A dynamic information scheduling method and system based on a rule engine

Abstract

The application relates to the technical field of information scheduling, and discloses a dynamic information scheduling method and system based on a rule engine, which comprises the following steps: acquiring a scheduling rule set and constructing a multi-dimensional feature vector; performing linear weighting by using a preset initial weight coefficient, performing information entry sorting, obtaining a sorting index sequence, generating a preliminary scheduling instruction and executing the preliminary scheduling instruction to obtain scheduled information entries; acquiring user interaction feedback data, combining an urgency decay function to calculate a current urgency residual value, performing utility conversion to obtain an actual utility value; acquiring a rule comprehensive evaluation value to perform deviation calculation and obtain a predicted deviation; if the predicted deviation is greater than a preset deviation threshold, the initial weight coefficient is iteratively updated to obtain an iterative weight coefficient; and a new round of scheduling instruction generation is performed to obtain a new round of scheduling instruction. The method can consider multiple scheduling targets or scheduling rules and continuously output appropriate scheduling instructions in dynamic changes.

Description

Technical Field

[0001] This invention relates to the field of information scheduling, and in particular to a dynamic information scheduling method and system based on a rule engine. Background Technology

[0002] Currently, in the field of information scheduling, the key to improving overall system efficiency lies in how to efficiently allocate limited information resources to different demanders through big data analysis.

[0003] In one existing technology, scheduling methods rely on preset fixed rules or a single objective for decision-making. However, this method often appears rigid when facing complex and ever-changing information environments, struggling to simultaneously consider multiple conflicting scheduling objectives, such as high returns and rapid response. It typically only allows for simple trade-offs or fixed priorities, failing to dynamically balance according to real-time contexts. Furthermore, its technical limitation lies in the insufficient characterization of the multidimensionality and dynamic correlation of scheduling rules. Scheduling rules contain multidimensional attributes such as returns and urgency, and the relationships between these attributes constantly change with the business scenario. If rules are treated merely as isolated entries, the interactions between attributes cannot be quantified, let alone the optimal coordination point found at a global level.

[0004] In summary, existing technologies struggle to accommodate multiple scheduling objectives or rules, and cannot continuously output well-balanced scheduling instructions amidst dynamic changes. Summary of the Invention

[0005] This invention provides a dynamic information scheduling method and system based on a rule engine, which can take into account multiple scheduling objectives or scheduling rules and continuously output well-balanced scheduling instructions in dynamic changes.

[0006] Firstly, to address the aforementioned technical problems, this invention provides a dynamic information scheduling method based on a rule engine, comprising: Obtain the set of scheduling rules corresponding to the information entries to be scheduled, extract the rule attribute values ​​from the set of scheduling rules, construct a multi-dimensional feature vector based on the rule attribute values, and obtain the rule feature vector; A weighted comprehensive evaluation value for each rule is obtained by linearly weighting each dimension of the rule feature vector using a preset initial weight coefficient. Based on the weighted comprehensive evaluation value, the information items are sorted to obtain a sorted index sequence. A preliminary scheduling instruction is generated and executed based on the sorted index sequence to obtain the scheduled information items. The user interaction feedback data and execution timestamp of the scheduled information entry are obtained. Based on the execution timestamp and a preset urgency decay function, the residual value is calculated to obtain the current urgency residual value. Then, the utility value of the scheduled information entry is converted based on the current urgency residual value and the user interaction feedback data to obtain the actual utility value. Obtain the rule-based comprehensive evaluation value of the scheduled information entry, and calculate the deviation based on the rule-based comprehensive evaluation value and the actual utility value to obtain the prediction deviation; If the prediction deviation is greater than the preset deviation threshold, then the minimum prediction deviation is set as the target, a loss function model is constructed, and the initial weight coefficients are iteratively updated using the gradient descent method based on the loss function model to obtain the updated iterative weight coefficients. A new round of scheduling instructions is generated by combining the iterative weight coefficients, resulting in a new round of scheduling instructions used to drive the execution of the business system.

[0007] Secondly, the present invention provides a dynamic information scheduling system based on a rule engine, comprising: The rule feature extraction module is used to obtain the set of scheduling rules corresponding to the information entries to be scheduled, extract the rule attribute values ​​in the set of scheduling rules, construct a multi-dimensional feature vector based on the rule attribute values, and obtain the rule feature vector. The comprehensive evaluation calculation module is used to linearly weight each dimension of the rule feature vector using preset initial weight coefficients to obtain the weighted comprehensive evaluation value of each rule; The scoring instruction execution module is used to sort the information items according to the weighted comprehensive evaluation value, obtain a sorted index sequence, generate a preliminary scheduling instruction according to the sorted index sequence and execute it to obtain the scheduled information items. The feedback utility calculation module is used to obtain the user interaction feedback data and execution timestamp of the scheduled information item, and calculate the residual value based on the execution timestamp and a preset urgency decay function to obtain the current urgency residual value. Then, the module performs utility conversion on the scheduled information item based on the current urgency residual value and the user interaction feedback data to obtain the actual utility value. The prediction deviation calculation module is used to obtain the rule comprehensive evaluation value of the scheduled information entries, and to calculate the deviation based on the rule comprehensive evaluation value and the actual utility value to obtain the prediction deviation. The deviation weight optimization module is used to set the minimum prediction deviation as the target if the prediction deviation is greater than the preset deviation threshold, construct a loss function model, and use the gradient descent method to iteratively update the initial weight coefficients based on the loss function model to obtain the updated iterative weight coefficients. The secondary instruction generation module is used to generate a new round of scheduling instructions by combining the iterative weight coefficients, thereby obtaining a new round of scheduling instructions to drive the execution of the business system.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention constructs a multidimensional feature vector space and a dynamic weighted evaluation mechanism. Based on the vector space model and multi-attribute decision theory, it transforms the multidimensional attributes of scheduling rules into computable feature vectors and uses configurable weight coefficients for linear weighted fusion. At the same time, it eliminates attribute redundancy through orthogonalization and dimensionality reduction, thereby realizing the quantitative trade-off and unified ranking of multi-objective conflicts and breaking through the rigidity of decision-making with fixed rules or single objectives.

[0009] (2) This invention establishes a closed-loop optimization mechanism of prediction, execution, feedback and correction, collects user interaction feedback to calculate the actual utility value, compares it with the predicted evaluation value to obtain the deviation, and iteratively updates the weight coefficient based on the gradient descent method, so that the scheduling strategy continuously adapts to the user's real preferences and avoids the system from falling into a static suboptimal state.

[0010] (3) This invention uses a unified recalculation of existing and incremental information and a global priority fusion mechanism. Based on the feedback control closed-loop system design and rolling optimization theory, it uses optimized weight coefficients to recalculate the backlog of information to be scheduled and prioritizes it on the same scale as the newly added information. This ensures that the optimization results cover historical and real-time information in a timely manner, and achieves the technical effect of maintaining the continuity of scheduling decisions and the timeliness of response capabilities. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the dynamic information scheduling method based on a rule engine provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a dynamic information scheduling system based on a rule engine provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides a dynamic information scheduling method based on a rule engine, comprising the following steps: S11, obtain the set of scheduling rules corresponding to the information entry to be scheduled, extract the rule attribute values ​​in the set of scheduling rules, construct a multi-dimensional feature vector based on the rule attribute values, and obtain the rule feature vector; S12, each dimension of the rule feature vector is linearly weighted using a preset initial weight coefficient to obtain a weighted comprehensive evaluation value for each rule; S13, based on the weighted comprehensive evaluation value, sort the information items to obtain a sorted index sequence, and generate and execute a preliminary scheduling instruction based on the sorted index sequence to obtain the scheduled information items; S14, obtain the user interaction feedback data and execution timestamp of the scheduled information entry, and calculate the residual value based on the execution timestamp and the preset urgency decay function to obtain the current urgency residual value. Then, perform utility conversion on the scheduled information entry based on the current urgency residual value and the user interaction feedback data to obtain the actual utility value. S15, obtain the rule comprehensive evaluation value of the scheduled information entry, and calculate the deviation based on the rule comprehensive evaluation value and the actual utility value to obtain the prediction deviation; S16 If the prediction deviation is greater than the preset deviation threshold, then the minimum prediction deviation is set as the target, a loss function model is constructed, and the initial weight coefficients are iteratively updated using the gradient descent method based on the loss function model to obtain the updated iterative weight coefficients. S17, Combine the iterative weight coefficients to generate a new round of scheduling instructions, resulting in a new round of scheduling instructions to drive the execution of the business system.

[0014] In step S11, the set of scheduling rules corresponding to the information entry to be scheduled is obtained, and the rule attribute values ​​in the set of scheduling rules are extracted. A multi-dimensional feature vector is constructed based on the rule attribute values ​​to obtain the rule feature vector.

[0015] It is worth noting that this step ensures that all rules entering the decision-making process have clear attribute quantification values ​​through four progressive levels of processing: validity screening, weighted correction, vector construction, and dimensionality reduction optimization. These attribute values ​​accurately reflect the actual importance of the rules in the current business environment, while eliminating redundant correlations between attributes through mathematical methods.

[0016] The scheduling rule set is a set of preset decision conditions for the information items to be scheduled. It contains multiple scheduling rules, each defining the execution logic for information scheduling under specific conditions. These scheduling rules are pre-configured by the business system based on factors such as information type, recipient, and timeliness requirements. For example, if a company has 10 notification messages to be sent, involving different departments and priorities, it can obtain the associated scheduling rule set, containing 20 rules, such as those based on time, user role, and notification type.

[0017] In one implementation, rule attribute values ​​are extracted from the scheduling rule set, and a multi-dimensional feature vector is constructed based on the rule attribute values ​​to obtain the rule feature vector, including: Obtain resource consumption data, and perform validity filtering based on the resource consumption data and the scheduling rule set to obtain a valid scheduling rule set; Extract the initial benefit attribute value and initial urgency attribute value of each rule in the effective scheduling rule set, and perform weighted correction on the initial benefit attribute value and initial urgency attribute value according to historical scheduling frequency data to obtain the corrected benefit attribute value and corrected urgency attribute value. Obtain the category identifier and construct a multi-dimensional feature vector by combining the corrected benefit attribute value and the corrected urgency attribute value to obtain the initial feature vector; The initial feature vector is orthogonalized and dimensionality reduced to obtain the rule feature vector describing the attribute of each rule.

[0018] It's worth noting that resource consumption data is a quantitative indicator of the system resources required to execute a single scheduling rule, including but not limited to server load and network bandwidth, obtained through real-time system collection or based on historical execution log statistics. Validity filtering is a process of filtering scheduling rules based on preset resource consumption thresholds to eliminate rules that cannot be executed under the current system load due to excessive resource requirements. The resource consumption threshold is dynamically calculated based on the current remaining system resource capacity or preset based on historical peak loads. For example, regarding server load, the resource consumption threshold is set as the number of load units corresponding to a single application server's CPU utilization not exceeding 60%, or the number of memory units corresponding to memory usage not exceeding 70% of the total physical memory. Regarding network bandwidth, the resource consumption threshold is set as the remaining available bandwidth value after deducting the currently occupied bandwidth from 80% of the total dedicated line bandwidth, or a single rule bandwidth limit set based on historical traffic statistics during peak business periods, such as 10Mbps. The resource consumption threshold is dynamically refreshed during system operation; when a cluster expansion or contraction event is detected, the threshold value is automatically recalculated and updated.

[0019] During validity screening, each rule in the scheduling rule set needs to be traversed, and its corresponding resource consumption data extracted. For example, rule R1, which pushes to all users immediately, requires 8 units of server load and 15Mbps of network bandwidth. This resource consumption data is compared with the resource consumption threshold, which is 5 units of server load and 10Mbps of network bandwidth. If the value is greater than the threshold, the rule is marked as temporarily unavailable and removed. If the resource consumption data of a rule is less than or equal to the threshold, the rule is included in the set of valid scheduling rules, ensuring that all subsequent calculations are based on the executable subset of rules.

[0020] In this embodiment, the rule attribute values ​​are continuous numerical values ​​ranging from 1.0 to 10.0, including but not limited to revenue attribute values ​​and urgency attribute values. For attribute value extraction, the system reads values ​​from the attribute fields of the rule attribute database based on the same unique rule identifier. The values ​​are predetermined according to information timeliness requirements, event urgency levels, or response time limits agreed upon in the service level agreement. The initial revenue attribute value is a quantitative indicator of the business value expected to be generated by executing the scheduling rule, including at least one of user conversion rate, expected revenue amount, information dissemination scope, or business goal achievement. The initial urgency attribute value is a quantitative indicator of the time sensitivity of executing the scheduling rule, and the time sensitivity is determined based on information timeliness, event urgency levels, or SLA (Service Level Agreement) requirements. Historical scheduling frequency data is a statistical value of the number of times the scheduling rule is triggered within a preset historical time window, reflecting the actual usage activity of the rule. A higher frequency indicates that the rule has been verified as important or frequently used in business practice.

[0021] Specifically, weighted correction is a calculation process that dynamically adjusts the original attribute values ​​using historical scheduling frequency data. For example, it extracts the trigger execution records of each rule in the set of valid scheduling rules from the system log database within the last 30 days, statistically obtains historical scheduling frequency data, and sets up a frequency correction coefficient mapping table. This table maps different frequency ranges to different correction coefficients, with higher frequencies corresponding to larger correction coefficients. Then, the initial benefit attribute value, the initial urgency attribute value, and the corresponding historical scheduling frequency correction coefficient are multiplied to obtain the corrected benefit attribute value and the corrected urgency attribute value.

[0022] The frequency correction coefficient mapping table divides the frequency values ​​into four intervals: low frequency [0, 5), low-to-medium frequency [5, 15), medium-to-high frequency [15, 30), and high frequency [30, +∞), corresponding to correction coefficients of 0.90, 1.00, 1.10, and 1.20, respectively. Taking the annual welfare notification of rule R5 as an example, its historical scheduling frequency value within the past 30 days is 2, falling within the interval [0, 5), corresponding to a correction coefficient of 0.90. Its original benefit attribute value is 9.0, and its original urgency attribute value is 3.0. Therefore, the corrected benefit attribute value is 9.0 × 0.90 = 8.1, and the corrected urgency attribute value is 3.0 × 0.90 = 2.7.

[0023] Furthermore, the category identifier is a label used to distinguish the business type to which the scheduling rule belongs, including but not limited to emergency notifications, marketing promotions, system alarms, and daily information. The initial feature vector is a multi-dimensional vector directly concatenated from the corrected revenue attribute value, the corrected urgency attribute value, the category identifier, and more dimensional attribute values ​​such as resource consumption. For example, the category identifier is a discrete label, including four business types: emergency notifications, marketing promotions, system alarms, and daily information. The category identifier needs to be converted into a numerical vector using a one-hot encoding method. For a simple example, the resource consumption data for rule R3 is 3.2 (a low load value confirmed by validity screening), the corrected benefit attribute value is 9.2, the corrected urgency attribute value is 8.05, the category is emergency notification, and the one-hot encoding identifier is [1,0,0,0], which are concatenated to form a 7-dimensional initial feature vector [3.2, 9.2, 8.05, 1, 0, 0, 0]; the resource consumption data for rule R4 is 1.5, the corrected benefit attribute value is 4.75, the corrected urgency attribute value is 3.8, the category is daily information, and the one-hot encoding identifier is [0,0,1,0], which are concatenated to form a 7-dimensional initial feature vector [1.5, 4.75, 3.8, 0, 0, 1, 0].

[0024] In one example, the rule feature vector is a final multi-dimensional vector obtained after orthogonalization and dimensionality reduction, accurately describing the comprehensive characteristics of each rule in terms of benefit, urgency, and category attributes, with each dimension being independent and information-rich. When using Principal Component Analysis (PCA) for feature optimization based on orthogonalization and dimensionality reduction, the initial feature vectors of the effective scheduling rule set are selected as samples, and the projection matrix is ​​calculated online. Specifically, the initial feature vector set is standardized by a min-max linear transformation, and the covariance matrix of the standardized feature vectors is calculated. The covariance matrix is ​​then decomposed into eigenvalues, and the top k feature vectors are selected in descending order of eigenvalues ​​to form the projection matrix. The selection of k satisfies either a cumulative contribution rate of 85% or a minimum eigenvalue of not less than 0.5, whichever comes first. The initial feature vectors are projected onto the k-dimensional subspace spanned by the projection matrix, resulting in rule feature vectors with independent dimensions and concentrated information. The projection matrix is ​​calculated online based on the current batch of samples to ensure adaptive matching between feature extraction and the current data distribution.

[0025] The value of 0.5 is an empirical threshold that can be adjusted based on the actual data distribution or actual needs.

[0026] In step S12, each dimension of the rule feature vector is linearly weighted using preset initial weight coefficients to obtain a weighted comprehensive evaluation value for each rule, including: Construct a weight coefficient vector based on the preset initial weight coefficients; Obtain the feature vector to be calculated from the feature vector of the rule, which corresponds to the dimension of the weight coefficient vector; A linear transformation mapping is performed on the feature vector to be calculated to obtain a standardized feature vector; The weighted comprehensive evaluation value is obtained by performing a dot product operation between the standardized feature vector and the weight coefficient vector.

[0027] It is worth noting that this step introduces preset initial weight coefficients to linearly weight and fuse multiple dimensions of the rule feature vector, enabling rules with different attributes to be compared on a unified evaluation scale.

[0028] The initial weight coefficients are importance-weighted parameters pre-set for each dimension of the rule feature vector. These are configured by the business system based on current operating strategies or historical experience data, and range from 0 to 1. The sum of all dimension coefficients in the same weight coefficient vector is 1. In this embodiment, resource consumption and business urgency are used as examples. The resource consumption dimension represents the system resource overhead required to execute the rule in the rule feature vector. The business urgency dimension represents the time sensitivity of an information item from its creation to its expiration and the risk of business loss due to delayed execution. It represents the time sensitivity and business criticality of executing the rule in the rule feature vector. For example, if a delay of 1 minute in processing an information item may cause service unavailability in the production environment, it has high business urgency. The weight coefficient vector is a vector composed of the initial weight coefficients of each dimension arranged in a fixed order, and its number of dimensions is equal to the number of dimensions of the rule feature vector. For example, the timeliness of notification is often more important than resource consumption. Based on the principle of business priority, the weight coefficient of the resource consumption dimension is set to 0.4, and the weight coefficient of the business urgency dimension is set to 0.6, thus forming a weight coefficient vector [0.4, 0.6].

[0029] In this embodiment, the feature vector to be calculated is a sub-vector extracted from the regular feature vector that corresponds to the dimension of the weight coefficient vector, and contains the feature values ​​of each dimension participating in the weighted calculation. For example, the weight coefficient vector has two dimensions: resource consumption dimension and business urgency dimension. Then, the first dimension value 2.85, corresponding to the principal component of resource consumption, and the second dimension value 1.92, corresponding to the principal component of business urgency, are extracted from the regular feature vector generated in step S11 to form the feature vector to be calculated [2.85, 1.92].

[0030] Furthermore, the standardized feature vector is a feature vector obtained after linear transformation mapping. The values ​​of each dimension of the vector are all within a uniform numerical range, usually the closed interval [0,1]. Linear transformation mapping is a mathematical process of standardizing the numerical range of the feature vector to be calculated. For all feature vectors to be calculated, the historical maximum and minimum values ​​in the historical data of the corresponding dimension are extracted. For any original feature value, the standardized value is calculated using the maximum-minimum linear transformation formula, linearly mapping the original value to the closed interval [0,1]. Specifically, for the resource consumption dimension, if the historical maximum value is 10, the historical minimum value is 2, and the original value of a certain rule is 8, then the standardized value is (8-2) / (10-2)=0.75; for the business urgency dimension, if the historical minimum value is 9, the historical minimum value is 3, and the original value of a certain rule is 9, then the standardized value is (9-3) / (9-3)=1.00. After transformation, the values ​​of each dimension are on a uniform scale and stored in the standardized feature vector library, which can be directly used for weighted summation in subsequent steps.

[0031] Specifically, the weighted comprehensive evaluation value is a single scalar value obtained through a dot product operation. It quantifies the relative priority of the rule under the current weight configuration. The larger the value, the higher the comprehensive evaluation of the rule, and the more likely it should be scheduled for execution first. For example, the standardized feature vector and the weight coefficient vector are multiplied together by their corresponding dimensions and then summed. For instance, if the standardized feature vector is [0.13, 0.18] and the weight coefficient vector is [0.3, 0.7], then the weighted comprehensive evaluation value obtained by the dot product operation is 0.13 × 0.3 + 0.18 × 0.7 = 0.039 + 0.126 = 0.165.

[0032] In step S13, the information entries are sorted according to the weighted comprehensive evaluation value to obtain a sorted index sequence, and a preliminary scheduling instruction is generated and executed according to the sorted index sequence to obtain the scheduled information entries.

[0033] It is worth noting that this step establishes a global priority sequence by sorting the weighted comprehensive evaluation values ​​in descending order, and extracts the top high-priority items based on system resource constraints to form a candidate scheduling set, thereby generating, packaging and executing scheduling instructions, and finally generating scheduled information items as the data basis for subsequent feedback optimization.

[0034] The sorted index sequence is an ordered data structure formed by arranging elements in descending order based on weighted comprehensive evaluation values. Each element in the sequence contains a unique identifier for an information item and its corresponding evaluation value. Elements in the sequence are arranged from highest to lowest evaluation value. For example, a quicksort algorithm is used to sort the elements in descending order based on their evaluation values, with higher evaluation values ​​listed first and lower evaluation values ​​listed last. A sequential index is assigned to the elements in the sorted sequence, establishing a mapping relationship from evaluation values ​​to the unique identifiers of information items, thus forming the sorted index sequence.

[0035] In one implementation, a preliminary scheduling instruction is generated and executed based on the sorted index sequence to obtain scheduled information entries, including: Using a preset scheduling window capacity, the head index data is extracted from the sorted index sequence to obtain a candidate scheduling set; Based on the candidate scheduling set, instructions are generated to obtain preliminary scheduling instructions; The preliminary scheduling instructions are packaged to obtain a scheduling instruction package; The scheduling instruction packet is parsed, and the instructions in the scheduling instruction packet are executed one by one. The successfully sent information entries are marked to obtain the scheduled information entries.

[0036] It is worth noting that the scheduling window capacity is a pre-set upper limit on the number of information entries that the system is allowed to process simultaneously within a single scheduling cycle, based on historical load data and hardware resource configuration; the candidate scheduling set is a subset of the information entries to be executed formed after truncating the header index data. For example, when the candidate scheduling set is obtained by truncating the header index data, if the current scheduling window capacity is set to 3, the first 3 elements are truncated from the beginning of the sorted index sequence to obtain the corresponding unique identifiers of the information entries ID_FIN_2024022001, ID_SEC_2024022003, and ID_EXE_2024022002, which correspond to rules R3, R7, and R2, respectively, thus forming the candidate scheduling set.

[0037] In this embodiment, the preliminary scheduling instruction is a set of specific execution parameters generated for each information entry in the candidate scheduling set, including at least one of the following: information content identifier, target receiver identifier, push timing, transmission protocol, and priority flag. This instruction guides the scheduling instruction execution module to complete information delivery. For example, for ID_FIN_2024022001, the system queries the information content database to obtain the financial settlement notification text, queries the receiver list to obtain the identifiers of 120 employees in the finance department, queries the scheduling rules to obtain the execution parameters for immediate push + SMS reinforcement, and fills the instruction template to generate the preliminary scheduling instruction. This instruction includes the information content identifier MSG_001, the receiver list [UID_1001 to UID_1120], the push timing as immediate, the transmission protocol as APP push + SMS, and the priority flag as high.

[0038] Furthermore, the scheduling instruction packet is a structured data unit formed after instruction packaging, containing complete information on all instructions to be executed within a scheduling cycle; the scheduling instruction execution module is a functional unit responsible for parsing the scheduling instruction packet and implementing specific information delivery, including a message push engine, a channel manager, and a delivery status monitor, used to deliver the information content to the designated receiving end and provide feedback on the execution result; the scheduled information entry is the information entry successfully delivered by the scheduling instruction execution module.

[0039] For example, multiple instructions in the initial scheduling instruction set are arranged in a fixed order. The cyclic redundancy check (CRC) code of the overall data stream is calculated as the verification field. A protocol header containing the version number, timestamp, batch sequence number, and instruction quantity is added, and these are combined to form a scheduling instruction packet. The scheduling instruction execution module receives the scheduling instruction packet, parses the protocol header to obtain batch information, and extracts the initial scheduling instructions one by one. For each instruction, the information text is read according to the information content identifier, the delivery target is determined according to the receiver identifier, the delivery channel is selected according to the transmission protocol, and the information push operation is performed. The delivery process is monitored, and the sending status, confirmation receipt, and exception information are recorded. For successfully delivered information entries, they are marked as scheduled, and the execution timestamp, batch identifier, and association rule identifier are recorded to form a scheduled information entry record, which is stored in the scheduling log database.

[0040] In step S14, the user interaction feedback data and execution timestamp of the scheduled information entry are obtained, and the residual value is calculated based on the execution timestamp and a preset urgency decay function to obtain the current urgency residual value. Then, the utility value of the scheduled information entry is converted based on the current urgency residual value and the user interaction feedback data to obtain the actual utility value.

[0041] It should be noted that this step dynamically calculates the actual business value of scheduled information items by collecting user interaction feedback data and combining it with an urgency decay mechanism over time, generating an actual utility value as the optimization target for subsequent weight optimization algorithms. The user interaction feedback data consists of quantifiable behavioral records of users at the information receiving end regarding scheduled information items, including two core metrics: click-through rate and dwell time.

[0042] In one implementation, user interaction feedback data and execution timestamps of the scheduled information entries are obtained, and residual values ​​are calculated based on the execution timestamps and a preset urgency decay function to obtain the current urgency residual value, including: Obtain user interaction feedback data for the scheduled information entries, wherein the user interaction feedback data includes click-through rate and dwell time; Obtain the execution timestamp of the scheduled information entry, and calculate the time difference between the execution timestamp and the current system time to obtain the duration of the scheduled information entry; Calculate the real-time attenuation coefficient by combining the preset urgency attenuation function with the duration span; Obtain the initial urgency attribute value, and calculate the residual value by combining the initial urgency attribute value and the real-time decay coefficient to obtain the current urgency residual value.

[0043] It is worth noting that the click-through rate is the proportion of users who click on the information item to the total number of users who receive it; the dwell time is defined as the average time interval between when a user opens the information content and when they leave the content page; and the duration span is the length of the time interval between when a scheduled information item is pushed and when it is currently calculated, measured in hours or minutes, reflecting the natural length of time that the information has elapsed since it was delivered.

[0044] For example, the execution timestamp is the moment when the information entry scheduling push is completed. In the calculation of the duration span, the execution timestamp of the entry needs to be extracted from the scheduling log record as 1708401605, which is February 20, 2024, 08:15:05. The system time at the current utility calculation time is 1708405205, which is the same day, 09:15:05. The calculation time difference is 3600 seconds, which, when converted to hours, gives a duration span of 1.0 hour.

[0045] Specifically, the urgency decay function is a mathematical function that describes how the urgency attribute of information decreases over time. It maps the duration of information to a real-time decay coefficient between 0 and 1, used to quantify the proportion of information whose timeliness value is retained. In one example, the urgency decay function is based on an exponential decay model. in, This is the real-time attenuation coefficient; The duration of existence; The attenuation rate parameter is set by the business system based on the information type; for example, emergency notifications. =0.15 / hour, regular news =0.05 / hour. Substitute the duration into the decay function to calculate the real-time decay coefficient, the coefficient ranges from (0,1], and decreases as time increases.

[0046] Furthermore, the current urgency residual value is the remaining quantified urgency value after the initial urgency attribute value has been discounted by the real-time attenuation coefficient. The current urgency residual value can be obtained by multiplying the initial urgency attribute value of the association rule for this information entry in step S11 by the current real-time attenuation coefficient.

[0047] In one implementation, the utility value of the scheduled information entry is calculated based on the current urgency residual value and the user interaction feedback data to obtain the actual utility value, including: If the click-through rate is not less than a preset click-through rate threshold, a positive utility increment is generated by combining the dwell time with the current urgency residual value; if the click-through rate is less than the preset click-through rate threshold, a negative utility reduction is generated by combining the dwell time with the current urgency residual value. Obtain the basic evaluation value of the scheduled information entry, and update the basic evaluation value according to the positive utility increment and the negative utility decrement to obtain the actual utility value.

[0048] It's worth noting that the preset click-through rate (CTR) threshold is a critical percentage for determining user engagement, used to distinguish between high-response and low-response information. It's set based on the average CTR of similar information over historical periods, typically ranging from 5% to 15%. Positive utility increment is a positive adjustment to the utility value of high-CTR information items, used to enhance the perceived actual utility of high-value information; negative utility reduction is a negative adjustment to the utility value of low-CTR information items, used to reduce the perceived actual utility of low-value information. For example, when using a CTR threshold-based utility increment direction determination mechanism, a preset CTR threshold P is set. th Compare the actual click-through rate P with the threshold P th Compare, if P≥P th If P is positive, the user response is considered positive, and the program proceeds to the positive utility increment calculation branch; if P <P th If the user's response is negative, the calculation branch for negative utility reduction will be entered.

[0049] In this embodiment, the utility increment calculation process requires setting three parameters: a positive adjustment coefficient, a preset click-through rate threshold, and a negative adjustment coefficient. This establishes a numerical scaling mechanism between the predicted evaluation value and the actual utility value. The positive adjustment coefficient aims to establish a quantitative mapping relationship between prediction deviation and weight adjustment magnitude, amplifying the utility incentive of high-response information. Simultaneously, it ensures that the positive increment does not exceed the reasonable fluctuation limit of the basic evaluation value, which is set to 0.2 here. The negative adjustment coefficient is set to 0.3, slightly higher than the positive coefficient, reflecting the current business orientation of suppressing interference from low-value information, while ensuring that the absolute value of the maximum negative reduction does not exceed 0.6, preventing excessive penalties from causing weight optimization oscillations. The preset click-through rate threshold is typically set to 1.2 times the average click-through rate of similar information in the same historical period to distinguish the top 20% of high-response information.

[0050] For the positive utility increment branch, if the current urgency residual value is 6.93, and the baseline dwell time is set to 30 seconds, the dwell time multiplier k = 48 / 30 = 1.6 is calculated. With a positive adjustment coefficient β = 0.2, the positive utility increment is calculated to be 0.2 × 1.6 × 6.93 ≈ 2.22. For the negative utility increment branch, if the current urgency residual value is 3.35, and the click-through rate (CTR) of 6.5% is lower than the preset CTR threshold of 12%, the CTR deviation is calculated to be (12% - 6.5%) / 12% ≈ 0.458. With a negative adjustment coefficient γ = 0.3, the negative utility reduction is calculated to be -0.3 × 0.458 × 3.35 ≈ -0.46. The setting of the dwell time baseline value requires extracting dwell time data from high-CTR samples in historically scheduled information entries and calculating the arithmetic mean or median of these samples as the dwell time baseline value.

[0051] It's worth noting that the preset click-through rate (CTR) threshold is used to differentiate between information items with positive and negative user responses. Its setting is based on the average CTR statistics of similar information items from the same historical period, and configured in conjunction with business objectives. For example, CTR data for scheduled information items of the same type as the current information item can be extracted from the past 30 days, their arithmetic mean calculated, and 1.2 times this average can be used as the preset CTR threshold to ensure that the threshold reflects the average response level of the current user group. If historical data is insufficient, a fixed value such as 10% can be temporarily set and dynamically updated as sufficient data accumulates.

[0052] The dwell time benchmark is used to quantify the depth of user attention to high-response information. It is set based on the dwell time data of scheduled information items with a click-through rate (CTR) not lower than a preset CTR threshold. Specifically, all samples meeting the CTR criteria are selected from historically scheduled information items, and the arithmetic mean or median of the dwell time of these samples is calculated as the dwell time benchmark. This benchmark can be recalculated periodically with each business cycle to adapt to changes in user behavior.

[0053] Preset positive and negative adjustment coefficients are used to control the magnitude of utility increments and decrements, reflecting the intensity of incentives or inhibitions for information with different response levels. The positive adjustment coefficient must be set to ensure that the positive utility increment does not exceed a certain percentage of the base evaluation value. For example, the coefficient value can be determined through historical data simulation, ensuring that the actual utility value of most high-response information increases within a reasonable range; a common range is 0.1 to 0.3. The negative adjustment coefficient, on the other hand, needs to consider the penalty for low-response information. It is usually slightly higher than the positive coefficient to highlight the inhibitory effect, but should avoid causing drastic fluctuations in the actual utility value. A value between 0.2 and 0.5 can be selected through cross-validation. These coefficients can be preset based on experience during the initial stage of system operation and then fine-tuned by monitoring the stability of weight updates or business feedback.

[0054] All preset parameters can be dynamically adjusted according to changes in business needs. For example, historical data can be re-statistically analyzed weekly or monthly to update preset click-through rate thresholds and dwell time benchmarks, so that scheduling strategies can continuously adapt to the latest user behavior patterns.

[0055] In one example, the base evaluation value is the weighted comprehensive evaluation value of step S12 corresponding to the scheduled information item, which serves as the benchmark starting point for utility conversion. After positive increment or negative decrement adjustment, the actual utility value is obtained. The actual utility value comprehensively reflects the final quantitative result of the real business value of the scheduled information item at a specific time, driving the adaptive optimization of the weight coefficient.

[0056] For example, for a positive utility increment of 2.22, the weighted comprehensive evaluation value, i.e., the basic evaluation value of 0.851, is obtained from the record in step S12, and the actual utility value is calculated as 0.851 + 2.22 = 3.071. For a negative utility decrease of -0.46, the weighted comprehensive evaluation value, i.e., the basic evaluation value of 0.42, is obtained from the record in step S12, and the actual utility value is 0.42 - 0.46 = -0.04. After boundary truncation, the actual utility value is 0. It should be noted that an actual utility value of 0 indicates that the user response to this item is insufficient, and the timeliness value has significantly decreased after 2.5 hours, and it is comprehensively identified as a low-value scheduling.

[0057] In step S15, it is necessary to obtain the rule comprehensive evaluation value of the scheduled information entry, and calculate the prediction deviation based on the rule comprehensive evaluation value and the actual utility value, including: Obtain the feature attribute value vector and rule weight vector of the scheduled information entries; The feature attribute value vector and the rule weight vector are weighted and summed to obtain the comprehensive rule evaluation value; The prediction bias is obtained by calculating the difference between the comprehensive evaluation value and the actual utility value based on the rules.

[0058] It is worth noting that this step, by recalculating the rule-based comprehensive evaluation value of the scheduled information entries and performing a difference operation with the actual utility value of the same information entry, yields a prediction deviation value. The purpose is to establish a numerical difference metric between the prediction evaluation value generated in step S12 and the actual utility value generated in step S14, providing a clear optimization direction and objective for subsequent adaptive adjustment of weight coefficients. The difference operation uses a signed difference form, i.e., the actual utility value minus the rule-based comprehensive evaluation value. The sign of the difference indicates the direction of prediction deviation: positive indicates an underestimation, and negative indicates an overestimation. The absolute value indicates the magnitude of the deviation. The signed difference directly serves the gradient calculation of the loss function model in step S16, providing clear optimization guidance for the gradient descent method.

[0059] In one example, the feature attribute value vector is a multi-dimensional standardized feature vector describing the scheduling rules associated with the scheduled information entries. It originates from the standardized feature vector in step S12 and is stored in the standardized feature vector library. Step S15 retrieves the vector directly from the standardized feature vector library using the unique identifier of the rule associated with the scheduled information entries, without recalculation. The rule weight vector is a vector composed of the currently effective weight coefficients of each dimension. The method for obtaining the rule comprehensive evaluation value is the same as the weighted comprehensive evaluation value calculated for the information entries to be scheduled in step S12. The value reflects the prediction and determination of the priority of the information entries based on the weight configuration at that time before scheduling execution. However, unlike step S12, step S15 uses the weight coefficients that are most recently effective at the time of calculation. Step S15 and step S12 are in different execution stages of the same scheduling cycle. The weight coefficients used by both are the currently effective weight coefficients determined at the beginning of this cycle; if this cycle is the first run, they are the initial weight coefficients; if the previous cycle has been optimized, they are the iterative weight coefficients. The necessity of recalculation lies in the fact that even if the weighting coefficients remain unchanged, the execution of step S15 serves as a numerical verification, confirming that the calculation results of step S12 have not experienced data drift during storage and transmission.

[0060] Specifically, a scheduled information entry is represented by a vector of feature attribute values. Each notification corresponds to multiple feature dimensions, such as resource consumption and business urgency, as well as additional dimensions such as notification type importance score, sending department level weight, content keyword urgency tag, target audience matching degree, and historical average response tendency. These features together form the column vector of the matrix, while the rule weight vector contains the importance coefficient corresponding to each feature.

[0061] For example, when performing a weighted summation of a device maintenance reminder notification, i.e., a scheduled information item, its feature attribute value vector might be [0.75, 0.90, 0.65, 0.82, 0.70]; the rule weight vector is initially set to [0.30, 0.25, 0.20, 0.15, 0.10]. Multiplying the two values ​​one by one and then summing them, we get the comprehensive rule evaluation value as 0.75×0.30+0.90×0.25+0.65×0.20+0.82×0.15+0.70×0.10=0.773.

[0062] Furthermore, the prediction bias is the absolute value or signed difference between the rule's overall evaluation value and the actual utility value. The absolute value of the prediction bias is used to trigger a judgment, while the signed value is used to indicate the direction of optimization. For example, when calculating the difference, if a maintenance reminder is pushed 4 hours later, the click-through rate reaches 22%, the average dwell time is 51 seconds, and the actual utility value is calculated to be 0.841 after the urgency decays; the rule's overall evaluation value is 0.773; then the prediction bias is 0.841 - 0.773 = 0.068.

[0063] In step S16, if the prediction deviation is greater than a preset deviation threshold, the minimum prediction deviation is set as the target, a loss function model is constructed, and the initial weight coefficients are iteratively updated using the gradient descent method based on the loss function model to obtain the updated iterative weight coefficients.

[0064] It is worth noting that when the prediction deviation exceeds an acceptable range, the optimization calculation of the weight coefficients is automatically triggered, enabling the multi-attribute decision model of the scheduling system to continuously converge towards the actual user preferences. This step involves constructing a loss function model with the optimization objective of minimizing the prediction deviation, and using gradient descent to iteratively optimize the weight coefficients, ultimately obtaining the updated iterative weight coefficients.

[0065] The preset deviation threshold is a critical value used to determine whether the current weight configuration needs optimization, and it is set according to the acceptable prediction error range for the business. For example, if the prediction deviation of the currently scheduled information entry is 0.068, and the preset threshold is set to 0.05, then the deviation exceeds the threshold, triggering weight adjustment, i.e., constructing a loss function model to iteratively update the initial weight coefficients.

[0066] Furthermore, the loss function model is a mathematical function in the form of mean squared error, with weight coefficients as independent variables and prediction bias as the dependent variable. It quantifies the overall deviation between the predicted and actual values ​​under a given weight configuration, and its minimum point corresponds to the optimal weight configuration. In one example, when constructing the loss function model, the currently scheduled information items are selected as the training sample set. The currently scheduled information items are the set of the latest collected scheduled information items that the system has scheduled according to a fixed time period, such as 1 hour, through step S13 and after completing step S14 feedback data. This ensures the timeliness of the training samples and the stability of the batch size, avoiding optimization lag caused by overfitting of a single sample or long-term stagnation. The number of samples is denoted as N. For the i-th sample, the loss function is: in, For predicted evaluation values; This is the current weight coefficient vector; It is a vector of feature attribute values; This represents the actual utility value. The loss function is the weight vector. The quadratic convex function has a globally unique minimum point, which guarantees the convergence of the gradient descent method.

[0067] In one implementation, the initial weight coefficients are iteratively updated using gradient descent based on the loss function model to obtain updated iterative weight coefficients, including: Calculate the partial derivatives based on the loss function model to obtain the gradient vector of the loss function at the current iteration position; The weight iteration increment is determined by combining the preset gradient descent step size with the gradient vector of the loss function, and the current weight coefficient is corrected in reverse by using the weight iteration increment to obtain an intermediate set of weight coefficients. The weight iteration increment is converged. If the absolute value of the weight iteration increment is not less than a preset convergence threshold, the iterative update operation continues. If the absolute value of the weight iteration increment is less than the preset convergence threshold, the set of intermediate weight coefficients that finally converges is locked to obtain the iterative weight coefficients.

[0068] It's worth noting that the gradient vector of the loss function is a vector composed of the partial derivatives of the loss function with respect to the weight coefficients of each dimension. The direction of the vector indicates the direction of the fastest growth of the loss function, and its opposite direction is the direction of the fastest descent of the weight adjustment. The current weight coefficients are the weight vector values ​​used in this gradient descent iteration. For the first iteration update operation in step S16, the current weight coefficients are the preset initial weight coefficients. For subsequent iteration updates (i.e., this is the a-th iteration), the current weight coefficients are the set of intermediate weight coefficients generated in the (a-1)-th iteration. When calculating the partial derivatives, the partial derivatives of the loss function with respect to each component of the weight coefficient vector are taken. The partial derivative in the j-th dimension is... in, Let be the j-th dimension component of the feature vector of the i-th sample. Combining the partial derivatives of all dimensions yields the gradient vector of the loss function. The meaning of the j-th dimension of the gradient vector is that if this component is positive, then w increases. j This will lead to an increase in the loss function; if the component is negative, then w will increase. j This will lead to a reduction in the loss function.

[0069] Furthermore, the weight iteration increment is the weight adjustment vector obtained by multiplying the gradient descent step size by the gradient vector of the loss function. The gradient descent step size is the magnitude coefficient by which the weight coefficients are adjusted in the opposite direction of the gradient in a single iteration. The value range is typically from 0.01 to 0.1, set by the system based on historical convergence speed experience. The intermediate weight coefficient set is a temporary weight configuration obtained after a single iteration update. This set is continuously updated during the iteration process until the convergence criterion is met, at which point it is determined as the final output. When performing reverse correction on the current weight coefficients, the weight iteration increment is calculated. Then update the weight coefficient vector. The update operation shifts the new weight configuration along the descent direction of the loss function, thus... As an intermediate set of weight coefficients, it enters the next iteration until the weight iteration increment is determined to converge.

[0070] Specifically, convergence determination is the operation to check whether the iterative process has reached the termination condition. The preset convergence threshold is a critical value for determining whether the iterative process has converged, typically a small positive value, such as 0.001 or 0.0001. The updated iterative weight coefficients are the final weight configuration obtained after iterative optimization using the gradient descent method, replacing the initial weight coefficients and applied to subsequent scheduling cycles to achieve dynamic updates of the scheduling strategy. For example, the Euclidean norm of the weight iteration increment is calculated. The criterion is used as a convergence criterion and compared with a preset convergence threshold. If the criterion is greater than or equal to the convergence threshold, it indicates that the weight configuration is still changing significantly. The intermediate set of weight coefficients is used as the current weight vector, and the iteration returns to the gradient calculation step to continue. If the criterion is less than the convergence threshold, it indicates that the weight configuration has stabilized. The current intermediate set of weight coefficients is locked as the updated iterative weight coefficients, and the iteration process is terminated.

[0071] In step S17, a new round of scheduling instructions is generated by combining the iterative weight coefficients to obtain a new round of scheduling instructions for driving the execution of the business system, including: Obtain the unscheduled information entries from the previous steps to obtain the existing information entries; The recalculated comprehensive evaluation value of the existing information item is calculated using the iterative weight coefficients. Obtain newly added information entries and calculate the first comprehensive evaluation value of the newly added information entries using the iterative weight coefficients; A priority sequence is constructed based on the recalculated comprehensive evaluation value and the first comprehensive evaluation value, and instructions are generated based on the priority sequence to obtain a new round of scheduling instructions.

[0072] It is worth noting that this step applies the updated weight coefficients obtained from iterative optimization to the actual scheduling decision, driving the business system to execute a new round of information scheduling that has been adaptively optimized. By distinguishing between existing information items and new information items, the evaluation values ​​are recalculated and initially calculated using a unified iterative weight coefficient, a global priority sequence that integrates the two is constructed, and an executable new round of scheduling instructions is generated accordingly.

[0073] The iterative weight coefficients are the updated weight coefficients obtained after optimization using the gradient descent method in step S16. The preceding steps are the technical steps that have been completed before the current scheduling cycle, including parts of steps S11 to S13. Existing information entries are those that have entered the scheduling queue in the preceding steps but have not yet been scheduled; they are retained in the scheduling buffer pool awaiting the next round of priority evaluation. The recalculated comprehensive evaluation value is the evaluation value obtained by re-weighting the regular feature vectors of the existing information entries using the iterative weight coefficients. It is worth noting that the calculation process for the recalculated comprehensive evaluation value is consistent with the calculation logic of step S12, and for simplified calculation, the standardized feature vectors from step S12 can be directly used for linear weighting with the iterative weight coefficients.

[0074] Furthermore, newly added information entries are those newly generated and entering the scheduling system during this scheduling cycle, participating in the scheduling decision-making process for the first time. The first comprehensive evaluation value is the evaluation value obtained by first weighting the rule feature vector of the newly added information entry using iterative weight coefficients. Specifically, for each newly added information entry, the feature vector construction process in step S11 needs to be executed, which involves extracting rule attribute values, performing weighted corrections, constructing an initial feature vector, performing orthogonalization and dimensionality reduction processing, and generating the rule feature vector corresponding to the newly added information entry. This process ensures that newly added entries and existing entries are represented in the same feature space, possessing a comparable mathematical basis.

[0075] The priority sequence is an ordered data structure formed by merging the comprehensive evaluation values ​​of existing and newly added information items and arranging them in descending order, establishing a global relative priority relationship for all currently scheduled information. The new round of scheduling instructions is a set of instructions generated based on the priority sequence, used to drive the business system to perform information push operations. The construction of the priority sequence and instruction generation are logically consistent with step S13, using a quicksort algorithm to sort the global evaluation value set in descending order, with higher evaluation values ​​placed first, resulting in a sorted index sequence, i.e., the priority sequence. Elements are selected sequentially from the starting position of the priority sequence, with the number of selections equal to the scheduling window capacity, forming the candidate scheduling set for this round.

[0076] For each information entry in the candidate scheduling set, a preliminary scheduling instruction is generated by filling in the instruction template based on its information content identifier, receiver identifier, and association rule parameters. Multiple preliminary scheduling instructions are packaged together, with protocol headers, checksums, and batch identifiers added to form a scheduling instruction package. This package is then sent to the scheduling instruction execution module, driving the business system to complete the information push operation. Simultaneously, a record of the scheduled information entries is generated and stored in the scheduling log database, serving as feedback data input for the next round of steps S14 to S16.

[0077] In summary, this invention discloses a dynamic information scheduling method based on a rule engine. By constructing a multi-dimensional feature vector space and a dynamic weighted evaluation mechanism, and based on the vector space model and multi-attribute decision theory, the multi-dimensional attributes (benefit, urgency, etc.) of scheduling rules are transformed into computable feature vectors. Configurable weight coefficients are used for linear weighted fusion, and attribute redundancy is eliminated through orthogonalization and dimensionality reduction. This achieves quantitative trade-offs and unified ranking of multi-objective conflicts, overcoming the rigidity of decision-making with fixed rules or single objectives. A closed-loop optimization mechanism of prediction, execution, feedback, and correction is established. User interaction feedback (click rate, dwell time) is collected to calculate the actual utility value, which is compared with the predicted evaluation value to obtain the deviation. The weight coefficients are iteratively updated based on gradient descent, enabling the scheduling strategy to continuously adapt to users' true preferences and preventing the system from falling into a static suboptimal state. Through a unified recalculation of existing and incremental information and a global priority fusion mechanism, based on feedback control closed-loop system design and rolling optimization theory, the optimized weight coefficients are used to recalculate the backlog of information to be scheduled and prioritized on the same scale as newly added information. Ensure that optimization results are promptly covered by historical and real-time information, thereby achieving the technical effect of maintaining the continuity of scheduling decisions and the ability to respond in a timely manner.

[0078] Reference Figure 2 The second embodiment of the present invention provides a dynamic information scheduling system based on a rule engine, comprising: The rule feature extraction module is used to obtain the set of scheduling rules corresponding to the information entries to be scheduled, extract the rule attribute values ​​in the set of scheduling rules, construct a multi-dimensional feature vector based on the rule attribute values, and obtain the rule feature vector. The comprehensive evaluation calculation module is used to linearly weight each dimension of the rule feature vector using preset initial weight coefficients to obtain the weighted comprehensive evaluation value of each rule; The scoring instruction execution module is used to sort the information items according to the weighted comprehensive evaluation value, obtain a sorted index sequence, generate a preliminary scheduling instruction according to the sorted index sequence and execute it to obtain the scheduled information items. The feedback utility calculation module is used to obtain the user interaction feedback data and execution timestamp of the scheduled information item, and calculate the residual value based on the execution timestamp and a preset urgency decay function to obtain the current urgency residual value. Then, the module performs utility conversion on the scheduled information item based on the current urgency residual value and the user interaction feedback data to obtain the actual utility value. The prediction deviation calculation module is used to obtain the rule comprehensive evaluation value of the scheduled information entries, and to calculate the deviation based on the rule comprehensive evaluation value and the actual utility value to obtain the prediction deviation. The deviation weight optimization module is used to set the minimum prediction deviation as the target if the prediction deviation is greater than the preset deviation threshold, construct a loss function model, and use the gradient descent method to iteratively update the initial weight coefficients based on the loss function model to obtain the updated iterative weight coefficients. The secondary instruction generation module is used to generate a new round of scheduling instructions by combining the iterative weight coefficients, thereby obtaining a new round of scheduling instructions to drive the execution of the business system.

[0079] It should be noted that the rule engine-based dynamic information scheduling system provided in this embodiment of the invention is used to execute all the process steps of the rule engine-based dynamic information scheduling method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0080] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0081] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A dynamic information scheduling method based on a rule engine, characterized in that, include: Obtain the set of scheduling rules corresponding to the information entries to be scheduled, extract the rule attribute values ​​from the set of scheduling rules, construct a multi-dimensional feature vector based on the rule attribute values, and obtain the rule feature vector; A weighted comprehensive evaluation value for each rule is obtained by linearly weighting each dimension of the rule feature vector using a preset initial weight coefficient. Based on the weighted comprehensive evaluation value, the information items are sorted to obtain a sorted index sequence. A preliminary scheduling instruction is generated and executed based on the sorted index sequence to obtain the scheduled information items. The user interaction feedback data and execution timestamp of the scheduled information entry are obtained. Based on the execution timestamp and a preset urgency decay function, the residual value is calculated to obtain the current urgency residual value. Then, the utility value of the scheduled information entry is converted based on the current urgency residual value and the user interaction feedback data to obtain the actual utility value. Obtain the rule-based comprehensive evaluation value of the scheduled information entry, and calculate the deviation based on the rule-based comprehensive evaluation value and the actual utility value to obtain the prediction deviation; If the prediction deviation is greater than the preset deviation threshold, then the minimum prediction deviation is set as the target, a loss function model is constructed, and the initial weight coefficients are iteratively updated using the gradient descent method based on the loss function model to obtain the updated iterative weight coefficients. A new round of scheduling instructions is generated by combining the iterative weight coefficients, resulting in a new round of scheduling instructions used to drive the execution of the business system.

2. The dynamic information scheduling method based on a rule engine according to claim 1, characterized in that, The step of extracting rule attribute values ​​from the scheduling rule set and constructing a multi-dimensional feature vector based on the rule attribute values ​​to obtain the rule feature vector includes: Obtain resource consumption data, and perform validity filtering based on the resource consumption data and the scheduling rule set to obtain a valid scheduling rule set; Extract the initial benefit attribute value and initial urgency attribute value of each rule in the effective scheduling rule set, and perform weighted correction on the initial benefit attribute value and initial urgency attribute value according to historical scheduling frequency data to obtain the corrected benefit attribute value and corrected urgency attribute value. Obtain the category identifier and construct a multi-dimensional feature vector by combining the corrected benefit attribute value and the corrected urgency attribute value to obtain the initial feature vector; The initial feature vector is orthogonalized and dimensionality reduced to obtain the rule feature vector describing the attribute of each rule.

3. The dynamic information scheduling method based on a rule engine according to claim 1, characterized in that, The step of linearly weighting each dimension of the rule feature vector using preset initial weight coefficients to obtain a weighted comprehensive evaluation value for each rule includes: Construct a weight coefficient vector based on the preset initial weight coefficients; Obtain the feature vector to be calculated from the feature vector of the rule, which corresponds to the dimension of the weight coefficient vector; A linear transformation mapping is performed on the feature vector to be calculated to obtain a standardized feature vector; The weighted comprehensive evaluation value is obtained by performing a dot product operation between the standardized feature vector and the weight coefficient vector.

4. The dynamic information scheduling method based on a rule engine according to claim 1, characterized in that, The step of generating and executing preliminary scheduling instructions based on the sorted index sequence to obtain scheduled information entries includes: Using a preset scheduling window capacity, the head index data is extracted from the sorted index sequence to obtain a candidate scheduling set; Based on the candidate scheduling set, instructions are generated to obtain preliminary scheduling instructions; The preliminary scheduling instructions are packaged to obtain a scheduling instruction package; The scheduling instruction packet is parsed, and the instructions in the scheduling instruction packet are executed one by one. The successfully sent information entries are marked to obtain the scheduled information entries.

5. The dynamic information scheduling method based on a rule engine according to claim 1, characterized in that, The process of obtaining user interaction feedback data and execution timestamps of the scheduled information entries, and calculating the current urgency residual value based on the execution timestamps and a preset urgency decay function, includes: Obtain user interaction feedback data for the scheduled information entries, wherein the user interaction feedback data includes click-through rate and dwell time; Obtain the execution timestamp of the scheduled information entry, and calculate the time difference between the execution timestamp and the current system time to obtain the duration of the scheduled information entry; Calculate the real-time attenuation coefficient by combining the preset urgency attenuation function with the duration span; Obtain the initial urgency attribute value, and calculate the residual value by combining the initial urgency attribute value and the real-time decay coefficient to obtain the current urgency residual value.

6. The dynamic information scheduling method based on a rule engine according to claim 5, characterized in that, The step of performing utility conversion on the scheduled information entries based on the current urgency residual value and the user interaction feedback data to obtain the actual utility value includes: If the click-through rate is not less than a preset click-through rate threshold, a positive utility increment is generated by combining the dwell time with the current urgency residual value; if the click-through rate is less than the preset click-through rate threshold, a negative utility reduction is generated by combining the dwell time with the current urgency residual value. Obtain the basic evaluation value of the scheduled information entry, and update the basic evaluation value according to the positive utility increment and the negative utility decrement to obtain the actual utility value.

7. The dynamic information scheduling method based on a rule engine according to claim 1, characterized in that, The step of obtaining the rule-based comprehensive evaluation value of the scheduled information entries, and calculating the deviation based on the rule-based comprehensive evaluation value and the actual utility value to obtain the prediction deviation includes: Obtain the feature attribute value vector and rule weight vector of the scheduled information entries; The feature attribute value vector and the rule weight vector are weighted and summed to obtain the comprehensive rule evaluation value; The prediction bias is obtained by calculating the difference between the comprehensive evaluation value and the actual utility value based on the rules.

8. The dynamic information scheduling method based on a rule engine according to claim 1, characterized in that, The step of iteratively updating the initial weight coefficients using gradient descent based on the loss function model to obtain updated iterative weight coefficients includes: Calculate the partial derivatives based on the loss function model to obtain the gradient vector of the loss function at the current iteration position; The weight iteration increment is determined by combining the preset gradient descent step size with the gradient vector of the loss function, and the current weight coefficient is corrected in reverse by using the weight iteration increment to obtain an intermediate set of weight coefficients. The weight iteration increment is converged. If the absolute value of the weight iteration increment is not less than a preset convergence threshold, the iterative update operation continues. If the absolute value of the weight iteration increment is less than the preset convergence threshold, the set of intermediate weight coefficients that finally converges is locked to obtain the iterative weight coefficients.

9. The dynamic information scheduling method based on a rule engine according to claim 1, characterized in that, The step of generating a new round of scheduling instructions by combining the iterative weight coefficients to obtain a new round of scheduling instructions for driving the execution of the business system includes: Obtain the unscheduled information entries from the previous steps to obtain the existing information entries; The recalculated comprehensive evaluation value of the existing information item is calculated using the iterative weight coefficients. Obtain newly added information entries and calculate the first comprehensive evaluation value of the newly added information entries using the iterative weight coefficients; A priority sequence is constructed based on the recalculated comprehensive evaluation value and the first comprehensive evaluation value, and instructions are generated based on the priority sequence to obtain a new round of scheduling instructions.

10. A dynamic information scheduling system based on a rule engine, characterized in that, include: The rule feature extraction module is used to obtain the set of scheduling rules corresponding to the information entries to be scheduled, extract the rule attribute values ​​in the set of scheduling rules, construct a multi-dimensional feature vector based on the rule attribute values, and obtain the rule feature vector. The comprehensive evaluation calculation module is used to linearly weight each dimension of the rule feature vector using preset initial weight coefficients to obtain the weighted comprehensive evaluation value of each rule; The scoring instruction execution module is used to sort the information items according to the weighted comprehensive evaluation value, obtain a sorted index sequence, generate a preliminary scheduling instruction according to the sorted index sequence and execute it to obtain the scheduled information items. The feedback utility calculation module is used to obtain the user interaction feedback data and execution timestamp of the scheduled information item, and calculate the residual value based on the execution timestamp and a preset urgency decay function to obtain the current urgency residual value. Then, the module performs utility conversion on the scheduled information item based on the current urgency residual value and the user interaction feedback data to obtain the actual utility value. The prediction deviation calculation module is used to obtain the rule comprehensive evaluation value of the scheduled information entries, and to calculate the deviation based on the rule comprehensive evaluation value and the actual utility value to obtain the prediction deviation. The deviation weight optimization module is used to set the minimum prediction deviation as the target if the prediction deviation is greater than the preset deviation threshold, construct a loss function model, and use the gradient descent method to iteratively update the initial weight coefficients based on the loss function model to obtain the updated iterative weight coefficients. The secondary instruction generation module is used to generate a new round of scheduling instructions by combining the iterative weight coefficients, thereby obtaining a new round of scheduling instructions to drive the execution of the business system.

Images