Remote digital service resource recommendation method and system based on artificial intelligence mining

Abstract

The invention provides a remote digital service resource recommendation method and system based on artificial intelligence mining, and the method comprises the steps: firstly obtaining historical interaction data, including interaction records, service resource attribute information and user behavior sequences, of a target user in a plurality of remote service scenes; and generating a dynamic feature vector representing user preference distribution and a service demand evolution path. And inputting the candidate service resource prediction set into a multi-modal recommendation model fusing time sequence behavior analysis and a cross-scene semantic association model, mining a potential matching mode, and outputting the candidate service resource prediction set of the current scene. Generating a hierarchical recommendation list according to the real-time scene adaptation degree, and dividing a service resource priority cluster according to an adaptation degree threshold value; and finally, dynamically adjusting model parameter weights based on user real-time feedback data, updating dynamic feature vector generation logic, and forming a closed-loop optimization link, thereby effectively improving the accuracy and adaptability of remote digital service resource recommendation.

Description

Technical Field

[0001] The present application relates to the field of cloud office technology, and more specifically, 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 more and more widespread. How to accurately recommend appropriate service resources to users has become a key issue in improving service quality and user experience.

[0003] In the early days, traditional service resource recommendation methods were mainly based on simple rules or static user portraits. For example, some recommendation systems only recommend similar services based on the types of services that users have 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 with changes in time, scenarios, and their own behavior.

[0004] With the development of technology, some recommendation systems have begun to try to use user behavior data, but most of them are limited to single-dimensional analysis. For example, they only focus on users' click behavior to build recommendation strategies, ignoring the rich attribute information of the service resources themselves and the complex connections between different service scenarios. As a result, the recommendation results often cannot accurately meet the real needs of users in specific scenarios, resulting in a large gap between the recommended service resources and user expectations, and poor user experience. Summary of the invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present application, an embodiment of the present application provides a remote digital service resource recommendation method based on artificial intelligence mining, the method comprising: Acquire historical interaction data of a target user, wherein the historical interaction data includes interaction records between the target user and service resources in multiple remote service scenarios, service resource attribute information, and user behavior sequences; Generate 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; The dynamic feature vector is input into a pre-trained multimodal recommendation model, and a predicted set of candidate service resources for the target user in the current service scenario is output. The multimodal recommendation model mines potential matching patterns between user behavior and service resources by fusing a temporal behavior analysis model and a cross-scenario semantic association model; 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; 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.

[0006] 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.

[0007] 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.

[0008] 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.

[0009] 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.

[0010] 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. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the execution flow of the remote digital service resource recommendation method based on artificial intelligence mining provided in an embodiment of the present application.

[0012] Figure 2 It is a schematic diagram of the hardware architecture of the remote digital service system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0013] The present application will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a 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 is introduced in detail below.

[0014] Step S110, acquiring historical interaction data of the target user, wherein the historical interaction data includes interaction records between the target user and service resources in multiple remote service scenarios, service resource attribute information, and user behavior sequences.

[0015] In this embodiment, taking the remote office scenario as an example, the target user may be a project manager. In his daily work, he will be involved in the interaction of multiple remote service scenarios. For example, he often uses video conferencing services to communicate on projects with team members distributed in different regions. The interaction record of each use of the video conferencing service includes 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 image quality, screen sharing function, and an upper limit on the number of participants. The user behavior sequence may be manifested as follows: he first opens the video conferencing software, checks whether the participants are present, and then turns on the screen sharing function to display the project documents, adjusts the image quality during the meeting to adapt to the network conditions, etc.

[0016] He also uses project management tool services, and the interaction records include operations such as creating project tasks, assigning tasks to team members, and setting task deadlines. The service resource attributes of project management tools include task classification functions and task progress tracking functions. The user behavior sequence may be to first create a project framework task, then assign different subtasks to members based on their expertise, and constantly check the task progress during the process.

[0017] 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 cloud storage services 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 based on the importance and confidentiality of the files. These interaction records with various service resources in multiple remote service scenarios, service resource attribute information, and user behavior sequences together constitute the historical interaction data of the target user.

[0018] Step S120: generating a dynamic feature vector of the target user based on the historical interaction data, wherein the dynamic feature vector is used to characterize the preference distribution and service demand evolution path of the target user in different service scenarios.

[0019] For example, for this project manager, extracting user behavior time series segments from historical interaction data. Taking video conferencing services as an example, the interval between consecutive service requests may reflect his project communication rhythm. For example, he will start video conferences every Monday, Wednesday, and Friday morning, which reflects his tendency to conduct concentrated communication during specific time periods on weekdays. In terms of the frequency of switching between service types, he may switch to the project management tool to check the task updates immediately after the video conference. This frequent switching behavior indicates his need for collaborative use of different services. In terms of the depth of service resource usage, his frequent adjustments to the image quality and screen sharing functions of video conferences show that he is highly dependent on these functions.

[0020] The pre-trained spatiotemporal attention network is used to extract multi-granular features from these user behavior time series fragments to generate an initial behavior feature matrix. The user behavior time series fragments are divided into multiple time windows, for example, one week as a time window. In each time window, the local behavior subsequence contains all video conferencing related operations in that week. The convolution kernel sliding operation is performed on each local behavior subsequence. The use of multi-scale convolution kernel groups can capture behavior patterns of different time spans, such as capturing the difference between his video conferencing usage pattern at the key node of the project and at ordinary times. The local spatiotemporal features are input into the 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 his video conferencing usage habits in the early stage of the project and the possible association in the later stage of the project. When constructing the spatiotemporal attention weight matrix, the behavior similarity between any two time windows is calculated, such as calculating the cosine similarity of the distribution of participants and meeting duration in video conferences in different weeks. The initial attention score matrix is ​​generated according to the behavioral similarity and sparsely processed. The attention connections exceeding the preset attention score are retained and then added to the position encoding matrix. Finally, the spatiotemporal attention weight matrix is ​​generated through the normalized exponential function to dynamically adjust the contribution of each time window to the final feature representation, thereby generating the initial behavioral feature matrix.

[0021] Then, the service resource attribute information and the initial behavior feature matrix are fused to generate cross-modal fusion features. For example, the attributes of the video conferencing service, such as high-definition picture quality and screen sharing function, are embedded into the user behavior feature space through the cross-modal alignment algorithm. The scene switching marks of the target user in multiple service scenarios are weighted by scene perception to generate a scenario feature vector. The scene switching time sequence in the scene switching mark is extracted. Assuming that he will switch from the video conferencing scene to the project management tool scene when the project stage is switched, the scene switching trigger event type may be the advancement of the project stage. According to the scene association strength corresponding to the scene switching trigger event type, the initial scene weight vector is generated. For example, in the key decision-making stage of the project, the weight of the video conferencing scene will be higher. The time decay factor is calculated based on the time interval between adjacent scene switches in the scene switching time sequence. As time goes by, the scene weight of the previous project stage will gradually decay. The time decay factor is element-wise multiplied with the initial scene weight vector, and then dynamically adjusted through the scene attention mechanism to generate a dynamic scene weight distribution. The dynamic scene weight distribution is tensor-multiplied with the cross-modal fusion feature to generate a weighted scene association feature matrix. The scene-related feature matrix is ​​integrated with the current scene context features collected in real time. For example, if the current project is in the final stage, more document sorting and review work is required. The corresponding current scene context features are the focus on the completion of tasks in cloud storage services and project management tools. The enhanced scene feature tensor is verified for scene consistency to ensure that the orthogonality between feature subspaces corresponding to different scene types is reasonable, and the conflicting weights in the dynamic scene weight distribution are corrected. The cross-modal fusion features are weighted twice according to the corrected dynamic scene weight distribution to generate an optimized scene-aware feature projection. Finally, the feature dimension compression algorithm is used to reduce the dimension to the preset dimensional space to generate a scenario-based feature vector. This vector contains his long-term preference characteristics, such as the preference for efficient communication tools, and the short-term demand characteristics for cloud storage services and project management tools in the final stage of the current project.

[0022] Step S130, input the dynamic feature vector into a pre-trained multimodal recommendation model, and output a predicted set of candidate service resources for the target user in the current service scenario. The multimodal recommendation model mines potential matching patterns between user behavior and service resources by integrating a temporal behavior analysis model and a cross-scenario semantic association model.

[0023] For example, for this project manager, the current service scenario is assumed to be that the project is about to end, and the project results need to be summarized and presented. The previously generated dynamic feature vector is input into the pre-trained multimodal recommendation model. The temporal behavior analysis model in the multimodal recommendation model will analyze his previous behavior patterns in similar project stages. For example, when the previous project was about to end, his frequency of using document organization tools and presentation production tools increased. The cross-scenario semantic association model will mine the semantic associations between different scenarios. For example, the video conferencing service he used when communicating the project plan with the customer may be related to the presentation of project results, because it may be necessary to show the customer a presentation of the project results.

[0024] Through the integration of these models, potential matching patterns between user behavior and service resources are mined, thereby outputting a set of candidate service resource predictions. This set may include some document management tools suitable for summarizing project results, presentation production software with good presentation effects, and cloud storage service upgrade options that can safely store the final results of the project. These service resources are predicted based on the analysis of his historical behavior and semantic associations in different scenarios, and can meet his potential needs in the service scenario where the current project is about to end.

[0025] Step S140, 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 correlation strength between the service resource attributes and the contextual features of the current service scenario, and the hierarchical recommendation list divides the service resources into multiple priority clusters according to the adaptability threshold.

[0026] 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 the current scenario requires the final version of the project document to be determined and jointly reviewed by team members, then the co-occurrence frequency of these attribute labels and the contextual features of the current scenario is high, resulting in a higher scenario matching score.

[0027] The service resource availability index is generated based on the real-time available status of the service resources and the load capacity of the service provider. Assuming that the server where the document management tool is located has recently been maintained and upgraded, the response delay is low, and it can withstand concurrent access by a large number of users, then its service resource availability index is high.

[0028] The scenario matching score and service resource availability index are integrated, and a weighted harmonic average algorithm is used to generate a comprehensive recommendation priority. For example, if the scenario matching score of a document management tool is high and its availability index is also good, its comprehensive recommendation priority will be high.

[0029] The candidate service resources are sorted according to the comprehensive recommendation priority, and the sorting results are divided into multiple recommendation levels. For example, the document management tools and presentation production software with the highest comprehensive recommendation priority are classified into the first recommendation level. The display strategy corresponding to this level may be to recommend them in a prominent position on the project manager's office software interface, and the interactive method may be to directly pop up a prompt box to inform of its functional advantages. The cloud storage service upgrade option with a slightly lower comprehensive recommendation priority is classified into the second recommendation level. The display strategy may be to display it in a secondary position on the software interface, and the interactive method may be to display it in detail when the user actively views storage-related services.

[0030] Diversity control is performed on the service resources of each recommendation level. The type similarity matrix of service resources in the recommendation level is calculated. For example, document management tools and presentation software have certain functional differences, but they both belong to the type of tools related to project results organization and display. A diversity constraint optimization objective function is constructed, the goal is to maximize the type difference of service resources in the recommendation level and minimize the deviation from user expectations. A greedy algorithm is used to iteratively select service resources. Each iteration selects service resources that can maximize the current diversity index and do not violate the comprehensive recommendation priority constraint. Assume that in the first recommendation level, in addition to document management tools and presentation software, an online collaboration platform is also considered. This platform can facilitate the final communication and adjustment of team members in the final stage. Post-process and filter the selected service resource set to remove candidate resources whose similarity with 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 suboptimal but high diversity contribution alternative resource will be added. For example, if there is a presentation tool with a special display effect but a slightly lower comprehensive recommendation priority, it can be added. Dynamically adjust the diversity weight coefficient. If the project manager rejects multiple presentation software-related recommendations in succession, the corresponding type of diversity penalty item will be automatically added to increase the diversity of recommendations.

[0031] 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 multimodal recommendation model, and update the generation logic of the dynamic feature vector to form a closed-loop optimization link.

[0032] When the project manager is presented with the hierarchical recommendation list, 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 his favorites, it indicates that he is relatively satisfied with these two service resources. In terms of implicit feedback signals, he stays in the document management tool for a long time and interacts with the online collaboration platform more frequently, which also reflects his attention to these service resources.

[0033] Construct a feedback enhancement feature vector and perform nonlinear fusion of explicit feedback signals and implicit feedback signals. For example, the rating, favorite mark, service stay time, and interaction operation frequency are fused through a specific nonlinear function to generate a feedback enhancement feature vector.

[0034] The feedback enhancement feature vector is input into the parameter adjustment network, and the model weight update is output. The parameter adjustment network uses a meta-learning framework to predict the impact of different feedback modes on model parameters. For example, if the project manager gives a high score to the document management tool, the parameter adjustment network will analyze that this feedback mode has a greater impact on the parameters related to the document management tool recommendation in the model, and output the corresponding model weight update.

[0035] Perform knowledge distillation on the embedding layer of the multimodal recommendation model, and retain the feature mapping relationships corresponding to the confidence prediction results with confidence greater than the set confidence. For example, in the previous recommendation, the recommendation of document management tools is based on a certain feature mapping relationship. If the confidence of this recommendation is high, then this feature mapping relationship is retained when adjusting the model.

[0036] A model version rollback mechanism is established. When it is detected that the recommendation performance has degraded due to multiple batches of feedback data, the model parameters are automatically restored to the historical stable version. For example, if the service resources such as document management tools and presentation software recommended later no longer meet the needs of the project manager, resulting in a decrease in his satisfaction, when this happens repeatedly, the model automatically restores the parameters of the previous version that can accurately recommend.

[0037] By processing the real-time feedback data of the target users on the hierarchical recommendation list, dynamically adjusting the parameter weights of the multimodal recommendation model and updating the generation logic of the dynamic feature vector, a closed-loop optimization link is formed, which enables the recommendation system to continuously adapt to the changing needs of the target users in remote office scenarios and improve the accuracy and effectiveness of the recommendations.

[0038] 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.

[0039] 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.

[0040] 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.

[0041] 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.

[0042] In a possible implementation, step S120 includes: Step S121, extracting user behavior time series segments from the historical interaction data, wherein the user behavior time series segments include continuous service request intervals, service type switching frequency, and service resource usage depth.

[0043] In this embodiment, in the remote office scenario, the continuous service request interval reflects the work rhythm of the project manager. For example, when he uses the video conferencing service during the project cycle, the time interval between each meeting request may have a certain regularity, such as a video conference once a week in the early stage of the project. As the project advances to the critical stage, the meeting request interval is shortened to once every three days. The change in this interval reflects his demand for communication frequency at different project stages. In terms of the frequency of switching service types, after completing the video conference, he often quickly switches to the project management tool to check the progress of the task. Sometimes he switches between video conferencing, project management tools, and cloud storage services many times a day. This frequency reflects his demand for collaborative work between different services. The depth of service resource use also has obvious characteristics. Taking the project management tool as an example, he will deeply use the functions such as task allocation and progress tracking during the project execution stage, and frequently adjust the priority and deadline of the task. The deep use of the function reflects his dependence on the function of the project management tool. These continuous service request intervals, service type switching frequency, and service resource use depth together constitute the user behavior time series fragment.

[0044] Step S122, performing multi-granularity feature extraction on the user behavior time series segments through a pre-trained spatiotemporal attention network to generate an initial behavior feature matrix. The spatiotemporal attention network identifies key service interaction nodes by capturing the dependencies of user behaviors in time and space dimensions.

[0045] 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.

[0046] 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.

[0047] 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.

[0048] Step S124, performing scene-aware weighting on the cross-modal fusion features according to the scene switching marks of the target user in multiple service scenes, and generating a scene-based feature vector.

[0049] The scene switching timing sequence in the scene switching mark contains the scene type identifier and scene switching trigger event type triggered by the project manager in the continuous time unit. For example, when the project enters the execution stage from the planning stage, 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 tracking progress). The scene switching trigger event type is the transition of the project stage. The initial scene weight vector is generated according to the scene association strength corresponding to the scene switching trigger event type. In the project execution stage, the weight of the project management tool scene is relatively high because the demand for task management is the greatest in this stage. The time decay factor is calculated based on the time interval between adjacent scene switches in the scene switching timing sequence. As time goes by, the weight of the related scenes in the previous project planning stage will gradually decay. The time decay factor is element-wise multiplied with the initial scene weight vector, and then dynamically adjusted through the scene attention mechanism. For example, when the project is nearing the end, although the weight of the document editing scene has been reduced due to time decay, 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 contextual features of the current scene (such as the need to edit the project summary document). The dynamic scene weight distribution is multiplied by the cross-modal fusion feature tensor to generate a weighted scene-related feature matrix, in which each feature channel corresponds to the preference expression of a specific scene type. The scene-related feature matrix is ​​fused with the current scene context features collected in real time. At the project closing stage, the current scene context features may include the review, organization and storage requirements of the final project results, and these features are fused with the scene-related feature matrix. The enhanced scene feature tensor is verified for scene consistency to ensure that the orthogonality between the feature subspaces corresponding to different scene types is reasonable, and the conflicting weights in the dynamic scene weight distribution are corrected. For example, if there is confusion between the document editing scene and the project management tool scene in some features, adjustments are made through scene consistency verification. The cross-modal fusion features are weighted twice according to the corrected dynamic scene weight distribution to generate an optimized scene-aware feature projection, and finally the feature dimension compression algorithm is used to reduce the dimension to the preset dimensional space to generate a scenario-based feature vector. This scenario-based feature vector contains 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 project results at the current project closing stage.

[0050] Step S125, dynamically updating the scenario feature vector based on the gated loop unit to generate the dynamic feature vector, wherein the dynamic feature vector includes the user's long-term preference features and the short-term demand features of the current service scenario.

[0051] In this embodiment, the gated loop unit can be dynamically updated according to the previous state information and the current input information (i.e., the scenario-based feature vector). In the remote office scenario, as the project progresses, the needs of the project manager will continue to change. For example, at different stages of the project, his demand for video conferencing services may gradually change from simple communication to project results display, and this change in demand will be reflected in the scenario-based feature vector. The gated loop unit dynamically adjusts the scenario-based feature vector based on the previously accumulated behavior patterns and preference information of the project manager. It can organically integrate the long-term preference characteristics of the project manager (such as the consistent demand for efficient communication and task management) with the short-term demand characteristics of the current service scenario (such as the demand for results display and storage in the final stage of the project) according to 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 needs and preference patterns of the project manager in the remote office scenario.

[0052] In a possible implementation, step S122 includes: Step S1221, dividing the user behavior time series segment into multiple time windows, and the behavior data in each time window forms a local behavior subsequence.

[0053] In this embodiment, for the project manager, during the remote office 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. In each such time window, all the behavioral data related to the project manager and various remote services are included, and these behavioral data together constitute a local behavioral subsequence. For example, in a two-week time window, his behavioral data in the video conferencing service may include initiating three video conferences, the participants of each meeting, the duration of the meeting, whether the screen sharing function is used, etc.; the behavioral data in the project management tool include how many new tasks are created, which tasks have their deadlines adjusted, and the distribution of tasks; the behavioral data in the cloud storage service includes which project files are uploaded, the size of the files, the access permissions set, etc. These behavioral data related to different services in a specific time window fully record the work behavior pattern of the project manager during that time period.

[0054] Step S1222, performing a convolution kernel sliding operation on each local behavior subsequence to extract local spatiotemporal features, wherein the convolution kernel sliding operation uses a multi-scale convolution kernel group to capture behavior patterns of different time spans.

[0055] For the local behavior subsequences in each time window mentioned above, a multi-scale convolution kernel group is used to perform convolution kernel sliding operations. Taking the video conferencing service as an example, the smaller-scale convolution kernel can capture the detailed patterns of the project manager's video conferencing behavior in a short period of time (such as within a day) during the sliding process. For example, he held two consecutive video conferences on a certain day. The first meeting was mainly to discuss project technical problems with the technical team, and the second meeting was to communicate the project promotion plan with the marketing team. The smaller-scale convolution kernel can capture the rapid changes in the types of participants and the key content of the two meetings. The larger-scale convolution kernel can analyze the video conference behavior pattern from a longer time span (such as the entire two-week time window). For example, in these two weeks, the overall duration trend of video conferences is gradually increasing, which may mean that the project has entered the key discussion and decision-making stage. The larger-scale convolution kernel can identify this long-span behavior pattern change. Through the convolution kernel sliding operation of this multi-scale convolution kernel group, the local spatiotemporal features of each local behavior subsequence in different time spans can be fully extracted. These features cover rich information about the behavior patterns of project managers in different services.

[0056] Step S1223, inputting the local spatiotemporal features into a bidirectional long short-term memory network to generate a temporal context encoding, wherein the bidirectional long short-term memory network captures the implicit dependencies between time windows through forward and backward paths.

[0057] The local spatiotemporal features extracted from each local behavior subsequence are input into the bidirectional long short-term memory network. Taking the task allocation behavior in the project management tool as an example, there is a certain implicit dependency between the task allocations in different time windows (every two weeks). From the forward path of the bidirectional long short-term memory network, in the early time window of the project, the project manager may assign some basic tasks to junior team members according to the preliminary plan of the project. As the project progresses, in the subsequent time windows, the forward path can analyze how the subsequent task allocation is adjusted based on the progress of the previous tasks according to the previous task allocation, for example, some more complex tasks are assigned to experienced team members according to the completion of the previous basic tasks. From the backward path, when the task allocation in the later stage of the project is observed, it can be reversely inferred whether the previous task allocation is reasonable and whether it has laid a good foundation for the development of the later tasks. This bidirectional long short-term memory network can fully capture the implicit dependencies between different time windows in the task allocation behavior through forward and backward paths, and the same is true for other service-related behaviors such as video conferencing and cloud storage. By performing such processing on the local spatiotemporal features of all service-related behaviors, a temporal context encoding containing implicit dependencies between time windows is generated.

[0058] Step S1224, constructing a spatiotemporal attention weight matrix, which dynamically adjusts the contribution of each time window to the final feature representation by calculating the correlation strength of behavioral features in different time windows.

[0059] Step S1225, weighted aggregation is performed on the temporal context encoding according to the spatiotemporal attention weight matrix to generate the initial behavior feature matrix.

[0060] For example, for the temporal context encoding previously generated by the bidirectional long short-term memory network, the encoding part corresponding to each time window has its importance weight, which is determined by the spatiotemporal attention weight matrix. For example, in the time window of a certain project stage, since the behavior patterns of the video conferencing service and project management tool service in this stage have a high similarity with other key stages (determined by the spatiotemporal attention weight matrix), the weight of the temporal context encoding of this time window in the weighted aggregation process will be relatively high. Through this weighted aggregation method, the temporal context encodings of all time windows are combined according to their respective weights, and finally an initial behavior feature matrix is ​​generated. This initial behavior feature matrix integrates the behavior information in different time windows, and reasonably weights them according to the correlation strength between time windows, laying the foundation for further analysis of the behavioral characteristics of project managers.

[0061] In a possible implementation, step S1224 includes: Step S1224-1, calculating the behavior similarity between any two time windows, wherein the behavior similarity measures the matching degree of the service type distribution and the service interaction intensity in the two time windows by using a cosine similarity algorithm.

[0062] Step S1224-2, generating an initial attention score matrix based on the behavior similarity, and performing sparse processing on the initial attention score matrix to retain attention connections that exceed a preset attention score.

[0063] Step S1224-3, add the sparse attention score matrix to the position coding matrix to generate an enhanced attention matrix, where the position coding matrix is ​​used to characterize the relative position relationship of the time window in the overall time sequence.

[0064] Step S1224-4, normalize the enhanced attention matrix through a normalized exponential function to generate the spatiotemporal attention weight matrix.

[0065] For example, taking video conferencing services and project management tool services as examples, in different time windows, the service type distribution of video conferencing services may include different types of participant combinations (such as technical personnel-led, marketing personnel-led, etc.) and different meeting purposes (such as technical discussions, project progress reports, etc.), and 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.), and the service interaction intensity can be measured by indicators such as the number of task creation and task adjustment frequency. The cosine similarity algorithm is used to calculate the matching degree of these service type distributions and service interaction intensity between any two time windows to obtain the behavioral similarity. The initial attention score matrix is ​​generated based on the behavioral similarity, in which each element represents the initial attention score between the two time windows. Then, the initial attention score matrix is ​​sparsely processed to retain the attention connections that exceed the preset attention score. For example, if the behavioral similarity between the two time windows is low, lower than the preset attention score, then the corresponding attention connection is removed, and only those attention connections with high behavioral similarity indicating a strong correlation between the two time windows are retained. The sparse attention score matrix is ​​added 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 multiple time window sequences throughout the project cycle, earlier time windows and later time windows have different positional relationships, and the position encoding matrix can reflect this difference. After addition, an enhanced attention matrix is ​​generated, and finally the enhanced attention matrix is ​​normalized by a normalized exponential function to generate a spatiotemporal attention weight matrix. This spatiotemporal attention weight matrix can dynamically adjust the contribution of each time window in the final feature representation according to the strength of association between different time windows.

[0066] In a possible implementation, step S124 includes: Step S1241: extracting a scene switching timing sequence from the scene switching mark, wherein the scene switching timing sequence includes a scene type identifier and a scene switching triggering event type triggered by a user within a continuous time unit.

[0067] In this embodiment, in the remote office activities of the project manager, the scene type identifier may include a video conferencing scene, a project management tool usage scene, a cloud storage service scene, etc. From the perspective of the entire cycle of the project, within the continuous time unit, the scene switching trigger event type is closely related to the progress stage of the project. For example, in the early stage of the project, when the project team members are determined, the scene is switched from a separate document editing scene (such as writing a project plan) to a video conferencing scene. At this time, the scene switching trigger event type is that the project team is formed and team communication is required. As the project advances to the execution stage, the scene is switched from the video conferencing scene to the project management tool usage scene. The scene switching trigger event type is that the project starts to execute, and tasks need to be assigned and progress tracked. These scene switching records constitute a scene switching timing sequence in chronological order, which contains the specific scene type identifier of each scene switch and the event type that triggers the scene switch. This information fully reflects the interactive conversion of the project manager with different service scenes at different stages of the project.

[0068] Step S1242, generating an initial scene weight vector according to the scene association strength corresponding to the scene switching trigger event type, wherein the weight value of each dimension in the initial scene weight vector reflects the cumulative influence of the corresponding scene type in the historical interaction.

[0069] For project managers, different scene switching trigger event types correspond to different scene association strengths. For example, in the project execution phase, the event type of switching from the video conferencing scene to the project management tool usage scene has a high scene association strength for the project management tool usage scene at this stage due to the high reliance on task management during project execution. When the project is nearing the end, the scene association strength for the cloud storage service scene is relatively high under this event type, which switches from the project management tool usage scene to the cloud storage service scene (used to store the final results of the project). Based on these different scene association strengths, an initial scene weight vector is generated for each scene type. For example, during the entire project cycle, if the video conferencing scene frequently appears in the project communication stage and has an important impact on the project progress, then in the initial scene weight vector, the weight value corresponding to the video conferencing scene will be higher, which reflects the cumulative impact of the video conferencing scene in historical interactions. Similarly, for the project management tool usage scene and the cloud storage service scene, their weight values ​​are also determined according to their respective importance and impact in historical interactions.

[0070] Step S1243: Calculate a time decay factor based on the time interval between adjacent scene switches in the scene switching timing sequence, wherein the time decay factor is used to dynamically adjust the decay rate of the contribution of the historical scene type to the current scene perception weighting.

[0071] Step S1244, performing an element-by-element product operation on the time attenuation factor and the initial scene weight vector to generate an updated scene weight vector, wherein the scene type closer to the current time in the updated scene weight vector has a smaller attenuation amplitude of the corresponding weight value.

[0072] On the timeline of a project, the time intervals between adjacent scene switches are different. For example, at the beginning of a project, the time interval between switching from a document editing scene to a video conference scene is short, while during the execution of a project, the time interval between switching from a project management tool usage scene for task adjustment to a video conference scene for the next large-scale team communication may be longer. The time decay factor is calculated based on these time intervals. For scene switches with shorter time intervals, the corresponding historical scene type has a slower decay rate of contribution to the current scene perception weighting, while for scene switches with longer time intervals, the contribution of the historical scene type has a faster decay rate. For example, if the document editing scene switches to the video conference scene in a short time, the weight of the document editing scene in the subsequent scene perception weighting will decay relatively slowly, because the two scenes are closer in time and may have a closer logical relationship. On the contrary, for scene switches after a long time interval, such as switching between a scene in the early stage of a project and a scene in the later stage of a project, the weight of the early scene in the later scene perception weighting will decay rapidly.

[0073] For example, when a project is nearing its end, the weight values ​​of the project management tool usage scenarios that have been frequently used recently have a smaller attenuation due to the effect of the time decay factor. However, some scenarios in the early stage of the project have a larger attenuation after the weight values ​​are adjusted by the time decay factor because they are far away from the current time. In this way, the updated scenario weight vector can better reflect the relative importance of different scenario types at the current time.

[0074] 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 weight of historical scenes that are highly correlated with the current scene by analyzing the semantic correlation between the contextual features of the current service scene and the historical scene types.

[0075] In the project closing phase, the contextual features of the current service scenario may include requirements for reviewing, organizing, and finally storing project results. The scenario attention mechanism analyzes the semantic relevance between the current scenario and the historical scenario types. For example, the cloud storage service scenario has a high semantic relevance to the current project results storage requirements, while the video conferencing scenario, although very important in the project communication phase, has a relatively low semantic relevance to the current scenario in the project closing phase. Through the scenario attention mechanism, the weight of the cloud storage service scenario in the updated scenario weight vector is 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 phase (such as checking and improving the final task), 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 contextual features of the current scenario, so that the scenario weight is more in line with the needs of the current scenario.

[0076] Step S1246, performing a tensor multiplication operation on the dynamic scene weight distribution and the cross-modal fusion feature to generate a weighted scene-related feature matrix, wherein each feature channel in the scene-related feature matrix corresponds to a preference expression of a specific scene type.

[0077] Taking the cross-modal fusion features of project managers as an example, it includes fusion features related to the functional attributes of video conferencing services (such as high-definition image 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 behavioral characteristics of project managers in the same feature space. When the dynamic scene weight distribution is multiplied by the cross-modal fusion features, the weight of each scene type will weight the corresponding cross-modal fusion features. For example, if the cloud storage service scenario has a higher weight in the dynamic scene weight distribution, then in the weighted scene-related feature matrix, the feature channel related to the cloud storage service attributes will get more weight. This feature channel expresses the project manager's preference for cloud storage services in specific scenarios, such as the preference for secure and large-capacity cloud storage in the final stage of the project. Similarly, for video conferencing scenarios and project management tool usage scenarios, their corresponding feature channels will also be weighted according to their respective weights, thereby generating a scene-related feature matrix that expresses the preference of each feature channel for a specific scene type.

[0078] Step S1247, fusing the scene-related feature matrix with the current scene context features collected in real time to generate an enhanced scene feature tensor, wherein the enhanced scene feature tensor retains cross-scene common patterns and current scene-specific features through a gated fusion unit.

[0079] At the project closing stage, the current scene context features collected in real time include the specific type of the current project results (such as documents, reports, presentations, etc.), the importance level of the results, related security requirements, etc. The scene-related 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-scenario common patterns and which are current scene-specific features, and effectively retain them in the enhanced scene feature tensor. For example, in the project management process, some basic logic of task management (such as the order of tasks, dependencies, etc.) are cross-scenario common patterns, which exist at any stage of the project. At the project closing stage, the specific security requirements for project results (such as the advanced encryption requirements for confidential files) are current scene-specific features. The gated fusion unit can ensure that these different types of features are reasonably retained during the fusion process, thereby generating an enhanced scene feature tensor containing rich information.

[0080] Step S1248, performing scene consistency verification on the enhanced scene feature tensor, wherein the scene consistency verification corrects the conflict weights in the dynamic scene weight distribution by detecting the degree of orthogonality between feature subspaces corresponding to different scene types.

[0081] Among the different scenario types involved in the entire project cycle, each scenario type has a corresponding feature subspace in the enhanced scenario feature tensor. For example, the feature subspace corresponding to the video conferencing scenario contains behavioral features and service attribute features related to the meeting, and the feature subspace corresponding to the project management tool usage scenario contains features related to task management, etc. By detecting the degree of orthogonality between the feature subspaces corresponding to these different scenario types, it can be found whether there is a weight conflict. For example, if there is confusion in some features between the cloud storage service scenario and the project management tool usage scenario at the project closing stage (possibly due to some overlap in task association between the two at the project closing stage), this situation can be detected through scenario consistency verification. Then, the conflicting weights in the dynamic scenario weight distribution are corrected according to the degree of orthogonality. If it is found that the feature subspace of the cloud storage service scenario is non-orthogonal to the feature subspace of the project management tool usage scenario on a certain feature (i.e., there is confusion), the weight of the cloud storage service scenario on the weight related to the feature may be appropriately reduced to ensure the independence and rationality between the feature subspaces corresponding to different scenario types.

[0082] Step S1249, performing secondary weighting on the cross-modal fusion features according to the corrected dynamic scene weight distribution to generate an optimized scene-aware feature projection, reducing the dimension of the optimized scene-aware feature projection to a preset dimensional space through a feature dimension compression algorithm to generate the scenario-based feature vector. The scenario-based feature vector retains the linear combination relationship between the cross-scene shared feature components and the scene-specific feature components during the dimensionality reduction process.

[0083] After the previous step corrects the dynamic scene weight distribution, the corrected weight distribution is applied 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 project management tool usage scenario) on certain key features is adjusted, then in the secondary weighting process, the cross-modal fusion features related to the project management tool usage scenario will be weighted according to the new weight. The optimized scene-aware feature projection generated in this way can more accurately reflect the behavioral preferences and demand characteristics of the project manager in the current scenario, because it comprehensively considers multiple factors such as the scene switching history, the current scene context features, and the relationship between different scene types.

[0084] Finally, due to the involvement of multiple service scenarios and a large amount of feature information, the optimized scenario-aware feature projection may have a higher dimension. Through the feature dimension compression algorithm, it is reduced to the preset dimensional space. In this process, it is necessary to ensure that the linear combination relationship between the cross-scenario shared feature components and the scenario-specific feature components is preserved. For example, throughout the project cycle, the linear combination relationship between the basic logic of task allocation (cross-scenario shared feature components) and the demand for specific results storage in the project closing stage (scenario-specific feature components) will not be destroyed in the dimensionality reduction process. The scenario-based feature vector generated in this way can accurately represent the behavioral characteristics of the project manager in the current scenario with a lower dimension, which includes both long-term cross-scenario general preferences and specific demand characteristics in the current scenario, providing an effective feature representation for subsequent operations such as recommendations.

[0085] In a possible implementation, the step of training the multimodal recommendation model includes: Step S210, constructing a multi-task learning framework, wherein the multi-task learning framework includes a service click rate prediction task, a service completion prediction task, and a service satisfaction prediction task.

[0086] For example, for the target user of project managers in a remote office scenario, the service click rate prediction task involves predicting the possibility of the project manager clicking on a specific service resource (such as a specific function in the video conferencing software, a specific task operation entry in the project management tool, etc.). For example, in the video conferencing software, the probability of the project manager clicking on the screen sharing function is predicted based on the historical behavior and current project requirements of the project manager. The service completion prediction task focuses on whether the project manager can successfully complete the relevant tasks after using the service resources. For example, in the project management tool, it is predicted whether he can complete the allocation, tracking and final project progress management of all tasks as planned. The service satisfaction prediction task aims to evaluate the project manager's satisfaction with the service resources used (such as the security of cloud storage services, the stability of video conferencing services, etc.). By building such a multi-task learning framework, the multimodal recommendation model can be optimized from multiple perspectives to better meet the needs of project managers in remote office.

[0087] Step S220, extracting entity relationship features from the service resource knowledge graph, wherein the entity relationship features include a service provider association network, a service category hierarchy, and a service quality evaluation link.

[0088] In the remote office scenario, in terms of the service provider association network, taking video conferencing services as an example, different video conferencing service providers may have cooperative or competitive relationships. Some providers may specialize in providing high-security, high-concurrency video conferencing services for large enterprises, while others may focus on simplicity and ease of use for small teams. The associations between these providers constitute part of the service provider association network. In terms of the service category hierarchy, there are different levels of classification for project management tool services. 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 functional 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. Users' evaluation of cloud storage services will form a link. For example, user feedback data shows that a cloud storage service performs well in data encryption, but there are occasional delays in access speed. These evaluation information are interrelated to form a service quality evaluation link.

[0089] Step S230, performing heterogeneous feature alignment on the entity relationship features and the user behavior features to generate a joint embedding representation, wherein the heterogeneous feature alignment aggregates attribute information of adjacent nodes through a graph neural network.

[0090] The user behavior characteristics of the project manager, such as the frequency of his participation in video conferences and the frequency of task adjustment in the project management tool, are aligned with the entity relationship characteristics extracted from the service resource knowledge graph. The graph neural network achieves this process by aggregating the attribute information of adjacent nodes. Taking a video conference service provider node in the service provider association network as an example, its adjacent nodes may include other service providers that cooperate with it, user groups that use the provider's services, etc. The graph neural network aggregates the attribute information of these adjacent nodes, such as the technical advantages of the cooperating provider, the size of the user group, and the industry distribution. At the same time, the user behavior characteristics of the project manager 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 behavioral habits of the project manager and the various attributes and relationships of the service resources.

[0091] Step S240, performing multi-task learning based on the joint embedding representation under the multi-task learning framework. During the multi-task learning process, a dynamic gradient allocation strategy is used to adjust the loss weights of the service click-through rate prediction task, the service completion prediction task, and the service satisfaction prediction task. The dynamic gradient allocation strategy automatically adjusts the weight coefficient according to the convergence speed of the service click-through rate prediction task, the service completion prediction task, and the service satisfaction prediction task in the current training batch.

[0092] For example, in a training batch, the service click rate prediction task may converge faster for predicting the function clicks of project managers in video conferencing software, which means that the model has learned a more effective pattern on this task. However, the service completion prediction task, such as predicting whether the project manager can complete the task on time in the project management tool, may converge slower and require more training to optimize. The service satisfaction prediction task, such as predicting the project manager's satisfaction with the cloud storage service, also has its own convergence. According to these different convergence speeds, the dynamic gradient allocation strategy automatically adjusts the loss weights of each task. If the service click rate prediction task converges quickly, its loss weight may be appropriately reduced, while for the service completion prediction task and service satisfaction prediction task that converge more slowly, their loss weights may be increased to encourage the model to pay more attention to and optimize these tasks, thereby improving the overall multi-task learning effect.

[0093] Step S250, enhancing the generalization ability of the multimodal recommendation model through an adversarial training mechanism, wherein the adversarial training mechanism perturbs the user feature space by generating adversarial samples and forces the multimodal recommendation model to maintain prediction consistency after the perturbation.

[0094] For the user feature space of project managers, the adversarial training mechanism will generate adversarial samples. For example, for the user features of project managers in video conferencing services (such as frequency of attendance, requirements for image 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 attendance frequency or adjusting the degree of image quality requirements. These adversarial samples are then input into the multimodal recommendation model, which needs to maintain the consistency of predictions for tasks such as service click-through rate, service completion, and service satisfaction under such perturbed user feature input. If the model predicts that the project manager has a high click-through rate for a certain video conferencing function under the original user features, it should also make similar predictions when facing adversarial samples. In this way, through continuous adversarial training, the multimodal recommendation model can better adapt to various possible changes in user features, thereby enhancing its generalization ability.

[0095] The steps of constructing the service resource knowledge graph include: Step S310: Collect metadata information of service resources, where the metadata information includes service description text, service provider qualifications, and service history evaluation records.

[0096] Taking video conferencing services as an example, the service description text may include information such as the functions supported by the video conferencing service (such as high-definition image quality, multi-person screen sharing, etc.), applicable scenarios (such as large team meetings, one-on-one communication, etc.). The qualifications of the service provider may include the provider's technical strength, security certification status, etc. The service history evaluation record reflects other users' experience of using the video conferencing service, such as whether it is stable and easy to operate. For project management tool services and cloud storage services, metadata information in these aspects is also collected.

[0097] Step S320: extracting entity nodes and relationship edges from the metadata information, wherein the entity nodes include service type, service provider and service quality label, and the relationship edges include service type affiliation, service provider cooperation relationship and service quality association relationship.

[0098] Still taking the video conferencing service as an example, the service type entity node clarifies that it belongs to the video conferencing type of service. The service provider entity node is the company or organization that provides video conferencing services. The service quality label entity node may include labels such as high-definition, stable, and secure. The service type affiliation relationship edge indicates that the video conferencing service belongs to the large category of remote office services. The service provider partnership may be reflected in the partnership between the video conferencing service provider and other related service providers (such as cooperating with certain network equipment providers to optimize network transmission). The service quality association relationship indicates the association between the service quality label and the service provider and the service type. For example, a service provider uses its advanced technology to ensure that the video conferencing service has high-definition and stable quality.

[0099] Step S330, predict and fill in the missing relationship edges based on the knowledge graph completion algorithm, and the knowledge graph completion algorithm uses a relationship path reasoning model to mine implicit associations between entities.

[0100] Taking the project management tool service as an example, there may be implicit associations 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 type of enterprise using the tool, the relationship between the enterprise type and the industry standard, etc.), and then predict and fill in the missing relationship edges. This will make the service resource knowledge graph more complete and better reflect the complex relationships between service resources.

[0101] Step S340, dynamically update the knowledge graph. When it is detected that a new service resource is online or the attributes of an old service resource are changed, an incremental graph update process is triggered, and historical version snapshots are retained to support time series analysis.

[0102] For example, when a new video conferencing service is launched, metadata information such as new features (such as new interactive features) and qualifications of new service providers needs to be integrated into the knowledge graph. At the same time, if the properties of an old project management tool service change, such as adding new task management functions or adjusting the charging model, the knowledge graph also needs to be updated in a timely manner. During the update process, historical version snapshots are retained to support time-series analysis of service resources. For example, the development and changes of a service resource at different time points and the impact of these changes on users such as project managers can be analyzed, so as to better provide comprehensive and accurate service resource information for the multimodal recommendation model.

[0103] In a possible implementation, step S140 includes: Step S141, calculating the scenario matching score of each service resource in the candidate service resource prediction set, wherein the scenario matching score is determined by analyzing the co-occurrence frequency of the service resource attribute tag and the context feature of the current service scenario.

[0104] For example, for the target user, the project manager, in the current service scenario of project closing, the candidate service resource prediction set may include document management tools, presentation creation tools, cloud storage services, 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 contextual features of the current service scenario include the need to finalize and review the project documents, ensure the security of the documents, and facilitate the final collaborative modification of team members. If the version management function of the document management tool is often used in the previous project closing stage, and the multi-person collaborative editing function also matches the current team members' need to make the final improvement of the documents, then the co-occurrence frequency of these attribute tags and the contextual features of the current scenario is high, resulting in a higher scenario matching score. For the presentation creation tool, its attribute tags such as template richness, animation effect addition, presentation sharing function, etc., if these functions match the need to make high-quality and easy-to-share presentations in the current scenario when the project results are presented, they will also get the corresponding scenario matching score. The attribute labels of the cloud storage service, such as storage space size, data encryption method, access permission settings, etc., will also calculate the corresponding scenario matching score if they match the needs for secure storage of project results and convenient access by team members at the end of the current project.

[0105] Step S142: Generate a service resource availability index based on the real-time available status of the service resource and the load capacity of the service provider. The service resource availability index is dynamically calculated by monitoring the response delay and concurrent processing capability of the service instance.

[0106] Still taking these candidate service resources as an example, for document management tools, if its server has been recently maintained and upgraded, the response delay 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 is good and the service provider has sufficient load capacity, so its service resource availability index is high. For presentation creation tools, if it can still respond quickly to operation instructions when a large number of users are making presentations at the same time (such as multiple projects in the company are preparing for presentations at the same time), there is no obvious lag, and the service provider has sufficient resources to cope with this concurrency, its service resource availability index will also be at a good level. If a cloud storage service responds quickly during data upload and download, and does not overload even when project team members access the stored project results at the same time, it means that its service resource availability index is good.

[0107] Step S143, integrating the scenario matching score and the service resource availability index to generate a comprehensive recommendation priority, wherein the comprehensive recommendation priority uses a weighted harmonic average algorithm to balance the matching degree and availability.

[0108] For example, if a document management tool has a high scenario matching score and a good service resource availability index, its comprehensive recommendation priority will be relatively high after calculation using the weighted harmonic mean algorithm. If a presentation creation tool has a slightly lower scenario matching score but a very high service resource availability index, or vice versa, it will also get a reasonable comprehensive recommendation priority after the weighted harmonic mean algorithm. Cloud storage services are also calculated based on their scenario matching scores and service resource availability indexes to obtain the corresponding comprehensive recommendation priority.

[0109] Step S144 , sorting the candidate service resources according to the comprehensive recommendation priority, and dividing the sorting results into multiple recommendation levels, each recommendation level corresponding to a different display strategy and interaction method.

[0110] Service resources with the highest priority for comprehensive recommendation, such as document management tools, may be ranked first and classified as the first recommendation level. The display strategy for this level can be displayed in the form of a striking prompt box on the main interface of the project manager's office software, and the interactive method can be to directly provide a quick entry to the relevant function page of the document management tool, such as directly entering the final review page of the project document. Presentation creation tools with a slightly lower priority for comprehensive recommendation may be classified as the second recommendation level. The display strategy can be displayed in the form of a scroll bar in a secondary position on the office software interface, and the interactive method can be to display its functional features and instructions in detail when the project manager clicks on the relevant area. If the cloud storage service has a lower priority for comprehensive recommendation, it may be classified as the third recommendation level. The display strategy can be displayed in a specific functional area of ​​the software interface (such as a section related to data storage), and the interactive method can be to highlight its advantages when the project manager actively searches for storage-related services.

[0111] Step S145 , performing diversity control on the service resources of each recommended level, so that the distribution of the service resources in the same level in terms of type, provider and price range meets a preset diversity threshold.

[0112] In the first recommendation level, it is assumed that in addition to document management tools, there are also candidate resources such as similar document collaboration platforms. By performing semantic embedding vector analysis on their functional description texts (such as document management tools emphasize version management and permission settings, and document collaboration platforms emphasize real-time collaboration and online editing), the semantic embedding vector distance between them is calculated to construct a type similarity matrix. If the functional descriptions of two service resources are very close in semantics, then their vector distance is small and the similarity value in the type similarity matrix is ​​high.

[0113] In a possible implementation, step S145 includes: Step S1451 , calculating a type similarity matrix of service resources in the recommendation level, wherein the type similarity matrix is ​​determined by the semantic embedding vector distance of the function description text of the service resources.

[0114] Step S1452: construct a diversity constraint optimization objective function, wherein the objective function maximizes the type difference of service resources within the recommendation hierarchy and minimizes the deviation from user expectations.

[0115] For example, in the first recommendation level, in order to maximize the type differentiation, try to choose service resources with different functions. For example, document management tools focus on the overall management of documents, while if there is a tool that focuses on document format conversion and optimization, it can increase the type differentiation. At the same time, the user's expected deviation should be considered, and service resources that are too far away from the project manager's needs in the current scenario should not be selected. For example, although there are some advanced document encryption tools, if document encryption is no longer a key requirement at the end of the current project, choosing this tool will increase the user's expected deviation.

[0116] Step S1453, using a greedy algorithm to iteratively select service resources, each iteration selecting a service resource that can maximize the current diversity index and does not violate the comprehensive recommendation priority constraint, to generate a set of selected service resources.

[0117] For example, in the selection process of the first recommendation level, the document management tool may be selected at the beginning because it has a high overall recommendation priority and is the core tool for document management. Then, in subsequent iterations, the objective function is optimized according to the type similarity matrix and diversity constraints. If a tool with unique functions (such as the ability to quickly generate a project document summary report) that does not violate the overall recommendation priority constraints is found, it will be added to the selected service resource collection.

[0118] Step S1454, post-processing and filtering the selected service resource set, removing candidate resources whose similarity with the selected resources exceeds a threshold, and supplementing suboptimal alternative resources with high diversity contribution.

[0119] If the selected document management tool and a new candidate document management tool are very similar in function, exceeding the set similarity threshold, the new candidate resource will be removed. Then if there is an alternative resource that has a slightly lower overall recommendation priority but has special functions (such as voice annotation of documents) and can increase diversity, it will be added.

[0120] Step S1455, dynamically adjust the diversity weight coefficient, and automatically increase the diversity penalty item of the corresponding type when it is detected that the user continuously rejects the same type of service resources.

[0121] If a project manager rejects several document management-related tool recommendations in a row, the system will automatically increase the diversity penalty item for document management-type service resources. In this way, in subsequent recommendations, other types of service resources will be more likely to be selected. For example, in the first recommendation level, more presentation creation tools or other non-document management service resources related to project closing will be considered, thereby increasing the diversity of recommendations and better meeting the needs of project managers in remote office scenarios.

[0122] In a possible implementation, step S150 includes: Step S151, extracting explicit feedback signals and implicit feedback signals from the real-time feedback data, wherein the explicit feedback signals include user ratings and service favorite marks, and the implicit feedback signals include service stay duration and interactive operation frequency.

[0123] In this embodiment, for the target user, the project manager, the user rating in the explicit feedback signal is his direct evaluation of the recommended service resources. For example, a document management tool is recommended in the hierarchical recommendation list, and the project manager rates it based on his own experience. If he thinks that the document management tool's document organization and permission management functions at the end of the project are very in line with the requirements, he may give it 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 presentation creation tool to the favorite mark, it shows that he is quite satisfied with the tool, probably because the tool has rich templates and is easy to create animation effects, which makes it suitable for project achievement presentations.

[0124] In terms of implicit feedback signals, the length of time a project manager stays in a service reflects the time a project manager spends on a service resource. For example, when using a cloud storage service, if he stays in the service for a long time, it may mean that he is performing detailed storage operations on the project results, such as categorizing and storing multiple files, setting different access permissions, etc., which indicates that he has a high demand for the functions of the cloud storage service. The frequency of interactive operations is also an important implicit feedback signal. Taking the project management tool as an example, if the project manager has a high frequency of interactive operations in the tool, frequently adjusting the task status, checking the task progress, etc., this shows that he is highly dependent on the project management tool and that the tool plays an important role in the project management process.

[0125] Step S152: construct a feedback enhancement feature vector, wherein the feedback enhancement feature vector is generated by nonlinearly fusing the explicit feedback signal with the implicit feedback signal.

[0126] For example, for the feedback data of project managers, the nonlinear fusion process will comprehensively consider factors such as user ratings, service favorite marks, service stay time, and interaction 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 stay time is short and the interaction frequency is low, it may be adjusted through a nonlinear function so that the feedback enhancement feature vector can fully reflect the project manager's comprehensive attitude towards the recommended service resources. This fusion method is not a simple linear combination, but a specific nonlinear function is used to capture the complex relationship between explicit feedback signals and implicit feedback signals, thereby generating a feedback enhancement feature vector that can accurately represent the overall situation of user feedback.

[0127] Step S153: input the feedback enhancement feature vector into a parameter adjustment network and output a model weight update amount. The parameter adjustment network uses a meta-learning framework to predict the influence of different feedback modes on model parameters.

[0128] For example, the meta-learning framework can analyze the feedback patterns of project managers. For example, if the feedback pattern of the project manager is that the document management tool has a low score and a short service stay, while the interaction frequency of other types of tools (such as presentation making tools) is high, the parameter adjustment network can predict based on this feedback pattern that the parameters related to the document management tool recommendation in the multimodal recommendation model need to be adjusted significantly, while the parameters related to the presentation making tool recommendation need to be appropriately optimized. In this way, the model weight update amount is accurately output according to different feedback patterns to adjust the parameter weights of the multimodal recommendation model so that it can better adapt to the changes in the needs of project managers.

[0129] Step S154, performing knowledge distillation on the embedding layer of the multimodal recommendation model, and retaining feature mapping relationships corresponding to confidence prediction results whose confidence is greater than a set confidence.

[0130] In multimodal recommendation models, the embedding layer is crucial for processing the feature representation of various service resources. For example, when embedding features of document management tools, presentation creation tools, and cloud storage services, the knowledge distillation process will screen these feature representations. If the confidence prediction result for a feature of a document management tool (such as the feature representation of the version management function) is high, that is, the model has a high degree of grasp of the feature, then the corresponding feature mapping relationship will be retained. However, some feature representations with lower confidence may be discarded in the knowledge distillation process. Doing so can reduce the complexity of the model while retaining feature information that has an important impact on the recommendation results, thereby improving the efficiency and accuracy of the model.

[0131] Step S155, establish a model version rollback mechanism, and automatically restore the model parameters to the historical stable version when it is detected that multiple consecutive batches of feedback data have caused the recommendation performance to degrade.

[0132] As the project progresses, the needs of the project manager and the feedback on service resources will continue to change. If multiple batches of feedback data show that the recommendation performance of the multimodal recommendation model for project managers is gradually declining, for example, the recommended service resources are increasingly mismatched with the actual needs of the project manager, user ratings continue to decline, and the service stay time is reduced. At this time, the model version rollback mechanism will be automatically triggered to restore the model parameters to the parameters of a previous historical stable version. This historical stable version may be at a certain stage in the early stage of the project, when the recommendation results of the model can better meet the needs of the project manager. By restoring the parameters to this version, the recommendation performance can be avoided from further decline, ensuring the stability and reliability of the model to a certain extent.

[0133] The training step of the parameter adjustment network includes: Step S410: construct a meta-training task set, where each meta-training task in the meta-training task set simulates a model parameter adjustment requirement under a user feedback distribution mode.

[0134] For example, in a remote office scenario, there may be multiple user feedback distribution patterns. For example, in one pattern, the project manager gives low scores to all recommended service resources, and the service stay time and interaction frequency are very low. This may mean that the recommended service resources do not meet the needs at all. The corresponding meta-training task is to simulate how to adjust the parameters of the multimodal recommendation model in this case to improve the accuracy of the recommendation. Another pattern may be that the project manager only has high scores and frequent interactions for service resources of a certain type (such as presentation making tools), but pays less attention to other types of service resources. The meta-training task in this pattern is to adjust the model parameters based on this feedback distribution, so that the recommendation model can recommend more service resources such as presentation making tools that meet user preferences.

[0135] Step S420, in the meta-training stage, a model-agnostic meta-learning algorithm is used to enable the parameter adjustment network to adapt to feedback modes that have not been learned.

[0136] For example, the feedback patterns of project managers may be very complex and diverse, and some feedback patterns may not have appeared in previous training data. Model-agnostic meta-learning algorithms enable the parameter adjustment network to make reasonable parameter adjustments based on the simulations in the meta-training task set when facing these new feedback patterns. For example, when a new feedback pattern emerges, that is, the project manager has a unique feedback behavior 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 for this new feedback pattern based on the adjustment experience of various feedback patterns previously simulated, thereby making the parameter adjustment network have stronger generalization capabilities.

[0137] Step S430, enabling a fast update path and a slow update path, wherein the fast update path generates temporary model parameters according to the feedback data of the current batch, and the slow update path optimizes basic parameters based on long-term feedback trends.

[0138] For example, for feedback data from project managers, the fast update path will respond promptly based on the feedback from the current batch. For example, if the project manager gives a very low score to a newly recommended document review tool in the current batch, the fast update path will quickly generate temporary model parameters and adjust the parts related to the document review tool recommendation to improve the next recommendation result as soon as possible. The slow update path, on the other hand, pays more attention to long-term feedback trends, such as the overall preference trend of project managers for different types of service resources (such as document management tools, presentation creation tools, etc.) throughout the project cycle. If the long-term trend shows that project managers are paying more and more attention to a certain function (such as the secure storage function of documents), the slow update path will optimize 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.

[0139] Step S440: introducing a gradient similarity constraint term so that the parameter adjustment direction maintains semantic consistency with the original optimization goal of the recommendation model.

[0140] For a multimodal recommendation model, its original optimization goal may be to improve the accuracy and satisfaction of recommendations. When adjusting parameters based on the feedback data from project managers, the gradient similarity constraint ensures that the direction of parameter adjustment is consistent with this original optimization goal. For example, if the original optimization goal is to improve the project manager's satisfaction with the recommended service resources, then when adjusting parameters, there should be no situation that contradicts this goal. If a parameter adjustment causes the recommendation result to deviate further from the needs of the project manager and violates the goal of improving satisfaction, the gradient similarity constraint will restrict it to ensure that the parameter adjustment direction is in the direction of improving satisfaction, thereby maintaining the semantic consistency between the parameter adjustment and the original optimization goal of the recommendation model.

[0141] Step S450, evaluating the effectiveness of parameter adjustment through an adversarial verification mechanism, and triggering a manual intervention alarm when the performance gain of the adjusted recommendation model on the verification set is lower than a threshold.

[0142] For example, the validation set may contain some previous feedback data from project managers or simulated similar feedback data. After the parameters of the multimodal recommendation model are adjusted according to the feedback from the project manager, the effectiveness of the adjustment is evaluated through the adversarial verification mechanism. If the performance gain of the adjusted model on the validation set (such as the improvement in user ratings, the increase in service stay time, and other indicators) is lower than the set threshold, this indicates that the parameter adjustment may not have achieved the expected effect. At this point, the manual intervention alarm is triggered, and manual intervention can further analyze the cause. It may be that there are special circumstances in the feedback data or the structure of the model itself needs to be adjusted, so as to take corresponding measures to improve the recommendation performance of the model and ensure that the model can make effective parameter adjustments based on the feedback from the project manager to better adapt to the changing needs in the remote office scenario.

[0143] Figure 2 The hardware structure of the remote digital service system 100 provided in the embodiment of the present application for implementing the above-mentioned remote digital service resource recommendation method based on artificial intelligence mining is shown as follows: Figure 2 As 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 .

[0144] In one possible design, the remote digital service system 100 can be a single server or a server group. The server group can be centralized or distributed (for example, the remote digital service system 100 can be a distributed system). In some embodiments, the remote digital service system 100 can be local or remote. For example, the remote digital service system 100 can access information and / or data stored in a machine-readable storage medium 120 via a network. For another example, the remote digital service system 100 can 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 can be implemented on a server. By way of example only, the server can 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.

[0145] 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 data and / or instructions used by the remote digital service system 100 to execute or use to complete the exemplary methods described in this application.

[0146] During the specific implementation process, one or more processors 110 execute 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 embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0147] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned remote digital service system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0148] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are preset. When the processor executes the computer executable instructions, the remote digital service resource recommendation method based on artificial intelligence mining as described above is implemented.

[0149] It should be noted that in order to simplify the description disclosed in the present application and thereby facilitate the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, various features are sometimes grouped into one embodiment, drawing or description thereof.

Claims

1. A remote digital service resource recommendation method based on artificial intelligence mining, characterized in that: The method comprises: Acquire historical interaction data of a target user, wherein the historical interaction data includes interaction records between the target user and service resources in multiple remote service scenarios, service resource attribute information, and user behavior sequences; Generate 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; The dynamic feature vector is input into a pre-trained multimodal recommendation model, and a predicted set of candidate service resources for the target user in the current service scenario is output. The multimodal recommendation model mines potential matching patterns between user behavior and service resources by fusing a temporal behavior analysis model and a cross-scenario semantic association model; 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; 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.

2. The method for recommending remote digital service resources based on artificial intelligence mining according to claim 1, characterized in that: The generating the dynamic feature vector of the target user based on the historical interaction data includes: Extracting user behavior time series segments from the historical interaction data, wherein the user behavior time series segments include continuous service request intervals, service type switching frequency, and service resource usage depth; Perform multi-granularity feature extraction on the user behavior time series segments through a pre-trained spatiotemporal attention network to generate an initial behavior feature matrix. The spatiotemporal attention network identifies key service interaction nodes by capturing the dependencies of user behaviors in time and space dimensions; Fusion of 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 attribute into the user behavior feature space through a cross-modal alignment algorithm; According to the scene switching marks of the target user in multiple service scenes, scene-aware weighting is performed on the cross-modal fusion features to generate a scene-based feature vector; The scenario-based feature vector is dynamically updated 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.

3. The method for recommending remote digital service resources based on artificial intelligence mining according to claim 2 is characterized in that: The pre-trained spatiotemporal attention network is used to extract multi-granular features from the user behavior time series segments to generate an initial behavior feature matrix, including: Dividing the user behavior time series segment into multiple time windows, the behavior data in each time window forms a local behavior subsequence; Performing a convolution kernel sliding operation on each local behavior subsequence to extract local spatiotemporal features, wherein the convolution kernel sliding operation uses a multi-scale convolution kernel group to capture behavior patterns of different time spans; Inputting the local spatiotemporal features into a bidirectional long short-term memory network to generate a temporal context encoding, wherein the bidirectional long short-term memory network captures implicit dependencies between time windows through forward and backward paths; Constructing a spatiotemporal attention weight matrix, which dynamically adjusts the contribution of each time window to the final feature representation by calculating the correlation strength of the behavior features in different time windows; The temporal context encoding is weightedly aggregated according to the spatiotemporal attention weight matrix to generate the initial behavior feature matrix.

4. The method for recommending remote digital service resources based on artificial intelligence mining according to claim 3 is characterized in that: The constructing of the spatiotemporal attention weight matrix includes: Calculate the behavior similarity between any two time windows, where the behavior similarity measures the matching degree of service type distribution and service interaction intensity in the two time windows through a cosine similarity algorithm; Generating an initial attention score matrix according to the behavior similarity, and performing sparse processing on the initial attention score matrix to retain attention connections exceeding a preset attention score; The sparse attention score matrix is ​​added to the position encoding matrix to generate an enhanced attention matrix, where the position encoding matrix is ​​used to characterize the relative position relationship of the time window in the overall time series; The enhanced attention matrix is ​​normalized by a normalized exponential function to generate the spatiotemporal attention weight matrix.

5. The method for recommending remote digital service resources based on artificial intelligence mining according to claim 2, characterized in that: The step of performing scene-aware weighting on the cross-modal fusion features according to scene switching marks of the target user in multiple service scenes to generate a scene-based feature vector comprises: Extracting a scene switching timing sequence from the scene switching mark, wherein the scene switching timing sequence includes a scene type identifier and a scene switching triggering event type triggered by a user within a continuous time unit; Generate an initial scene weight vector according to the scene association strength corresponding to the scene switching trigger event type, wherein the weight value of each dimension in the initial scene weight vector reflects the cumulative influence of the corresponding scene type in the historical interaction; Calculating a time decay factor based on a time interval between adjacent scene switches in the scene switching time sequence, wherein the time decay factor is used to dynamically adjust a decay rate of a contribution of the historical scene type to the current scene perception weighting; Performing an element-by-element product operation on the time attenuation factor and the initial scene weight vector to generate an updated scene weight vector, wherein the attenuation amplitude of the corresponding weight value of the scene type closer to the current time in the updated scene weight vector is smaller; Dynamically adjusting the updated scene weight vector through a scene attention mechanism to generate a dynamic scene weight distribution, wherein the scene attention mechanism enhances the weight of historical scenes that are highly associated with the current scene by analyzing the semantic correlation between the contextual features of the current service scene and the historical scene types; Performing a tensor multiplication operation on the dynamic scene weight distribution and the cross-modal fusion feature to generate a weighted scene association feature matrix, wherein each feature channel in the scene association feature matrix corresponds to a preference expression of a specific scene type; Fusion of the scene-related feature matrix and the current scene context features collected in real time to generate an enhanced scene feature tensor, wherein the enhanced scene feature tensor retains cross-scene common patterns and current scene-specific features through a gated fusion unit; Performing scene consistency verification on the enhanced scene feature tensor, wherein the scene consistency verification corrects conflicting weights in the dynamic scene weight distribution by detecting the degree of orthogonality between feature subspaces corresponding to different scene types; Performing secondary weighting on the cross-modal fusion features according to the corrected dynamic scene weight distribution to generate an optimized scene-aware feature projection; The optimized scene perception feature projection is reduced in dimension to a preset dimensional space through a feature dimension compression algorithm to generate the scene-based feature vector. The scene-based feature vector retains the linear combination relationship between the cross-scene shared feature components and the scene-specific feature components during the dimensionality reduction process.

6. The method for recommending remote digital service resources based on artificial intelligence mining according to claim 1, characterized in that: The training steps of the multimodal recommendation model include: Construct a multi-task learning framework, wherein the multi-task learning framework includes a service click rate prediction task, a service completion prediction task, and a service satisfaction prediction task; Extracting entity relationship features from the service resource knowledge graph, wherein the entity relationship features include a service provider association network, a service category hierarchy, and a service quality evaluation link; Performing heterogeneous feature alignment on the entity relationship features and the user behavior features to generate a joint embedding representation, wherein the heterogeneous feature alignment aggregates attribute information of adjacent nodes through a graph neural network; Under the multi-task learning framework, multi-task learning is performed based on the joint embedding representation. During the multi-task learning process, a dynamic gradient allocation strategy is used to adjust the loss weights of the service click rate prediction task, the service completion prediction task, and the service satisfaction prediction task. The dynamic gradient allocation strategy automatically adjusts the weight coefficient according to the convergence speed of the service click rate prediction task, the service completion prediction task, and the service satisfaction prediction task in the current training batch; The generalization ability of the multimodal recommendation model is enhanced through an adversarial training mechanism, wherein the adversarial training mechanism perturbs the user feature space by generating adversarial samples and forces the multimodal recommendation model to maintain prediction consistency after the perturbation; The steps of constructing the service resource knowledge graph include: Collect metadata information of service resources, including service description text, service provider qualifications and service history evaluation records; Extracting entity nodes and relationship edges from the metadata information, wherein the entity nodes include service types, service providers, and service quality labels, and the relationship edges include service type affiliation, service provider cooperation, and service quality association; Predict and fill in missing relationship edges based on a knowledge graph completion algorithm, wherein the knowledge graph completion algorithm uses a relationship path reasoning model to mine implicit associations between entities; The knowledge graph is dynamically updated. When it is detected that a new service resource is online or the attributes of an old service resource are changed, an incremental graph update process is triggered, and historical version snapshots are retained to support time series analysis.

7. The method for recommending remote digital service resources based on artificial intelligence mining according to claim 1, characterized in that: The step of generating a hierarchical recommendation list according to the real-time scene adaptability of each service resource in the candidate service resource prediction set includes: Calculating a scenario matching score for each service resource in the candidate service resource prediction set, wherein the scenario matching score is determined by analyzing the co-occurrence frequency of the service resource attribute tag and the context feature of the current service scenario; Generate a service resource availability indicator based on the real-time available status of the service resource and the load capacity of the service provider, wherein the service resource availability indicator is dynamically calculated by monitoring the response delay and concurrent processing capacity of the service instance; The scenario matching score and the service resource availability index are integrated to generate a comprehensive recommendation priority, wherein the comprehensive recommendation priority uses a weighted harmonic average algorithm to balance matching degree and availability; Sorting the candidate service resources according to the comprehensive recommendation priority, and dividing the sorting results into multiple recommendation levels, each recommendation level corresponding to a different display strategy and interaction method; Diversity control is performed on the service resources of each recommended level so that the distribution of service resources in the same level in terms of type, provider and price range meets the preset diversity threshold.

8. The method for recommending remote digital service resources based on artificial intelligence mining according to claim 7 is characterized in that: The diversity control of service resources at each recommendation level includes: Calculating a type similarity matrix of service resources in the recommendation level, wherein the type similarity matrix is ​​determined by the semantic embedding vector distance of the function description text of the service resources; Constructing a diversity constraint optimization objective function, wherein the objective function maximizes the type difference of service resources within the recommendation hierarchy and minimizes the deviation from user expectations; A greedy algorithm is used to iteratively select service resources. Each iteration selects the service resources that can maximize the current diversity index and do not violate the comprehensive recommendation priority constraints to generate a set of selected service resources. Post-processing and filtering the selected service resource set, removing candidate resources whose similarity with the selected resources exceeds a threshold, and supplementing suboptimal alternative resources with high diversity contribution; The diversity weight coefficient is adjusted dynamically. When it is detected that users continuously reject the same type of service resources, the diversity penalty item of the corresponding type is automatically increased.

9. The method for recommending remote digital service resources based on artificial intelligence mining according to claim 1, characterized in that: The dynamically adjusting the parameter weights of the multimodal recommendation model based on the real-time feedback data of the target user on the hierarchical recommendation list includes: Extracting explicit feedback signals and implicit feedback signals from the real-time feedback data, wherein the explicit feedback signals include user ratings and service favorite marks, and the implicit feedback signals include service stay duration and interactive operation frequency; Constructing a feedback enhancement feature vector, wherein the feedback enhancement feature vector is generated by nonlinearly fusing the explicit feedback signal with an implicit feedback signal; Inputting the feedback enhancement feature vector into a parameter adjustment network and outputting a model weight update amount, wherein the parameter adjustment network uses a meta-learning framework to predict the influence of different feedback modes on model parameters; Performing knowledge distillation on the embedding layer of the multimodal recommendation model, retaining feature mapping relationships corresponding to confidence prediction results having confidence greater than a set confidence; Establish a model version rollback mechanism to automatically restore model parameters to the historical stable version when it is detected that multiple batches of feedback data have caused the recommendation performance to degrade; The training step of the parameter adjustment network includes: Constructing a meta-training task set, wherein each meta-training task in the meta-training task set simulates a model parameter adjustment requirement under a user feedback distribution mode; In the meta-training phase, a model-agnostic meta-learning algorithm is used to enable the parameter adjustment network to adapt to unlearned feedback patterns; Enable a fast update path and a slow update path, wherein 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; A gradient similarity constraint is introduced to ensure that the parameter adjustment direction maintains semantic consistency with the original optimization goal of the recommendation model. The effectiveness of parameter adjustment is evaluated through an adversarial verification mechanism. When the performance gain of the adjusted recommendation model on the verification set is lower than the threshold, a manual intervention alarm is triggered.

10. 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 as described in any one of claims 1 to 9.

Images