Marketing scheduling optimization method and system based on AI algorithm

Abstract

The application provides a marketing scheduling optimization method and system based on an AI algorithm, relates to the technical field of marketing scheduling, and comprises the following steps: acquiring state data and resource occupation data of multi-channel marketing activities, calculating state fluctuation characteristics, and encoding and splicing to obtain activity characteristic representation; calculating competition relationship based on target audience overlap degree and solving activity evaluation vector; calculating trend sequence, constructing conflict propagation path based on competition relationship, and marking connectivity, combining path construction with time identifier and channel identifier as node attributes to construct scheduling dependency graph; traversing conflict nodes along the path to determine the conflict influence domain, aggregating the activity evaluation vector, and iteratively solving the resource allocation scheme in combination with the total resource constraint, calculating the execution priority according to the resource allocation scheme and the trend sequence, and generating a scheduling execution sequence to trigger resource allocation.

Description

Technical Field

[0001] This invention relates to the field of marketing scheduling technology, and in particular to a marketing scheduling optimization method and system based on AI algorithms. Background Technology

[0002] In current multi-channel marketing campaign scheduling and management practices, a common approach is to use scheduling methods based on static budget allocation or fixed rule engines. This typically involves pre-configuring fixed resource ratios for different channels based on historical data, empirical formulas, or manually set priorities. During campaign execution, operations staff primarily rely on periodically reviewing reports and manually adjusting resource allocation. To address the issue of overlapping target audiences across multiple channels, the conventional approach is to deduplicate the audience or establish independent budget pools to avoid internal competition.

[0003] Existing technologies cannot adjust resource allocation in a timely manner, easily leading to wasted funds or missed conversion opportunities. Furthermore, existing technologies lack a systematic analysis of the competitive relationships between multi-channel activities; simply deduplicating audiences cannot accurately quantify the intensity of conflict between activities, let alone identify the propagation paths of conflict arising from competition. When multiple activities simultaneously compete for the same target audience, internal competition and friction often occur, exacerbating resource conflicts and making it difficult for scheduling decisions to avoid or mitigate resource congestion within the conflict's impact domain.

[0004] At the scheduling and execution level, traditional methods usually pre-determine the hard priorities of each activity, but they cannot adaptively adjust when the total amount of resources fluctuates dynamically or the execution status of activities changes abruptly. They also lack consideration for trend sequences and can only achieve local optimization. They cannot iteratively solve the globally optimal resource allocation scheme under time alignment and resource constraints, resulting in limited overall marketing efficiency. Summary of the Invention

[0005] This invention provides a marketing scheduling optimization method and system based on AI algorithms, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of this invention provides a marketing scheduling optimization method based on AI algorithms, comprising: Acquire execution status data and resource usage data of multi-channel marketing activities, calculate the fluctuation amplitude corresponding to the execution status data to obtain status fluctuation features, encode and concatenate the status fluctuation features and the resource usage data to obtain activity feature representation, calculate the competitive relationship based on the target audience overlap between the multi-channel marketing activities, and combine the activity feature representation to obtain the activity evaluation vector; The gradient change between activity evaluation vectors at adjacent time points is calculated to obtain a trend sequence. Based on the competitive relationship, a conflict propagation path between channels is constructed and connectivity is marked. The time identifier corresponding to the trend sequence and the channel identifier corresponding to the activity feature representation are used as node attributes. The topology is constructed by combining the conflict propagation path and the corresponding connectivity mark to obtain a scheduling dependency graph. The conflict nodes in the scheduling dependency graph are traversed and the associated nodes are traced along the conflict propagation path to determine the conflict influence domain. The activity evaluation vector of each node in the conflict influence domain is aggregated and calculated, and the resource allocation scheme is obtained by iteratively solving the problem in combination with the preset total resource constraint. The resource allocation scheme and the trend sequence are time-series aligned and the execution priority is calculated. The scheduling execution sequence is generated according to the execution priority and the resource allocation action is triggered.

[0007] In one alternative implementation, The process involves acquiring execution status data and resource usage data for multi-channel marketing activities, calculating the fluctuation amplitude corresponding to the execution status data to obtain status fluctuation characteristics, and encoding and concatenating the status fluctuation characteristics and resource usage data to obtain activity feature representations, including: Data transmission channels are established with monitoring nodes of various channels through data acquisition interfaces to obtain execution status data and resource usage data. Conversion index sequences and engagement index sequences of marketing activities in different channels within multiple time slices are extracted from the execution status data. The conversion index sequences are sampled by sliding window and the range within the window is calculated to obtain the conversion fluctuation. The engagement index sequences are subjected to difference operation between adjacent time slices to obtain the engagement change. The conversion fluctuation is then normalized and spliced ​​together to obtain the status fluctuation characteristics. Extract the computational and transmission resources used by marketing activities in each channel from the resource usage data to construct a resource usage vector. Perform sparse encoding on the resource usage vector and identify the resource usage pattern. Based on the resource usage pattern, determine the resource type identifier and concatenate it with the resource usage vector to obtain extended resource features. The state fluctuation features and the extended resource features are arranged according to the channel dimension to obtain the fluctuation feature matrix and the resource feature matrix, and the mutual information is calculated for each pair to obtain the correlation matrix. Based on the correlation matrix, a dimension selection mask is generated, and the fluctuation feature matrix and the resource feature matrix are subjected to dimension filtering and concatenation to obtain the activity feature representation.

[0008] In one alternative implementation, Based on the overlap of target audiences among the multi-channel marketing campaigns, the competitive relationship is calculated, and combined with the campaign feature representation, the campaign evaluation vector is obtained, including: Obtain the target audience identifier set of each channel's marketing activities and perform an intersection operation to obtain a shared audience set. Calculate the ratio of the shared audience set to the target audience identifier set to obtain an overlap matrix. Perform asymmetric decomposition on the overlap matrix to extract the dominant competitive direction vector and combine it with the values ​​in the overlap matrix to construct a competitive intensity matrix to obtain the competitive relationship. Construct a directed competitive graph based on the competitive relationship. The activity features are represented and arranged into a feature matrix. The feature matrix is ​​projected onto the spectral embedding space of the directed competition graph and the geodesic distance from each node to the competition source node is calculated. Based on the geodesic distance, a competition position code is generated. Multi-hop message passing operation is performed on the feature matrix and the directed competition graph to obtain multi-order competition transmission features. The mutual information between different orders of the multi-order competition transmission features is calculated and an inter-order dependency matrix is ​​constructed. Based on the inter-order dependency matrix, the multi-order competition transmission features are adaptively aggregated to obtain competition-aware features. Extract the sum of incoming edge strength and the sum of outgoing edge strength from the directed competition graph and calculate the competition situation factor. Based on the competition situation factor and the competition position encoding, calculate the situation adjustment tensor and combine it with the competition perception feature and the activity feature representation to obtain the activity evaluation vector.

[0009] In one alternative implementation, Calculating the gradient change between activity evaluation vectors at adjacent time points to obtain a trend sequence, and constructing conflict propagation paths between channels based on the competitive relationship and performing connectivity labeling includes: Extract the activity evaluation vectors of adjacent time moments and calculate the vector difference to obtain the gradient change vector. Perform sliding accumulation on the gradient change vector within a time window and perform singular value decomposition on the accumulation result to extract the dominant singular vector. Based on the dominant singular vector, project and reconstruct the gradient change vector and extract the temporal change pattern of the projection coefficient to obtain the trend acceleration feature. Concatenate the gradient change vector with the trend acceleration feature to obtain the trend sequence. Extract the competition intensity matrix corresponding to the competition relationship and mark the active propagation edge. Construct a directed propagation graph based on the active propagation edge and calculate the in-degree difference of each node to identify the conflict source node and the conflict sink node. Starting from the conflict source node, perform a depth-first traversal of the directed propagation graph and record all paths to the conflict sink node to obtain a set of candidate conflict propagation paths. The candidate conflict propagation path set is filtered based on path length and competition intensity to obtain conflict propagation paths. The starting node and ending node corresponding to each conflict propagation path are extracted, and the temporal characteristics of the trend sequence on the corresponding activities of the starting node and the ending node are determined. Based on the temporal characteristics, connectivity labeling is performed to obtain connectivity labeling results.

[0010] In one alternative implementation, Using the time identifier corresponding to the trend sequence and the channel identifier corresponding to the activity feature representation as node attributes, and combining them with the conflict propagation path and the corresponding connectivity marker, a topology is constructed to obtain a scheduling dependency graph, including: Extract the time identifier corresponding to each time slice from the trend sequence, extract the channel identifier corresponding to each marketing activity from the activity feature representation and perform a Cartesian product operation with the time identifier to obtain a candidate node set, filter the nodes in the candidate node set that exist in the conflict propagation path based on the connectivity labeling result and label the node type to obtain a valid node set, extract the trend sequence and the activity feature representation corresponding to each node in the valid node set and concatenate them to obtain a node attribute vector; Extract the nodes contained in each conflict propagation path and construct a set of directed edges based on the relationship between adjacent nodes. Extract the competitive intensity between marketing activities corresponding to adjacent nodes from the competitive relationship. Calculate the weight of each edge in the set of directed edges based on the competitive intensity and assign values ​​in a hierarchical manner to obtain edge attributes with different dependency levels. The effective node set and the node attribute vector are used as the node set and node attributes of the topology graph, and the directed edge set and the edge attributes are used as the edge set and edge attributes of the topology graph. The topology graph is sorted and the topology number of each node is marked. The node priority is calculated based on the topology number and marked in the node attribute to obtain the scheduling dependency graph.

[0011] In one alternative implementation, Traversing the scheduling dependency graph to identify conflicting nodes and tracing related nodes along the conflict propagation path to determine the conflict influence domain, the activity evaluation vector of each node in the conflict influence domain is aggregated and calculated, and combined with a preset total resource constraint, an iterative solution is obtained to obtain a resource allocation scheme, including: Traverse each node in the scheduling dependency graph and extract edge attributes. Based on the edge attributes, sum the dependency levels corresponding to the incoming and outgoing edges of each node and calculate the imbalance degree. Filter based on the imbalance degree to obtain a set of conflicting nodes. Extract the propagation path segment containing any node in the set of conflicting nodes from the conflict propagation path and trace bidirectionally along the propagation path segment to obtain a set of associated nodes. Merge the set of conflicting nodes and the set of associated nodes to obtain the conflict influence domain. Extract the activity evaluation vectors corresponding to each node in the conflict influence domain and calculate the influence domain aggregate vector. Calculate the discreteness of the dimensional components in the influence domain aggregate vector and construct a discreteness vector. Based on the discreteness vector, identify high-discreteness dimensions as resource competition dimensions and normalize the activity evaluation vector components corresponding to the resource competition dimensions to obtain the standardized competition intensity. Using the standardized competition intensity as the initial resource allocation coefficient, the deviation between the total resource consumption corresponding to the resource allocation coefficient and the preset total resource constraint is calculated. The resource allocation coefficient is corrected based on the deviation, and the correction is repeated until the deviation meets the preset convergence condition. The resource allocation coefficient that meets the convergence condition is mapped to each node in the conflict influence domain to obtain the resource allocation scheme.

[0012] In one alternative implementation, Based on the resource allocation scheme and the trend sequence, time-series alignment is performed and execution priorities are calculated. Then, a scheduling execution sequence is generated according to the execution priorities, and resource allocation actions are triggered, including: Extract the resource allocation coefficients corresponding to each node in the conflict influence domain from the resource allocation scheme, extract the time identifiers corresponding to each node from the scheduling dependency graph, extract the trend time points and trend change rates corresponding to each time slice from the trend sequence, perform time axis projection on the time identifiers and the trend time points to determine the time offset, and correct the time identifiers to obtain aligned time identifiers, calculate the time proximity between the aligned time identifiers and the trend time points, and combine the trend change rate to determine the time sequence matching degree; An initial priority score is calculated based on the resource allocation coefficient and the timing matching degree. The sum of the inbound edge dependency levels of each node is extracted from the scheduling dependency graph as the dependency weight. A weighted priority score is calculated based on the initial priority score and the dependency weight. The execution priority is obtained by normalizing the weighted priority score. According to the execution priority, the channel identifier and alignment time identifier corresponding to each node in the conflict influence domain are combined to obtain the scheduling execution sequence. The scheduling execution sequence is traversed and the corresponding resource allocation coefficients are extracted. The resource type identifier corresponding to the resource occupancy data is extracted from the activity feature representation. Based on the resource allocation coefficient and the resource type identifier, a resource allocation instruction is generated and the resource allocation action is triggered.

[0013] A second aspect of this invention provides a marketing scheduling optimization system based on AI algorithms, comprising: The activity evaluation unit obtains execution status data and resource usage data of multi-channel marketing activities, calculates the fluctuation amplitude corresponding to the execution status data to obtain status fluctuation features, encodes and concatenates the status fluctuation features and the resource usage data to obtain activity feature representation, calculates the competitive relationship based on the target audience overlap between the multi-channel marketing activities, and combines the activity feature representation to solve for the activity evaluation vector. The scheduling dependency unit is used to calculate the gradient change between activity evaluation vectors at adjacent times to obtain a trend sequence, construct conflict propagation paths between channels based on the competition relationship and mark the connectivity, take the time identifier corresponding to the trend sequence and the channel identifier corresponding to the activity feature representation as node attributes, and construct a topology structure by combining the conflict propagation path and the corresponding connectivity mark to obtain a scheduling dependency graph. The scheduling execution unit is used to traverse the conflict nodes in the scheduling dependency graph and trace the associated nodes along the conflict propagation path to determine the conflict influence domain. It aggregates and calculates the activity evaluation vector of each node in the conflict influence domain and iteratively solves the resource allocation scheme in combination with the preset total resource constraint. Based on the resource allocation scheme and the trend sequence, it performs time-series alignment and calculates the execution priority. Based on the execution priority, it generates a scheduling execution sequence and triggers the resource allocation action.

[0014] A third aspect of the present invention provides an electronic device, comprising: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] In this invention, state fluctuation characteristics are obtained by calculating the fluctuation amplitude of execution state data, and combined with resource occupancy data encoding and splicing to form activity feature representation, thereby achieving accurate quantification of multi-dimensional features of marketing activities. Based on the joint solution of competitive relationships calculated by target audience overlap and activity feature representation, activity evaluation vectors are generated, providing accurate activity value measurement basis for subsequent scheduling decisions. By constructing a trend sequence through the gradient change of activity evaluation vectors at adjacent time points, combined with the conflict propagation path and connectivity markers generated by competitive relationships, a scheduling dependency graph is constructed, dynamically displaying the temporal evolution and channel conflict topology between marketing activities. The conflict propagation path can trace the dependency relationship and influence range between conflict nodes, and the connectivity markers clearly identify the transmission association of conflicts between channels, enabling the scheduling system to grasp the global conflict situation in real time. This provides structured support for subsequent conflict influence domain identification and resource allocation, improving the interpretability and controllability of multi-activity collaborative scheduling, and avoiding scheduling decision deviations caused by information silos. By traversing the conflict nodes in the scheduling dependency graph and tracing the associated nodes along the conflict propagation path, the conflict influence domain is determined. By aggregating and calculating the activity evaluation vectors within the influence domain, combined with the preset total resource constraints, the resource allocation scheme is iteratively solved to achieve the optimal allocation of global resources. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the marketing scheduling optimization method based on AI algorithms according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the conflict node resource allocation optimization process of the marketing scheduling optimization method based on AI algorithms, as described in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] Figure 1 This is a flowchart illustrating the marketing scheduling optimization method based on AI algorithms according to an embodiment of the present invention. Figure 1 As shown, the method includes: Acquire execution status data and resource usage data of multi-channel marketing activities, calculate the fluctuation amplitude corresponding to the execution status data to obtain status fluctuation features, encode and concatenate the status fluctuation features and the resource usage data to obtain activity feature representation, calculate the competitive relationship based on the target audience overlap between the multi-channel marketing activities, and combine the activity feature representation to obtain the activity evaluation vector; The gradient change between activity evaluation vectors at adjacent time points is calculated to obtain a trend sequence. Based on the competitive relationship, a conflict propagation path between channels is constructed and connectivity is marked. The time identifier corresponding to the trend sequence and the channel identifier corresponding to the activity feature representation are used as node attributes. The topology is constructed by combining the conflict propagation path and the corresponding connectivity mark to obtain a scheduling dependency graph. The conflict nodes in the scheduling dependency graph are traversed and the associated nodes are traced along the conflict propagation path to determine the conflict influence domain. The activity evaluation vector of each node in the conflict influence domain is aggregated and calculated, and the resource allocation scheme is obtained by iteratively solving the problem in combination with the preset total resource constraint. The resource allocation scheme and the trend sequence are time-series aligned and the execution priority is calculated. The scheduling execution sequence is generated according to the execution priority and the resource allocation action is triggered.

[0021] In one alternative implementation, The process involves acquiring execution status data and resource usage data for multi-channel marketing activities, calculating the fluctuation amplitude corresponding to the execution status data to obtain status fluctuation characteristics, and encoding and concatenating the status fluctuation characteristics and resource usage data to obtain activity feature representations, including: Data transmission channels are established with monitoring nodes of various channels through data acquisition interfaces to obtain execution status data and resource usage data. Conversion index sequences and engagement index sequences of marketing activities in different channels within multiple time slices are extracted from the execution status data. The conversion index sequences are sampled by sliding window and the range within the window is calculated to obtain the conversion fluctuation. The engagement index sequences are subjected to difference operation between adjacent time slices to obtain the engagement change. The conversion fluctuation is then normalized and spliced ​​together to obtain the status fluctuation characteristics. Extract the computational and transmission resources used by marketing activities in each channel from the resource usage data to construct a resource usage vector. Perform sparse encoding on the resource usage vector and identify the resource usage pattern. Based on the resource usage pattern, determine the resource type identifier and concatenate it with the resource usage vector to obtain extended resource features. The state fluctuation features and the extended resource features are arranged according to the channel dimension to obtain the fluctuation feature matrix and the resource feature matrix, and the mutual information is calculated for each pair to obtain the correlation matrix. Based on the correlation matrix, a dimension selection mask is generated, and the fluctuation feature matrix and the resource feature matrix are subjected to dimension filtering and concatenation to obtain the activity feature representation.

[0022] During real-time monitoring of multi-channel marketing campaigns, a stable data transmission channel is established with monitoring nodes across various marketing channels through pre-configured data acquisition interfaces. Monitoring nodes are deployed in the data aggregation layer across different channels, responsible for collecting various status data and resource consumption information generated during campaign execution in real time. The data transmission channel employs a polling mechanism, sending data requests to each monitoring node at fixed time intervals to obtain the latest execution status data and resource usage data. Execution status data includes metrics reflecting campaign effectiveness such as user click-through rate, conversion rate, page dwell time, and interaction counts, while resource usage data covers metrics reflecting system resource consumption such as server CPU utilization, memory usage, network bandwidth consumption, and database query counts.

[0023] The acquired execution status data is divided into multiple consecutive time slices, with each slice's duration adjustable to 5 minutes, 10 minutes, or other granularities depending on business needs. For each marketing channel, conversion metric sequences and engagement metric sequences are extracted from its execution status data. The conversion metric sequence records the channel's conversion rate within each time slice, while the engagement metric sequence records user activity levels. To capture the fluctuation characteristics of conversion metrics, a fixed-length sliding window moves across the conversion metric sequence, with the window length potentially set to 5 time slices. At each window position, the difference between the maximum and minimum conversion rate within the window is calculated; this difference represents the conversion fluctuation, reflecting the degree of fluctuation in the conversion metric during that time period. The sliding window moves sequentially in steps of single time slices until the entire conversion metric sequence is traversed, ultimately yielding a sequence of conversion fluctuation values.

[0024] For the participation index sequence, the change in participation is calculated using a difference operation between adjacent time slices. The participation value of the current time slice is subtracted from the participation value of the previous time slice to obtain the change in participation at that moment. A positive value indicates an increase in participation, while a negative value indicates a decrease. The absolute value reflects the degree of change. After performing the difference operation on all adjacent time slices, a complete sequence of changes in participation is obtained.

[0025] After obtaining the conversion fluctuation sequence and the engagement change sequence, both are normalized to map their numerical ranges to between 0 and 1, eliminating the influence of dimensional differences. Normalization employs a maximum-minimum standardization method, subtracting the minimum value from each value in the sequence and then dividing by the sequence range. The normalized conversion fluctuation sequence and engagement change sequence are then concatenated element-wise in chronological order to form a comprehensive state fluctuation feature vector. This vector simultaneously contains fluctuation information of conversion indicators and trends in engagement, comprehensively depicting the dynamic characteristics of the marketing campaign's execution state.

[0026] Resource usage data is processed by extracting the computational and transmission resources used by each marketing channel. Computational resources include the number of CPU cores, bytes of memory usage, and bytes of storage space usage, while transmission resources include uplink bandwidth usage, downlink bandwidth usage, and the number of data packets transmitted. The computational and transmission resource values ​​for a specific channel are combined to form a resource usage vector. Since resource usage data is often high-dimensional and sparse, the resource usage vector is sparsely encoded. Sparse encoding uses a threshold filtering method, setting elements in the resource usage vector with values ​​less than a preset threshold to zero, preserving significant resource usage characteristics, thereby reducing data dimensionality and highlighting key resource consumption patterns.

[0027] By performing cluster analysis or pattern matching on the sparsified resource usage vector, the resource usage pattern of the current channel is identified. The resource usage pattern reflects the typical characteristics of the channel's resource usage; for example, compute-intensive patterns mainly consume CPU resources, while transmission-intensive patterns mainly consume network bandwidth. Based on the identified resource usage pattern, a resource type identifier is assigned to the channel. This identifier can be a one-hot encoded vector used to distinguish different resource consumption types. The resource type identifier vector is then concatenated with the original resource usage vector element-wise to form an extended resource feature vector. This extended resource feature vector contains not only the specific numerical value of resource usage but also the type information of the resource consumption pattern.

[0028] Arrange the state fluctuation feature vectors of all channels in channel order to form a fluctuation feature matrix, where each row of the matrix corresponds to the state fluctuation feature of a channel. Similarly, arrange the extended resource feature vectors of all channels in the same channel order to form a resource feature matrix. To evaluate the correlation between state fluctuation features and resource occupancy features, calculate the mutual information for each pair of corresponding rows in the fluctuation feature matrix and resource feature matrix. Mutual information measures the statistical dependence between two variables; a higher value indicates a stronger correlation. For example, for the ... Each channel has a fluctuation characteristic vector denoted as... The resource feature vector is denoted as Calculate the mutual information between the two. Organize the mutual information values ​​of all channels into a correlation matrix. The diagonal elements of this matrix are the mutual information values ​​between the state fluctuations and resource consumption of each channel.

[0029] A dimension selection mask is generated based on the correlation matrix to filter out dimensions that contribute significantly to the representation of activity features. The generation rule for the dimension selection mask is as follows: for each column of the fluctuation feature matrix and each column of the resource feature matrix, the correlation between that column and its corresponding correlation degree is calculated. Columns with a correlation higher than a preset threshold are marked as retained dimensions, and columns with a correlation lower than the threshold are marked as discarded dimensions. The dimension selection mask is a binary vector, with 1 for retained dimensions and 0 for discarded dimensions.

[0030] Dimensionality selection masks are used to filter the dimensions of the fluctuation feature matrix and resource feature matrix, retaining only columns with a mask value of 1 and removing columns with a mask value of 0, resulting in dimensionality-reduced fluctuation feature matrix and resource feature matrix. These dimensionality-reduced fluctuation feature matrix and resource feature matrix are then concatenated column-wise, that is, the fluctuation feature vector and resource feature vector of each channel are joined end-to-end to form a longer comprehensive feature vector. This is the activity feature representation.

[0031] In one alternative implementation, Based on the overlap of target audiences among the multi-channel marketing campaigns, the competitive relationship is calculated, and combined with the campaign feature representation, the campaign evaluation vector is obtained, including: Obtain the target audience identifier set of each channel's marketing activities and perform an intersection operation to obtain a shared audience set. Calculate the ratio of the shared audience set to the target audience identifier set to obtain an overlap matrix. Perform asymmetric decomposition on the overlap matrix to extract the dominant competitive direction vector and combine it with the values ​​in the overlap matrix to construct a competitive intensity matrix to obtain the competitive relationship. Construct a directed competitive graph based on the competitive relationship. The activity features are represented and arranged into a feature matrix. The feature matrix is ​​projected onto the spectral embedding space of the directed competition graph and the geodesic distance from each node to the competition source node is calculated. Based on the geodesic distance, a competition position code is generated. Multi-hop message passing operation is performed on the feature matrix and the directed competition graph to obtain multi-order competition transmission features. The mutual information between different orders of the multi-order competition transmission features is calculated and an inter-order dependency matrix is ​​constructed. Based on the inter-order dependency matrix, the multi-order competition transmission features are adaptively aggregated to obtain competition-aware features. Extract the sum of incoming edge strength and the sum of outgoing edge strength from the directed competition graph and calculate the competition situation factor. Based on the competition situation factor and the competition position encoding, calculate the situation adjustment tensor and combine it with the competition perception feature and the activity feature representation to obtain the activity evaluation vector.

[0032] Obtain the target audience identifier set for each marketing channel, represented by user ID, unique device identifier, or customer number, etc. For any two channels... and Users simultaneously targeted by two channels are extracted through set intersection operations, forming a shared audience set. A shared audience set reflects the level of competition among different marketing campaigns in terms of audience reach. Competition for attention resources arises when two channels attempt to push marketing content to the same users.

[0033] Calculation Channel target audience set With shared audience set The proportion obtained from channels For channels overlap ,in This indicates the number of elements in the set. Iterate through all channel pairs and construct the overlap matrix. Matrix elements Indicates channel For channels The degree of audience reach dependence. This matrix typically exhibits asymmetric characteristics because of the channel... For channels Overlap ratio and channels For channels The differences in the overlap ratio reflect the directionality of the competitive relationship.

[0034] Perform nonnegative matrix decomposition on the overlap matrix, and then... Decomposed into ,in and Let them represent the row basis matrix and the column basis matrix, respectively. To determine the rank of the decomposition. This is achieved through calculation. and Difference measurement between Identify the channels that dominate in a competitive relationship. When At that time, it indicates the channel. More as the exertor of competition; when At that time, it indicates the channel. More often, they act as the bearers of competition. They will use the difference vector... Normalization yields the dominant competition direction vector .

[0035] By combining the numerical values ​​in the overlap matrix and the dominant direction vector, a competition intensity matrix is ​​constructed. Its elements The exponential function amplifies the impact of differences in dominant direction, thus numerically enhancing the competitive effect of the dominant party. The competition intensity matrix clearly depicts the distribution of competitive pressure among channels; larger matrix element values ​​indicate higher channel strength. For channels The stronger the competitive pressure, the better. A directed competition graph is constructed based on the competition intensity matrix. , where the node set For each marketing channel, there is a directed edge The weights are set to The direction of the edge is from the party initiating the competition to the party receiving the competition.

[0036] The activity characteristics of each channel are arranged in channel order to form a feature matrix. ,in The feature dimension is used to calculate the Laplacian matrix of the directed competition graph. , Given a degree matrix, the eigenvalues ​​of the Laplacian matrix are obtained by performing eigenvalue decomposition. The eigenvectors corresponding to the smallest eigenvalues ​​constitute the basis of the spectral embedding space. Projecting the feature matrix onto the spectral embedding space yields... The projection preserves the geometric information of the competition relationships in the graph structure.

[0037] In a directed competition graph, the source node of the competition is defined as a node whose out-degree is much greater than its in-degree, satisfying the condition... ,in The preset threshold is used for each non-source node. Calculate the geodesic distance from the node to all competing source nodes, which is the shortest path length along the directed edges. Dijkstra's algorithm is used to calculate the geodesic distance. ,in This represents the set of competing source nodes. The geodesic distance is converted into a high-dimensional vector using a sinusoidal position coding function to obtain the competing position code. This encoding captures the relative position of a node in a contention propagation chain.

[0038] Execute multi-hop message passing operations, in the... When jumping, node Feature update to ,in For nodes The set of neighboring nodes, For the first Learnable transformation matrix for jumps, This is the activation function. Execute. By skipping message passing, we can obtain the characteristics of multi-order competitive propagation. Different levels of features capture the competitive influence of different propagation distances.

[0039] Calculate the first Rank and first Mutual information between first-order features Estimating through variational lower bound A contrastive learning approach is employed, treating features of different orders within the same node as positive sample pairs and features of different nodes as negative sample pairs. Mutual information is estimated using a neural network discriminator. All order pairs are traversed to construct the inter-order dependency matrix. ,element Indicates the first Rank and number The mutual information strength between features of order.

[0040] Adaptive aggregation of multi-order features based on inter-order dependency matrix, and calculation of aggregation weights. This weight reflects the first The strength of the dependency between first-order features and other first-order features. Aggregation yields competitive perceptual features. This feature integrates competitive information from different communication distances to fully depict the impact of marketing campaigns on the competitive network.

[0041] Extract each node from the directed competition graph. The sum of the strength of the incoming edges Sum of outgoing edge strength Calculate the competitive situation factor. The range of values ​​for this factor is: Positive values ​​indicate that a node is in a competitive advantageous position, while negative values ​​indicate that it is in a competitive disadvantageous position. The competitive position factor quantifies the relative position of a channel within the overall competitive landscape.

[0042] Based on the competitive situation factor and competitive position encoding, a situation adjustment tensor is constructed through outer product operation. ,in The scalar situation factors are extended to vector form. The situation conditioning tensor is ternarily fused with the competition-aware features and the original activity features to calculate the activity evaluation vector. ,in For nodes The original activity characteristics, These are the corresponding competition-aware features. It represents the Hadamah accumulation. and This is a learnable fusion matrix. The activity evaluation vector comprehensively considers the activity's own characteristics, the transmission influence in the competitive network, and its position in the competitive landscape, providing a comprehensive evaluation basis for subsequent scheduling decisions.

[0043] In one alternative implementation, Calculating the gradient change between activity evaluation vectors at adjacent time points to obtain a trend sequence, and constructing conflict propagation paths between channels based on the competitive relationship and performing connectivity labeling includes: Extract the activity evaluation vectors of adjacent time moments and calculate the vector difference to obtain the gradient change vector. Perform sliding accumulation on the gradient change vector within a time window and perform singular value decomposition on the accumulation result to extract the dominant singular vector. Based on the dominant singular vector, project and reconstruct the gradient change vector and extract the temporal change pattern of the projection coefficient to obtain the trend acceleration feature. Concatenate the gradient change vector with the trend acceleration feature to obtain the trend sequence. Extract the competition intensity matrix corresponding to the competition relationship and mark the active propagation edge. Construct a directed propagation graph based on the active propagation edge and calculate the in-degree difference of each node to identify the conflict source node and the conflict sink node. Starting from the conflict source node, perform a depth-first traversal of the directed propagation graph and record all paths to the conflict sink node to obtain a set of candidate conflict propagation paths. The candidate conflict propagation path set is filtered based on path length and competition intensity to obtain conflict propagation paths. The starting node and ending node corresponding to each conflict propagation path are extracted, and the temporal characteristics of the trend sequence on the corresponding activities of the starting node and the ending node are determined. Based on the temporal characteristics, connectivity labeling is performed to obtain connectivity labeling results.

[0044] The dynamic changes of the campaign evaluation vector over time can reflect key turning points and trend evolution patterns in the marketing campaign execution process. (Targeting specific moments...) and time The activity evaluation vector is used to obtain the gradient change vector by calculating the vector difference. The gradient change vector describes the instantaneous direction and intensity of change in the activity evaluation over consecutive time intervals. To capture the cumulative effect of trends over a longer time span, a time window is introduced. Perform a sliding accumulation operation on the gradient change vector to construct the gradient change matrix. The column vectors of the matrix correspond to the gradient change vectors at each time step within the window. For the gradient change matrix... Perform singular value decomposition, decomposing it into three parts: a left singular matrix, a singular value diagonal matrix, and a right singular matrix. Select the left singular vector corresponding to the largest singular value in the singular value diagonal matrix as the dominant singular vector. The dominant singular vector represents the main direction of gradient change within the window.

[0045] By using the gradient change vector at each time step Projected onto the dominant singular vector Above, calculate the projection coefficient. The projection coefficients quantify the magnitude of the gradient change component in the dominant direction at the current moment. For the projection coefficient sequence... Perform time series analysis and calculate the second-order difference of adjacent projection coefficients. The second-order difference describes the acceleration characteristics of the trend change. After normalizing the second-order difference sequence, the trend acceleration features are obtained. This feature can identify trend inflection points and acceleration or deceleration patterns. The original gradient change vector... Trend acceleration characteristics at corresponding times A concatenation operation is performed to create an enhanced trend representation. The enhanced trend representations from all time points are then arranged chronologically to form a complete trend sequence. Each trend vector It contains information on instantaneous changes and acceleration changes.

[0046] Competition relationships are represented by a competition intensity matrix. Quantitative description, matrix elements Indicates channel For channels The intensity of competitive influence. This is determined by setting an active propagation threshold. The competition intensity matrix is ​​binarized when... Time marker edge Active propagation edges are identified. A directed propagation graph is constructed based on the labeled active propagation edges. , where the node set Edge collection It consists of active propagation edges. The in-degree is calculated for each node in the directed propagation graph. and out-degree Calculate the in-degree difference When the difference between in and out degrees is significantly positive, that is... (in (A threshold for identifying source nodes) is used to mark a node as a conflict source node, indicating that the activity has a strong competitive impact on other activities. When the in-degree difference is significantly negative, i.e. (in (Identify a threshold for the sink node), and mark the node as a conflicting sink node, indicating that the activity is affected by the concentrated influence of multiple competing activities.

[0047] Starting from each conflict source node, perform a depth-first traversal of the directed propagation graph, recording the sequence of visited nodes and edges traversed during the traversal. When the traversal reaches the conflict sink node, add the current traversal path as a candidate conflict propagation path to the candidate set. Continue traversing until all reachable paths originating from the conflict source node have been fully explored, ultimately obtaining the set of candidate conflict propagation paths. This set contains all possible propagation paths from the conflict source to the conflict sink, but may include redundant or weak propagation paths. The path length is calculated for each path in the candidate set. and path cumulative competition intensity This cumulative intensity characterizes the overall impact of the conflict propagating along the path. A path effectiveness score is defined. ,in This is a path length penalty coefficient used to avoid propagation attenuation caused by excessively long paths. Candidate paths are sorted in descending order of effectiveness score, and paths with scores exceeding a set threshold are selected. The path serves as the final conflict propagation path.

[0048] For each conflict propagation path, extract its starting node (corresponding to a conflict source activity) and ending node (corresponding to a conflict confluence activity), and obtain the temporal characteristics of the marketing activities corresponding to the starting and ending nodes in the trend sequence. Specifically, extract the trend sequence segment of the activity at the starting node. Trend sequence fragments of terminating node activities By calculating the dynamic time warping distance between two time segments. The temporal alignment similarity of the trend change patterns of the starting and ending nodes is evaluated. When the dynamic temporal warping distance is less than the connectivity threshold... When the trend changes of the starting and ending nodes on the conflict propagation path are strongly correlated, the path is marked as "strongly connected". When the dynamic time warp distance is greater than the connectivity threshold but less than the weak connectivity threshold... When the dynamic time warp distance exceeds the weak connectivity threshold, the path is marked as "weakly connected." When the dynamic time warp distance exceeds the weak connectivity threshold, although there is competition between paths, the temporal correlation is weak, and it is marked as "indirectly connected." Through the above marking process, each conflict propagation path is assigned a connectivity label, forming a connectivity marking result. This marking result not only records the structural path of conflict propagation but also incorporates the correlation characteristics of temporal evolution, providing crucial topological and temporal composite information for the subsequent construction of the scheduling dependency graph and the identification of conflict influence domains.

[0049] During the labeling process, the impact of abnormal fluctuations in the trend sequence on connectivity determination must also be considered. Time-series features are smoothed using an exponentially weighted moving average. (in To eliminate short-term noise interference (using a smoothing coefficient), the stability of connectivity labels is ensured. Ultimately, all conflict propagation paths and their corresponding connectivity label results are integrated to form a complete knowledge graph describing the conflict propagation structure and temporal relationships among multi-channel marketing campaigns.

[0050] In one alternative implementation, Using the time identifier corresponding to the trend sequence and the channel identifier corresponding to the activity feature representation as node attributes, and combining them with the conflict propagation path and the corresponding connectivity marker, a topology is constructed to obtain a scheduling dependency graph, including: Extract the time identifier corresponding to each time slice from the trend sequence, extract the channel identifier corresponding to each marketing activity from the activity feature representation and perform a Cartesian product operation with the time identifier to obtain a candidate node set, filter the nodes in the candidate node set that exist in the conflict propagation path based on the connectivity labeling result and label the node type to obtain a valid node set, extract the trend sequence and the activity feature representation corresponding to each node in the valid node set and concatenate them to obtain a node attribute vector; Extract the nodes contained in each conflict propagation path and construct a set of directed edges based on the relationship between adjacent nodes. Extract the competitive intensity between marketing activities corresponding to adjacent nodes from the competitive relationship. Calculate the weight of each edge in the set of directed edges based on the competitive intensity and assign values ​​in a hierarchical manner to obtain edge attributes with different dependency levels. The effective node set and the node attribute vector are used as the node set and node attributes of the topology graph, and the directed edge set and the edge attributes are used as the edge set and edge attributes of the topology graph. The topology graph is sorted and the topology number of each node is marked. The node priority is calculated based on the topology number and marked in the node attribute to obtain the scheduling dependency graph.

[0051] After obtaining the trend sequence from the real-time data stream of multi-channel marketing scheduling, the trend sequence needs to be discretized into analytical units with clear time attributes. A sliding time window mechanism is used to slice the trend sequence, dividing the continuous time series into several time slices, each time slice corresponding to a time identifier. ,in This is a time slice index, with a value range of [value range missing]. Indicates the total number of time slices. The generation of time stamps follows timestamp rules, using a Unix timestamp and the time slice length. The combination of these elements ensures that each time slice has a unique and traceable time attribute.

[0052] The activity feature representation contains rich information about multi-channel marketing activities, with channel identifiers being the core attribute that distinguishes different marketing channels. The process involves iterating through the set of activity feature representations and extracting the channel identifier field from each activity feature vector. ,in For the active index, the value range is: This indicates the total number of activities participating in the scheduling. Channel identifiers typically use an enumeration coding method to map different types of marketing channels, such as SMS channels, email channels, push channels, and outbound call channels, to integer codes, establishing a mapping relationship between channel identifiers and activity entities.

[0053] Constructing the node space of the scheduling dependency graph requires considering both time and channel dimensions. This involves the extracted set of time identifiers. With channel identifier set Perform the Cartesian product operation to generate a set of candidate nodes. The result of the Cartesian product operation is This set contains all possible time-channel combinations, with a cardinality of . This Cartesian product construction method ensures the completeness of the scheduling decision space, providing sufficient candidate nodes for subsequent screening operations.

[0054] Conflict propagation path analysis reveals the transmission relationship of resource competition between marketing campaigns, while connectivity labeling quantifies the strength of this transmission relationship. Based on the connectivity labeling results, the candidate node set is filtered for effectiveness, traversing each candidate node. Check whether the node appears in any conflict propagation path. The specific determination process is as follows: Let the set of conflict propagation paths be... ,in The total number of paths, for candidate nodes If a path exists Make If a candidate node is found to be a valid node, then the selection process is complete, resulting in a set of valid nodes. .

[0055] Node type labeling is a crucial step in distinguishing the functional roles of nodes in a conflict propagation network. Based on the node's position in the conflict propagation path, effective nodes are divided into three categories: conflict source nodes, relay nodes, and conflict sink nodes. Conflict source nodes are the starting points in the conflict propagation path, typically corresponding to marketing campaigns with a sudden increase in resource demand or a significant increase in priority. Relay nodes are located in the middle of the path, responsible for transmitting competitive pressure. Conflict sink nodes are the endpoints of the path, typically corresponding to activities with limited resources or forced downgrades in execution priority. Node type labeling uses enumerated values. To represent, where For node indexing.

[0056] The construction of node attribute vectors requires integrating temporal evolution features and activity resource features. For each node in the valid node set... Its corresponding time identifier is Channel identifier is Extract the trend vector of the time slice corresponding to the node from the trend sequence. Extract the feature vector of the activity corresponding to the node from the activity feature representation. The node attribute vector is generated through concatenation, and the calculation formula is as follows: The semicolon indicates vector concatenation. The dimension of this attribute vector is... The last dimension is the numerical encoding of the node type.

[0057] The conflict propagation path contains the sequential dependencies between nodes, which are expressed by directed edges. Traverse the set of conflict propagation paths. For each path ,in For path The number of nodes included is used to extract adjacent node pairs in the path. For each pair of adjacent nodes in the path... ,in Construct a directed edge The direction of the edges represents the causal order of conflict propagation. After extracting all adjacent node pairs in all conflict propagation paths, a set of directed edges is obtained. .

[0058] Competition intensity quantifies the fierce competition for resources between marketing campaigns; this intensity value can be used as a weight attribute for directed edges. For directed edges... Extract source node Corresponding activity index and target node Corresponding activity index Query the competition intensity between these two activities from the competition matrix. This value is directly used as the initial weight of the edge. .

[0059] The introduction of dependency levels reflects the varying importance of edges at different locations along the conflict propagation path. A hierarchical assignment strategy based on path depth is adopted for directed edges. Its path The position index in is Define dependency level functions ,in The maximum number of dependency levels is preset. This is a rounding operation. The function assigns lower weights to edges at the beginning of the path and higher weights to edges closer to the end, reflecting the cumulative effect of conflict impact as it propagates. The final edge weight is adjusted based on the initial weight and dependency level, calculated using the following formula: ,in This is the level gain coefficient, used to adjust the strength of the influence of the dependency level on the edge weights.

[0060] The construction of the topology graph will include the set of effective nodes. As a node set Set of node attribute vectors As a node attribute, the set of directed edges As an edge set Set the edge weights As an edge attribute. Formalized as a topological graph. .

[0061] Topological sorting is used to identify the linear execution order of nodes in a directed acyclic graph, satisfying the direction constraints of all edges. Kahn's algorithm is used to perform topological sorting, first calculating the in-degree of each node. Add all nodes with an in-degree of zero to the initial queue. Iterate through the nodes in the queue, removing one node from the queue each time. Assign a topology number to it The sequence numbers are incremented according to the processing order. After processing all nodes... Then, traverse all its outgoing edges. , target node Decrement the in-degree by one. If the in-degree becomes zero after the decrease, then... Add to the queue. Repeat the above process until the queue is empty and all nodes are assigned a topology number.

[0062] Node priority is calculated based on a combination of topology index and node type. For nodes... Its basic priority is This value is normalized to Within a given range, smaller topology numbers correspond to higher priorities. Priority adjustments are made based on node type, with the priority boosting factor for conflict-generating nodes being [value missing]. The coefficient of the relay propagation node is The coefficient of the conflict sink node is , usually satisfy The final priority of a node is calculated as follows: The calculated priority value is appended to the node attribute vector, and the updated node attributes are: This completes the construction of the scheduling dependency graph, which is formally represented as follows: .

[0063] Figure 2 This is a flowchart illustrating the conflict node resource allocation optimization process of the marketing scheduling optimization method based on AI algorithms, as described in an embodiment of the present invention.

[0064] In one alternative implementation, Traversing the scheduling dependency graph to identify conflicting nodes and tracing related nodes along the conflict propagation path to determine the conflict influence domain, the activity evaluation vector of each node in the conflict influence domain is aggregated and calculated, and combined with a preset total resource constraint, an iterative solution is obtained to obtain a resource allocation scheme, including: Traverse each node in the scheduling dependency graph and extract edge attributes. Based on the edge attributes, sum the dependency levels corresponding to the incoming and outgoing edges of each node and calculate the imbalance degree. Filter based on the imbalance degree to obtain a set of conflicting nodes. Extract the propagation path segment containing any node in the set of conflicting nodes from the conflict propagation path and trace bidirectionally along the propagation path segment to obtain a set of associated nodes. Merge the set of conflicting nodes and the set of associated nodes to obtain the conflict influence domain. Extract the activity evaluation vectors corresponding to each node in the conflict influence domain and calculate the influence domain aggregate vector. Calculate the discreteness of the dimensional components in the influence domain aggregate vector and construct a discreteness vector. Based on the discreteness vector, identify high-discreteness dimensions as resource competition dimensions and normalize the activity evaluation vector components corresponding to the resource competition dimensions to obtain the standardized competition intensity. Using the standardized competition intensity as the initial resource allocation coefficient, the deviation between the total resource consumption corresponding to the resource allocation coefficient and the preset total resource constraint is calculated. The resource allocation coefficient is corrected based on the deviation, and the correction is repeated until the deviation meets the preset convergence condition. The resource allocation coefficient that meets the convergence condition is mapped to each node in the conflict influence domain to obtain the resource allocation scheme.

[0065] After obtaining the scheduling dependency graph and conflict propagation paths, it is necessary to accurately locate the scope of conflict impact and solve for the optimal allocation scheme under resource constraints. This involves traversing all nodes in the scheduling dependency graph and, for each node... Extract the attribute information of all its associated edges, where This indicates the node's number in the dependency graph. Each edge carries a dependency level identifier, reflecting the strength of resource competition or temporal dependencies between different marketing campaigns. For nodes... The sum of the dependency levels of all its incoming edges is denoted as . The sum of the dependency levels of all outgoing edges is denoted as . Calculate the dependency imbalance of this node. A large imbalance indicates that the node is in an asymmetric position in the resource scheduling network, and is often a focal point or propagation node for resource conflicts. An imbalance threshold should be set. , will satisfy All nodes collected constitute the conflict node set. .

[0066] Based on the set of conflict nodes, locate the specific segments affecting the propagation of conflict within the conflict propagation path. Traverse all identified conflict propagation paths, and for each path... Check if its contained node sequence intersects with the set of conflicting nodes. If the path Include Nodes in If a node is encountered, a bidirectional tracing operation is performed starting from that node and proceeding along the path. Forward tracing involves moving backward along the path to the starting point or encountering a marked boundary node, while backward tracing involves moving forward along the path to the ending point or encountering a boundary node. All nodes traversed during the tracing process constitute the set of associated nodes for that path segment. This bidirectional tracing operation is performed on all path segments containing conflicting nodes, and all associated nodes obtained are then aggregated to form the set of associated nodes. Merging the set of conflicting nodes with the set of associated nodes yields the complete conflict influence domain. This influence domain includes nodes that directly cause conflicts as well as nodes that are indirectly affected through dependencies, providing a clear optimization scope for subsequent resource allocation.

[0067] For each node in the conflict influence domain, extract its corresponding activity evaluation vector. Assume the node... The corresponding activity evaluation vector is ,in This represents the number of dimensions of the evaluation vector. (This applies to all vectors within the influence domain.) The activity evaluation vectors of each node are aggregated to obtain the influence domain aggregate vector. , its first Dimensional components ,in This serves as the dimension index for the evaluation vector. This aggregated vector reflects the overall evaluation characteristics of activities within the conflict's influence domain. Further, the dispersion of each dimension of the aggregated vector is calculated, using the standard deviation as a measure of dispersion. For the dimensions... Calculate the dispersion Construct a discrete vector .

[0068] Sort the discrete vectors in descending order and select the top-ranked ones. The dimensions as a set of resource competition dimensions ,in The number of competition dimensions is pre-defined. Highly discrete dimensions indicate significant differences in evaluation values ​​among different nodes in that dimension, often corresponding to resource types with large differences in resource demand or intense competition. For each resource competition dimension... Extract the activity evaluation vector components of all nodes in the conflict influence domain along this dimension, and then evaluate the nodes. The amount Normalization is performed. The max-min normalization method is used. Normalized components Represents a node In the dimension of resource competition The intensity of standardized competition.

[0069] The intensity of standardized competition is used as the initial coefficient for resource allocation. For nodes... In terms of competition The initial resource allocation coefficient on is denoted as Assume the system targets dimensions. The total resource budget is Calculate the total resource consumption for this dimension under the current allocation coefficient. ,in Represents a node In dimensions The unit resource baseline requirement. Define resource consumption deviation. .when This indicates that resources are over-allocated. This indicates that resources are not being fully utilized.

[0070] Resource allocation coefficients are adjusted based on deviations. A proportional adjustment strategy is adopted, if... Then the allocation coefficients for all nodes are reduced proportionally. .like Then, the allocation coefficient is increased proportionally according to the competition intensity of the nodes, and the incremental allocation factor is calculated. Update allocation coefficients After correction, the total resource consumption and deviation are recalculated, and the iteration number is recorded. Set the convergence condition as follows: ,in The preset tolerance threshold, or the maximum number of iterations, is reached. Repeat the deviation calculation and coefficient correction process until the deviation meets the convergence condition.

[0071] Once the allocation coefficients for all resource competition dimensions converge, the results for each node are obtained. In various competitive dimensions The final resource allocation coefficient on These coefficients are mapped back to the marketing activities corresponding to the nodes to generate specific resource allocation plans. For nodes... Corresponding activities In the competitive dimension The amount of resources allocated is Summarize the resource allocation results across all competitive dimensions to form a complete resource allocation plan. ,in This represents the set of activities involved in the conflict's impact domain. While satisfying the overall global resource constraints, this solution achieves differentiated resource allocation based on the competitive characteristics of the activity evaluation vectors, effectively mitigating resource conflicts between multi-channel marketing activities and providing a resource guarantee foundation for subsequent scheduling and execution.

[0072] In one alternative implementation, Based on the resource allocation scheme and the trend sequence, time-series alignment is performed and execution priorities are calculated. Then, a scheduling execution sequence is generated according to the execution priorities, and resource allocation actions are triggered, including: Extract the resource allocation coefficients corresponding to each node in the conflict influence domain from the resource allocation scheme, extract the time identifiers corresponding to each node from the scheduling dependency graph, extract the trend time points and trend change rates corresponding to each time slice from the trend sequence, perform time axis projection on the time identifiers and the trend time points to determine the time offset, and correct the time identifiers to obtain aligned time identifiers, calculate the time proximity between the aligned time identifiers and the trend time points, and combine the trend change rate to determine the time sequence matching degree; An initial priority score is calculated based on the resource allocation coefficient and the timing matching degree. The sum of the inbound edge dependency levels of each node is extracted from the scheduling dependency graph as the dependency weight. A weighted priority score is calculated based on the initial priority score and the dependency weight. The execution priority is obtained by normalizing the weighted priority score. According to the execution priority, the channel identifier and alignment time identifier corresponding to each node in the conflict influence domain are combined to obtain the scheduling execution sequence. The scheduling execution sequence is traversed and the corresponding resource allocation coefficients are extracted. The resource type identifier corresponding to the resource occupancy data is extracted from the activity feature representation. Based on the resource allocation coefficient and the resource type identifier, a resource allocation instruction is generated and the resource allocation action is triggered.

[0073] After obtaining the resource allocation scheme, scheduling dependency graph, and trend sequence, they need to be integrated and transformed into an executable scheduling sequence. From the resource allocation scheme... Extracting the conflict influence domain Each node Corresponding resource allocation coefficient ,in Iterate through all resource contention dimensions. Simultaneously extract each node from the scheduling dependency graph. Binding time stamp This time marker reflects the node's scheduled execution time in the original time plan. From the trend sequence... Extract key trend inflection points, identify trend moments by detecting the rate of change of the trend vector's magnitude, and calculate the change in trend intensity between adjacent moments. ,when At that time, Mark as trend moment ,in Indicates the first The sequence number of each trend moment. This is the threshold for trend detection. For each trend time point, the rate of change of the trend is calculated. The rate of change quantifies the sensitivity of the trend at that moment.

[0074] The time axis projection process requires the time markers of the nodes to be... Trend Moments Alignment is required. Since time markers originate from discrete time slices, while trend points are derived from continuous time series analysis, there is a difference in time scale between the two. A unified time axis needs to be constructed. This maps the two time representations to a common reference frame. For nodes... Time stamp Calculate its time distance from all trend time points. Find the most recent trend point The time offset is defined as follows: This offset reflects the time difference between the node's original execution time and the most recent trend inflection point. To synchronize node execution time with trend changes, the time marker is corrected. The correction strategy employs an adaptive adjustment mechanism. The time node is scheduled later than the trend reversal point, when The time node is scheduled earlier than the trend reversal point. A correction factor is introduced. ,in This is a time scale parameter. The alignment time identifier is calculated as follows: This correction causes the node time markers to align with the most recent trend point, with the degree of alignment determined by the offset.

[0075] Calculating the alignment time proximity requires quantifying the degree of matching between the aligned time markers and the trend time points. Defining time proximity. ,in This is the proximity bandwidth parameter. The proximity is expressed as a Gaussian kernel function, applied when the alignment time marker falls exactly at the trend time point. The greater the deviation, the lower the similarity. This is combined with the rate of change of the trend. Determine the time series matching degree ,in This is the trend sensitivity coefficient. This matching degree considers not only the degree of time alignment but also the influence of the intensity of trend changes. At moments of drastic trend change, the time series matching degree is more sensitive to time deviations. When the rate of trend change is large, even with high time proximity, the time series matching degree will be penalized if it is not precisely aligned to the trend's exact moment.

[0076] The initial priority score is determined by a combination of resource allocation and timing matching. For nodes... Its overall resource allocation level across all resource dimensions is calculated through weighted aggregation. Resource dimension importance weights are introduced. ,satisfy The resource sufficiency of a node is defined as follows: This indicator reflects the overall resource guarantee level obtained by a node. The initial priority score is calculated as follows: This score multiplies resource sufficiency by timing match, ensuring that priority reflects both resource availability and timing suitability. Nodes with high resource sufficiency but low timing match, or nodes with high timing match but insufficient resource allocation, will have their initial priority scores suppressed.

[0077] Extract the sum of the ingress edge dependency levels of each node from the scheduling dependency graph as the dependency weight. For each node... All its incoming edges The dependency level carried is denoted as The sum of the inbound dependency levels is This sum quantifies the degree to which a node is depended on by other nodes. A higher total dependency level indicates a greater impact of the node's execution on downstream nodes, thus warranting a higher scheduling priority. The weighted priority score is adjusted by incorporating dependency weights into the initial priority. A dependency weighting factor is used. Using a logarithmic form avoids numerical instability caused by excessively large dependency weights. The weighted priority score is calculated as follows: This calculation prioritizes nodes with higher total dependency levels, ensuring that critical dependent nodes execute first and preventing downstream nodes from experiencing scheduling blockages due to unmet prerequisites.

[0078] Normalization converts weighted priority scores into standardized execution priorities. This applies to the conflict impact domain. The set of weighted priority scores of all nodes in the set. Calculate its maximum value and minimum value Execution priority is obtained through max-min normalization. This ensures that priority values ​​are distributed between 0 and 1, facilitating unified processing of subsequent scheduling decisions. Normalized execution priority. The original relative order was maintained, while the influence of dimensions and numerical ranges was eliminated.

[0079] Generating the scheduling execution sequence requires sorting and organizing nodes according to execution priority. All nodes in the conflict influence domain are then sorted according to execution priority. Sort in descending order to obtain the node sequence. ,in For each sorted node Extract its corresponding channel identifier and alignment time marker Combined to form scheduling tuples ,in This represents the index of the node within the sequence. The scheduling execution sequence is defined as follows: The sequence records the execution channel, execution time, and priority order of each node to be scheduled in the form of triples.

[0080] Traversing the scheduling execution sequence Each tuple in Extract nodes Corresponding resource allocation coefficient The resource type identifiers corresponding to resource usage data are extracted from the activity feature representation. Specifically, the activity feature representation contains the demand descriptions of each activity on different resource types, and the set of resource type identifiers is denoted as... ,in Representing resource dimensions The corresponding specific resource types, such as "computing resources," "bandwidth resources," and "human resources," etc. For nodes... In terms of resource type The actual allocation amount is ,in This represents the baseline demand for resources per unit.

[0081] Generating resource allocation instructions requires converting resource allocation coefficients into executable resource operation commands. This involves considering the positions within the scheduling sequence. nodes Construct resource allocation instructions The instruction contains the following field: Target Activity Identifier Implementation Channel Identification Execution time Resource type and allocation mapping table Execution priority and dependent preconditions Resource allocation instructions are encapsulated in a structured format to ensure accurate parsing and execution by downstream systems. Resource allocation actions are triggered by sending instructions to the resource management module. In practice, the resource management module reserves or allocates corresponding resources from the resource pool based on the resource type and allocation amount in the instruction, and binds the resources to the execution channel corresponding to the target activity at the specified execution time. For instructions with pre-existing dependencies, the resource management module maintains a dependency monitoring mechanism, which automatically triggers the resource allocation action for the current instruction after the pre-existing activity is executed. The entire scheduling and execution sequence is processed sequentially according to priority, with high-priority nodes receiving resource allocation and scheduling execution first, ensuring that key activities receive priority protection in a resource-competitive environment. At the same time, a time-series alignment mechanism keeps the execution of activities consistent with the rhythm of marketing trend changes, maximizing the overall marketing effect.

[0082] A second aspect of this invention provides a marketing scheduling optimization system based on AI algorithms, comprising: The activity evaluation unit obtains execution status data and resource usage data of multi-channel marketing activities, calculates the fluctuation amplitude corresponding to the execution status data to obtain status fluctuation features, encodes and concatenates the status fluctuation features and the resource usage data to obtain activity feature representation, calculates the competitive relationship based on the target audience overlap between the multi-channel marketing activities, and combines the activity feature representation to solve for the activity evaluation vector. The scheduling dependency unit is used to calculate the gradient change between activity evaluation vectors at adjacent times to obtain a trend sequence, construct conflict propagation paths between channels based on the competition relationship and mark the connectivity, take the time identifier corresponding to the trend sequence and the channel identifier corresponding to the activity feature representation as node attributes, and construct a topology structure by combining the conflict propagation path and the corresponding connectivity mark to obtain a scheduling dependency graph. The scheduling execution unit is used to traverse the conflict nodes in the scheduling dependency graph and trace the associated nodes along the conflict propagation path to determine the conflict influence domain. It aggregates and calculates the activity evaluation vector of each node in the conflict influence domain and iteratively solves the resource allocation scheme in combination with the preset total resource constraint. Based on the resource allocation scheme and the trend sequence, it performs time-series alignment and calculates the execution priority. Based on the execution priority, it generates a scheduling execution sequence and triggers the resource allocation action.

[0083] A third aspect of the present invention provides an electronic device, comprising: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0084] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0085] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions 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 marketing scheduling optimization method based on AI algorithms, characterized in that, include: Acquire execution status data and resource usage data of multi-channel marketing activities, calculate the fluctuation amplitude corresponding to the execution status data to obtain status fluctuation features, encode and concatenate the status fluctuation features and the resource usage data to obtain activity feature representation, calculate the competitive relationship based on the target audience overlap between the multi-channel marketing activities, and combine the activity feature representation to obtain the activity evaluation vector; The gradient change between activity evaluation vectors at adjacent time points is calculated to obtain a trend sequence. Based on the competitive relationship, a conflict propagation path between channels is constructed and connectivity is marked. The time identifier corresponding to the trend sequence and the channel identifier corresponding to the activity feature representation are used as node attributes. The topology is constructed by combining the conflict propagation path and the corresponding connectivity mark to obtain a scheduling dependency graph. The conflict nodes in the scheduling dependency graph are traversed and the associated nodes are traced along the conflict propagation path to determine the conflict influence domain. The activity evaluation vector of each node in the conflict influence domain is aggregated and calculated, and the resource allocation scheme is obtained by iteratively solving the problem in combination with the preset total resource constraint. The resource allocation scheme and the trend sequence are time-series aligned and the execution priority is calculated. The scheduling execution sequence is generated according to the execution priority and the resource allocation action is triggered.

2. The method according to claim 1, characterized in that, The process involves acquiring execution status data and resource usage data for multi-channel marketing activities, calculating the fluctuation amplitude corresponding to the execution status data to obtain status fluctuation characteristics, and encoding and concatenating the status fluctuation characteristics and resource usage data to obtain activity feature representations, including: Data transmission channels are established with monitoring nodes of various channels through data acquisition interfaces to obtain execution status data and resource usage data. Conversion index sequences and engagement index sequences of marketing activities in different channels within multiple time slices are extracted from the execution status data. The conversion index sequences are sampled by sliding window and the range within the window is calculated to obtain the conversion fluctuation. The engagement index sequences are subjected to difference operation between adjacent time slices to obtain the engagement change. The conversion fluctuation is then normalized and spliced ​​together to obtain the status fluctuation characteristics. Extract the computational and transmission resources used by marketing activities in each channel from the resource usage data to construct a resource usage vector. Perform sparse encoding on the resource usage vector and identify the resource usage pattern. Based on the resource usage pattern, determine the resource type identifier and concatenate it with the resource usage vector to obtain extended resource features. The state fluctuation features and the extended resource features are arranged according to the channel dimension to obtain the fluctuation feature matrix and the resource feature matrix, and the mutual information is calculated for each pair to obtain the correlation matrix. Based on the correlation matrix, a dimension selection mask is generated, and the fluctuation feature matrix and the resource feature matrix are subjected to dimension filtering and concatenation to obtain the activity feature representation.

3. The method according to claim 1, characterized in that, Based on the overlap of target audiences among the multi-channel marketing campaigns, the competitive relationship is calculated, and combined with the campaign feature representation, the campaign evaluation vector is obtained, including: Obtain the target audience identifier set of each channel's marketing activities and perform an intersection operation to obtain a shared audience set. Calculate the ratio of the shared audience set to the target audience identifier set to obtain an overlap matrix. Perform asymmetric decomposition on the overlap matrix to extract the dominant competitive direction vector and combine it with the values ​​in the overlap matrix to construct a competitive intensity matrix to obtain the competitive relationship. Construct a directed competitive graph based on the competitive relationship. The activity features are represented and arranged into a feature matrix. The feature matrix is ​​projected onto the spectral embedding space of the directed competition graph and the geodesic distance from each node to the competition source node is calculated. Based on the geodesic distance, a competition position code is generated. Multi-hop message passing operation is performed on the feature matrix and the directed competition graph to obtain multi-order competition transmission features. The mutual information between different orders of the multi-order competition transmission features is calculated and an inter-order dependency matrix is ​​constructed. Based on the inter-order dependency matrix, the multi-order competition transmission features are adaptively aggregated to obtain competition-aware features. Extract the sum of incoming edge strength and the sum of outgoing edge strength from the directed competition graph and calculate the competition situation factor. Based on the competition situation factor and the competition position encoding, calculate the situation adjustment tensor and combine it with the competition perception feature and the activity feature representation to obtain the activity evaluation vector.

4. The method according to claim 1, characterized in that, Calculating the gradient change between activity evaluation vectors at adjacent time points to obtain a trend sequence, and constructing conflict propagation paths between channels based on the competitive relationship and performing connectivity labeling includes: Extract the activity evaluation vectors of adjacent time moments and calculate the vector difference to obtain the gradient change vector. Perform sliding accumulation on the gradient change vector within a time window and perform singular value decomposition on the accumulation result to extract the dominant singular vector. Based on the dominant singular vector, project and reconstruct the gradient change vector and extract the temporal change pattern of the projection coefficient to obtain the trend acceleration feature. Concatenate the gradient change vector with the trend acceleration feature to obtain the trend sequence. Extract the competition intensity matrix corresponding to the competition relationship and mark the active propagation edge. Construct a directed propagation graph based on the active propagation edge and calculate the in-degree difference of each node to identify the conflict source node and the conflict sink node. Starting from the conflict source node, perform a depth-first traversal of the directed propagation graph and record all paths to the conflict sink node to obtain a set of candidate conflict propagation paths. The candidate conflict propagation path set is filtered based on path length and competition intensity to obtain conflict propagation paths. The starting node and ending node corresponding to each conflict propagation path are extracted, and the temporal characteristics of the trend sequence on the corresponding activities of the starting node and the ending node are determined. Based on the temporal characteristics, connectivity labeling is performed to obtain connectivity labeling results.

5. The method according to claim 1, characterized in that, Using the time identifier corresponding to the trend sequence and the channel identifier corresponding to the activity feature representation as node attributes, and combining them with the conflict propagation path and the corresponding connectivity marker, a topology is constructed to obtain a scheduling dependency graph, including: Extract the time identifier corresponding to each time slice from the trend sequence, extract the channel identifier corresponding to each marketing activity from the activity feature representation and perform a Cartesian product operation with the time identifier to obtain a candidate node set, filter the nodes in the candidate node set that exist in the conflict propagation path based on the connectivity labeling result and label the node type to obtain a valid node set, extract the trend sequence and the activity feature representation corresponding to each node in the valid node set and concatenate them to obtain a node attribute vector; Extract the nodes contained in each conflict propagation path and construct a set of directed edges based on the relationship between adjacent nodes. Extract the competitive intensity between marketing activities corresponding to adjacent nodes from the competitive relationship. Calculate the weight of each edge in the set of directed edges based on the competitive intensity and assign values ​​in a hierarchical manner to obtain edge attributes with different dependency levels. The effective node set and the node attribute vector are used as the node set and node attributes of the topology graph, and the directed edge set and the edge attributes are used as the edge set and edge attributes of the topology graph. The topology graph is sorted and the topology number of each node is marked. The node priority is calculated based on the topology number and marked in the node attribute to obtain the scheduling dependency graph.

6. The method according to claim 1, characterized in that, Traversing the scheduling dependency graph to identify conflicting nodes and tracing related nodes along the conflict propagation path to determine the conflict influence domain, the activity evaluation vector of each node in the conflict influence domain is aggregated and calculated, and combined with a preset total resource constraint, an iterative solution is obtained to obtain a resource allocation scheme, including: Traverse each node in the scheduling dependency graph and extract edge attributes. Based on the edge attributes, sum the dependency levels corresponding to the incoming and outgoing edges of each node and calculate the imbalance degree. Filter based on the imbalance degree to obtain a set of conflicting nodes. Extract the propagation path segment containing any node in the set of conflicting nodes from the conflict propagation path and trace bidirectionally along the propagation path segment to obtain a set of associated nodes. Merge the set of conflicting nodes and the set of associated nodes to obtain the conflict influence domain. Extract the activity evaluation vectors corresponding to each node in the conflict influence domain and calculate the influence domain aggregate vector. Calculate the discreteness of the dimensional components in the influence domain aggregate vector and construct a discreteness vector. Based on the discreteness vector, identify high-discreteness dimensions as resource competition dimensions and normalize the activity evaluation vector components corresponding to the resource competition dimensions to obtain the standardized competition intensity. Using the standardized competition intensity as the initial resource allocation coefficient, the deviation between the total resource consumption corresponding to the resource allocation coefficient and the preset total resource constraint is calculated. The resource allocation coefficient is corrected based on the deviation, and the correction is repeated until the deviation meets the preset convergence condition. The resource allocation coefficient that meets the convergence condition is mapped to each node in the conflict influence domain to obtain the resource allocation scheme.

7. The method according to claim 1, characterized in that, Based on the resource allocation scheme and the trend sequence, time-series alignment is performed and execution priorities are calculated. Then, a scheduling execution sequence is generated according to the execution priorities, and resource allocation actions are triggered, including: Extract the resource allocation coefficients corresponding to each node in the conflict influence domain from the resource allocation scheme, extract the time identifiers corresponding to each node from the scheduling dependency graph, extract the trend time points and trend change rates corresponding to each time slice from the trend sequence, perform time axis projection on the time identifiers and the trend time points to determine the time offset, and correct the time identifiers to obtain aligned time identifiers, calculate the time proximity between the aligned time identifiers and the trend time points, and combine the trend change rate to determine the time sequence matching degree; An initial priority score is calculated based on the resource allocation coefficient and the timing matching degree. The sum of the inbound edge dependency levels of each node is extracted from the scheduling dependency graph as the dependency weight. A weighted priority score is calculated based on the initial priority score and the dependency weight. The execution priority is obtained by normalizing the weighted priority score. According to the execution priority, the channel identifier and alignment time identifier corresponding to each node in the conflict influence domain are combined to obtain the scheduling execution sequence. The scheduling execution sequence is traversed and the corresponding resource allocation coefficients are extracted. The resource type identifier corresponding to the resource occupancy data is extracted from the activity feature representation. Based on the resource allocation coefficient and the resource type identifier, a resource allocation instruction is generated and the resource allocation action is triggered.

8. A marketing scheduling optimization system based on AI algorithms, used to implement the method of any one of claims 1-7, characterized in that, include: The activity evaluation unit obtains execution status data and resource usage data of multi-channel marketing activities, calculates the fluctuation amplitude corresponding to the execution status data to obtain status fluctuation features, encodes and concatenates the status fluctuation features and the resource usage data to obtain activity feature representation, calculates the competitive relationship based on the target audience overlap between the multi-channel marketing activities, and combines the activity feature representation to solve for the activity evaluation vector. The scheduling dependency unit is used to calculate the gradient change between activity evaluation vectors at adjacent times to obtain a trend sequence, construct conflict propagation paths between channels based on the competition relationship and mark the connectivity, take the time identifier corresponding to the trend sequence and the channel identifier corresponding to the activity feature representation as node attributes, and construct a topology structure by combining the conflict propagation path and the corresponding connectivity mark to obtain a scheduling dependency graph. The scheduling execution unit is used to traverse the conflict nodes in the scheduling dependency graph and trace the associated nodes along the conflict propagation path to determine the conflict influence domain. It aggregates and calculates the activity evaluation vector of each node in the conflict influence domain and iteratively solves the resource allocation scheme in combination with the preset total resource constraint. Based on the resource allocation scheme and the trend sequence, it performs time-series alignment and calculates the execution priority. Based on the execution priority, it generates a scheduling execution sequence and triggers the resource allocation action.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images