommendation model that integrates time-series behavior analysis and cross-scenario semantic association models to mine potential matching patterns and output a prediction set of candidate service resources for the current scenario. A hierarchical recommendation list is generated according to the real-time scenario fitness, and service resource priority clusters are divided according to the fitness threshold. Finally, based on the user's real-time feedback data, the model parameter weights are dynamically adjusted, and the generation logic of the dynamic feature vector is updated to form a closed-loop optimization link, effectively improving the accuracy and adaptability of remote digital service resource recommendation.

Description

Technical Field

[0001] This application relates to the field of cloud office technology. Specifically, it relates to a method and system for recommending remote digital service resources based on artificial intelligence mining. Background Art

[0002] In today's digital age, the application of remote digital service resources is becoming increasingly widespread. How to accurately recommend suitable service resources for users has become a key issue in improving service quality and user experience.

[0003] In the early stage, traditional service resource recommendation methods were mainly based on simple rules or static user portraits. For example, some recommendation systems only made similar recommendations based on the service types that users had used before. This method is too simple and mechanical and cannot take into account the diverse needs of users in different situations. It does not fully recognize that users' service needs are not static but will evolve dynamically over time, scenarios, and their own behaviors.

[0004] With the development of technology, some recommendation systems have begun to try to use user behavior data, but most are limited to single-dimensional analysis. For example, only focusing on users' click behaviors to build recommendation strategies, ignoring the rich attribute information of service resources themselves and the complex connections between different service scenarios. This results in the recommendation results often being unable to accurately meet the real needs of users in specific scenarios, making the recommended service resources have a large gap with users' expectations and poor user experience. Summary of the Invention

[0005] In view of the problems mentioned above, in combination with the first aspect of this application, embodiments of this application provide a method for recommending remote digital service resources based on artificial intelligence mining. The method includes:

[0006] Obtain the historical interaction data of the target user. The historical interaction data includes the interaction records of the target user with service resources in multiple remote service scenarios, service resource attribute information, and user behavior sequences;

[0007] Generate a dynamic feature vector of the target user based on the historical interaction data. The dynamic feature vector is used to represent the preference distribution and service demand evolution path of the target user in different service scenarios;

[0008] Input the dynamic feature vector into a pre-trained multi-modal recommendation model, and output a predicted set of candidate service resources for the target user in the current service scenario. The multi-modal recommendation model mines the potential matching patterns between user behaviors and service resources by integrating a time-series behavior analysis model and a cross-scenario semantic association model;

[0009] generating a hierarchical recommendation list according to the real-time scenario adaptability of each service resource in the candidate service resource prediction set, wherein the real-time scenario adaptability is determined by analyzing the dynamic association strength between the service resource attribute and the context feature of the current service scenario, and the hierarchical recommendation list divides the service resources into a plurality of priority clusters according to the adaptability threshold;

[0010] Based on the real-time feedback data of the target user on the hierarchical recommendation list, the parameter weights of the multimodal recommendation model are dynamically adjusted, and the generation logic of the dynamic feature vector is updated to form a closed-loop optimization link.

[0011] On the other hand, an embodiment of the present application also provides a remote digital service system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0012] Based on the above aspects, the embodiments of the present application can comprehensively and meticulously portray user portraits by acquiring rich historical interaction data of target users, covering interaction records, service resource attribute information, and user behavior sequences in multiple remote service scenarios. Compared with traditional methods that only rely on a single dimension or a small amount of data, the depth and accuracy of understanding of user preferences and needs are greatly improved, so that the generated dynamic feature vector can accurately characterize the target user's preference distribution and service demand evolution path in different service scenarios, and capture the subtle trends of user needs changing over time and in different scenarios.

[0013] Next, the time series behavior analysis model and the cross-scenario semantic association model are integrated to form a multimodal recommendation model, which breaks the limitation of traditional recommendation models that only analyze the relationship between user behavior and service resources from a single perspective, and can deeply explore the complex and potential matching patterns between user behavior and service resources. Traditional models often have difficulty handling the diversity and relevance of data in different scenarios, while this multimodal recommendation model, with its unique fusion architecture, effectively integrates the behavioral changes in the time series and the semantic associations between different scenarios, significantly improving the accuracy and comprehensiveness of the recommendation, and the output candidate service resource prediction set is more in line with the real needs of users in the current service scenario.

[0014] Then, generate a hierarchical recommendation list based on the real-time scene adaptability. Determine the real-time scene adaptability by analyzing the dynamic association strength between the service resource attributes and the context features of the current service scene, and divide the service resources into multiple priority clusters according to the adaptability threshold. This hierarchical recommendation method not only considers the matching between the service resources and the user needs, but also combines the real-time characteristics of the current service scene. Different from the traditional fixed recommendation strategy, it can flexibly adjust the recommendation order according to the scene changes, and preferentially display the service resources that best meet the scene requirements and user preferences, greatly improving the efficiency of users to discover suitable service resources and enhancing the user experience.

[0015] In addition, based on the real-time feedback data of the target user on the hierarchical recommendation list, dynamically adjust the parameter weights of the multi-modal recommendation model, and update the generation logic of the dynamic feature vector to form a closed-loop optimization link, enabling the recommendation system to have self-learning and adaptive capabilities, and being able to continuously optimize the recommendation effect according to the actual feedback of users. Traditional recommendation systems often lack this real-time dynamic optimization mechanism. As time goes by and user behaviors change, the recommendation accuracy is prone to decline. However, through the closed-loop optimization link, this method can continuously keep up with the changes in user needs, continuously improve the recommendation performance, and ensure that the recommendation results always maintain high accuracy and effectiveness. Brief Description of the Drawings

[0016] Figure 1 It is a schematic flowchart of the execution process of the remote digital service resource recommendation method based on artificial intelligence mining provided by an embodiment of the present application.

[0017] Figure 2 It is a schematic hardware architecture diagram of the remote digital service system provided by an embodiment of the present application. Detailed Embodiments

[0018] The present application will be specifically described below in conjunction with the accompanying drawings of the specification. Figure 1 It is a schematic flowchart of the remote digital service resource recommendation method based on artificial intelligence mining provided by an embodiment of the present application. The remote digital service resource recommendation method based on artificial intelligence mining will be introduced in detail below.

[0019] Step S110, obtain the historical interaction data of the target user, where the historical interaction data includes the interaction records of the target user with service resources in multiple remote service scenarios, service resource attribute information, and user behavior sequences.

[0020] In this embodiment, taking the telecommuting scenario as an example, the target user may be a project manager. In his daily work, he is involved in the interaction of multiple remote service scenarios. For example, he often uses video conferencing services to communicate with team members in different regions about projects. The interaction records of each use of the video conferencing service include the time of the meeting, the participants, the duration of the meeting, etc. In terms of service resource attribute information, the video conferencing service may have different functional attributes, such as high-definition picture quality, screen sharing function, upper limit of the number of participants, etc. The user behavior sequence may be that he first opens the video conferencing software, checks whether the participants have arrived, then enables the screen sharing function to display project documents, and adjusts the picture quality during the meeting to adapt to the network conditions, etc.

[0021] He also uses project management tool services. The interaction records include operations such as creating project tasks, assigning tasks to team members, setting task deadlines, etc. The service resource attributes of the project management tool include task classification function, task progress tracking function, etc. His user behavior sequence may be to first create a project framework task, and then assign different subtasks according to the expertise of the members, and continuously check the task progress during this period.

[0022] In addition, he occasionally uses cloud storage services to store project-related documents and materials. The interaction records include file upload time, file size, access permission settings, etc. The attribute information of the cloud storage service involves storage space size, data encryption method, etc. His behavior sequence is to first organize the files to be stored, then upload them to the cloud storage, and set different access permissions according to the importance and confidentiality of the files. These interaction records, service resource attribute information, and user behavior sequences with various service resources in multiple remote service scenarios together constitute the historical interaction data of the target user.

[0023] Step S120, generating a dynamic feature vector of the target user based on the historical interaction data, where the dynamic feature vector is used to characterize the preference distribution and service demand evolution path of the target user in different service scenarios.

[0024] For example, for this project manager, extract user behavior time series segments from the historical interaction data. Taking the video conferencing service as an example, the continuous service request interval may reflect his project communication rhythm. For example, he starts video conferences every Monday, Wednesday, and Friday morning, which reflects his preference for centralized communication during specific time periods on weekdays. In terms of the service type switching frequency, he may switch to the project management tool to check task updates immediately after the video conference. This frequent switching behavior indicates his demand for the collaborative use of different services. In terms of the depth of service resource use, his frequent adjustment of the picture quality and screen sharing function of the video conference shows a relatively high degree of dependence on these functions.

[0025] Multi-granularity feature extraction is performed on these user behavior time series segments through a pre-trained spatio-temporal attention network to generate an initial behavior feature matrix. The user behavior time series segments are divided into multiple time windows, for example, one week as a time window. Within each time window, the local behavior subsequence contains all video conference-related operations within that week. A convolutional kernel sliding operation is performed on each local behavior subsequence. Using a multi-scale convolutional kernel group can capture behavior patterns with different time spans, such as capturing the differences in the usage patterns of video conferences at key project nodes compared to normal times. The local spatio-temporal features are input into a bidirectional long short-term memory network. The forward and backward paths of the bidirectional long short-term memory network can capture the implicit dependencies between time windows, such as the possible associations between his video conference usage habits in the early stage of the project and those in the later stage. When constructing the spatio-temporal attention weight matrix, the behavior similarity between any two time windows is calculated, such as calculating the cosine similarity in aspects such as the distribution of participants and the meeting duration in video conferences of different weeks. An initial attention score matrix is generated based on the behavior similarity and undergoes sparsification processing. Attention connections with attention scores exceeding the preset threshold are retained, and then added to the position encoding matrix. Finally, a spatio-temporal attention weight matrix is generated through the softmax function to dynamically adjust the contribution degree of each time window to the final feature representation, thereby generating the initial behavior feature matrix.

[0026] Next, fuse the service resource attribute information and the initial behavior feature matrix to generate cross-modal fusion features. For example, attributes such as high-definition video quality and screen sharing function of the video conferencing service are embedded into the user behavior feature space through a cross-modal alignment algorithm. Generate a scene-aware weighted scene feature vector according to the scene switching marks of the target user in multiple service scenarios. Extract the scene switching time sequence in the scene switching mark. Suppose he will switch from the video conferencing scene to the project management tool scene when the project phase changes, and the scene switching trigger event type may be the advancement of the project phase. Generate an initial scene weight vector according to the scene association strength corresponding to the scene switching trigger event type. For example, in the critical decision-making phase of the project, the weight of the video conferencing scene will be relatively high. Calculate the time decay factor based on the time interval between adjacent scene switches in the scene switching time sequence. As time goes by, the scene weights of the previous project phases will gradually decay. Perform an element-wise multiplication operation on the time decay factor and the initial scene weight vector, and then dynamically adjust it through a scene attention mechanism to generate a dynamic scene weight distribution. Perform a tensor multiplication operation on the dynamic scene weight distribution and the cross-modal fusion features to generate a weighted scene association feature matrix. Fuse the scene association feature matrix with the currently collected current scene context features. For example, when the current project is in the closing phase and more document collation and review work are required, the corresponding current scene context features are the attention to the task completion status in the cloud storage service and the project management tool. Perform scene consistency verification on the enhanced scene feature tensor to ensure that the orthogonality degree between the feature subspaces corresponding to different scene types is reasonable, and correct the conflicting weights in the dynamic scene weight distribution. Perform secondary weighting on the cross-modal fusion features according to the corrected dynamic scene weight distribution to generate an optimized scene-aware feature projection. Finally, reduce its dimension to a preset dimension space through a feature dimension compression algorithm to generate a scene feature vector, which contains his long-term preference features, such as the preference for efficient communication tools, and the short-term demand features for the cloud storage service and the project management tool in the current project closing phase.

[0027] Step S130: Input the dynamic feature vector into a pre-trained multi-modal recommendation model, and output a candidate service resource prediction set of the target user in the current service scenario. The multi-modal recommendation model mines the potential matching patterns between user behaviors and service resources by fusing a time-series behavior analysis model and a cross-scene semantic association model.

[0028] For example, for this project manager, the current service scenario assumes that the project is about to end, and it is necessary to summarize and present the project results. The previously generated dynamic feature vectors are input into the pre-trained multi-modal recommendation model. The time-series behavior analysis model in the multi-modal recommendation model will analyze his behavior patterns in previous similar project phases. For example, when previous projects were coming to an end, his usage frequency of document sorting tools and presentation-making tools increased. The cross-scenario semantic association model will mine the semantic associations between different scenarios. For example, there may be an association between the video conferencing service he used when communicating the project plan with the client and the project result presentation, because it may be necessary to show the client the presentation of the project results.

[0029] By fusing these models, potential matching patterns between user behaviors and service resources are mined, and then a candidate service resource prediction set is output. This set may include some document management tools suitable for summarizing project results, presentation-making software with good presentation effects, and cloud storage service upgrade options that can securely store the final project results, etc. These service resources are predicted based on the analysis of his historical behaviors and semantic associations in different scenarios, and can meet his potential needs in the current service scenario where the project is about to end.

[0030] Step S140, generate a hierarchical recommendation list according to the real-time scenario fitness of each service resource in the candidate service resource prediction set. The real-time scenario fitness is determined by analyzing the dynamic association strength between the service resource attributes and the context features of the current service scenario. The hierarchical recommendation list divides the service resources into multiple priority clusters according to the fitness threshold.

[0031] In the scenario where the current project is about to end, each service resource in the candidate service resource prediction set is analyzed. Taking the document management tool as an example, its scenario matching score is calculated. If the document management tool has attributes such as version management function and multi-person collaborative editing function, and in the current scenario, it is necessary to determine the final version of the project documents and conduct joint reviews by team members, then the co-occurrence frequency of these attribute tags and the context features of the current scenario is relatively high, resulting in a relatively high scenario matching score.

[0032] Generate a service resource availability indicator based on the real-time availability status of the service resource and the load capacity of the service provider. Assume that the server where the document management tool is located has recently undergone maintenance and upgrade, with low response latency and the ability to withstand a large number of concurrent user accesses, then its service resource availability indicator is relatively high.

[0033] Fuse the scenario matching score and the service resource availability indicator, and use the weighted harmonic mean algorithm to generate a comprehensive recommendation priority. For example, if the document management tool has a relatively high scenario matching score and a good availability indicator, then its comprehensive recommendation priority will be relatively high.

[0034] Sort the candidate service resources according to the comprehensive recommendation priority, and divide the sorting results into multiple recommendation levels. For example, the document management tool and the presentation software with the highest comprehensive recommendation priority are divided into the first recommendation level. The display strategy corresponding to this level may be to recommend them at a prominent position in the project manager's office software interface, and the interaction method may be to directly pop up a prompt box to inform its functional advantages. The cloud storage service upgrade option with a slightly lower comprehensive recommendation priority is divided into the second recommendation level. The display strategy may be to display it at a secondary position in the software interface, and the interaction method may be to display it in detail only when the user actively views the storage-related services.

[0035] Control the diversity of service resources for each recommendation level. Calculate the type similarity matrix of service resources within the recommendation level. For example, although the document management tool and the presentation software have certain differences in functions, they both belong to the tool types related to project result collation and display. Construct a diversity constraint optimization objective function, with the goal of maximizing the type difference degree of service resources within the recommendation level and minimizing the deviation degree from user expectations. Use the greedy algorithm to iteratively select service resources, and each iteration selects the service resource that can maximize the current diversity index and does not violate the comprehensive recommendation priority constraint. Suppose in the first recommendation level, in addition to the document management tool and the presentation software, an online collaboration platform is also considered. This platform can facilitate team members to conduct final communication and adjustment in the final stage. Post-process and filter the selected service resource set, remove the candidate resources whose similarity to the selected resources exceeds the threshold. For example, if there is an alternative tool with highly similar functions to the presentation software, it will be removed, and a sub-optimal but highly diverse contribution alternative resource will be added. For example, if there is a presentation tool with special display effects but a slightly lower comprehensive recommendation priority, it can be added. Dynamically adjust the diversity weight coefficient. If the project manager continuously rejects multiple recommendations related to the presentation software, automatically increase the diversity penalty term for the corresponding type to increase the diversity of recommendations.

[0036] Step S150: Based on the real-time feedback data of the target user on the hierarchical recommendation list, dynamically adjust the parameter weights of the multi-modal recommendation model, and update the generation logic of the dynamic feature vector to form a closed-loop optimization link.

[0037] When presenting the hierarchical recommendation list to the project manager, he will generate real-time feedback data. Explicit feedback signals and implicit feedback signals are extracted from the real-time feedback data. In terms of explicit feedback signals, if he rates the document management tool and adds the presentation software to the favorites, it indicates that he is relatively satisfied with these two service resources. In terms of implicit feedback signals, the long duration of his stay in the service of the document management tool and the high frequency of interaction with the online collaboration platform also reflect his degree of attention to these service resources.

[0038] Construct a feedback-enhanced feature vector by non-linearly fusing the explicit feedback signal and the implicit feedback signal. For example, fuse the rating, favorite marking, service stay duration, and interaction frequency through a specific non-linear function to generate a feedback-enhanced feature vector.

[0039] Input the feedback-enhanced feature vector into the parameter adjustment network to output the update amount of the model weights. The parameter adjustment network uses a meta-learning framework to predict the influence intensity of different feedback patterns on the model parameters. For example, if the project manager rates the document management tool highly, then the parameter adjustment network will analyze that this feedback pattern has a greater impact on the parameters related to the recommendation of the document management tool in the model, and thus output the corresponding update amount of the model weights.

[0040] Perform knowledge distillation on the embedding layer of the multi-modal recommendation model, and retain the feature mapping relationship corresponding to the confidence prediction result with a confidence greater than the set confidence. For example, in previous recommendations, the recommendation of the document management tool was based on a certain feature mapping relationship. If the confidence of this recommendation is relatively high, then this feature mapping relationship is retained when adjusting the model.

[0041] Establish a model version rollback mechanism. When it is detected that the feedback data of consecutive batches causes the recommendation performance to decline, automatically restore to the model parameters of the historical stable version. For example, if the subsequent recommended service resources such as the document management tool and the presentation software no longer meet the needs of the project manager, resulting in a decrease in his satisfaction, when this situation occurs continuously, the model automatically restores to the parameters of the version that could accurately recommend before.

[0042] Through this processing of the real-time feedback data of the target user on the hierarchical recommendation list, dynamically adjust the parameter weights of the multi-modal recommendation model, and update the generation logic of the dynamic feature vector, forming a closed-loop optimization link, enabling the recommendation system to continuously adapt to the changing needs of the target user in the remote work scenario and improving the accuracy and effectiveness of the recommendation.

[0043] Based on the above steps, the embodiment of the present application can comprehensively and meticulously portray user portraits by acquiring rich historical interaction data of target users, covering interaction records, service resource attribute information, and user behavior sequences in multiple remote service scenarios. Compared with traditional methods that only rely on a single dimension or a small amount of data, the depth and accuracy of understanding of user preferences and needs are greatly improved, so that the generated dynamic feature vector can accurately characterize the target user's preference distribution and service demand evolution path in different service scenarios, and capture the subtle trends of user needs changing over time and in different scenarios.

[0044] Next, the time series behavior analysis model and the cross-scenario semantic association model are integrated to form a multimodal recommendation model, which breaks the limitation of traditional recommendation models that only analyze the relationship between user behavior and service resources from a single perspective, and can deeply explore the complex and potential matching patterns between user behavior and service resources. Traditional models often have difficulty handling the diversity and relevance of data in different scenarios, while this multimodal recommendation model, with its unique fusion architecture, effectively integrates the behavioral changes in the time series and the semantic associations between different scenarios, significantly improving the accuracy and comprehensiveness of the recommendation, and the output candidate service resource prediction set is more in line with the real needs of users in the current service scenario.

[0045] Then, a hierarchical recommendation list is generated based on the real-time scene adaptability. The real-time scene adaptability is determined by analyzing the dynamic correlation strength between the service resource attributes and the contextual features of the current service scene, and the service resources are divided into multiple priority clusters according to the adaptability threshold. This hierarchical recommendation method not only considers the matching of service resources with user needs, but also combines the real-time characteristics of the current service scene. Different from the traditional fixed recommendation strategy, it can flexibly adjust the recommendation order according to the scene changes, and give priority to the service resources that best meet the scene needs and user preferences, which greatly improves the efficiency of users in discovering suitable service resources and improves the user experience.

[0046] In addition, based on the target user's real-time feedback data on the hierarchical recommendation list, the parameter weights of the multimodal recommendation model are dynamically adjusted, and the generation logic of the dynamic feature vector is updated to form a closed-loop optimization link, so that the recommendation system has self-learning and adaptive capabilities, and can continuously optimize the recommendation effect according to the user's actual feedback. Traditional recommendation systems often lack this real-time dynamic optimization mechanism. As time goes by and user behavior changes, the accuracy of recommendations tends to decrease. This method can continuously keep up with changes in user needs through a closed-loop optimization link, continuously improve recommendation performance, and ensure that the recommendation results always maintain high accuracy and effectiveness.

[0047] In a possible implementation, step S120 includes:

[0048] Step S121: Extract the user behavior time series segment from the historical interaction data. The user behavior time series segment includes the continuous service request interval, the service type switching frequency, and the service resource usage depth.

[0049] In this embodiment, in the remote working scenario, the continuous service request interval reflects the working rhythm pattern of the project manager. For example, when using the video conferencing service during the project cycle, there may be a certain pattern in the time interval between each meeting request. For instance, at the initial stage of the project, a video conference is held once a week. As the project progresses to the critical stage, the meeting request interval shortens to once every three days. This change in the interval reflects his demand for communication frequency at different project stages. In terms of the service type switching frequency, after finishing a video conference, he often quickly switches to the project management tool to check the task progress. Sometimes, he switches between the video conferencing service, the project management tool, and the cloud storage service multiple times within a day. This frequency reflects the degree of his need for the collaborative work of different services. The service resource usage depth also has obvious characteristics. Taking the project management tool as an example, during the project execution stage, he deeply uses functions such as task assignment and progress tracking, frequently adjusts the task priorities and deadlines. The in-depth use of the functions reflects his degree of dependence on the functions of the project management tool. These continuous service request intervals, service type switching frequencies, and service resource usage depths together constitute the user behavior time series segment.

[0050] Step S122: Perform multi-granularity feature extraction on the user behavior time series segment through a pre-trained spatio-temporal attention network to generate an initial behavior feature matrix. The spatio-temporal attention network identifies key service interaction nodes by capturing the dependency relationships of user behavior in the time and space dimensions.

[0051] For example, the user behavior sequence segments are divided into multiple time windows according to the time dimension, such as one month as a time window. In each local time window of a month, all remote office-related behavior operations of the project manager in that month are included. These local behavior subsequences carry the behavior pattern information of different time periods. The convolution kernel sliding operation is performed on each local behavior subsequence, and a multi-scale convolution kernel group is used. Convolution kernels of different scales can capture behavior patterns of different time spans. For example, a smaller-scale convolution kernel can capture the changes in the operation mode of the project manager within a certain day, while a larger-scale convolution kernel can reflect the overall trend of his use of video conferencing within a month, such as the change in the duration of video conferences at key nodes of the project. The local spatiotemporal features are input into the bidirectional long short-term memory network, and the forward and backward paths of the network can capture the implicit dependencies between time windows. For example, from the early stage to the late stage of the project, there is a correlation between the composition of the participants in his video conference and the key content of the meeting. The forward path can infer the possible development direction of the focus of the later meeting from the composition of the personnel in the early stage, and the backward path can reversely infer the rationality of the composition of the personnel in the early stage from the focus of the later meeting. When constructing the spatiotemporal attention weight matrix, the behavioral similarity between any two time windows is calculated. Taking the video conferencing service as an example, the cosine similarity of the distribution of participants, meeting duration, whether screen sharing is enabled, etc. in video conferences in different months is calculated. The initial attention score matrix is ​​generated based on the behavioral similarity, and the connections with low scores and large differences in behavioral patterns are sparsely processed. The attention connections that exceed the preset attention score are retained and then added to the position encoding matrix that represents the relative position relationship of the time window in the overall time series. Finally, the spatiotemporal attention weight matrix is ​​generated through the normalized exponential function, so as to dynamically adjust the contribution of each time window to the final feature representation according to the importance of different time windows, and finally generate the initial behavioral feature matrix.

[0052] Step S123, fusing the service resource attribute information and the initial behavior feature matrix to generate a cross-modal fusion feature, wherein the cross-modal fusion feature embeds the service resource attributes into the user behavior feature space through a cross-modal alignment algorithm.

[0053] Taking the project management tool as an example, its service resource attributes include task classification function, task progress tracking function, task dependency setting function, etc. These service resource attributes are embedded into the user behavior feature space through the cross-modal alignment algorithm. For example, the frequent use of the task classification function by the project manager is fused with the attribute of the task classification function itself, so that the two are interrelated in the same feature space. If the project manager often creates different task groups according to the task type, then this behavior and the attribute of the task classification function will be reflected in the cross-modal fusion feature, thereby realizing the organic combination of service resource attributes and user behavior characteristics.

[0054] Step S124: Perform scene-aware weighting on the cross-modal fusion features according to the scene switching marks of the target user in multiple service scenarios to generate a scene-based feature vector.

[0055] The scene switching time sequence in the scene switching mark includes the scene type identifiers triggered by the project manager in consecutive time units and the scene switching trigger event types. For example, when the project enters the execution phase from the planning phase, he switches from the document editing scene (such as writing a project plan document) to the project management tool scene (starting to assign tasks and track progress), and the scene switching trigger event type is the conversion of the project phase. Generate an initial scene weight vector according to the scene association strength corresponding to the scene switching trigger event type. In the project execution phase, the weight of the project management tool scene is relatively high because the demand for task management is the greatest in this phase. Calculate the time decay factor based on the time interval between adjacent scene switches in the scene switching time sequence. As time goes by, the weights of the scenes related to the previous project planning phase will gradually decay. Perform an element-wise multiplication operation on the time decay factor and the initial scene weight vector, and then dynamically adjust it through the scene attention mechanism. For example, when the project is approaching the end, although the weight of the document editing scene has decreased due to time decay before, if the project summary document needs to be edited at this time, the scene attention mechanism will enhance the weight of the document editing scene according to the context features of the current scene (such as the editing requirements of the project summary document). Perform a tensor multiplication operation on the dynamic scene weight distribution and the cross-modal fusion features to generate a weighted scene association feature matrix, where each feature channel corresponds to the preference expression of a specific scene type. Fuse the scene association feature matrix with the current scene context features collected in real time. In the project closing phase, the current scene context features may include the review, sorting, and storage requirements of the project final results, and fuse these features with the scene association feature matrix. Perform scene consistency verification on the enhanced scene feature tensor to ensure that the orthogonality degree between the feature subspaces corresponding to different scene types is reasonable, and correct the conflicting weights in the dynamic scene weight distribution. For example, if there is confusion in some features between the document editing scene and the project management tool scene, adjust it through scene consistency verification. Perform secondary weighting on the cross-modal fusion features according to the corrected dynamic scene weight distribution to generate an optimized scene-aware feature projection, and finally reduce its dimension to a preset dimension space through a feature dimension compression algorithm to generate a scene-based feature vector. This scene-based feature vector includes both the long-term preference features of the project manager, such as the preference for the efficient task management function of the project management tool, and the short-term demand features related to the review and storage of the project results in the current project closing phase.

[0056] Step S125: Dynamically update the scenario-based feature vector based on a gated recurrent unit to generate the dynamic feature vector, where the dynamic feature vector includes the user's long-term preference features and the short-term demand features of the current service scenario.

[0057] In this embodiment, the gated recurrent unit can perform dynamic updates based on previous state information and current input information (i.e., the scenario-based feature vector). In the telecommuting scenario, as the project progresses, the needs of the project manager will change continuously. For example, at different stages of the project, his demand for video conferencing services may gradually change from simple communication to project result presentation, and this change in demand will be reflected in the scenario-based feature vector. The gated recurrent unit dynamically adjusts the scenario-based feature vector according to the previously accumulated behavior patterns and preference information of the project manager. It can organically integrate the project manager's long-term preference features (such as the consistent demand for efficient communication and task management) with the short-term demand features of the current service scenario (such as the demand for result presentation and storage in the project closing stage) based on the long-term development trend of the project and the special needs of the current stage, and generate a dynamic feature vector, which can accurately reflect the changing demand and preference patterns of the project manager in the telecommuting scenario.

[0058] In a possible implementation manner, step S122 includes:

[0059] Step S1221: Divide the user behavior time series segment into multiple time windows, and the behavior data within each time window forms a local behavior subsequence.

[0060] In this embodiment, for the project manager, during the telecommuting process, his work cycle can be divided into multiple time windows according to a certain time scale. For example, every two weeks in a project cycle can be used as a time window. Within each such time window, all the behavior data related to various remote services of the project manager is included, and these behavior data together form a local behavior subsequence. For example, within a two-week time window, his behavior data in the video conferencing service may include initiating three video conferences, the participants in each conference, the duration of each conference, whether the screen sharing function is used, etc.; the behavior data in the project management tool includes how many new tasks are created, which task deadlines are adjusted, the task assignment situation, etc.; the behavior data in the cloud storage service includes which project files are uploaded, the size of the files, the set access permissions, etc. The behavior data related to different services within a specific time window completely records the work behavior pattern of the project manager during this time period.

[0061] Step S1222: Perform a convolutional kernel sliding operation on each local behavior subsequence to extract local spatio-temporal features. The convolutional kernel sliding operation uses a multi-scale convolutional kernel group to capture behavior patterns with different time spans.

[0062] For each local behavior subsequence within each of the aforementioned time windows, perform a convolutional kernel sliding operation using a multi-scale convolutional kernel group. Taking a video conferencing service as an example, during the sliding process, a convolutional kernel with a smaller scale can capture the detailed patterns of the project manager's video conferencing behavior within a shorter time period (such as within a single day). For instance, within a single day, he conducted two video conferences in a row. The first conference was mainly to discuss project technical problems with the technical team, and the second conference was to communicate the project promotion plan with the marketing team. The convolutional kernel with a smaller scale can capture the rapid change patterns in aspects such as the types of participants and the key content of the two conferences. On the other hand, a convolutional kernel with a larger scale can analyze the video conferencing behavior patterns from a longer time span (such as the entire two-week time window). For example, within these two weeks, the overall duration trend of the video conferences was gradually increasing, which may indicate that the project has entered a crucial discussion and decision-making stage. The convolutional kernel with a larger scale can identify this change in behavior patterns over a long time span. Through this convolutional kernel sliding operation with a multi-scale convolutional kernel group, it is possible to comprehensively extract the local spatio-temporal features of each local behavior subsequence under different time spans, and these features cover rich information about the project manager's behavior patterns in different services.

[0063] Step S1223: Input the local spatio-temporal features into a bidirectional long short-term memory network to generate a temporal context encoding. The bidirectional long short-term memory network captures the implicit dependencies between time windows through forward and backward paths.

[0064] For the local spatio-temporal features extracted from each local behavior subsequence, they are input into a bidirectional long short-term memory network. Taking the task assignment behavior in a project management tool as an example, there are certain implicit dependencies in the task assignment situations within different time windows (every two weeks). From the forward path of the bidirectional long short-term memory network, in the time window at the early stage of the project, the project manager may assign some basic tasks to junior team members according to the preliminary project plan. As the project progresses, in subsequent time windows, the forward path can analyze how the subsequent task assignment is adjusted based on the progress of the previous tasks according to the previous task assignment situation. For example, according to the completion of the previous basic tasks, some more complex tasks are assigned to experienced team members. From the backward path, when observing the task assignment situation in the later stage of the project, it can be inferred reversely whether the previous task assignment is reasonable and whether it has laid a good foundation for the subsequent tasks. Such a bidirectional long short-term memory network can comprehensively capture the implicit dependencies in the task assignment behavior between different time windows through the forward and backward paths, and the same is true for other service-related behaviors such as video conferencing and cloud storage. By processing the local spatio-temporal features of all service-related behaviors in this way, a temporal context encoding containing the implicit dependencies between time windows is generated.

[0065] Step S1224, construct a spatio-temporal attention weight matrix, and the spatio-temporal attention weight matrix dynamically adjusts the contribution degree of each time window to the final feature representation by calculating the correlation intensity of behavior features within different time windows.

[0066] Step S1225, perform weighted aggregation on the temporal context encoding according to the spatio-temporal attention weight matrix to generate the initial behavior feature matrix.

[0067] For example, for the temporal context encoding previously generated by the bidirectional long short-term memory network, each encoding part corresponding to a time window has its importance weight, which is determined by the spatio-temporal attention weight matrix. For example, in the time window of a certain project stage, since the behavior patterns of the video conferencing service and the project management tool service in this stage have a high similarity with those in other key stages (determined by the spatio-temporal attention weight matrix), the weight of the temporal context encoding of this time window will be relatively high in the weighted aggregation process. Through this weighted aggregation method, the temporal context encodings of all time windows are combined according to their respective weights, and finally the initial behavior feature matrix is generated. This initial behavior feature matrix synthesizes the behavior information within different time windows and is reasonably weighted according to the correlation intensity between time windows, laying a foundation for further analyzing the behavior features of the project manager.

[0068] In a possible implementation manner, step S1224 includes:

[0069] Step S1224-1, calculate the behavior similarity between any two time windows, where the behavior similarity measures the matching degree of the service type distribution and service interaction intensity within two time windows through the cosine similarity algorithm.

[0070] Step S1224-2, generate an initial attention score matrix based on the behavior similarity, and perform sparsification processing on the initial attention score matrix to retain the attention connections exceeding the preset attention score.

[0071] Step S1224-3, add the sparsified attention score matrix to the position encoding matrix to generate an enhanced attention matrix, where the position encoding matrix is used to represent the relative position relationship of the time window in the overall time sequence.

[0072] Step S1224-4, perform normalization processing on the enhanced attention matrix through the softmax function to generate the spatio-temporal attention weight matrix.

[0073] For example, taking video conferencing services and project management tool services as examples, within different time windows, the service type distribution of video conferencing services may include different combinations of participants (such as dominated by technical personnel, dominated by marketing personnel, etc.) and different meeting purposes (such as technical discussions, project progress reports, etc.). The service interaction intensity can be measured by indicators such as meeting duration and number of participants. For project management tool services, the service type distribution may be reflected in the proportion of different types of tasks (such as technical development tasks, market research tasks, etc.). The service interaction intensity can be measured by indicators such as the number of task creations and the frequency of task adjustments. Calculate the matching degree of these service type distributions and service interaction intensities between any two time windows through the cosine similarity algorithm to obtain the behavior similarity. Generate an initial attention score matrix based on the behavior similarity. In this matrix, each element represents the initial attention score between two time windows. Then, perform sparsification processing on the initial attention score matrix, retaining the attention connections that exceed the preset attention score. For example, if the behavior similarity between two time windows is low and lower than the preset attention score, then the corresponding attention connection is removed, and only those attention connections with high behavior similarity, indicating a strong association between the two time windows, are retained. Add the sparsified attention score matrix to the position encoding matrix, which is used to characterize the relative position relationship of the time window in the overall time series. For example, in a sequence of multiple time windows throughout the project cycle, earlier time windows and later time windows have different position relationships, and the position encoding matrix can reflect this difference. After addition, an enhanced attention matrix is generated. Finally, perform normalization processing on the enhanced attention matrix through the softmax function to generate a spatio-temporal attention weight matrix. This spatio-temporal attention weight matrix can dynamically adjust the contribution degree of each time window in the final feature representation according to the association strength between different time windows.

[0074] In a possible implementation manner, step S124 includes:

[0075] Step S1241, extract the scene switching time sequence in the scene switching mark, and the scene switching time sequence includes the scene type identifier triggered by the user within consecutive time units and the scene switching trigger event type.

[0076] In this embodiment, in the remote work activities of the project manager, the scenario type identifiers may include video conferencing scenarios, project management tool usage scenarios, cloud storage service scenarios, etc. From the perspective of the entire project cycle, within consecutive time units, the scenario switching trigger event types are closely related to the project's progress stage. For example, at the initial stage of the project, when the project team members are determined, the scenario switches from a single document editing scenario (such as writing a project plan) to a video conferencing scenario. At this time, the scenario switching trigger event type is that the project team formation is completed and team communication is required. As the project progresses to the execution stage, the scenario switches from a video conferencing scenario to a project management tool usage scenario. The scenario switching trigger event type is that the project starts to execute and task allocation and progress tracking are required. The records of these scenario switches form a scenario switching time sequence in chronological order, which contains the specific scenario type identifiers of each scenario switch and the event types that trigger the scenario switch. These information completely reflect the interaction and conversion of the project manager with different service scenarios at different stages of the project.

[0077] Step S1242: Generate an initial scenario weight vector according to the scenario association strength corresponding to the scenario switching trigger event type. The weight value of each dimension in the initial scenario weight vector reflects the cumulative influence degree of the corresponding scenario type in historical interactions.

[0078] For the project manager, different scenario switching trigger event types correspond to different scenario association strengths. For example, during the project execution stage, for the event type of switching from a video conferencing scenario to a project management tool usage scenario, due to the high dependence on task management during project execution, the scenario association strength of the project management tool usage scenario is relatively high at this stage. When the project is approaching the end, for the event type of switching from a project management tool usage scenario to a cloud storage service scenario (for storing the final project results), the scenario association strength of the cloud storage service scenario is relatively high under this event type. Based on these different scenario association strengths, an initial scenario weight vector is generated for each scenario type. For example, throughout the project cycle, if the video conferencing scenario frequently appears during the project communication stage and has an important impact on the project progress, then in the initial scenario weight vector, the weight value corresponding to the video conferencing scenario will be relatively high, and this weight value reflects the cumulative influence degree of the video conferencing scenario in historical interactions. Similarly, for the project management tool usage scenario and the cloud storage service scenario, etc., their weight values are also determined according to their respective importance and influence degree in historical interactions.

[0079] Step S1243: Calculate a time decay factor based on the time interval between adjacent scenario switches in the scenario switching time sequence. The time decay factor is used to dynamically adjust the contribution decay rate of historical scenario types to the current scenario perception weighting.

[0080] Step S1244: Perform an element-wise multiplication operation on the time decay factor and the initial scene weight vector to generate an updated scene weight vector, where the weight value decay amplitude of the scene type closer to the current time in the updated scene weight vector is smaller.

[0081] On the time axis of the project, the time intervals between adjacent scene switches are different. For example, in the early stage of the project, the time interval from the document editing scene to the video conferencing scene is shorter, while during the project execution, the time interval from a project management tool usage scene for task adjustment to the next video conferencing scene for large-scale team communication may be longer. Calculate the time decay factor based on these time intervals. For scene switches with shorter time intervals, the contribution decay rate of the corresponding historical scene type to the current scene perception weighting is slower, while for scene switches with longer time intervals, the contribution decay rate of its historical scene type is faster. For example, if the document editing scene switches to the video conferencing scene in a short time, then the weight decay of the document editing scene in the subsequent scene perception weighting is relatively slow because these two scenes are closer in time and may have a closer logical relationship. Conversely, for scene switches after a long time interval, such as the switch between a scene in the early stage of the project and a scene in the later stage of the project, the weight of the early scene in the later scene perception weighting will decay rapidly.

[0082] For example, when the project is approaching the end, due to the effect of the time decay factor, the weight value decay amplitude of the project management tool usage scene that has been frequently used recently is smaller, while for some scenes in the early stage of the project, since they are far from the current time, their weight values decay significantly after being adjusted by the time decay factor. In this way, the updated scene weight vector can better reflect the relative importance of different scene types at the current time.

[0083] Step S1245: Dynamically adjust the updated scene weight vector through the scene attention mechanism to generate a dynamic scene weight distribution. The scene attention mechanism enhances the weights of historical scenes with high relevance to the current scene by analyzing the semantic correlation between the context features of the current service scene and historical scene types.

[0084] In the project closing stage, the context features of the current service scenario may include requirements such as the review, collation, and final storage of project outcomes. The scenario attention mechanism analyzes the semantic relevance between the current scenario and historical scenario types. For example, the cloud storage service scenario has a high semantic relevance to the current project outcome storage requirement, while the video conferencing scenario, although very important in the project communication stage, has a relatively low semantic relevance to the current scenario in the project closing stage. Through the scenario attention mechanism, the weight of the cloud storage service scenario in the updated scenario weight vector will be enhanced because it has a high correlation with the current scenario. For other historical scenario types, such as the project management tool usage scenario, if there is still a certain task association with the current scenario in the project closing stage (such as the inspection and improvement of final tasks), its weight will also be enhanced accordingly. In this way, a dynamic scenario weight distribution is generated, which can dynamically adjust the weights of different historical scenario types according to the context features of the current scenario, making the scenario weights more in line with the requirements of the current scenario.

[0085] Step S1246: Perform a tensor multiplication operation on the dynamic scenario weight distribution and the cross-modal fusion feature to generate a weighted scenario association feature matrix, where each feature channel in the scenario association feature matrix corresponds to the preference expression of a specific scenario type.

[0086] Taking the cross-modal fusion feature of the project manager as an example, it includes fusion features related to the functional attributes of video conferencing services (such as high-definition video quality, screen sharing function, etc.), the functional attributes of project management tool services (such as task classification, progress tracking, etc.), and the functional attributes of cloud storage services (such as storage space size, data encryption method, etc.). These features are interrelated with the behavior features of the project manager in the same feature space. When performing a tensor multiplication operation on the dynamic scenario weight distribution and the cross-modal fusion feature, the weight of each scenario type will weight the corresponding cross-modal fusion feature. For example, if the weight of the cloud storage service scenario in the dynamic scenario weight distribution is high, then in the weighted scenario association feature matrix, the feature channel related to the cloud storage service attribute will receive more weight, and what this feature channel expresses is the project manager's preference for the cloud storage service in a specific scenario, such as the preference for secure and large-capacity cloud storage in the project closing stage. Similarly, for the video conferencing scenario and the project management tool usage scenario, etc., their corresponding feature channels will also be weighted according to their respective weights, thus generating a scenario association feature matrix where each feature channel corresponds to the preference expression of a specific scenario type.

[0087] Step S1247: Fuse the scenario association feature matrix and the currently collected current scenario context feature to generate an enhanced scenario feature tensor, and the enhanced scenario feature tensor retains the cross-scenario common pattern and the current scenario exclusive feature through a gated fusion unit.

[0088] In the project closing stage, the current scene context features collected in real time include the specific types of current project outcomes (such as documents, reports, presentations, etc.), the importance levels of the outcomes, relevant security requirements, etc. The scene association feature matrix is fused with these current scene context features. The gated fusion unit plays a key role in this process. It can distinguish which are cross-scene common patterns and which are current scene-specific features, and effectively retain them in the enhanced scene feature tensor. For example, in the process of project management, some basic logics of task management (such as the sequence and dependency of tasks, etc.) are cross-scene common patterns and exist regardless of which stage of the project. In the project closing stage, specific security requirements for project outcomes (such as high-level encryption requirements for confidential documents) are current scene-specific features. The gated fusion unit can ensure that these different types of features are reasonably retained during the fusion process, thus generating an enhanced scene feature tensor containing rich information.

[0089] Step S1248, perform scene consistency verification on the enhanced scene feature tensor. The scene consistency verification corrects the conflicting weights in the dynamic scene weight distribution by detecting the degree of orthogonality between the feature subspaces corresponding to different scene types.

[0090] Among different scene types involved in the entire project cycle, each scene type has a corresponding feature subspace in the enhanced scene feature tensor. For example, the feature subspace corresponding to the video conferencing scene contains behavior features and service attribute features related to the conference, and the feature subspace corresponding to the project management tool usage scene contains features related to task management, etc. By detecting the degree of orthogonality between the feature subspaces corresponding to these different scene types, it can be found whether there are weight conflicts. For example, if in the project closing stage, there is some confusion between the cloud storage service scene and the project management tool usage scene in certain features (possibly due to some cross-over in task association during project closing), through scene consistency verification, this situation can be detected. Then, the conflicting weights in the dynamic scene weight distribution are corrected according to the degree of orthogonality. If it is found that the cloud storage service scene has a non-orthogonal situation (i.e., there is confusion) with the feature subspace of the project management tool usage scene in a certain feature, the weight of the cloud storage service scene in the weight related to this feature may be appropriately reduced to ensure the independence and rationality between the feature subspaces corresponding to different scene types.

[0091] Step S1249, perform secondary weighting on the cross-modal fusion features according to the corrected dynamic scene weight distribution to generate an optimized scene-aware feature projection, and reduce the dimension of the optimized scene-aware feature projection to a preset dimensional space through a feature dimension compression algorithm to generate the scene-based feature vector. The linear combination relationship between the cross-scene shared feature components and the scene-specific feature components is retained during the dimension reduction process of the scene-based feature vector.

[0092] After the correction of the dynamic scene weight distribution in the previous steps, the corrected weight distribution is applied again to the cross-modal fusion features for secondary weighting. For example, in the corrected dynamic scene weight distribution, if the weight of a certain scene type (such as the usage scenario of project management tools) is adjusted on some key features, then during the secondary weighting process, the cross-modal fusion features related to the usage scenario of project management tools will be weighted according to the new weight. The generated optimized scene-aware feature projection can more accurately reflect the behavior preferences and demand characteristics of the project manager in the current scene because it comprehensively considers various factors such as the scene switching history, the current scene context features, and the relationships between different scene types.

[0093] Finally, due to the involvement of multiple service scenarios and a large amount of feature information, the optimized scene-aware feature projection may have a relatively high dimension. Through the feature dimension compression algorithm, its dimension is reduced to a preset dimensional space. During this process, it is necessary to ensure that the linear combination relationship between the cross-scene shared feature components and the scene-specific feature components is retained. For example, during the entire project cycle, the linear combination relationship between the basic logic of task assignment (cross-scene shared feature components) and the requirement for storing specific results in the project closing stage (scene-specific feature components) will not be damaged during the dimension reduction process. The generated scene-based feature vector can accurately represent the behavior characteristics of the project manager in the current scene with a lower dimension, including both the long-term cross-scene general preferences and the specific demand characteristics in the current scene, providing an effective feature representation for subsequent operations such as recommendation.

[0094] In a possible implementation manner, the training steps of the multi-modal recommendation model include:

[0095] Step S210, construct a multi-task learning framework, and the multi-task learning framework includes a service click-through rate prediction task, a service completion rate prediction task, and a service satisfaction prediction task.

[0096] For example, in the case of the target user being a project manager in a remote work scenario, the service click-through rate prediction task involves predicting the likelihood of the project manager clicking on specific service resources (such as specific functions in video conferencing software, specific task operation entrances in project management tools, etc.). For example, in video conferencing software, based on the project manager's historical behavior and current project requirements, predict the probability of him clicking on the screen sharing function. The service completion rate prediction task focuses on whether the project manager can successfully complete relevant tasks after using the service resources. For example, in a project management tool, predict whether he can complete all task assignments, tracking, and final project progress management according to the plan. The service satisfaction prediction task aims to evaluate the satisfaction level of the project manager with the service resources used (such as the security of cloud storage services, the stability of video conferencing services, etc.). By constructing such a multi-task learning framework, the multi-modal recommendation model can be optimized from multiple perspectives to better meet the needs of project managers in remote work.

[0097] Step S220: Extract entity relationship features from the service resource knowledge graph. The entity relationship features include the service provider association network, the service category hierarchy, and the service quality evaluation link.

[0098] In the remote work scenario, in terms of the service provider association network, taking video conferencing services as an example, different video conferencing service providers may have cooperation or competition relationships. Some providers may specialize in providing video conferencing services with high security and high concurrency processing capabilities for large enterprises, while others may focus on the simplicity and ease of use for small teams. The association relationships between these providers constitute part of the service provider association network. In terms of the service category hierarchy, for project management tool services, there are different levels of classification. Basic project management tools may only provide simple task creation and tracking functions, while advanced project management tools may also include complex functions such as risk management and resource allocation optimization. These different levels of function classification constitute the service category hierarchy. In terms of the service quality evaluation link, taking cloud storage services as an example, the quality evaluation of cloud storage services may be related to factors such as data security, storage stability, and access speed. User evaluations of cloud storage services will form a link. For example, user feedback data shows that a certain cloud storage service performs well in data encryption but occasionally has delays in access speed. These evaluation information are interrelated to form the service quality evaluation link.

[0099] Step S230: Align the entity relationship features with the user behavior features to generate a joint embedding representation. The heterogeneous feature alignment aggregates the attribute information of adjacent nodes through a graph neural network.

[0100] For the user behavior characteristics of the project manager, such as his participation frequency in video conferences, task adjustment frequency in project management tools, etc., they are aligned with the entity relationship characteristics extracted from the service resource knowledge graph. The graph neural network realizes this process by aggregating the attribute information of adjacent nodes. Taking a certain video conferencing service provider node in the service provider association network as an example, its adjacent nodes may include other service providers it cooperates with, the user groups using the services of this provider, etc. The graph neural network will aggregate the attribute information of these adjacent nodes, such as the technical advantages of the cooperative providers, the scale and industry distribution of the user groups, etc. At the same time, for the user behavior characteristics of the project manager, they are also regarded as nodes in the graph and associated with the relevant nodes in the service resource knowledge graph. In this way, entity relationship characteristics of different structures and types are integrated with user behavior characteristics to generate a joint embedding representation. This joint embedding representation can comprehensively reflect the connection between the behavior habits of the project manager and various attributes and relationships of the service resources.

[0101] Step S240, perform multi-task learning based on the joint embedding representation under the multi-task learning framework. During the multi-task learning process, adopt a dynamic gradient allocation strategy to adjust the loss weights of the service click-through rate prediction task, service completion degree prediction task, and service satisfaction prediction task. The dynamic gradient allocation strategy automatically adjusts the weight coefficients according to the convergence speeds of the service click-through rate prediction task, service completion degree prediction task, and service satisfaction prediction task in the current training batch.

[0102] For example, in a certain training batch, the service click-through rate prediction task may have a relatively fast convergence speed for predicting the function clicks of the project manager in the video conferencing software, which means that the model has learned relatively effective patterns in this task. For the service completion degree prediction task, such as predicting whether the project manager can complete tasks on time in the project management tool, the convergence speed may be slower and more training is needed for optimization. The service satisfaction prediction task, such as predicting the satisfaction of the project manager with the cloud storage service, also has its own convergence situation. According to these different convergence speeds, the dynamic gradient allocation strategy will automatically adjust the loss weights of each task. If the service click-through rate prediction task has a fast convergence speed, its loss weight may be appropriately reduced, while for the service completion degree prediction task and service satisfaction prediction task with slower convergence, their loss weights may be increased to prompt the model to pay more attention to and optimize these tasks, thereby improving the overall multi-task learning effect.

[0103] Step S250, enhance the generalization ability of the multi-modal recommendation model through an adversarial training mechanism. The adversarial training mechanism perturbs the user feature space by generating adversarial samples and forces the multi-modal recommendation model to maintain prediction consistency after perturbation.

[0104] For the user feature space of the project manager, the adversarial training mechanism generates adversarial samples. For example, for the user features of the project manager in the video conferencing service (such as the participation frequency, requirements for video quality, etc.), the adversarial training mechanism may generate some adversarial samples that are slightly different from the original features, such as changing the distribution of the participation frequency or adjusting the degree of requirements for video quality. Then, these adversarial samples are input into the multi-modal recommendation model, and the model needs to maintain the prediction consistency for tasks such as service click-through rate, service completion rate, and service satisfaction under the input of such perturbed user features. If the model predicts a high click-through rate of the project manager for a certain video conferencing function under the original user features, it should also make a similar prediction when facing adversarial samples. In this way, through continuous adversarial training, the multi-modal recommendation model can better adapt to various possible changes in user features, thereby enhancing its generalization ability.

[0105] Among them, the construction steps of the service resource knowledge graph include:

[0106] Step S310, collect the metadata information of the service resources, and the metadata information includes service description text, service provider qualifications, and service historical evaluation records.

[0107] Taking the video conferencing service as an example, the service description text may include the functions supported by the video conferencing service (such as high-definition video quality, multi-person screen sharing, etc.), applicable scenarios (such as large team meetings, one-on-one communications, etc.). In terms of service provider qualifications, it may include the technical strength of the provider, security certification status, etc. The service historical evaluation records reflect the usage experience of other users for the video conferencing service, such as whether it is stable, whether it is easy to operate, etc. For project management tool services and cloud storage services, the same metadata information in these aspects is also collected.

[0108] Step S320, extract entity nodes and relationship edges from the metadata information, where the entity nodes include service types, service providers, and service quality labels, and the relationship edges include service type subordination relationships, service provider cooperation relationships, and service quality association relationships.

[0109] Still taking the video conferencing service as an example, the entity node of the service type clarifies that it belongs to the service of the video conferencing type. The entity node of the service provider is the company or organization that provides the video conferencing service. The entity node of the service quality label may include labels such as high definition, stability, and security. The relationship edge of the service type subordination relationship indicates that the video conferencing service belongs to the large category of remote office services. The service provider cooperation relationship may be reflected in the cooperation relationship between the video conferencing service provider and other related service providers (such as cooperation with certain network equipment providers to optimize network transmission). The service quality association relationship indicates the association between the service quality label, the service provider, and the service type. For example, a certain service provider ensures that the video conferencing service has high definition and stable quality with its advanced technology.

[0110] Step S330, predict and fill the missing relationship edges based on the knowledge graph completion algorithm, and the knowledge graph completion algorithm uses the relationship path reasoning model to mine the implicit associations between entities.

[0111] Taking the project management tool service as an example, there may be an implicit association between some service providers and specific industry standards, but this relationship edge is missing in the initially constructed knowledge graph. The relationship path reasoning model will mine this implicit association by analyzing other known relationship edges (such as the relationship between the service provider and the enterprise type using the tool, the relationship between the enterprise type and the industry standard, etc.), and then predict and fill the missing relationship edges. This can make the service resource knowledge graph more complete and better reflect the complex relationships between service resources.

[0112] Step S340, dynamically update the knowledge graph. When a new service resource goes online or the attributes of an old service resource change, trigger the incremental graph update process and retain the historical version snapshot to support time series analysis.

[0113] For example, when a new video conferencing service goes online, the metadata information such as its new functions (such as brand new interactive functions) and the qualifications of the new service provider need to be integrated into the knowledge graph. At the same time, if the attributes of an old project management tool service change, such as adding new task management functions or adjusting the charging mode, the knowledge graph also needs to be updated in a timely manner. During the update process, retain the historical version snapshot, which can support the time series analysis of service resources. For example, it is possible to analyze the development and changes of a service resource at different time points and the impact of these changes on users such as project managers, so as to better provide comprehensive and accurate service resource information for the multi-modal recommendation model.

[0114] In a possible implementation manner, step S140 includes:

[0115] Step S141: Calculate the scenario matching scores for each service resource in the candidate service resource prediction set. The scenario matching scores are determined by analyzing the co-occurrence frequency of service resource attribute tags and the context features of the current service scenario.

[0116] For example, for a target user who is a project manager, in the service scenario of the current project closing, the candidate service resource prediction set may include a document management tool, a presentation production tool, a cloud storage service, etc. Taking the document management tool as an example, its service resource attribute tags include version management, multi-person collaborative editing, document permission setting, etc. The context features of the current service scenario include the need to perform final sorting, review, ensure the security of project documents, and facilitate the final collaborative modification by team members. If the version management function of the document management tool was frequently used in the previous project closing stage, and the multi-person collaborative editing function also matches the current need of team members to finalize the document, then the co-occurrence frequency of these attribute tags and the context features of the current scenario is relatively high, resulting in a higher scenario matching score. For the presentation production tool, its attribute tags such as rich template, animation effect addition, and presentation sharing function, if these functions match the current need to produce high-quality and easy-to-share presentations during project result display, will also obtain corresponding scenario matching scores. For the cloud storage service, its attribute tags such as storage space size, data encryption method, and access permission setting, if they match the current need for secure storage of project results and convenient access by team members during project closing, will also calculate corresponding scenario matching scores.

[0117] Step S142: Generate service resource availability metrics based on the real-time availability status of the service resources and the load capacity of the service provider. The service resource availability metrics are dynamically calculated by monitoring the response latency and concurrent processing capacity of service instances.

[0118] Still taking these candidate service resources as examples, for the document management tool, if its server has recently been maintained and upgraded, the response latency is very low, and it can handle concurrent access requests from multiple team members at the same time, this indicates that its real-time availability status is good and the load capacity of the service provider is sufficient, so its service resource availability metric is relatively high. For the presentation production tool, if it can still quickly respond to operation instructions without obvious lag when a large number of users are producing presentations simultaneously (such as when multiple projects in the company are preparing for result display at the same time), and the service provider has sufficient resources to handle this concurrency situation, its service resource availability metric will also be at a good level. If the cloud storage service responds quickly during data upload and download, and there is no overload even when project team members access the stored project results simultaneously, this means that its service resource availability metric is good.

[0119] Step S143: Integrate the scenario matching score and the service resource availability metric to generate a comprehensive recommendation priority. The comprehensive recommendation priority uses the weighted harmonic mean algorithm to balance the matching degree and availability.

[0120] For example, if the scenario matching score of a document management tool is high and the service resource availability metric is also good, after calculation using the weighted harmonic mean algorithm, its comprehensive recommendation priority will be relatively high. If the scenario matching score of a presentation-making tool is slightly lower, but the service resource availability metric is very high, or vice versa, after the weighted harmonic mean algorithm, a reasonable comprehensive recommendation priority will also be obtained. The cloud storage service is also calculated based on its scenario matching score and service resource availability metric to obtain the corresponding comprehensive recommendation priority.

[0121] Step S144: Sort the candidate service resources according to the comprehensive recommendation priority, and divide the sorting results into multiple recommendation levels. Each recommendation level corresponds to different display strategies and interaction methods.

[0122] The service resource with the highest comprehensive recommendation priority, such as a document management tool, may be ranked at the top and divided into the first recommendation level. The display strategy for this level can be to display it in a prominent prompt box on the main interface of the project manager's office software, and the interaction method can be to directly provide a quick entry to enter the relevant function page of the document management tool, such as directly entering the final review page of the project documents. The presentation-making tool with a slightly lower comprehensive recommendation priority may be divided into the second recommendation level. The display strategy can be to display it in a scroll bar form at a secondary position on the office software interface, and the interaction method can be to display its function features and usage instructions in detail only when the project manager clicks on the relevant area. If the comprehensive recommendation priority of the cloud storage service is even lower, it may be divided into the third recommendation level. The display strategy can be to display it in a specific function area of the software interface (such as the section related to data storage), and the interaction method can be to highlight its advantages only when the project manager actively searches for storage-related services.

[0123] Step S145: Control the diversity of the service resources in each recommendation level so that the distribution of the service resources within the same level in terms of type, provider, and price range meets the preset diversity threshold.

[0124] In the first recommendation level, assume that in addition to the document management tool, there are also candidate resources such as similar document collaboration platforms. By performing semantic embedding vector analysis on their functional description texts (for example, the document management tool emphasizes version management and permission settings, and the document collaboration platform emphasizes real-time collaboration and online editing), calculate the semantic embedding vector distances between them, thereby constructing a type similarity matrix. If the functional descriptions of two service resources are very close semantically, then their vector distances are smaller, and the similarity values in the type similarity matrix are higher.

[0125] In a possible implementation manner, step S145 includes:

[0126] Step S1451, calculate the type similarity matrix of the service resources within the recommendation level, where the type similarity matrix is determined by the semantic embedding vector distances of the functional description texts of the service resources.

[0127] Step S1452, construct a diversity-constrained optimization objective function, where the objective function maximizes the type difference degree of the service resources within the recommendation level and minimizes the user expectation deviation degree.

[0128] For example, in the first recommendation level, in order to maximize the type difference degree, service resources with different functions should be selected as much as possible. For example, the document management tool focuses on the overall management of documents, and if there is a tool dedicated to document format conversion and optimization, it can increase the type difference degree. At the same time, the user expectation deviation degree should be considered, and service resources that are too different from the needs of the project manager in the current scenario cannot be selected. For example, although there are some advanced document encryption tools, if document encryption is no longer a key requirement in the current project closing stage, selecting this tool will increase the user expectation deviation degree.

[0129] Step S1453, use the greedy algorithm to iteratively select service resources, and each iteration selects the service resource that can maximize the current diversity metric and does not violate the comprehensive recommendation priority constraint, generating a set of selected service resources.

[0130] For example, in the selection process of the first recommendation level, at the beginning, the document management tool may be selected first because of its high comprehensive recommendation priority and being the core tool for document management. Then in subsequent iterations, according to the type similarity matrix and the diversity-constrained optimization objective function, a tool with unique functions (such as being able to quickly generate a project document summary report) and not violating the comprehensive recommendation priority constraint is found, and it is added to the set of selected service resources.

[0131] Step S1454, perform post-processing filtering on the set of selected service resources, remove candidate resources whose similarity to the selected resources exceeds the threshold, and supplement alternative resources that are sub-optimal but have a high diversity contribution.

[0132] If the selected document management tool and a new candidate document management tool are very similar in function and exceed the set similarity threshold, remove this new candidate resource. Then, if there is an alternative resource that, although having a slightly lower comprehensive recommendation priority, has special functions (such as being able to perform voice annotation of documents) and can increase diversity, supplement it.

[0133] Step S1455, dynamically adjust the diversity weight coefficient. When it is detected that the user continuously rejects service resources of the same type, automatically increase the diversity penalty term corresponding to that type.

[0134] If the project manager continuously rejects several tool recommendations related to document management, the system will automatically increase the diversity penalty term for service resources of the document management type. In subsequent recommendations, it will be more inclined to select service resources of other types. For example, in the first recommendation level, it will consider more presentation-making tools or other non-document management-related service resources related to project closure, so as to improve the diversity of recommendations and better meet the needs of the project manager in the remote work scenario.

[0135] In a possible implementation manner, step S150 includes:

[0136] Step S151, extract explicit feedback signals and implicit feedback signals from the real-time feedback data. The explicit feedback signals include user ratings and service favorite marks, and the implicit feedback signals include service stay duration and interaction operation frequency.

[0137] In this embodiment, for the target user of the project manager, the user rating in the explicit feedback signal is his direct evaluation of the recommended service resources. For example, when a document management tool is recommended in the hierarchical recommendation list, the project manager rates it based on his own usage experience. If he believes that the document management tool's functions for document sorting and permission management in the project closure stage very much meet the requirements, he may give a higher rating; conversely, if there are inconveniences in operation or the functions do not meet expectations, the rating will be lower. The service favorite mark is also an explicit feedback signal. If the project manager adds a certain presentation-making tool to the favorite mark, this indicates that he is relatively satisfied with this tool, perhaps because of features such as rich templates and convenient animation effect production, which are suitable for project result display.

[0138] In terms of implicit feedback signals, the service residence duration reflects the time the project manager spends on a certain service resource. For example, when using a cloud storage service, if his residence duration in this service is long, it may mean that he is performing detailed storage operations on the project results, such as classifying and storing multiple files, setting different access permissions, etc., indicating that he has a relatively high demand for the functions of the cloud storage service. The frequency of interaction operations is also an important implicit feedback signal. Taking a project management tool as an example, if the project manager has a high frequency of interaction operations in this tool, frequently adjusting task status, viewing task progress, etc., it shows that he has a relatively high degree of dependence on the project management tool, and this tool plays an important role in the project management process.

[0139] Step S152, construct a feedback enhancement feature vector, which is generated by non-linearly fusing the explicit feedback signal and the implicit feedback signal.

[0140] For example, for the feedback data of the project manager, the non-linear fusion process will comprehensively consider factors such as user ratings, service favorite marks, service residence duration, and interaction operation frequency. For example, high user ratings and service favorite marks may have a positive impact on the final feedback enhancement feature vector, but if the service residence duration is short and the interaction operation frequency is low, it may be adjusted through a non-linear function so that the feedback enhancement feature vector can comprehensively reflect the overall attitude of the project manager towards the recommended service resources. This fusion method is not a simple linear combination, but captures the complex relationship between the explicit feedback signal and the implicit feedback signal through a specific non-linear function, thereby generating a feedback enhancement feature vector that can accurately represent the overall situation of user feedback.

[0141] Step S153, input the feedback enhancement feature vector into a parameter adjustment network, and output the model weight update amount. The parameter adjustment network uses a meta-learning framework to predict the influence intensity of different feedback patterns on the model parameters.

[0142] For example, the meta-learning framework can analyze the feedback pattern of the project manager. For example, if the feedback pattern of the project manager is that the score for the document management tool is low and the service residence duration is short, while the interaction operation frequency for other types of tools (such as presentation-making tools) is high, the parameter adjustment network can predict according to this feedback pattern that a relatively large adjustment needs to be made to the parameters related to the recommendation of the document management tool in the multi-modal recommendation model, and appropriate optimization needs to be made to the parameters related to the recommendation of the presentation-making tool. In this way, the model weight update amount is accurately output according to different feedback patterns to adjust the parameter weights of the multi-modal recommendation model so that it can better adapt to the changing needs of the project manager.

[0143] Step S154, perform knowledge distillation on the embedding layer of the multi-modal recommendation model, and retain the feature mapping relationship corresponding to the confidence prediction result with a confidence greater than the set confidence.

[0144] In a multi-modal recommendation model, the embedding layer is crucial for processing the feature representations of various service resources. For example, when performing feature embedding on a document management tool, a presentation production tool, and a cloud storage service, etc., the knowledge distillation process will screen these feature representations. If for a certain feature of the document management tool (such as the feature representation of the version management function), its confidence prediction result is high, that is, the model has a greater grasp of this feature, then the corresponding feature mapping relationship will be retained. For some feature representations with low confidence, they may be discarded during the knowledge distillation process. This can reduce the complexity of the model, while retaining the feature information that has an important impact on the recommendation result, and improving the efficiency and accuracy of the model.

[0145] Step S155, establish a model version rollback mechanism. When it is detected that the recommendation performance decreases due to consecutive batches of feedback data, automatically restore the model parameters to the historical stable version.

[0146] As the project progresses, the requirements of the project manager and the feedback on service resources will change continuously. If consecutive batches of feedback data show that the recommendation performance of the multi-modal recommendation model for the project manager gradually decreases, for example, the recommended service resources are increasingly inconsistent with the actual needs of the project manager, the user rating continues to decrease, the service stay time decreases, etc. At this time, the model version rollback mechanism will be automatically triggered to restore the model parameters to the parameters of a certain historical stable version. This historical stable version may be at a certain stage in the early stage of the project, when the recommendation result of the model could better meet the needs of the project manager. By restoring to the parameters of this version, the further decrease of the recommendation performance can be avoided, and the stability and reliability of the model to a certain extent can be ensured.

[0147] Among them, the training steps of the parameter adjustment network include:

[0148] Step S410, construct a meta-training task set, and each meta-training task in the meta-training task set simulates the model parameter adjustment requirements under a user feedback distribution pattern.

[0149] For example, in a telecommuting scenario, there may be multiple user feedback distribution patterns. For example, in one pattern, the project manager gives low ratings to all recommended service resources, and both the service stay duration and the frequency of interaction operations are very low. This may indicate that the recommended service resources do not meet the requirements at all. The corresponding meta-training task is to simulate how to adjust the parameters of the multi-modal recommendation model to improve the recommendation accuracy in this situation. In another pattern, the project manager may only have high ratings and frequent interaction operations for service resources of a specific type (such as presentation-making tools), and pay less attention to other types of service resources. In this pattern, the meta-training task is to adjust the model parameters according to this feedback distribution so that the recommendation model can recommend more service resources that meet the user preferences, such as presentation-making tools.

[0150] Step S420, in the meta-training stage, adopt a model-agnostic meta-learning algorithm to enable the parameter adjustment network to adapt to feedback patterns that have not been learned.

[0151] For example, the feedback pattern of the project manager may be very complex and diverse, and some feedback patterns may not have appeared in the previous training data. The model-agnostic meta-learning algorithm can enable the parameter adjustment network to still make reasonable parameter adjustments according to the simulation situations in the meta-training task set when facing these new feedback patterns. For example, when a new feedback pattern appears, that is, the project manager has unique feedback behaviors for a newly launched service resource (such as a new type of project collaboration platform), the model-agnostic meta-learning algorithm can make corresponding parameter adjustment decisions based on the adjustment experience of various previously simulated feedback patterns, so that the parameter adjustment network has stronger generalization ability.

[0152] Step S430, enable a fast update path and a slow update path. The fast update path generates temporary model parameters according to the feedback data of the current batch, and the slow update path optimizes the basic parameters based on the long-term feedback trend.

[0153] For example, for the feedback data of project managers, the fast update path reacts promptly based on the feedback of the current batch. For example, if a project manager gives a very low score to a newly recommended document review tool in the current batch, the fast update path quickly generates temporary model parameters and adjusts the part related to the recommendation of the document review tool to improve the next recommendation result as soon as possible. The slow update path, on the other hand, pays more attention to the long-term feedback trend. For example, throughout the project cycle, the overall preference trend of project managers for different types of service resources (such as document management tools, presentation production tools, etc.). If the long-term trend shows that project managers are increasingly emphasizing a certain function (such as the secure storage function of documents), the slow update path optimizes the basic parameters based on this long-term feedback trend to ensure that the model's recommendation results can meet the needs of project managers in the long run.

[0154] Step S440, introduce a gradient similarity constraint term to make the parameter adjustment direction semantically consistent with the original optimization goal of the recommendation model.

[0155] For a multi-modal recommendation model, its original optimization goal may be to improve the accuracy, satisfaction, etc. of the recommendation. When adjusting the parameters according to the feedback data of project managers, the gradient similarity constraint term ensures that the parameter adjustment direction is consistent with this original optimization goal. For example, if the original optimization goal is to improve the satisfaction of project managers with the recommended service resources, then when adjusting the parameters, there should be no situation that goes against this goal. If a parameter adjustment causes the recommendation result to deviate more from the needs of project managers and violates the goal of improving satisfaction, the gradient similarity constraint term will impose restrictions to ensure that the parameter adjustment direction is towards improving satisfaction, thus maintaining the semantic consistency between the parameter adjustment and the original optimization goal of the recommendation model.

[0156] Step S450, evaluate the effectiveness of the parameter adjustment through an adversarial verification mechanism. When the performance gain of the adjusted recommendation model on the validation set is lower than the threshold, trigger an artificial intervention alert.

[0157] For example, the validation set can contain some previous feedback data of project managers or simulated similar feedback data. When adjusting the parameters of the multi-modal recommendation model according to the feedback of project managers, use the adversarial verification mechanism to evaluate the effectiveness of the adjustment. If the performance gain of the adjusted model on the validation set (such as indicators like the increase in user ratings, the increase in service stay duration, etc.) is lower than the set threshold, this indicates that the parameter adjustment may not have achieved the expected effect. At this time, trigger an artificial intervention alert. The artificial intervention can further analyze the reasons, which may be special situations in the feedback data or the need to adjust the structure of the model itself, etc., and then take corresponding measures to improve the recommendation performance of the model, ensuring that the model can make effective parameter adjustments according to the feedback of project managers and better adapt to the changing needs in the remote work scenario.

[0158] Figure 2 shows the hardware structure diagram of the remote digital service system 100 provided by the embodiments of the present application for implementing the above-mentioned remote digital service resource recommendation method based on artificial intelligence mining, as Figure 2 shown, the remote digital service system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0159] In a possible design, the remote digital service system 100 may be a single server or a server group. The server group may be centralized or distributed (for example, the remote digital service system 100 may be a distributed system). In some embodiments, the remote digital service system 100 may be local or remote. For example, the remote digital service system 100 may access information and / or data stored in the machine-readable storage medium 120 via a network. For another example, the remote digital service system 100 may be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the remote digital service system 100 may be implemented on a server. By way of example only, the server may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc. or any combination thereof.

[0160] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store the data and / or instructions used by the remote digital service system 100 to execute or use to complete the exemplary methods described in the present application.

[0161] In a specific implementation process, one or more processors 110 execute the computer-executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the remote digital service resource recommendation method based on artificial intelligence mining in the above method embodiments. The processor 110, the machine-readable storage medium 120, and the communication unit 140 are connected through the bus 130, and the processor 110 may be used to control the transceiver actions of the communication unit 140.

[0162] For the specific implementation process of the processor 110, reference may be made to the various method embodiments executed by the above-mentioned remote digital service system 100. The implementation principles and technical effects are similar, and will not be elaborated here in this embodiment.

[0163] In addition, an embodiment of the present application further provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the above-mentioned remote digital service resource recommendation method based on artificial intelligence mining is implemented.

[0164] It should be noted that, in order to simplify the description of the present application disclosure and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, sometimes multiple features are merged into one embodiment, drawing or description thereof.

Claims

1. A method for recommending remote digital service resources based on artificial intelligence mining, characterized in that, The method includes: Obtaining the historical interaction data of the target user, where the historical interaction data includes the interaction records of the target user with service resources in multiple remote service scenarios, service resource attribute information, and user behavior sequences; Generating a dynamic feature vector of the target user based on the historical interaction data, where the dynamic feature vector is used to characterize the preference distribution and service demand evolution path of the target user in different service scenarios; Inputting the dynamic feature vector into a pre-trained multi-modal recommendation model to output a predicted set of candidate service resources for the target user in the current service scenario. The multi-modal recommendation model mines the potential matching patterns between user behavior and service resources by integrating a temporal behavior analysis model and a cross-scenario semantic association model; Generating a hierarchical recommendation list according to the real-time scenario fitness of each service resource in the predicted set of candidate service resources. The real-time scenario fitness is determined by analyzing the dynamic association strength between service resource attributes and the context features of the current service scenario. The hierarchical recommendation list divides service resources into multiple priority clusters according to a fitness threshold; Based on the real-time feedback data of the target user on the hierarchical recommendation list, dynamically adjusting the parameter weights of the multi-modal recommendation model and updating the generation logic of the dynamic feature vector to form a closed-loop optimization link; The generating the dynamic feature vector of the target user based on the historical interaction data includes: Extracting the user behavior time series segments from the historical interaction data, where the user behavior time series segments include continuous service request intervals, service type switching frequencies, and service resource usage depths; Performing multi-granularity feature extraction on the user behavior time series segments through a pre-trained spatio-temporal attention network to generate an initial behavior feature matrix. The spatio-temporal attention network identifies key service interaction nodes by capturing the dependencies of user behavior in the time and space dimensions; Fusing the service resource attribute information and the initial behavior feature matrix to generate a cross-modal fusion feature. The cross-modal fusion feature embeds service resource attributes into the user behavior feature space through a cross-modal alignment algorithm; Performing scene-aware weighting on the cross-modal fusion feature according to the scene switching marks of the target user in multiple service scenarios to generate a scene-based feature vector; Dynamically updating the scene-based feature vector based on a gated recurrent unit to generate the dynamic feature vector, where the dynamic feature vector contains the long-term preference features of the user and the short-term demand features of the current service scenario. The training steps of the multi-modal recommendation model include: Constructing a multi-task learning framework, where the multi-task learning framework includes a service click-through rate prediction task, a service completion rate prediction task, and a service satisfaction prediction task; Extracting entity relationship features from a service resource knowledge graph, where the entity relationship features include a service provider association network, a service category hierarchy, and a service quality evaluation link; Aligning the entity relationship features with user behavior features to generate a joint embedding representation. The heterogeneous feature alignment aggregates the attribute information of adjacent nodes through a graph neural network; Perform multi-task learning based on the joint embedding representation under the multi-task learning framework. During the multi-task learning process, adopt a dynamic gradient allocation strategy to adjust the loss weights of the service click-through rate prediction task, service completion rate prediction task, and service satisfaction prediction task. The dynamic gradient allocation strategy automatically adjusts the weight coefficients according to the convergence speeds of the service click-through rate prediction task, service completion rate prediction task, and service satisfaction prediction task in the current training batch; Enhance the generalization ability of the multi-modal recommendation model through an adversarial training mechanism. The adversarial training mechanism perturbs the user feature space by generating adversarial samples and forces the multi-modal recommendation model to maintain prediction consistency after perturbation.

2. The remote digital service resource recommendation method based on artificial intelligence mining according to claim 1, wherein The multi-granularity feature extraction of the user behavior time-series segment through a pre-trained spatio-temporal attention network to generate an initial behavior feature matrix includes: Divide the user behavior time-series segment into multiple time windows, and the behavior data within each time window forms a local behavior subsequence; Perform a convolutional kernel sliding operation on each local behavior subsequence to extract local spatio-temporal features. The convolutional kernel sliding operation uses a multi-scale convolutional kernel group to capture behavior patterns with different time spans; Input the local spatio-temporal features into a bidirectional long short-term memory network to generate a time-series context encoding. The bidirectional long short-term memory network captures the implicit dependencies between time windows through forward and backward paths; Construct a spatio-temporal attention weight matrix. The spatio-temporal attention weight matrix dynamically adjusts the contribution degree of each time window to the final feature representation by calculating the correlation strength of behavior features within different time windows; Perform weighted aggregation on the time-series context encoding according to the spatio-temporal attention weight matrix to generate the initial behavior feature matrix.

3. The remote digital service resource recommendation method based on artificial intelligence mining according to claim 2, wherein The construction of the spatio-temporal attention weight matrix includes: Calculate the behavior similarity between any two time windows. The behavior similarity measures the matching degree of the service type distribution and service interaction intensity within two time windows through the cosine similarity algorithm; Generate an initial attention score matrix according to the behavior similarity and perform sparsification processing on the initial attention score matrix, retaining the attention connections exceeding the preset attention score; Add the sparsified attention score matrix to the position encoding matrix to generate an enhanced attention matrix; Perform normalization processing on the enhanced attention matrix through a normalized exponential function to generate the spatio-temporal attention weight matrix.

4. The remote digital service resource recommendation method based on artificial intelligence mining according to claim 1, wherein The step of performing scene-aware weighting on the cross-modal fusion feature according to the scene switching marks of the target user in multiple service scenarios to generate a scene-based feature vector includes: Extract the scene switching time-series sequence in the scene switching mark. The scene switching time-series sequence contains the scene type identifiers triggered by the user within continuous time units and the scene switching trigger event types; Generate an initial scene weight vector according to the scene association strength corresponding to the scene switching trigger event type. Each dimension of the initial scene weight vector reflects the cumulative influence degree of the corresponding scene type in historical interactions; Calculate a time decay factor based on the time interval between adjacent scene switches in the scene switch timing sequence, where the time decay factor is used to dynamically adjust the contribution decay rate of the historical scene type to the current scene perception weighting; Perform an element-wise multiplication operation on the time decay factor and the initial scene weight vector to generate an updated scene weight vector, where the closer the scene type is to the current time in the updated scene weight vector, the smaller the attenuation amplitude of the corresponding weight value; Dynamically adjust the updated scene weight vector through a scene attention mechanism to generate a dynamic scene weight distribution, where the scene attention mechanism enhances the weights of historical scenes with high relevance to the current scene by analyzing the semantic correlation between the context features of the current service scene and the historical scene types; Perform a tensor multiplication operation on the dynamic scene weight distribution and the cross-modal fusion features to generate a weighted scene association feature matrix, where each feature channel in the scene association feature matrix corresponds to the preference expression of each scene type; Fuse the scene association feature matrix and the context features of the current scene collected in real time to generate an enhanced scene feature tensor, where the enhanced scene feature tensor retains the cross-scene common pattern and the current scene-specific features through a gated fusion unit; Perform scene consistency verification on the enhanced scene feature tensor, where the scene consistency verification corrects the conflicting weights in the dynamic scene weight distribution by detecting the degree of orthogonality between the feature subspaces corresponding to different scene types; Perform secondary weighting on the cross-modal fusion features according to the corrected dynamic scene weight distribution to generate an optimized scene perception feature projection; Reduce the dimension of the optimized scene perception feature projection to a preset dimensional space through a feature dimension compression algorithm to generate the sceneized feature vector, where the sceneized feature vector retains the linear combination relationship of the cross-scene shared feature components and the scene-specific feature components during the dimension reduction process.

5. The remote digital service resource recommendation method based on artificial intelligence mining according to claim 1, wherein, The construction steps of the service resource knowledge graph include: Collect the metadata information of the service resources, where the metadata information includes service description text, service provider qualifications, and service historical evaluation records; Extract entity nodes and relationship edges from the metadata information, where the entity nodes include service types, service providers, and service quality labels, and the relationship edges include service type subordination relationships, service provider cooperation relationships, and service quality association relationships; Predict and fill in the missing relationship edges based on a knowledge graph completion algorithm, where the knowledge graph completion algorithm uses a relationship path reasoning model to mine the implicit associations between entities; Dynamically update the knowledge graph. When a new service resource goes online or the attributes of an old service resource change, trigger an incremental graph update process and retain historical version snapshots to support time series analysis.

6. The remote digital service resource recommendation method based on artificial intelligence mining according to claim 1, characterized in that The generation of a hierarchical recommendation list according to the real-time scene adaptability of each service resource in the candidate service resource prediction set includes: Calculate the scene matching score of each service resource in the candidate service resource prediction set, where the scene matching score is determined by analyzing the co-occurrence frequency of the service resource attribute labels and the context features of the current service scene; Generate service resource availability metrics based on the real-time available status of service resources and the load capacity of service providers, where the service resource availability metrics are dynamically calculated by monitoring the response latency and concurrent processing capabilities of service instances; Fuse the scenario matching score and the service resource availability metrics to generate a comprehensive recommendation priority, where the comprehensive recommendation priority uses a weighted harmonic mean algorithm to balance the matching degree and availability; Sort the candidate service resources according to the comprehensive recommendation priority, and divide the sorting results into multiple recommendation levels, with each recommendation level corresponding to different display strategies and interaction methods; Control the diversity of service resources at each recommendation level so that the distribution of service resources within the same level in terms of type, provider, and price range meets a preset diversity threshold.

7. The remote digital service resource recommendation method based on artificial intelligence mining according to claim 6, wherein The controlling the diversity of service resources at each recommendation level includes: Calculate the type similarity matrix of service resources within the recommendation level, where the type similarity matrix is determined by the semantic embedding vector distance of the functional description text of service resources; Construct a diversity constraint optimization objective function that maximizes the type difference degree of service resources within the recommendation level and minimizes the deviation from user expectations; Use a greedy algorithm to iteratively select service resources, and each iteration selects the service resource that can maximize the current diversity metric and does not violate the comprehensive recommendation priority constraint to generate a set of selected service resources; Perform post-processing filtering on the set of selected service resources to remove candidate resources whose similarity to the selected resources exceeds the threshold; Dynamically adjust the diversity weight coefficient, and when it is detected that the user continuously rejects service resources of the same type, automatically increase the diversity penalty term for the corresponding type.

8. The remote digital service resource recommendation method based on artificial intelligence mining according to claim 1, wherein The dynamically adjusting the parameter weights of the multi-modal recommendation model based on the real-time feedback data of the target user on the hierarchical recommendation list includes: Extract explicit feedback signals and implicit feedback signals from the real-time feedback data, where the explicit feedback signals include user ratings and service collection marks, and the implicit feedback signals include service stay duration and interaction operation frequency; Construct a feedback-enhanced feature vector, which is generated by non-linearly fusing the explicit feedback signals and the implicit feedback signals; Input the feedback-enhanced feature vector into a parameter adjustment network to output a model weight update amount, where the parameter adjustment network uses a meta-learning framework to predict the influence intensity of different feedback modes on model parameters; Perform knowledge distillation on the embedding layer of the multi-modal recommendation model, and retain the feature mapping relationship corresponding to the confidence prediction result with a confidence greater than the set confidence; Establish a model version rollback mechanism, and when it is detected that the recommendation performance degrades due to consecutive batches of feedback data, automatically restore to the model parameters of the historical stable version; Among them, the training steps of the parameter adjustment network include: Construct a meta-training task set, where each meta-training task in the meta-training task set simulates the model parameter adjustment requirements under a user feedback distribution pattern; In the meta-training stage, use a model-agnostic meta-learning algorithm to enable the parameter adjustment network to adapt to feedback modes that have not been learned. Enable a fast update path and a slow update path, where the fast update path generates temporary model parameters based on the feedback data of the current batch, and the slow update path optimizes the basic parameters based on the long-term feedback trend; Introduce a gradient similarity constraint term to make the parameter adjustment direction semantically consistent with the original optimization objective of the recommendation model; Evaluate the effectiveness of parameter adjustment through an adversarial verification mechanism. When the performance gain of the adjusted recommendation model on the validation set is lower than the threshold, trigger an artificial intervention alert.

9. A remote digital service system, characterized in that, The remote digital service system includes a processor and a memory. The memory is connected to the processor. The memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the remote digital service resource recommendation method based on artificial intelligence mining according to any one of claims 1-8 above.

Images