Student affairs allocation and service evaluation method and system based on multimodal hosting

Abstract

The present invention provides a student affairs allocation and service evaluation method and system based on multimodal hosting, which relates to the technical field of intelligent affairs scheduling and service evaluation. The method comprises extracting semantic features from student affairs requests and fusing transaction type information to generate transaction feature vectors, constructing transaction time series association graphs, identifying processing dependencies, forming transaction families and their processing priority sequences; combining the skill vectors of processing personnel, and using adaptive dynamic programming to achieve the optimal matching path; finally, dynamic adjustment and feedback update are performed based on the matching results and service scores, which can effectively improve transaction processing efficiency and service matching accuracy.

Description

Technical Field

[0001] The present invention relates to intelligent affairs scheduling and service evaluation technology, and in particular to a student affairs allocation and service evaluation method and system based on multimodal hosting. Background Art

[0002] Current student hosting service management models generally rely on manual intervention or fixed-rule allocation strategies. These systems lack a deep understanding of transaction content and the ability to dynamically adjust, making them incapable of adapting to students' personalized and diverse hosting requests. Faced with the complex and ever-changing nature of transaction types and processing requirements, traditional systems struggle to efficiently extract transaction semantic features and achieve structured modeling of complete transaction information, impacting overall service response efficiency and processing accuracy.

[0003] Existing technologies generally employ static matching or simple priority mechanisms in task scheduling and allocation, failing to fully consider potential dependencies and resource competition between tasks. This leads to conflicts, duplication, and delays in task execution. Furthermore, the allocation of processing personnel capabilities fails to accurately align with task characteristics, lacking a matching mechanism based on multi-stage optimization. This makes it difficult to effectively balance service quality and resource utilization, hindering the refined upgrade of digital governance for managed services.

[0004] With the development of multimodal information fusion, graph structure modeling and intelligent scheduling algorithms, there is an urgent need to build a comprehensive transaction management method that can integrate transaction semantic characteristics, dependencies, processing capabilities and historical performance, support automatic clustering of student affairs, intelligent scheduling and service evaluation closed-loop feedback, so as to improve the intelligence level and service quality of managed transaction processing. Therefore, it is urgent to propose a student affairs allocation and service evaluation method based on multimodal hosting. Summary of the Invention

[0005] The embodiments of the present invention provide a student affairs allocation and service evaluation method and system based on multimodal hosting, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provided is a student affairs assignment and service evaluation method based on multimodal hosting, including:

[0008] Receive student transaction request data, extract transaction descriptions from the student transaction request data, perform semantic analysis to extract semantic features, and fuse the semantic features with transaction type identifiers to generate a transaction feature vector containing complete transaction information;

[0009] A transaction timing association graph is constructed based on transaction feature vectors. Recursive deep feature extraction is used to calculate the feature propagation between nodes in the transaction timing association graph, generating a processing dependency probability matrix. A transaction clustering evaluation function is constructed based on the processing dependency probability matrix. Transaction clustering evaluation functions are iteratively optimized to generate clusters of transactions. A processing priority sequence is generated for the transaction families based on the constraint strength in the processing dependency probability matrix.

[0010] A skill vector is constructed based on the professional skills of the processing personnel. The skill vector and the transaction feature vector in the transaction family are input into the adaptive dynamic programming algorithm. The optimal matching path is obtained through multi-stage decision optimization. A comprehensive matching score is generated based on the optimal matching path score and historical service scores.

[0011] The transaction family and processing priority sequence are assigned to the target processing personnel with the highest comprehensive matching score. The professional skill vector and historical service score of the target processing personnel are updated according to the processing evaluation score of the transaction family, which are used for matching calculation of subsequent transaction allocation.

[0012] In an optional embodiment,

[0013] Transaction descriptions are extracted from student transaction request data and semantic features are extracted through semantic analysis. The semantic features are then fused with transaction type identifiers to generate a transaction feature vector containing complete transaction information.

[0014] Extracting transaction descriptions from student transaction request data, performing multi-level decomposition of the transaction descriptions according to core information, attribute information, and environmental information, extracting transaction goals and requirement features from the core information, extracting transaction timeliness and urgency features from the attribute information, and extracting processing scenarios and resource requirement features from the environmental information to obtain multi-level transaction features;

[0015] performing a correlation analysis on the multi-level transaction features, calculating a feature recombination coefficient based on information completeness and information redundancy between the features, and combining transaction features whose complementarity is greater than a preset complementarity threshold according to the feature recombination coefficient to generate a transaction semantic feature;

[0016] A mapping rule base for transaction type identifiers is established, in which the correspondence and confidence between transaction type identifiers and transaction semantic features are stored. Feature matching is performed in the mapping rule base based on the transaction semantic features to obtain a probability distribution of transaction type identifiers. The transaction semantic features are combined with the transaction type identifier with the highest confidence to generate a transaction feature vector containing complete transaction information.

[0017] In an optional embodiment,

[0018] Based on the transaction feature vector, a transaction time series association graph is constructed. The feature propagation between nodes in the transaction time series association graph is calculated through recursive deep feature extraction. The generated processing dependency probability matrix includes:

[0019] Construct a transaction timing association graph based on the transaction feature vector, where nodes represent transactions and edges between nodes represent the timing association between transactions;

[0020] Establishing a conflict transaction tag library based on historical processing data, wherein the conflict transaction tag library records transaction combination types that cause resource contention, matching nodes in the transaction time sequence association graph with the conflict transaction tag library, identifying node pairs with resource contention, and generating a contention node set;

[0021] For each node pair in the competitive node set, the resource demand information of each node is extracted, the resource overlap between the node pairs is calculated, and the resource competition intensity is determined based on the resource overlap as the competition edge weight between the competing node pairs;

[0022] For the remaining node pairs in the transaction timing association graph, the time interval information and transaction feature vectors between the nodes are extracted, and the timing similarity between the node pairs is calculated as the association edge weight between the remaining node pairs;

[0023] The competitive edge weights and associated edge weights are combined to form complete edge weight information, which is used to guide the recursive propagation of node features. The node features are updated to obtain propagated node features. The processing dependency probability between nodes is calculated based on the propagated node features to generate a processing dependency probability matrix.

[0024] In an optional embodiment,

[0025] A transaction clustering evaluation function is constructed based on the processing dependency probability matrix. Transaction families are generated by iteratively optimizing the transaction clustering evaluation function. A processing priority sequence is generated for the transaction family based on the constraint strength in the processing dependency probability matrix. The sequence includes:

[0026] Constructing a transaction clustering evaluation function based on the processing dependency probability matrix, wherein the transaction clustering evaluation function includes intra-cluster processing dependency constraint items and inter-cluster processing dependency constraint items;

[0027] Extracting historical execution records from the processing dependency probability matrix, calculating the transaction execution time fluctuation coefficient based on the historical execution records, constructing an execution stability matrix and inputting it into the transaction clustering evaluation function, performing initial clustering on the transactions to obtain an initial transaction family set;

[0028] Calculating the execution time fluctuation coefficient differences of the transaction pairs in the initial transaction family set, determining the processing dependency constraint weights and fluctuation compensation coefficients of the transaction pairs in the transaction clustering evaluation function based on the execution time fluctuation coefficient differences, iteratively optimizing the transaction family set, and obtaining the final transaction family set when the evaluation values ​​converge;

[0029] Transactions with execution duration fluctuation coefficients less than a preset fluctuation threshold are selected from the final transaction family set to construct a benchmark execution sequence. The degree of interference of the remaining transactions on the benchmark execution sequence is calculated according to the constraint strength in the processing dependency probability matrix, and the optimal insertion position of the remaining transactions is determined. Based on the benchmark execution sequence and the optimal insertion position, a transaction family processing priority sequence that meets the processing dependency constraints is generated.

[0030] In an optional embodiment,

[0031] Generating a transaction family processing priority sequence that satisfies processing dependency constraints based on the benchmark execution sequence and the optimal insertion position includes:

[0032] An initial constraint edge set is constructed based on the benchmark execution sequence, and the transitive constraint strength between nodes in the initial constraint edge set is calculated using the shortest path algorithm. Transitive constraints with transitive constraint strength greater than a preset strength threshold are added to the initial constraint edge set to obtain a complete constraint edge set.

[0033] Dividing the complete constraint edge set into layers based on the in-degree and out-degree of the nodes to obtain a constraint graph including inter-layer constraint relations and intra-layer constraint relations;

[0034] Obtaining a deviation between the actual execution time and the expected execution time of the transaction, the number of constraint violations, and the degree of contention; calculating a performance score for the transaction based on the deviation, the number of violations, and the degree of contention; performing a weighted combination of the performance score and the base priority to obtain an updated transaction priority; and determining a set of affected nodes in the constraint graph that are affected by the updated priority;

[0035] sequentially adjusting the priority of each node in the influencing node set, calculating the dependency chain length between the nodes based on the constraint graph, extracting the optimized nodes whose dependency chain length is greater than a preset length threshold, reordering the optimized nodes to obtain multiple reordering schemes, and selecting the reordering scheme with the smallest processing delay as the target reordering scheme;

[0036] The target reordering scheme is compared with the historical priority sequence, and when the improvement exceeds a preset improvement threshold, the target reordering scheme is adopted to update the transaction family processing priority sequence.

[0037] In an optional embodiment,

[0038] A skill vector is constructed based on the professional skills of the processing personnel. The skill vector and the transaction feature vector in the transaction family are input into the adaptive dynamic programming algorithm. The optimal matching path is obtained through multi-stage decision optimization. The comprehensive matching score is generated based on the optimal matching path score and historical service scores. The score includes:

[0039] Obtain the professional qualification characteristics, work experience characteristics, and skill coverage characteristics of the processing personnel, and generate the professional skill vector of the processing personnel by weighted combination based on preset weights;

[0040] Calculating the degree of matching between the professional skill vector and the transaction feature vector in the transaction family to obtain a basic matching value, calculating the degree to which the skill item satisfies the transaction feature based on the basic matching value, and generating an initial state value of the state space;

[0041] Constructing a skill transfer matrix to represent the complementary relationship between skills, calculating a state transition probability based on the skill transfer matrix, and combining the state transition probability with an initial state value to construct a state space;

[0042] Obtain the service completion quality and processing timeliness of the processing personnel, calculate the historical service score, adjust the state value of the state space based on the historical service score to construct a state value function, perform policy iteration calculation on the state value function to obtain a state value sequence, determine the optimal state transition path based on the state value sequence, and generate the optimal matching path;

[0043] The state transfer benefit of the optimal matching path is calculated to obtain a path score, and the path score is combined with the historical service score according to the weight determined by the transaction characteristics to generate a comprehensive matching score.

[0044] In an optional embodiment,

[0045] Performing a policy iteration calculation on the state value function to obtain a state value sequence, determining an optimal state transition path based on the state value sequence, and generating an optimal matching path includes:

[0046] Obtain the instantaneous benefit and discount factor of the state, construct the Bellman equation, update the current state value by calculating the expected value after the state transition to obtain the initial state value sequence, calculate the state transition probability under the initial matching strategy based on the initial state value sequence, adjust the temperature parameter according to the value improvement of the historical matching data, and update the state transition probability based on the temperature parameter;

[0047] Starting from the current state, the state value of all transferable states is calculated based on the updated state transition probability. The transition state with the largest state value is selected and added to the matching path. The matching path is then judged to see if it meets any of the conditions: target state, path length limit, or state value benefit threshold. If so, the matching path is marked as a candidate matching path.

[0048] Calculating the state value of each state node in the candidate matching path that can be transferred in multiple steps in the future, selecting the transfer state with the largest multi-step state value to update the candidate matching path, and obtaining an optimized matching path;

[0049] The optimized matching path is compared with the historical optimal matching path. When the state value is greater, the historical optimal matching path is updated. The state value lower limit is set based on the historical optimal matching path. The branches below the state value lower limit are removed to obtain the final optimal matching path.

[0050] According to a second aspect of the embodiments of the present invention,

[0051] Provide a student affairs assignment and service evaluation system based on multimodal hosting, including:

[0052] The first unit is used to receive student transaction request data, extract transaction descriptions from the student transaction request data, perform semantic analysis to extract semantic features, and fuse the semantic features with transaction type identifiers to generate a transaction feature vector containing complete transaction information;

[0053] The second unit is used to receive student transaction request data, extract transaction descriptions from the student transaction request data, perform semantic analysis to extract semantic features, and fuse the semantic features with transaction type identifiers to generate a transaction feature vector containing complete transaction information;

[0054] The third unit is used to construct a skill vector based on the professional skills of the processing personnel. The skill vector and the transaction feature vector in the transaction family are input into the adaptive dynamic programming algorithm. The optimal matching path is obtained through multi-stage decision optimization. The comprehensive matching score is generated based on the optimal matching path score and the historical service score.

[0055] The fourth unit is used to assign the transaction family and processing priority sequence to the target processing personnel with the highest comprehensive matching score, and update the target processing personnel's professional skill vector and historical service score based on the processing evaluation score of the transaction family for matching calculation of subsequent transaction allocation.

[0056] According to a third aspect of the embodiments of the present invention,

[0057] An electronic device is provided, comprising:

[0058] processor;

[0059] a memory for storing processor-executable instructions;

[0060] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0061] According to a fourth aspect of the embodiments of the present invention,

[0062] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0063] In this embodiment, it is possible to achieve intelligent identification, dynamic clustering, precise allocation and continuous optimization of student affairs. By constructing transaction feature vectors and time-series association graphs, the system can not only identify potential dependencies between transactions, but also evaluate resource competition and processing priorities, solving the problems of frequent transaction processing conflicts and process fragmentation in the traditional model. An adaptive dynamic programming algorithm is used to match processing personnel and transaction families, so that allocation decisions take into account both capability adaptability and historical service performance, significantly improving the professionalism and stability of transaction processing. Compared with existing methods that rely on manual judgment or static rules, this application provides a scheduling strategy that integrates semantic understanding, graph structure analysis and path optimization, breaking through the contradiction between transaction complexity and service personalization, realizing the transformation of student affairs processing from extensive to intelligent, and improving the overall response efficiency and user satisfaction of the hosting service system. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 Schematic diagram of the flow of a method for student affairs assignment and service evaluation based on multimodal hosting according to an embodiment of the present invention;

[0065] Figure 2 Processing dependency probability matrix heat map for embodiments of the present invention;

[0066] Figure 3 A comparison diagram of the iterative convergence process of the state value function strategy in an embodiment of the present invention;

[0067] Figure 4 This is a structural diagram of a student affairs allocation and service evaluation system based on multimodal hosting according to an embodiment of the present invention. DETAILED DESCRIPTION

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0069] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0070] Figure 1 FIG. 1 is a flow chart of a method for student affairs assignment and service evaluation based on multimodal hosting according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0071] Receive student transaction request data, extract transaction descriptions from the student transaction request data, perform semantic analysis to extract semantic features, and fuse the semantic features with transaction type identifiers to generate a transaction feature vector containing complete transaction information;

[0072] A transaction timing association graph is constructed based on transaction feature vectors. Recursive deep feature extraction is used to calculate the feature propagation between nodes in the transaction timing association graph, generating a processing dependency probability matrix. A transaction clustering evaluation function is constructed based on the processing dependency probability matrix. Transaction clustering evaluation functions are iteratively optimized to generate clusters of transactions. A processing priority sequence is generated for the transaction families based on the constraint strength in the processing dependency probability matrix.

[0073] A skill vector is constructed based on the professional skills of the processing personnel. The skill vector and the transaction feature vector in the transaction family are input into the adaptive dynamic programming algorithm. The optimal matching path is obtained through multi-stage decision optimization. A comprehensive matching score is generated based on the optimal matching path score and historical service scores.

[0074] The transaction family and processing priority sequence are assigned to the target processing personnel with the highest comprehensive matching score. The professional skill vector and historical service score of the target processing personnel are updated according to the processing evaluation score of the transaction family, which are used for matching calculation of subsequent transaction allocation.

[0075] In an optional embodiment,

[0076] Transaction descriptions are extracted from student transaction request data and semantic features are extracted through semantic analysis. The semantic features are then fused with transaction type identifiers to generate a transaction feature vector containing complete transaction information.

[0077] Extracting transaction descriptions from student transaction request data, performing multi-level decomposition of the transaction descriptions according to core information, attribute information, and environmental information, extracting transaction goals and requirement features from the core information, extracting transaction timeliness and urgency features from the attribute information, and extracting processing scenarios and resource requirement features from the environmental information to obtain multi-level transaction features;

[0078] performing a correlation analysis on the multi-level transaction features, calculating a feature recombination coefficient based on information completeness and information redundancy between the features, and combining transaction features whose complementarity is greater than a preset complementarity threshold according to the feature recombination coefficient to generate a transaction semantic feature;

[0079] A mapping rule base for transaction type identifiers is established, in which the correspondence and confidence between transaction type identifiers and transaction semantic features are stored. Feature matching is performed in the mapping rule base based on the transaction semantic features to obtain a probability distribution of transaction type identifiers. The transaction semantic features are combined with the transaction type identifier with the highest confidence to generate a transaction feature vector containing complete transaction information.

[0080] For example, the first step is to extract transaction description information from the transaction request data submitted by students. Transaction request data can include multimodal data such as text descriptions, images, and audio. Transaction descriptions typically come from free text information submitted by students or parents on the platform, covering the purpose of the transaction application, specific demands, time requirements, and execution background. After preliminary word segmentation and syntactic analysis, the transaction description is divided into three categories: core information, attribute information, and environmental information. Core information mainly includes the request intent and target verb phrases, such as "applying to change the after-school program," "providing feedback on course quality issues," or "applying for early pick-up." Attribute information indicates the timeliness and urgency of the request, such as "processing before tomorrow afternoon" or "must be completed today." Environmental information reflects the context and constraints of the transaction, such as "no one to pick up the child because the parent is on a business trip" or "the student is not adapting to the current course content."

[0081] Feature extraction is performed on each of the three types of information. Core information uses syntactic dependency analysis to determine the subject-object structure and identify specific transaction goals and their corresponding demands, thereby forming transaction goal features and demand features. Attribute information uses descriptive time limits, time ranges, and qualifiers to extract the timeliness level of the transaction. For example, "immediately" and "today" are classified as high timeliness, while "before next week" is classified as medium timeliness. Furthermore, if the text contains words such as "must" and "urgent," the task is marked as high-urgency. Environmental information is analyzed using location words, causal association words, and contextual descriptive words to extract scenario elements and resource dependencies related to transaction processing. For example, information such as "lack of caregivers," "lack of stationery materials," and "need for psychological counseling" are abstracted into resource demand features and processing environment features.

[0082] After the extracted multi-level transaction features are stored in a unified structured manner, all features are compared pairwise to construct a feature association network. In this process, the information redundancy and complementary relationships between different transaction features are dynamically analyzed. If there is obvious complementarity between the contents described by two features, for example, one describes "class time conflict" and the other describes "desire to change the after-school course", it is considered a highly complementary relationship. The system assigns a combination priority to each set of features. Based on the set complementarity threshold, only feature pairs with complementarity greater than the threshold are retained and combined into a new transaction semantic feature unit. This combination integrates contextual information through semantic information reconstruction to maintain semantic consistency and integrity. For example, "applying for course adjustment due to family changes" is a fusion of the two original features "family changes" and "course adjustment needs".

[0083] A mapping rule base between transaction type identifiers and semantic features is established. In the rule base, each mapping rule contains a set of semantic features, the corresponding transaction type identifier, and its confidence. For example, in a certain rule, when the semantic features include "lack of caregivers" and "urgent need to arrange temporary custody", the corresponding transaction type is "emergency custody arrangement" and the confidence is high. The mapping rule base is constructed through historical data training and manual rule setting, covering common transaction scenarios in custody services, including routine adjustment transactions, emergency replacement transactions, feedback and suggestion transactions, course evaluation transactions, and other types. After receiving the structured transaction semantic features, the system automatically searches the rule base for the transaction type identifier that best matches the feature set, and returns a list of matching results sorted by confidence. The confidence is comprehensively evaluated based on the frequency of feature occurrence, matching accuracy, and rule weight.

[0084] The item with the highest confidence is selected from the matching results and used as the type identifier of the transaction. It is then jointly encoded with the semantic features of the transaction to generate a complete transaction feature vector.

[0085] The feature vector contains complete transaction information, preserving the original semantic features while assigning clear type attributes. This multi-layered feature extraction and fusion approach fully leverages information from multimodal data to improve the accuracy of transaction understanding. Furthermore, through dynamic updates to the rule base, the system continuously optimizes feature mapping and enhances transaction classification accuracy.

[0086] To calculate information completeness, we use a method based on semantic similarity. First, we establish a feature word vector space, representing each feature as a vector. By calculating the cosine similarity between these vectors, we determine the degree of semantic relevance between the features. When two features have low semantic similarity but strong information complementarity, this indicates that combining these two features can provide a more complete transaction description.

[0087] Information redundancy is calculated based on feature content overlap analysis. Keyword terms are extracted from the features and their recurrence frequency is calculated. When two features contain a large number of identical or similar terms, this indicates information redundancy. Based on a preset redundancy threshold, features are selected for merging or deletion.

[0088] The calculation of the feature recombination coefficient takes into account both information completeness and redundancy. When features are highly complementary and have low redundancy, the recombination coefficient is high, indicating that these features are suitable for combination. For example, the maintenance target feature and the timeliness feature have a strong complementary relationship, and their combination can better express the urgency of the task.

[0089] The mapping rule base for transaction type identifiers utilizes a dynamic update mechanism. After each transaction is processed, the system optimizes the rule base based on the actual processing results. If a feature template closely matches the actual situation, the confidence level of the corresponding rule is increased; otherwise, the confidence level is decreased. This adaptive learning mechanism enables continuous improvement of the rule base and enhances the accuracy of feature mapping.

[0090] During the feature fusion phase, the system uses a weighted combination approach to generate the final transaction feature vector. Different feature types are assigned different weights, with core information typically receiving a higher weight than contextual information. Weights are set based on the importance and reliability of the features and can be optimized and adjusted through historical data analysis.

[0091] In this embodiment, by decomposing transaction descriptions at multiple levels, the core objectives, attribute characteristics, and environmental conditions are systematically divided, thereby enhancing the ability to capture and distinguish multi-dimensional information of transactions. Combined with the correlation analysis and reorganization between features, the expression of semantic features is optimized, the interference of redundant information is reduced, and the accuracy and completeness of transaction semantics are improved. The mapping rule library is used to achieve dynamic matching and confidence assessment of transaction type identification and semantic features, effectively improving the accuracy and reliability of transaction classification. Ultimately, the generated transaction feature vector fully reflects the semantic information and type attributes of the transaction, facilitating subsequent intelligent scheduling, resource allocation, and decision support, significantly enhancing the intelligence level and response efficiency of the student affairs processing system.

[0092] In an optional embodiment,

[0093] Based on the transaction feature vector, a transaction time series association graph is constructed. The feature propagation between nodes in the transaction time series association graph is calculated through recursive deep feature extraction. The generated processing dependency probability matrix includes:

[0094] Construct a transaction timing association graph based on the transaction feature vector, where nodes represent transactions and edges between nodes represent the timing association between transactions;

[0095] Establishing a conflict transaction tag library based on historical processing data, wherein the conflict transaction tag library records transaction combination types that cause resource contention, matching nodes in the transaction time sequence association graph with the conflict transaction tag library, identifying node pairs with resource contention, and generating a contention node set;

[0096] For each node pair in the competitive node set, the resource demand information of each node is extracted, the resource overlap between the node pairs is calculated, and the resource competition intensity is determined based on the resource overlap as the competition edge weight between the competing node pairs;

[0097] For the remaining node pairs in the transaction timing association graph, the time interval information and transaction feature vectors between the nodes are extracted, and the timing similarity between the node pairs is calculated as the association edge weight between the remaining node pairs;

[0098] The competitive edge weights and associated edge weights are combined to form complete edge weight information, which is used to guide the recursive propagation of node features. The node features are updated to obtain propagated node features. The processing dependency probability between nodes is calculated based on the propagated node features to generate a processing dependency probability matrix.

[0099] Exemplarily, after receiving multiple student transactions encoded by feature vectors, a transaction timing association graph is constructed based on these transaction feature vectors. Each node in the graph represents an independent student transaction. The request content and context attributes are represented by the transaction feature vector carried inside the node, and the edges between transactions are used to represent the temporal sequence relationship formed during the execution of historical transactions. During the graph construction process, the system establishes an initial edge set based on the transaction submission time sequence, and combines the actual task chain relationship in the hosting service scenario. For example, "applying for course change" must be executed after "course scheduling is completed". Directional connections are established between transaction nodes with clear temporal dependencies to form a complete transaction timing association graph.

[0100] Based on historical processing records, a conflict transaction tag library is established. This library contains transaction type combinations that conflict due to overlapping resource allocations, such as "course change application" and "course schedule review," "temporary custody application" and "teacher leave adjustment." After the transaction time-series association graph is constructed, all node pairs in the graph are retrieved and compared with the conflict transaction tag library in sequence to identify node pairs with resource contention. Identified transaction node pairs are grouped into a competing node set, and resource characteristics are compared for each transaction pair in this set.

[0101] The resource demand information for each transaction node is derived from the resource fields contained in its feature vector. Key fields such as teacher demand, venue occupancy, and time period usage are extracted from this information to construct a unified resource vector. For any node pair, the resource vector contents are compared to identify reused resource items. The number of overlapping items and the overlap intensity are counted to generate a resource overlap evaluation for the node pair. The higher the resource overlap value, the greater the degree of contention for platform resources by the transaction combination. The edge weight of the competing edge is set accordingly. The edge weight value range is set by configuration parameters. For example, full overlap is recorded as the maximum weight, and partial overlap is assigned a proportional value.

[0102] For all remaining pairs of nodes in the graph, excluding competing nodes, the temporal correlation between their transactions is calculated. The commit time, expected completion time, and service coverage time period of the two transactions are extracted to obtain time interval information. Combining the semantic dimensions of the transaction feature vectors, the transaction target similarity, time period overlap, and processing window consistency are comprehensively calculated to form a temporal similarity score for each node pair. Temporal similarity takes into account both temporal proximity and transaction feature similarity. Temporal similarity serves as the edge weight associated with these non-competing node pairs. The competing edge weights and the associated edge weights are unified to form complete edge weight information. This weight information guides the propagation of node features within the graph. Feature propagation is performed recursively, with each node receiving feature information from neighboring nodes and updating the features based on the edge weights.

[0103] During feature propagation, edge weights determine the strength of feature transfer. Between pairs of nodes with large competing edge weights, feature transfer is suppressed, indicating that these transactions need to be processed at different times. Between pairs of nodes with large associated edge weights, feature transfer is more robust, facilitating the centralized processing of similar transactions.

[0104] Based on the propagated node features, the system calculates the processing dependency probabilities between nodes. The processing dependency probabilities reflect the processing order constraints between transactions. A higher dependency probability indicates that the outcome of a previous transaction is more likely to affect the processing of a subsequent transaction. Finally, the system organizes the processing dependency probabilities between all node pairs into a matrix, generating a processing dependency probability matrix. This processing dependency probability matrix provides an important basis for subsequent transaction clustering and prioritization. The probability values ​​in the matrix account for both the processing constraints imposed by resource contention and the temporal dependencies between transactions, guiding the system in generating reasonable transaction processing plans.

[0105] In this embodiment, by constructing a transaction timing association graph and introducing a historical conflict transaction tag library, transaction combinations that may involve resource competition can be effectively discovered, thereby improving the conflict avoidance capability during concurrent transaction processing. Furthermore, by combining resource overlap and timing similarity to construct differentiated edge weights, the modeling of the relationship between transactions can be made more hierarchical and targeted. Through a feature recursive propagation mechanism guided by complete edge weights, the evolution of inter-transaction dependencies can be dynamically captured, thereby generating a highly reliable processing dependency probability matrix, providing accurate dependency information support for transaction clustering, prioritization, and scheduling optimization, significantly improving the overall system's intelligence and response efficiency.

[0106] In an optional embodiment,

[0107] A transaction clustering evaluation function is constructed based on the processing dependency probability matrix. Transaction families are generated by iteratively optimizing the transaction clustering evaluation function. A processing priority sequence is generated for the transaction family based on the constraint strength in the processing dependency probability matrix. The sequence includes:

[0108] Constructing a transaction clustering evaluation function based on the processing dependency probability matrix, wherein the transaction clustering evaluation function includes intra-cluster processing dependency constraint items and inter-cluster processing dependency constraint items;

[0109] Extracting historical execution records from the processing dependency probability matrix, calculating the transaction execution time fluctuation coefficient based on the historical execution records, constructing an execution stability matrix and inputting it into the transaction clustering evaluation function, performing initial clustering on the transactions to obtain an initial transaction family set;

[0110] Calculating the execution time fluctuation coefficient differences of the transaction pairs in the initial transaction family set, determining the processing dependency constraint weights and fluctuation compensation coefficients of the transaction pairs in the transaction clustering evaluation function based on the execution time fluctuation coefficient differences, iteratively optimizing the transaction family set, and obtaining the final transaction family set when the evaluation values ​​converge;

[0111] Transactions with execution duration fluctuation coefficients less than a preset fluctuation threshold are selected from the final transaction family set to construct a benchmark execution sequence. The degree of interference of the remaining transactions on the benchmark execution sequence is calculated according to the constraint strength in the processing dependency probability matrix, and the optimal insertion position of the remaining transactions is determined. Based on the benchmark execution sequence and the optimal insertion position, a transaction family processing priority sequence that meets the processing dependency constraints is generated.

[0112] For example, it is first necessary to construct a transaction clustering evaluation function based on the transaction processing dependency probability matrix. The evaluation function contains two main parts: intra-family processing dependency constraint items and inter-family processing dependency constraint items. The intra-family processing dependency constraint items are used to measure the strength of the processing dependency relationship between transactions in the same transaction family, and the inter-family processing dependency constraint items are used to measure the strength of the processing dependency relationship between transactions in different transaction families. Each element in the processing dependency probability matrix represents the processing dependency probability between the corresponding two transactions, and the value range is 0 to 1. 0 means that the two transactions are completely independent and have no dependency, and 1 means that the two transactions have a mandatory processing dependency order. Taking 10 transactions as an example, in the generated processing dependency probability matrix, the processing dependency probability of transaction 1 and transaction 2 is 0.8, which means that there is an 80% probability that transaction 1 needs to rely on the processing result of transaction 2 when it is executed.

[0113] Next, historical execution record information is extracted from the processing dependency probability matrix, and the coefficient of fluctuation is calculated based on the historical execution time of each transaction. The execution time fluctuation coefficient reflects the degree of transaction execution stability, with a smaller coefficient indicating more stable execution time. This coefficient of fluctuation is calculated by taking the ratio of the standard deviation to the average execution time of each transaction in the historical execution records. For example, if a transaction has an average execution time of 10 seconds and a standard deviation of 2 seconds over the past 100 executions, its execution time fluctuation coefficient is 0.2. The execution stability matrix is ​​constructed based on the calculated execution time fluctuation coefficients of each transaction.

[0114] The execution stability matrix is ​​input into the transaction clustering evaluation function, and spectral clustering is used to perform initial clustering of transactions, resulting in an initial set of transaction families. The spectral clustering method first constructs a transaction similarity matrix, then calculates the matrix's eigenvalues ​​and eigenvectors. Finally, transactions are divided into different families based on the eigenvectors. The construction of the similarity matrix comprehensively considers both processing dependency probability and execution stability. Using the 10 transactions mentioned above as an example, spectral clustering can yield three initial transaction families, containing three, four, and three transactions, respectively.

[0115] For each pair of transactions in the initial transaction family, calculate the difference in their execution time fluctuation coefficients. A larger fluctuation coefficient difference indicates a greater difference in the execution stability of the two transactions. Based on the fluctuation coefficient difference, determine the processing dependency constraint weight and fluctuation compensation coefficient in the clustering evaluation function for each transaction pair. The processing dependency constraint weight is inversely proportional to the fluctuation coefficient difference; that is, the greater the difference in execution stability, the smaller the constraint weight. The fluctuation compensation coefficient is directly proportional to the fluctuation coefficient difference and is used to balance the impact of execution stability differences. For example, if the fluctuation coefficient difference between a transaction pair is 0.3, the processing dependency constraint weight can be set to 0.7 and the fluctuation compensation coefficient can be set to 0.3.

[0116] Based on the updated constraint weights and fluctuation compensation coefficients, the transaction family set is iteratively optimized. In each iteration, the evaluation value of the current transaction family is first calculated. Then, transactions are attempted to be migrated between different families, selecting the migration solution that reduces the evaluation value. The evaluation value calculation takes into account the processing dependency constraints of transactions within a family, the processing dependency constraints of transactions between families, and fluctuation compensation for execution stability. When the change in the evaluation value over multiple consecutive iterations is less than a preset threshold, the evaluation value is considered converged, and the final transaction family set is obtained.

[0117] From the final set of transaction families, transactions with execution duration fluctuation coefficients less than a preset fluctuation threshold are selected to construct a baseline execution sequence. Assuming the fluctuation threshold is set to 0.15, transactions with fluctuation coefficients less than this threshold are prioritized for inclusion in the baseline execution sequence. These transactions have relatively stable execution times and are suitable as benchmarks for task scheduling. For the remaining transactions not included in the baseline execution sequence, their degree of interference with the baseline execution sequence is calculated based on the constraint strength in the processing dependency probability matrix.

[0118] The interference level is calculated by taking into account the processing dependency probabilities between the transaction to be inserted and each transaction in the baseline sequence. This is calculated by taking a weighted sum of the processing dependency probabilities between the transaction to be inserted and each transaction in the baseline sequence, with the weights decreasing as the transaction's position in the baseline sequence progresses. For example, if the processing dependency probabilities of a transaction to be inserted and the first three transactions in the baseline sequence are 0.6, 0.4, and 0.2, respectively, and position weights of 0.5, 0.3, and 0.2 are used, the calculated interference level is 0.44.

[0119] Based on the calculated interference level, the optimal insertion position is determined for each transaction to be inserted. Choosing the optimal insertion position balances two objectives: minimizing interference with the baseline execution sequence and ensuring that processing dependency constraints are satisfied. This determination is made by attempting to insert the transaction into each position in the baseline sequence, calculating the overall interference level and constraint violation level after insertion, and selecting the position with the minimum weighted sum of the two as the optimal insertion position.

[0120] Finally, based on the baseline execution sequence and the determined optimal insertion position, a transaction family processing priority sequence that satisfies the processing dependency constraints is generated. This generation process is iterative, each time selecting the transaction with the least interference from the pending transactions and inserting it into the corresponding optimal position. For example, a transaction family contains eight transactions, three of which form the baseline execution sequence. The remaining five transactions are inserted into positions determined in ascending order of interference, ultimately resulting in a complete processing priority sequence.

[0121] Through the detailed steps described above, we achieved transaction clustering assessment and priority sequence generation based on a processing dependency probability matrix. This method fully considers the processing dependencies and execution stability characteristics between transactions, achieves a reasonable transaction family division through iterative optimization, and generates a processing priority sequence that satisfies the constraints based on the baseline sequence and interference level analysis. The entire implementation process is highly operational and practical.

[0122] In this embodiment, a precise balance can be achieved between transaction processing dependencies and execution stability, significantly improving the scientific nature and execution efficiency of transaction scheduling. By combining the processing dependency probability with the historical execution fluctuation characteristics, the constructed transaction clustering evaluation function can effectively identify the internal coupling relationship and external conflict risks between transactions, ensuring that the clustering results have high correlation and stability. By introducing a fluctuation compensation mechanism, dynamic adjustment of execution uncertainty is achieved, improving the convergence effect of clustering iteration. The final generated transaction family processing priority sequence takes into account both dependency strength and execution stability, effectively reducing the risks of transaction conflicts, resource congestion and processing delays, and improving the intelligence level and operational performance of the overall scheduling system.

[0123] Figure 2 For the embodiment of the present invention, the processing depends on the probability matrix heat map, such as Figure 2The figure shows the interdependence strengths among eight transactions (T1 to T8). This matrix forms the basis for constructing transaction clustering evaluation functions, where the matrix elements represent the probability of a row transaction's dependency on a column transaction. The figure clearly reveals several clusters of high dependencies: T1 has a strong dependency of 0.78-0.85 with T2 and T4; T2 has a strong correlation of 0.75-0.92 with T3 and T6; the dependency probabilities between T3 and T4, T5, and T8 range from 0.64-0.81; the dependency probability between T4 and T5 is as high as 0.88; the dependency probability between T5 and T6 and T7 ranges from 0.61-0.79; the dependency between T6 and T7 is 0.87; the dependency between T7 and T8 is 0.83; and T8 has dependencies of 0.76 and 0.74 with T1 and T6, respectively. This complex dependency network needs to be reasonably divided through the clustering evaluation function proposed in this technical solution. Compared with the traditional K-means algorithm or hierarchical clustering method, this solution can more accurately capture the processing dependency constraints between transactions.

[0124] In an optional embodiment,

[0125] Generating a transaction family processing priority sequence that satisfies processing dependency constraints based on the benchmark execution sequence and the optimal insertion position includes:

[0126] An initial constraint edge set is constructed based on the benchmark execution sequence, and the transitive constraint strength between nodes in the initial constraint edge set is calculated using the shortest path algorithm. Transitive constraints with transitive constraint strength greater than a preset strength threshold are added to the initial constraint edge set to obtain a complete constraint edge set.

[0127] Dividing the complete constraint edge set into layers based on the in-degree and out-degree of the nodes to obtain a constraint graph including inter-layer constraint relations and intra-layer constraint relations;

[0128] Obtaining a deviation between the actual execution time and the expected execution time of the transaction, the number of constraint violations, and the degree of contention; calculating a performance score for the transaction based on the deviation, the number of violations, and the degree of contention; performing a weighted combination of the performance score and the base priority to obtain an updated transaction priority; and determining a set of affected nodes in the constraint graph that are affected by the updated priority;

[0129] sequentially adjusting the priority of each node in the influencing node set, calculating the dependency chain length between the nodes based on the constraint graph, extracting the optimized nodes whose dependency chain length is greater than a preset length threshold, reordering the optimized nodes to obtain multiple reordering schemes, and selecting the reordering scheme with the smallest processing delay as the target reordering scheme;

[0130] The target reordering scheme is compared with the historical priority sequence, and when the improvement exceeds a preset improvement threshold, the target reordering scheme is adopted to update the transaction family processing priority sequence.

[0131] For example, when generating a transaction family processing priority sequence, we first need to construct an initial constraint edge set based on the benchmark execution sequence. These transactions form a benchmark execution sequence based on their submission time and urgency. First, we identify the constraint relationships between directly adjacent transactions and construct the initial constraint edge set.

[0132] A shortest path algorithm is used to calculate the transitive constraint strength between nodes in the initial constraint edge set. Transitive constraint strength represents the strength of indirect dependencies between transactions and is calculated by accumulating the direct constraint strengths along the path. When the strength of a transitive constraint exceeds a preset strength threshold, the system adds the transitive constraint to the initial constraint edge set. In this way, the system obtains a complete constraint edge set that includes direct constraints and important transitive constraints. When partitioning the complete constraint edge set into layers, the system analyzes the in-degree and out-degree of the nodes. The in-degree represents the number of constraint edges pointing to the node, while the out-degree represents the number of constraint edges emanating from the node. By comparing the in-degree and out-degree of nodes, the system can determine the dependency hierarchy of transactions.

[0133] Real-time data on transaction execution is obtained, including the deviation between the actual execution time and the expected time. For example, a maintenance task was expected to be completed in two hours, but actually took three hours, resulting in a deviation of one hour. Constraint breaches record the number of times a transaction failed to execute according to the constraint requirements, such as failing to complete within the specified time or affecting the normal execution of other transactions. The degree of competition reflects the intensity of the competition for resources between a transaction and other transactions. Based on this data, the system calculates the performance score of the transaction. The performance score takes into account three dimensions: time deviation, breach of contract, and resource competition, and is weighted according to the importance of each dimension. The lower the performance score, the worse the transaction execution effect, and its priority needs to be adjusted appropriately.

[0134] The performance score is weighted and combined with the transaction's base priority to create an updated priority. The base priority is primarily determined by the transaction's urgency and importance, while the performance score reflects the transaction's actual execution performance. This combined priority more comprehensively reflects the transaction's processing requirements.

[0135] When the priority of a transaction changes, the system needs to determine the set of affected nodes in the constraint graph. These nodes include other transactions that have direct constraints or strong transitive constraints with the transaction. The priority of each node in the affected node set is adjusted in turn, and the length of the dependency chain between nodes is calculated based on the constraint graph. The dependency chain length represents the number of constraint edges passed from one node to another. When the dependency chain length exceeds the preset length threshold, these nodes are marked as optimized nodes and require key optimization. When reordering the optimized nodes, multiple alternative reordering schemes are generated. Each scheme needs to satisfy the dependencies in the constraint graph while taking into account resource utilization efficiency and processing time. By simulating and calculating the processing delay that may be caused by each scheme, the scheme with the smallest delay is selected as the target reordering scheme.

[0136] Finally, the target reordering scheme is compared with the historical priority sequence. The optimization effect of the new scheme is evaluated by calculating the degree of improvement in metrics such as processing time and resource utilization. If the improvement exceeds a preset improvement threshold, the new reordering scheme is adopted to update the transaction family's processing priority sequence.

[0137] Existing technologies usually rely on static rules or simple processing order settings in transaction priority sorting, which makes it difficult to dynamically reflect the complex dependencies and execution performance differences between transactions, and easily lead to priority conflicts, resource congestion and low scheduling efficiency. The present application constructs an initial constraint edge set through the benchmark execution sequence and the optimal insertion position, and introduces the transitive constraint strength to complete it, so that the dependency expression between transactions is more complete. Further, through hierarchical division to form a constraint graph, combined with the actual execution performance of the transaction and the historical default behavior, the transaction priority is dynamically adjusted, and a set of affected nodes is constructed to achieve fine control of local priorities. On this basis, a dependency chain length determination and multi-scheme reordering mechanism are designed to optimize the processing path from a global perspective, and set an improved threshold to control the priority update condition to avoid system instability caused by excessive adjustment. This scheme takes improving the adaptability of priority sorting and scheduling stability as its starting point, significantly enhancing the system's responsiveness to complex dependencies and dynamic performance of transactions, and effectively improving the problems of priority rigidity and lack of dynamic optimization capabilities in the existing technology.

[0138] In an optional embodiment,

[0139] A skill vector is constructed based on the professional skills of the processing personnel. The skill vector and the transaction feature vector in the transaction family are input into the adaptive dynamic programming algorithm. The optimal matching path is obtained through multi-stage decision optimization. The comprehensive matching score is generated based on the optimal matching path score and historical service scores. The score includes:

[0140] Obtain the professional qualification characteristics, work experience characteristics, and skill coverage characteristics of the processing personnel, and generate the professional skill vector of the processing personnel by weighted combination based on preset weights;

[0141] Calculating the degree of matching between the professional skill vector and the transaction feature vector in the transaction family to obtain a basic matching value, calculating the degree to which the skill item satisfies the transaction feature based on the basic matching value, and generating an initial state value of the state space;

[0142] Constructing a skill transfer matrix to represent the complementary relationship between skills, calculating a state transition probability based on the skill transfer matrix, and combining the state transition probability with an initial state value to construct a state space;

[0143] Obtain the service completion quality and processing timeliness of the processing personnel, calculate the historical service score, adjust the state value of the state space based on the historical service score to construct a state value function, perform policy iteration calculation on the state value function to obtain a state value sequence, determine the optimal state transition path based on the state value sequence, and generate the optimal matching path;

[0144] The state transfer benefit of the optimal matching path is calculated to obtain a path score, and the path score is combined with the historical service score according to the weight determined by the transaction characteristics to generate a comprehensive matching score.

[0145] For example, the basic information of the processing personnel is structured and extracted to obtain their professional qualification characteristics, work experience characteristics and skill coverage characteristics in the hosting service scenario. Professional qualification characteristics include whether they hold a teacher qualification certificate, a psychological counseling certificate, interest-based teaching qualifications and other categories. Work experience characteristics include cumulative service years, the number of service positions over the years, the diversity of service types and other contents. Skill coverage characteristics cover the current course teaching ability, event organization ability, emergency response ability and communication and coordination ability. Weight parameters are preset for each type of feature. For example, the weight of qualification completeness is higher than the number of positions, and the weight of service years is slightly higher than the coverage of activity types. After all features are standardized, the professional skill vector of the processing personnel is generated according to the set weight combination. The dimensional structure of this vector is consistent with the transaction feature vector to ensure that the two are comparable.

[0146] For each task family currently being assigned, the feature vector of each task is extracted and matched against the professional skill vector of the handler. The consistency of the corresponding dimensions between the vectors is calculated to determine the basic matching value between each task and handler pair. A higher basic matching value indicates a higher potential for the handler to execute the task. The coverage of each skill item in the task feature vector is further analyzed. Based on the handler's ability and skill level, a skill satisfaction score is assigned to form the initial value for each state. This initial value represents the static matching effect without considering skill combination and transfer.

[0147] A skill transfer matrix is ​​constructed based on a multidimensional skill labeling system. This matrix reflects the complementary and substitutable relationships between different skills. For example, "classroom management skills" can partially replace "discipline control skills," while "course explanation skills" and "student emotional guidance" have synergistic effects. Each skill pair is assigned a transfer probability value. Using this matrix, transition paths between initial states are constructed. A higher transfer probability indicates a more logically feasible skill combination path in the current task processing scenario. State transition paths indicate the likelihood that a person will be able to proceed to another task after completing one task.

[0148] Combining the aforementioned migration probabilities with the initial state values, a complete state space is constructed. Each state represents the degree of suitability for handling a particular transaction. Transitions between states are guided by skill transfer relationships, entering the dynamic programming calculation phase. Processing personnel's historical service completion quality data is obtained, including post-service evaluation scores, transaction processing feedback content, and quantitative scores. Processing timeliness data includes information such as the average time it takes to complete a transaction and the overdue rate. After comprehensive normalization, this data forms a historical service score. This score is used as a moderating factor to weight the state values ​​in the state space, enhancing the value representation of high-scoring processors at certain transaction nodes and effectively reflecting their historical capabilities.

[0149] After constructing the state-value function, a policy iteration update operation is performed, traversing all reachable state paths and updating the state-value sequence. During this iteration, the transition path with the highest payoff is selected as the optimal strategy for each step. After a fixed number of rounds or when convergence conditions are met, the optimal state transition path is finally generated. This path represents the optimal solution for selecting task processing order and matching objects from multiple transactions. Each node in the path represents the execution goal of the transaction, and each edge represents the dynamic adaptation of capability migration.

[0150] After obtaining the optimal path, the benefits of each state transition along the path are calculated. Benefits are determined by the degree of match, skill transfer fluidity, and processing experience. The sum of all state transition benefits along the path constitutes the path score. Using the path score as the primary reference value, it is combined with the handler's historical service score and weighted by factors set by the current transaction. For example, some transactions prioritize processing quality while others prioritize processing speed. These two factors are weighted according to different transaction dimensions to generate a comprehensive matching score.

[0151] For example, a handler holds a teaching certificate and psychological counseling qualifications, has five years of experience, and has participated in both basic tutoring and interest-based course organization. His historical service rating averages 4.8 out of 5, with an average response time of 30 minutes, below the system average. The task family includes three tasks: "Student Temporary Tuition Application," "Course Content Adjustment," and "Student Emotional Abnormality Feedback." The system generates skill vectors and compares them against the task features, yielding basic matching scores of 0.78, 0.69, and 0.91, respectively. Based on the skill transfer matrix, the preferred path is "Emotional Abnormality Handling → Temporary Tuition → Course Content Adjustment." This path has a calculated score of 0.85 and a historical score of 0.92. Considering the 0.7 weighting for the quality priority of the "Student Emotion" task and the 0.3 weighting for the efficiency priority of the "Course Adjustment" task, the overall matching score for this handler in this task family is 0.88, indicating significant ranking priority.

[0152] Multiple processors can participate in the matching score calculation in parallel, with the final allocation decision based on each processor's comprehensive matching score for each transaction. When the ratio of processors to transaction volume is unbalanced, the system performs batch matching based on the optimal path, enabling intelligent scheduling and allocation of multiple transactions, avoiding resource waste and inefficient execution, and improving the responsiveness of managed services and student satisfaction. The overall mechanism supports business environments such as dynamic addition of processors, real-time addition of transactions, and real-time adjustment of policy parameters, and has excellent scalability and stability.

[0153] In this embodiment, by building a matching mechanism between multi-dimensional transaction features and the skills of processing personnel, the intelligent and precise allocation of transactions is achieved, which not only improves the efficiency of student transaction processing, but also significantly improves the rationality of transaction matching and service quality. The feature propagation and dependency analysis mechanism can be used to identify the logical sequence and resource conflicts between transactions, effectively avoiding situations where the processing order is chaotic or resource scheduling conflicts. The comprehensive matching scoring mechanism introduces historical service capabilities into the decision-making process, realizing the quantitative assessment and dynamic feedback of the service performance of processing personnel in the transaction allocation process, thereby supporting the priority scheduling of high-quality personnel and ensuring that student affairs receive timely, professional and personalized responses.

[0154] In an optional embodiment,

[0155] Performing a policy iteration calculation on the state value function to obtain a state value sequence, determining an optimal state transition path based on the state value sequence, and generating an optimal matching path includes:

[0156] Obtain the instantaneous benefit and discount factor of the state, construct the Bellman equation, update the current state value by calculating the expected value after the state transition to obtain the initial state value sequence, calculate the state transition probability under the initial matching strategy based on the initial state value sequence, adjust the temperature parameter according to the value improvement of the historical matching data, and update the state transition probability based on the temperature parameter;

[0157] Starting from the current state, the state value of all transferable states is calculated based on the updated state transition probability. The transition state with the largest state value is selected and added to the matching path. The matching path is then judged to see if it meets any of the conditions: target state, path length limit, or state value benefit threshold. If so, the matching path is marked as a candidate matching path.

[0158] Calculating the state value of each state node in the candidate matching path that can be transferred in multiple steps in the future, selecting the transfer state with the largest multi-step state value to update the candidate matching path, and obtaining an optimized matching path;

[0159] The optimized matching path is compared with the historical optimal matching path. When the state value is greater, the historical optimal matching path is updated. The state value lower limit is set based on the historical optimal matching path. The branches below the state value lower limit are removed to obtain the final optimal matching path.

[0160] For example, based on the state space constructed in the early stage, an initial state value sequence is generated for each handler in the context of a specific transaction allocation. The required parameters include the immediate benefit and historical feedback impact factor of each state. The immediate benefit is given based on the current adaptability of the handler to a certain student transaction, which is used to reflect the direct effectiveness of the handler in the task; the historical feedback impact factor is calculated based on the evaluation results such as the historical processing success rate, parent satisfaction and transaction resolution time, which represents the discounted impact of the handler's long-term performance when handling similar tasks. In this way, a recursive function for value updating is constructed, and the current state value is iteratively corrected based on the expected benefit after the transaction transfer, gradually forming a complete initial state value sequence. Each state value in the sequence reflects the cumulative advantage of the current processing path at a certain stage.

[0161] Based on the initial state value sequence, the state transition probability distribution under the initial matching strategy is constructed. This probability determines the preferred jump direction between different task nodes in each stage. The initial value of the transition probability is determined by the relative difference between the state values. To improve the dynamic adaptability of the matching path, the system introduces a temperature parameter to control the sensitivity of the jump probability. The temperature parameter is adjusted based on the magnitude of the value change during the matching strategy iteration in the historical transaction matching data. If the value improvement brought about by the strategy update in the historical matching is significant, the temperature parameter is lowered to improve the bias of the optimal jump path; if the historical improvement is small, the temperature parameter is increased to increase the diversity of path exploration. Based on the updated state transition probability, starting from the current transaction state, the system gradually jumps to the next stage and selects the transfer target node with the largest state value in the current stage to be included in the matching path.

[0162] During the matching path generation process, the current path is continuously evaluated to see if it meets the specified termination criteria. These criteria include whether all target transactions have been covered, whether the path length has reached a system-defined limit, or whether the current cumulative value exceeds a specified revenue threshold. If any of these criteria are met, the current path is marked as a candidate matching path. This path represents the optimal intermediate outcome of the current strategy in terms of feasibility, completeness, and effectiveness.

[0163] Based on the candidate paths, we further consider the multi-step state evolution trends in future stages. Taking each state node in the path as a starting point, we predict its potential state value over multiple rounds of state transitions. We compare the cumulative effect of the state values ​​across all achievable paths and select the transition path with the best future value as the optimization direction for the current node. We then gradually update the state transition plans in the candidate paths to form an optimized matching path. The optimized path is based not only on current state performance but also incorporates the potential for future evolution, enabling a comprehensive assessment of long-term matching effectiveness.

[0164] The optimized matching path is compared against the best matching path stored in the historical records. If the current path has a higher cumulative state value, the existing path in the historical records is replaced as the latest best matching path. During this replacement process, the lower limit of the state value of the path starting point is also updated, which serves as a filtering threshold for subsequent path construction. Path branches with a state value below this threshold are excluded from further matching calculations, ensuring a balance between accuracy and efficiency in the path calculation process.

[0165] For example, a handler A possesses skills in three categories: course scheduling, emergency response, and psychological counseling. Their historical service quality rating is 4.6, and their average transaction response time is 25 minutes. The task family currently awaiting assignment includes temporary custody application, reporting abnormal student behavior, and course content adjustment. These three tasks are mapped to three state nodes, and the handler's skill vector is matched with the three task feature vectors to generate immediate benefits of 0.82, 0.76, and 0.88, respectively. Combined with the historical ratings, the initial state values ​​are 0.78, 0.74, and 0.84, respectively. In the initial round of policy iteration, the course content adjustment is determined to have the highest value and is designated as the first-hop path node. Based on the state transition probabilities, a candidate path sequence is constructed: "course adjustment → temporary custody → abnormal behavior." In future multi-step predictions, the combination "abnormal behavior → course adjustment → temporary custody" is found to have the highest cumulative value. Based on this, the matching path is modified, and if the cumulative value of this path exceeds the historical optimal value, it is updated as the current optimal path. This optimal path is ultimately output and used to accurately assign the task to handler A, achieving phased scheduling of processing tasks and optimal resource utilization. Throughout the entire process, the system eliminates the need for an exhaustive search of all state combinations. Instead, it retains only high-value paths and dynamically adjusts the strategy, ensuring the stable convergence and practical feasibility of the allocation algorithm in large-scale transaction scenarios. This technical solution not only improves the quality of transaction allocation but also enables predictive optimization of the processing process and resource risk avoidance, effectively enhancing the intelligent decision-making capabilities of the multimodal hosting platform in transaction scheduling.

[0166] Figure 3 This is a comparison diagram of the state value function strategy iteration convergence process of the embodiment of the present invention, as shown in Figure 3 The figure shows a comparison of the state-value function convergence process of this solution with traditional Q-learning and SARSA algorithms. The figure clearly shows that the state-value function converges significantly faster in this solution than traditional methods, reaching a state value of 0.85 after only 40 iterations, while Q-learning and SARSA only reach 0.60 and 0.48, respectively, with the same number of iterations. After 100 iterations, the state value of this solution reaches 1.05, 46% and 81% higher than Q-learning's 0.72 and SARSA's 0.58, respectively. This advantage stems from the solution's innovative approach of dynamically adjusting the temperature parameter based on the magnitude of value improvement in historical matching data and its multi-step forward state value calculation. Traditional Q-learning algorithms tend to greedily select states with higher immediate rewards, while SARSA is more conservative and fails to effectively balance exploration and exploitation. This solution achieves faster and more efficient state value convergence through improved calculation of the Bellman equation and policy iteration of the state value sequence.

[0167] The existing technology generally adopts static rule matching or fixed weight scoring in the process of student affairs allocation, which makes it difficult to dynamically adapt to the differences in the capabilities of processing personnel and changes in transaction characteristics, resulting in untimely response to transactions and low allocation efficiency. In addition, it lacks the ability to optimize the potential dependencies between transactions and processing paths in multi-transaction concurrent scenarios. This application constructs a state space and a state value function, introduces a policy iteration calculation and path optimization mechanism, and realizes dynamic planning and global optimal search of processing paths. The starting point of the improvement is to break through the limitations of a single matching decision and introduce multi-stage value judgments so that the skill transfer ability of processing personnel and the execution order of transactions can jointly participate in the decision-making process. In path generation and optimization, by introducing dynamic adjustment means such as temperature parameters and state value lower limits, efficient path screening and value-oriented matching control are achieved, effectively avoiding the problems of large path randomness and non-convergence of policy iteration in existing solutions. Ultimately, a significant improvement in the matching accuracy, policy stability and resource utilization efficiency of transaction processing paths is achieved, making the allocation of managed services more intelligent, targeted and sustainable.

[0168] In another optional implementation, the education administration unit can manage the overall configuration of the managed model, including maintaining basic information for all schools, grades, classes, and students in the region. After completing the basic school organizational structure and student information entry, a unified interface is provided to support import operations, while also allowing subsequent manual adjustments and editing, enabling dynamic updating and centralized management of student data.

[0169] Administrators customize the service content configuration under the "1+X" model based on actual local needs, including the specific categories, service time periods, teachers, enrollment targets, service forms and class schedules of basic and extended hosting projects. In the hosting mode settings, you can flexibly set whether to open single-choice and multiple-choice items, whether to allow over-enrollment, and even the content of service activities from weekly to daily can be precisely configured. Taking a school as an example, the basic project settings include daily homework guidance and physical activities, and the extended projects include hard pen calligraphy, scientific experiments and music appreciation. Each hosting service sets an upper limit on the number of applicants. Once the number of students reaches the upper limit, the system automatically prohibits registration and prompts that the quota is full.

[0170] When each school has completed the release of the managed courses, parent users can submit registration information through a unified platform. The platform automatically identifies the school, grade, and class to which the student belongs based on the student ID number filled in by the parent, and automatically fills in the student information maintained on file for the user to confirm and complete the submission. After the registration information is submitted, the system enters the review stage according to the preset process. The review level and authority are uniformly configured by the Education Bureau and can be designated as the class teacher, school administrator, or Education Bureau administrator. The system supports automatic review functions and makes judgments based on historical registration data, student status, and project capacity. If the review requires manual participation, the relevant reviewers can view the pending applications after logging into the platform, and approve or reject them based on the matching of student information and courses. The review results are synchronized to the parent end in real time.

[0171] After the registration is approved, if the system has configured payment functionality, parents must complete the payment process. Hosting programs support binding to fee templates, matching fees by project. For example, basic hosting can be set at 200 yuan per month, while expanded courses like calligraphy classes can be set at 300 yuan per session. The system automatically generates a payment list after approval. Payments can be made through the unified payment interface, which is linked to the financial module. Fee information can be queried and exported by school, class, and student, forming an auditable and traceable record.

[0172] During the service process, student attendance, leave requests, and early departures are continuously collected. Teachers enter this data daily on the school management side. Daily service performance is summarized, and abnormalities (such as three consecutive days of absence) are automatically alerted. After each period of tutoring, the platform offers a review function, allowing students and parents to evaluate course content, teachers, service environment, time slots, and other dimensions. The system supports a combination of star ratings and text reviews, and the review opening and closing times can be set in the backend. Review results are automatically archived in the data center, providing a basis for decision-making in future service optimizations.

[0173] All data is automatically categorized in the backend to form multi-dimensional reports, including enrollment statistics, program saturation analysis, evaluation satisfaction analysis, and payment completion rate analysis. For example, in a certain month, a district offered 60 after-school care classes, with a total of 1,320 students enrolled. Of these, 1,214 students paid their fees, a payment rate of 92%. Of these, 61% participated in outreach programs. The average course satisfaction rating was 4.7 out of 5, and the evaluation coverage rate reached 89%.

[0174] We support the integration of off-campus institutions, including the institution name, legal representative, contact information, and course types. After registering on the platform, institutions submit a partnership application, which is reviewed and approved. Once approved, they can publish partner courses and set the available schools, grade levels, and course times. Students who choose courses from off-campus institutions will be subject to additional review and approval, and a separate payment item will be generated.

[0175] On the teacher side, school administrators enter teacher information, supporting both batch import from templates and manual addition of individual teachers. The system automatically counts teacher teaching hours each semester, and course schedules are configured and displayed using the course schedule management feature. Administrators can review teacher schedules and attendance records by school, class, and course to ensure controllable teaching execution.

[0176] Permissions are also configured for different roles, with atomic-level function switches controlling the visible and operational scope of the interface. Each account can be assigned multiple roles, allowing, for example, a school administrator to have both student information review and course management permissions. In the backend system, the binding relationship between roles and functional modules is clearly displayed, allowing for dynamic adjustment and delegated management. For example, if a school schedules staff changes during the holidays, the managed review permissions can be temporarily delegated to another account.

[0177] In terms of the evaluation mechanism, the results of each course and teacher's evaluation are fed back to assist in optimizing the curriculum for the next semester. For example, a certain extension course, "Science Experiments," received a 95% "very satisfied" rating this semester, with students frequently mentioning "highly interesting" and "rich hands-on experience" in their written reviews. This course will be prioritized for recommendation and expanded to more schools in the next semester.

[0178] Finally, key operational indicators are displayed through the data cockpit, and core data such as the overall number of applicants in the region, the total amount of custody fees paid, service coverage by school, the proportion of applicants for different types of custody projects, and the satisfaction level of each project are displayed on the management side. At the same time, separate statistics are collected on indicators such as the number of applicants for community custody projects, course evaluations, and the number of courses, which facilitates targeted resource investment and optimized allocation.

[0179] Figure 4 FIG. 1 is a structural diagram of a student affairs allocation and service evaluation system based on multimodal hosting according to an embodiment of the present invention. Figure 4 As shown, the system includes:

[0180] The first unit is used to receive student transaction request data, extract transaction descriptions from the student transaction request data, perform semantic analysis to extract semantic features, and fuse the semantic features with transaction type identifiers to generate a transaction feature vector containing complete transaction information;

[0181] The second unit is used to receive student transaction request data, extract transaction descriptions from the student transaction request data, perform semantic analysis to extract semantic features, and fuse the semantic features with transaction type identifiers to generate a transaction feature vector containing complete transaction information;

[0182] The third unit is used to construct a skill vector based on the professional skills of the processing personnel. The skill vector and the transaction feature vector in the transaction family are input into the adaptive dynamic programming algorithm. The optimal matching path is obtained through multi-stage decision optimization. The comprehensive matching score is generated based on the optimal matching path score and the historical service score.

[0183] The fourth unit is used to assign the transaction family and processing priority sequence to the target processing personnel with the highest comprehensive matching score, and update the target processing personnel's professional skill vector and historical service score based on the processing evaluation score of the transaction family for matching calculation of subsequent transaction allocation.

[0184] According to a third aspect of the embodiments of the present invention,

[0185] An electronic device is provided, comprising:

[0186] processor;

[0187] a memory for storing processor-executable instructions;

[0188] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0189] According to a fourth aspect of the embodiments of the present invention,

[0190] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0191] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0192] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A student affairs assignment and service evaluation method based on multimodal hosting, characterized by: include: Receive student transaction request data, extract transaction descriptions from the student transaction request data, perform semantic analysis to extract semantic features, and fuse the semantic features with transaction type identifiers to generate a transaction feature vector containing complete transaction information; A transaction timing association graph is constructed based on transaction feature vectors. Recursive deep feature extraction is used to calculate the feature propagation between nodes in the transaction timing association graph, generating a processing dependency probability matrix. A transaction clustering evaluation function is constructed based on the processing dependency probability matrix. Transaction clustering evaluation functions are iteratively optimized to generate clusters of transactions. A processing priority sequence is generated for the transaction families based on the constraint strength in the processing dependency probability matrix. A skill vector is constructed based on the professional skills of the processing personnel. The skill vector and the transaction feature vector in the transaction family are input into the adaptive dynamic programming algorithm. The optimal matching path is obtained through multi-stage decision optimization. A comprehensive matching score is generated based on the optimal matching path score and historical service scores. Assign the task family and processing priority sequence to the target handler with the highest comprehensive matching score. Update the target handler's professional skill vector and historical service score based on the task family's processing evaluation score for matching calculations in subsequent task assignments. A transaction clustering evaluation function is constructed based on the processing dependency probability matrix. Transaction families are generated by iteratively optimizing the transaction clustering evaluation function. A processing priority sequence is generated for the transaction family based on the constraint strength in the processing dependency probability matrix. The sequence includes: Constructing a transaction clustering evaluation function based on the processing dependency probability matrix, wherein the transaction clustering evaluation function includes intra-cluster processing dependency constraint items and inter-cluster processing dependency constraint items; Extracting historical execution records from the processing dependency probability matrix, calculating the transaction execution time fluctuation coefficient based on the historical execution records, constructing an execution stability matrix and inputting it into the transaction clustering evaluation function, performing initial clustering on the transactions to obtain an initial transaction family set; Calculating the execution time fluctuation coefficient differences of the transaction pairs in the initial transaction family set, determining the processing dependency constraint weights and fluctuation compensation coefficients of the transaction pairs in the transaction clustering evaluation function based on the execution time fluctuation coefficient differences, iteratively optimizing the transaction family set, and obtaining the final transaction family set when the evaluation values ​​converge; Transactions with execution duration fluctuation coefficients less than a preset fluctuation threshold are selected from the final transaction family set to construct a benchmark execution sequence. The degree of interference of the remaining transactions on the benchmark execution sequence is calculated according to the constraint strength in the processing dependency probability matrix, and the optimal insertion position of the remaining transactions is determined. Based on the benchmark execution sequence and the optimal insertion position, a transaction family processing priority sequence that meets the processing dependency constraints is generated.

2. The method according to claim 1, characterized in that Transaction descriptions are extracted from student transaction request data and semantic features are extracted through semantic analysis. The semantic features are then fused with transaction type identifiers to generate a transaction feature vector containing complete transaction information. Extracting transaction descriptions from student transaction request data, performing multi-level decomposition of the transaction descriptions according to core information, attribute information, and environmental information, extracting transaction goals and requirement features from the core information, extracting transaction timeliness and urgency features from the attribute information, and extracting processing scenarios and resource requirement features from the environmental information to obtain multi-level transaction features; performing a correlation analysis on the multi-level transaction features, calculating a feature recombination coefficient based on information completeness and information redundancy between the features, and combining transaction features whose complementarity is greater than a preset complementarity threshold according to the feature recombination coefficient to generate a transaction semantic feature; A mapping rule base for transaction type identifiers is established, in which the correspondence and confidence between transaction type identifiers and transaction semantic features are stored. Feature matching is performed in the mapping rule base based on the transaction semantic features to obtain a probability distribution of transaction type identifiers. The transaction semantic features are combined with the transaction type identifier with the highest confidence to generate a transaction feature vector containing complete transaction information.

3. The method according to claim 1, characterized in that Based on the transaction feature vector, a transaction time series association graph is constructed. The feature propagation between nodes in the transaction time series association graph is calculated through recursive deep feature extraction. The generated processing dependency probability matrix includes: Construct a transaction timing association graph based on the transaction feature vector, where nodes represent transactions and edges between nodes represent the timing association between transactions; Establishing a conflict transaction tag library based on historical processing data, wherein the conflict transaction tag library records transaction combination types that cause resource contention, matching nodes in the transaction time sequence association graph with the conflict transaction tag library, identifying node pairs with resource contention, and generating a contention node set; For each node pair in the competitive node set, the resource demand information of each node is extracted, the resource overlap between the node pairs is calculated, and the resource competition intensity is determined based on the resource overlap as the competition edge weight between the competing node pairs; For the remaining node pairs in the transaction timing association graph, the time interval information and transaction feature vectors between the nodes are extracted, and the timing similarity between the node pairs is calculated as the association edge weight between the remaining node pairs; The competitive edge weights and associated edge weights are combined to form complete edge weight information, which is used to guide the recursive propagation of node features. The node features are updated to obtain propagated node features. The processing dependency probability between nodes is calculated based on the propagated node features to generate a processing dependency probability matrix.

4. The method according to claim 1, wherein Generating a transaction family processing priority sequence that satisfies processing dependency constraints based on the benchmark execution sequence and the optimal insertion position includes: An initial constraint edge set is constructed based on the benchmark execution sequence, and the transitive constraint strength between nodes in the initial constraint edge set is calculated using the shortest path algorithm. Transitive constraints with transitive constraint strength greater than a preset strength threshold are added to the initial constraint edge set to obtain a complete constraint edge set. Dividing the complete constraint edge set into layers based on the in-degree and out-degree of the nodes to obtain a constraint graph including inter-layer constraint relations and intra-layer constraint relations; Obtaining a deviation between the actual execution time and the expected execution time of the transaction, the number of constraint violations, and the degree of contention; calculating a performance score for the transaction based on the deviation, the number of violations, and the degree of contention; performing a weighted combination of the performance score and the base priority to obtain an updated transaction priority; and determining a set of affected nodes in the constraint graph that are affected by the updated priority; sequentially adjusting the priority of each node in the influencing node set, calculating the dependency chain length between the nodes based on the constraint graph, extracting the optimized nodes whose dependency chain length is greater than a preset length threshold, reordering the optimized nodes to obtain multiple reordering schemes, and selecting the reordering scheme with the smallest processing delay as the target reordering scheme; The target reordering scheme is compared with the historical priority sequence, and when the improvement exceeds a preset improvement threshold, the target reordering scheme is adopted to update the transaction family processing priority sequence.

5. The method according to claim 1, wherein A skill vector is constructed based on the professional skills of the processing personnel. The skill vector and the transaction feature vector in the transaction family are input into the adaptive dynamic programming algorithm. The optimal matching path is obtained through multi-stage decision optimization. The comprehensive matching score is generated based on the optimal matching path score and historical service scores. The score includes: Obtain the professional qualification characteristics, work experience characteristics, and skill coverage characteristics of the processing personnel, and generate the professional skill vector of the processing personnel by weighted combination based on preset weights; Calculating the degree of matching between the professional skill vector and the transaction feature vector in the transaction family to obtain a basic matching value, calculating the degree to which the skill item satisfies the transaction feature based on the basic matching value, and generating an initial state value of the state space; Constructing a skill transfer matrix to represent the complementary relationship between skills, calculating a state transition probability based on the skill transfer matrix, and combining the state transition probability with an initial state value to construct a state space; Obtain the service completion quality and processing timeliness of the processing personnel, calculate the historical service score, adjust the state value of the state space based on the historical service score to construct a state value function, perform policy iteration calculation on the state value function to obtain a state value sequence, determine the optimal state transition path based on the state value sequence, and generate the optimal matching path; The state transfer benefit of the optimal matching path is calculated to obtain a path score, and the path score is combined with the historical service score according to the weight determined by the transaction characteristics to generate a comprehensive matching score.

6. The method according to claim 5, characterized in that Performing a policy iteration calculation on the state value function to obtain a state value sequence, determining an optimal state transition path based on the state value sequence, and generating an optimal matching path includes: Obtain the instantaneous benefit and discount factor of the state, construct the Bellman equation, update the current state value by calculating the expected value after the state transition to obtain the initial state value sequence, calculate the state transition probability under the initial matching strategy based on the initial state value sequence, adjust the temperature parameter according to the value improvement of the historical matching data, and update the state transition probability based on the temperature parameter; Starting from the current state, the state value of all transferable states is calculated based on the updated state transition probability. The transition state with the largest state value is selected and added to the matching path. The matching path is then judged to see if it meets any of the conditions: target state, path length limit, or state value benefit threshold. If so, the matching path is marked as a candidate matching path. Calculating the state value of each state node in the candidate matching path that can be transferred in multiple steps in the future, selecting the transfer state with the largest multi-step state value to update the candidate matching path, and obtaining an optimized matching path; The optimized matching path is compared with the historical optimal matching path. When the state value is greater, the historical optimal matching path is updated. The state value lower limit is set based on the historical optimal matching path. The branches below the state value lower limit are removed to obtain the final optimal matching path.

7. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images