A multi-channel advertising performance prediction system integrating meta-learning

Abstract

This invention belongs to the field of computer data processing and artificial intelligence technology, and relates to a multi-channel advertising performance prediction system integrating meta-learning. The system includes: a time-series graph construction module, which constructs a dynamic time-series graph using event sequences; a signature generation module, which generates a causal graph signature containing a graph topology and probability matrix from the input time-series graph; an operator activation module, which activates basic signal processing operators based on the signature and extracts learnable meta-parameters; a model construction module, which generates a computational graph and an advertising performance prediction model; a performance prediction module, which generates advertising performance prediction results; a sampling trigger module, which calculates uncertainty, filters out abnormal samples, and generates true labels; and a meta-knowledge update module, which calculates the loss value based on the prediction results and labels and updates the meta-parameters and network weights. This invention solves the problem that existing advertising performance prediction models have fixed architectures and parameters when facing heterogeneous channels, making it difficult to dynamically adapt to unique data structures and behavioral patterns.

Description

Technical Field

[0001] This invention belongs to the technical field of computer data processing and artificial intelligence, and relates to a multi-channel advertising performance prediction system integrating meta-learning. Background Technology

[0002] In computational advertising, predicting ad performance across different channels is crucial for achieving precise ad targeting and optimized budget allocation. Current advertising channels are highly heterogeneous; for example, the user groups, interaction patterns, and behavioral data distributions of platforms such as social media, search engines, and short videos all differ significantly. This data heterogeneity requires ad performance prediction models to not only capture common behavioral patterns but also adapt to the unique data characteristics of specific channels.

[0003] Existing advertising effectiveness prediction solutions mainly fall into three categories. The first category employs a general deep learning model, building a unified architecture based on fully connected networks or attention mechanisms, and training it holistically on datasets from multiple heterogeneous channels to learn universal user behavior representations. The second category involves engineers independently designing and training a dedicated prediction model for each advertising channel to fit the specific channel's data distribution characteristics. The third category attempts to introduce categorical features representing channel identity into a general model, aiming to enable a single model to distinguish data from different channels.

[0004] However, the aforementioned technical solutions have certain limitations when dealing with complex channel heterogeneity. For the first type of solution, which uses a general model, its fixed underlying model structure easily leads to the averaging of unique patterns across different channels during the learning process, limiting the model's predictive performance on each specific channel. For the second type of solution, which independently builds its own dedicated model, this approach typically requires high development and maintenance costs and severs the data connections between different channels, making it difficult to transfer and reuse knowledge extracted from one channel to other channels, thus reducing the efficiency of model scaling. For the third type of solution, which introduces identity category features, it essentially still uses a fixed model architecture to process data from all channels, and the dynamic adaptability of the model structure still has room for improvement. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a multi-channel advertising performance prediction system integrating meta-learning.

[0006] A multi-channel advertising performance prediction system integrating meta-learning includes: The timing diagram construction module is used to acquire the raw event sequence data stream and construct a dynamic timing diagram using the raw event sequence data stream. The signature generation module is used to input dynamic time series graphs into a preset time series causal graph builder to generate causal graph signatures that include graph topology and probability matrices. The operator activation module is used to input the causal graph signature into the preset meta-algorithm configurator, activate the basic signal processing operators in the preset meta-algorithm library, and extract the learnable meta-parameters within the basic signal processing operators. The model building module connects basic signal processing operators based on the graph topology to generate a dynamic computation graph. Specific instance parameters are then injected into the dynamic computation graph to generate an advertising effectiveness prediction model. The performance prediction module is used to obtain standard input feature vectors and input them into the advertising performance prediction model to generate advertising performance prediction results; The sampling trigger module is used to calculate the prediction uncertainty of the advertising effectiveness prediction results, filter out abnormal prediction samples based on the prediction uncertainty, label them, and generate real effectiveness label data. The meta-knowledge update module calculates the meta-loss value based on the advertising effectiveness prediction results of the abnormal prediction samples and the actual effectiveness label data. It then uses the meta-loss value to update the learnable meta-parameters and the network weights of the time-series causal graph builder.

[0007] A further aspect of this invention involves constructing a dynamic time series graph using the original event sequence data stream, comprising the following steps: Parse the original event sequence data stream, extract user exposure event data, click event data, and conversion event data, and extract the timestamp and user identification corresponding to each event; Cluster user exposure event data, click event data, and conversion event data using user identity identifiers to generate a time-series behavior trajectory sequence; Abstract each event in the temporal behavior trajectory sequence into a graph node, and calculate the difference in timestamps between adjacent event nodes; Determine if the difference is less than the preset maximum time interval threshold. If the difference is less than the maximum time interval threshold, establish a directed edge between adjacent event nodes and generate a dynamic time sequence graph.

[0008] A further aspect of the present invention generates a causal graph containing a graph topology and a probability matrix, comprising the following steps: The graph neural network in the temporal causal graph builder is invoked to aggregate node features of the dynamic temporal graph and extract the time delay distribution features and path propagation features represented by directed edges; Input the time delay distribution characteristics and path propagation characteristics into the probability calculation unit in the temporal causal graph builder to calculate the directed propagation probability between each graph node; A probability matrix is ​​generated using the directed transitive probabilities between nodes in the graph. Extract the graph topology of the dynamic time series graph, concatenate the graph topology with the probability matrix, and generate a causal graph signature.

[0009] A further aspect of this invention involves extracting learnable meta-parameters within basic signal processing operators, including the following steps: Analyze the graph topology in the causal graph signature and calculate the temporal dependency strength and feature cross frequency corresponding to the graph topology. The matching weights of each operator in the meta-algorithm library are calculated by linearly weighting the temporal dependency strength and the feature cross frequency. Compare the matching weights with the preset activation thresholds, and filter out target operators whose matching weights exceed the activation thresholds; The target operator is activated as a basic signal processing operator, and the learnable meta-parameters inherent in the basic signal processing operator are extracted.

[0010] A further aspect of this invention involves generating an advertising effectiveness prediction model, comprising the following steps: Map the graph topology to computation graph node connection rules; The basic signal processing operators are connected in series and parallel according to the node connection rules of the computation graph to generate a dynamic computation graph with a specific information flow path; Perform the Hadamard product operation between the probability matrix and the learnable meta-parameters, and use the result of the Hadamard product operation as the specific instance parameter; Specific instance parameters are injected into the corresponding operator nodes in the dynamic computation graph to generate an advertising effectiveness prediction model.

[0011] A further aspect of the present invention generates advertising effectiveness prediction results, comprising the following steps: Collect the current context features and real-time interaction history sequence of the user to be predicted, and perform standardization processing on the current context features and real-time interaction history sequence to generate a standard input feature vector; The standard input feature vector is input into the advertising effectiveness prediction model, and the feature mapping and attenuation operations are performed sequentially through the basic signal processing operators along the information flow path of the dynamic computation graph to generate the target feature vector. Normalization mapping is performed on the target feature vector to generate a two-dimensional probability vector representing the predicted click-through rate and the predicted conversion rate; the two-dimensional probability vector is used as the result of advertising effectiveness prediction.

[0012] A further aspect of this invention generates real performance label data, comprising the following steps: The advertising effectiveness prediction results are decomposed into two independent Bernoulli distributions corresponding to click events and conversion events; Calculate the Shannon entropy of two independent Bernoulli distributions independently, and then calculate the prediction uncertainty by arithmetically averaging the two Shannon entropies. Determine whether the prediction uncertainty is higher than the preset uncertainty threshold. If the prediction uncertainty is higher than the uncertainty threshold, mark the corresponding user sample to be predicted as an abnormal prediction sample. Anomaly prediction samples are sent to the data acquisition system for automated alignment and verification of offline attribution logs, generating real performance label data.

[0013] A further aspect of this invention involves updating the learnable meta-parameters and the network weights of the time-series causal graph builder using meta-loss values, including the following steps: Input the advertising effectiveness prediction results of the abnormal prediction samples and the actual effectiveness label data into the binary cross-entropy loss function to calculate the cross-entropy loss gradient of the current prediction task. The cross-entropy loss gradient is used to perform an internal loop update on the specific instance parameters in the advertising effectiveness prediction model to obtain temporary adaptation parameters. Extract the reserved validation set, apply the advertising effectiveness prediction model with temporary adaptation parameters to the validation set for prediction calculation, and generate meta-loss value; The meta-loss value is backpropagated along the dynamic computation graph to the meta-algorithm library and the temporal causal graph builder, updating the learnable meta-parameters and the network weights of the temporal causal graph builder.

[0014] A further aspect of this invention involves linearly weighting the temporal dependency strength and feature crossover frequency to calculate the matching weights of each operator in the meta-algorithm library, including the following steps: Extract the preset first hyperparameter, second hyperparameter, third hyperparameter, and fourth hyperparameter; Multiply the first hyperparameter by the temporal dependency strength to calculate the first matching weight for generating the time decay operator; The second hyperparameter is multiplied by the feature cross frequency to calculate the second matching weight for generating the feature cross attention operator; The product of the third hyperparameter and the temporal dependency strength is added to the product of the fourth hyperparameter and the feature cross frequency to generate the third matching weight of the information flow gating operator. The first matching weight, the second matching weight, and the third matching weight are used as the matching weights of the corresponding operators.

[0015] A further aspect of this invention utilizes the cross-entropy loss gradient to perform an internal loop update on the specific instance parameters in the advertising effectiveness prediction model to obtain temporary adaptation parameters, including the following steps: Extract the preset internal loop learning rate; multiply the internal loop learning rate with the cross-entropy loss gradient to generate a parameter adjustment step size tensor; The difference between the specific instance parameters and the parameter adjustment step tensor is calculated to generate an updated parameter matrix; the updated parameter matrix is ​​then used as the temporary adaptation parameters.

[0016] In summary, the present invention has the following beneficial technical effects: 1. This invention improves upon the inadequacy of traditional single-channel prediction models for heterogeneous channel data by dynamically constructing prediction models for specific channels. Specifically, the system constructs a dynamic time-series graph from the original event sequence and extracts a causal graph signature. This signature guides the meta-algorithm configurator to activate suitable basic operators. Subsequently, the operators are reorganized according to the graph topology, and the operator parameters are instantiated through Hadamard product operations, thereby generating a customized micro-prediction model for the data characteristics of the current channel. This architecture design enables the model to adapt to the unique patterns of different advertising channels without changing the overall prediction framework.

[0017] 2. This invention introduces an active sampling and meta-update closed loop based on prediction uncertainty, enabling the system's self-iteration and knowledge accumulation. The uncertainty of the prediction results is quantified by calculating the information entropy of the output probability distribution. For high-uncertainty samples with information entropy exceeding a preset threshold, the system triggers an active sampling mechanism, submitting them to the underlying attribution verification pipeline. Based on real terminal playback data collected in the system's backend, the system automatically verifies and obtains the true label. After determining the true performance label, the system not only uses this sample to update the model parameters for the current task but also uses the backpropagation meta-update mechanism to update the learnable meta-parameters in the meta-algorithm library and the network weights of the temporal causal graph builder. This allows the system to learn from prediction errors and generalize knowledge, continuously optimizing its model-building capabilities, thereby enhancing the overall prediction system's generalization performance and long-term evolution capabilities when facing new tasks and channels in the future.

[0018] 3. This invention achieves a deep characterization of channel data characteristics by abstracting user behavior sequences across heterogeneous channels into dynamic time-series graphs and further generating causal graph signatures containing graph topology and probability matrices. This scheme leverages graph neural networks to aggregate node features, capture the temporal delay distribution and multi-step path propagation features in user behavior sequences, and quantifies them into directed propagation probabilities between nodes, forming a probability matrix. This approach encodes channel data characteristics into standardized signatures rich in structural and causal information. These signatures not only guide the dynamic construction of subsequent model architectures but also participate in the dynamic generation of model parameters. This enables the model to directly perceive and utilize the inherent logic of user behavior in specific channels, such as strong temporal dependencies or strong feature interactions, thereby achieving end-to-end adaptive mapping from the underlying data structure to the model structure and parameters. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention.

[0020] Figure 1 This is a schematic diagram of the framework in the embodiments of this application.

[0021] Figure 2 This is a flowchart illustrating an embodiment of this application. Detailed Implementation

[0022] The following is in conjunction with the appendix Figure 1 - Figure 2 A preferred description of the present invention is provided below.

[0023] See attached document Figure 1 - Figure 2 This invention proposes a multi-channel advertising performance prediction system integrating meta-learning, comprising the following modules: The system comprises the following modules: a time-series graph construction module, which acquires the original event sequence data stream and constructs a dynamic time-series graph; a signature generation module, which inputs the dynamic time-series graph into a preset time-series causal graph builder to generate a causal graph signature containing the graph topology and probability matrix; an operator activation module, which inputs the causal graph signature into a preset meta-algorithm configurator to activate basic signal processing operators in a preset meta-algorithm library and extract learnable meta-parameters from the basic signal processing operators; a model construction module, which connects the basic signal processing operators according to the graph topology to generate a dynamic computation graph, injects specific instance parameters into the dynamic computation graph, and generates an advertising effectiveness prediction model; an effectiveness prediction module, which acquires the standard input feature vector and inputs it into the advertising effectiveness prediction model to generate advertising effectiveness prediction results; a sampling trigger module, which calculates the prediction uncertainty of the advertising effectiveness prediction results, filters abnormal prediction samples based on the prediction uncertainty, labels them, and generates real effectiveness label data; and a meta-knowledge update module, which calculates the meta-loss value based on the advertising effectiveness prediction results of abnormal prediction samples and the real effectiveness label data, and uses the meta-loss value to update the learnable meta-parameters and the network weights of the time-series causal graph builder.

[0024] In one embodiment of the present invention, the timing graph construction module is configured to perform the following steps: The original event sequence data stream is parsed to extract user exposure event data, click event data, and conversion event data, and to extract the timestamps and user identifiers corresponding to each event. The user exposure event data, click event data, and conversion event data are clustered using the user identifiers to generate a time-series behavior trajectory sequence. Each event in the time-series behavior trajectory sequence is abstracted as a graph node, and the difference in timestamps between adjacent event nodes is calculated. It is then determined whether the difference is less than a preset maximum time interval threshold. If the difference is less than the maximum time interval threshold, directed edges are established between adjacent event nodes to generate a dynamic time-series graph.

[0025] Specifically, unstructured user behavior logs from different advertising channels are transformed into structured graph data. The data preprocessing server subscribes to and retrieves raw event sequence data streams from specified target heterogeneous channels through standard data interfaces, such as the Kafka message queue distributed stream processing platform or the RESTful API for presentation layer state transitions. It should be noted that target heterogeneous channels refer to advertising traffic channels with significant differences in data sources, user groups, and interaction patterns, such as social media feeds, search engine advertising channels, or short video application channels. This raw event sequence data stream is typically a JSON-formatted log record, where each record contains an event type field, a user identification field, and an event timestamp field. The parsing module of the data preprocessing server parses each log record, extracting user exposure event data, click event data, and conversion event data, and caching the corresponding UNIX timestamp and user identification for each event.

[0026] The data aggregation module uses extracted user identifiers as keys to group cached event data, grouping events belonging to the same user identifier together. Based on this, the system applies a preset time window to segment the event set for each user, thereby generating a time-series behavioral trajectory sequence for a specific user within the preset time window. This trajectory sequence is a list of events arranged in ascending order of timestamps. In this embodiment, the preset time window is used to define the duration of a single analysis session, typically ranging from 30 to 120 minutes. This value is set based on industry statistical analysis of average user online session duration, aiming to fully capture a complete user decision-making process. The time-series behavioral trajectory sequence is a list of events sorted by timestamp, with the format as follows: .

[0027] Finally, the graph construction module traverses the temporal behavior trajectory sequence. For each event in the sequence, the system creates a corresponding graph node in memory, which carries attributes such as event type and timestamp. Simultaneously, the module establishes connections between adjacent event nodes in the temporal behavior trajectory sequence based on a preset temporal proximity rule. Specifically, for any two consecutive event nodes in the sequence, the system calculates the difference between their timestamps. If this difference is less than a preset maximum time interval threshold, a directed edge is established between the node representing the preceding event and the node representing the following event, with the direction of the edge determined by the chronological order of the events. It should be understood that the temporal proximity rule is used to determine whether two consecutive events belong to the same micro-intention and are consecutive behaviors. The condition for establishing a directed edge can be expressed as: for any two adjacent events in the temporal behavior trajectory sequence... and Their corresponding timestamps are respectively and In the representation, if and only if the following conditions are met, nodes and representatives nodes Establish a directed edge between them .

[0028]

[0029] In the formula, Indicates the timestamp of the later event. Indicates the timestamp of the previous event. This represents the preset maximum time interval threshold. Specifically, the maximum time interval threshold in the time proximity rule... This is a key parameter used to distinguish between continuous and interrupted operations, and its value is usually set between 30 seconds and 600 seconds. This range is based on user behavior research, which assumes that when performing continuous operations within an application, the interval between actions is usually within a few minutes. If this threshold is exceeded, it is believed that the user's attention may have shifted, and the direct causal relationship between the two events is weakened.

[0030] By performing the above operations on the temporal behavior trajectory sequences of all users, a dynamic temporal graph that can represent the overall flow of user behavior is finally generated. This graph is serialized and stored in a graph database for subsequent steps. In summary, a dynamic temporal graph is a directed graph data structure whose node set consists of all events and whose edge set consists of adjacent event pairs that satisfy the temporal proximity rule.

[0031] For example, suppose the data preprocessing server obtains data from a short video application channel, belonging to user ID... The server retrieves the original event sequence data stream and sets a preset time window of 60 minutes and a maximum time interval threshold of 300 seconds. Within the set time window, the server extracts the following four event data: exposure event E1, with a timestamp of 1678886400 seconds; click event E2, with a timestamp of 1678886410 seconds; conversion event E3, with a timestamp of 1678886500 seconds; and another exposure event E4, with a timestamp of 1678886820 seconds. The server first sorts these events in ascending order of timestamps to generate a time-series behavior trajectory sequence [E1, E2, E3, E4]. Subsequently, the graph construction module begins to build a dynamic time-series graph. It first creates four graph nodes for E1, E2, E3, and E4. Then, it checks the time difference between adjacent nodes. The timestamp difference between E2 and E1 is 1678886410 - 1678886400 = 10 s, which is less than the maximum time interval threshold of 300 s. Therefore, a directed edge is established between nodes E1 and E2. The timestamp difference between E3 and E2 is 1678886500 - 1678886410 = 90 s, which is less than 300 s. Therefore, a directed edge is established between nodes E2 and E3. The timestamp difference between E4 and E3 is 1678886820 - 1678886500 = 320 s, which is greater than 300 s. Therefore, no edge is established between nodes E3 and E4. Finally, for the user... The generated dynamic sequence graph contains four nodes {E1, E2, E3, E4} and two directed edges {(E1, E2), (E2, E3)}.

[0032] In one embodiment of the present invention, the signature generation module is configured to perform the following steps: The graph neural network in the temporal causal graph builder is invoked to aggregate node features of the dynamic temporal graph, extracting the time delay distribution features and path propagation features represented by directed edges; the time delay distribution features and path propagation features are input into the probability calculation unit in the temporal causal graph builder to calculate the directed propagation probability between each graph node; the directed propagation probability between each graph node is used to generate a probability matrix; the graph topology structure of the dynamic temporal graph is extracted, and the graph topology structure is concatenated with the probability matrix to generate the causal graph signature.

[0033] Specifically, after acquiring the dynamic time series graph, a pre-defined time-series causal graph builder is invoked to process the graph data, extracting its deep causal relationships and generating standardized signatures. It should be noted that the time-series causal graph builder is a pre-trained deep learning model, whose core component is a graph neural network.

[0034] First, the graph neural network module within the temporal causal graph builder processes the input dynamic temporal graph. Specifically, it converts the topological structure of the dynamic temporal graph into an adjacency matrix and generates an initial feature vector for each node in the graph based on its event type. In this embodiment, the initial feature vector for each node can be generated using one-hot encoding, with the vector dimension equal to the total number of event types. The number of layers in the graph neural network is typically set between 2 and 4 to balance the receptive field size and the problem of oversmoothing.

[0035] Graph neural networks, such as graph convolutional networks, update node feature representations by performing multi-layer neighborhood aggregation operations. In each aggregation layer, each node collects the feature vectors of its immediate neighbors and combines them with its own feature vector, performing linear transformations and non-linear activations through a learnable weight matrix to generate a new feature vector for that layer. This iterative aggregation process ensures that each node's final feature vector not only contains its own attributes but also encodes structural information within its multi-order neighborhood. This information is a comprehensive representation of the time delay distribution characteristics and path propagation characteristics represented by directed edges. The time delay distribution characteristic refers to the statistical distribution pattern of time required to transition from one event type to another in a user behavior sequence; this characteristic is implicitly learned by the graph neural network during the aggregation process. The path propagation characteristic refers to the typical multi-step transition paths followed by user behavior in the graph, such as the sequence pattern from exposure to click to conversion; this characteristic is captured through the multi-layered structure of the graph neural network.

[0036] Building upon this, the probability calculation unit within the builder uses the updated node feature vectors to quantify the strength of causal relationships between nodes. For any two nodes in the graph, their corresponding final feature vectors are extracted, and this pair of vectors is input into a feedforward neural network classifier or directly performed as a dot product and then activated by the Sigmoid function to calculate the directed transit probability from one node to another. It should be understood that the directed transit probability... The calculation process can be represented by the following formula, which describes the process from node... To the node The possibility of conveying behavior.

[0037]

[0038] In the formula, and These represent the nodes after processing by the graph neural network. and The final eigenvector. It is a learnable weight matrix used to capture the interaction patterns between feature vectors. This represents the Sigmoid activation function, which maps the calculated result to a probability range of 0 to 1. In the formula... The calculation result is a scalar, with superscript. Represents the transpose of a vector or matrix.

[0039] By performing this calculation on all node pairs, a probability matrix is ​​generated, which constitutes the probability matrix characterizing the causal chain of user behavior across heterogeneous channels for this target. Here, the probability matrix is ​​a... The probability matrix, where It refers to the number of nodes in the dynamic time sequence graph and the elements in the matrix. That is, the directed transitive probability. Finally, the signature generation module encapsulates the graph topology of the dynamic temporal graph, typically represented as a sparse adjacency matrix, with the newly generated probability matrix into a standardized data structure, generating a causal graph signature for subsequent architecture routing instructions, and temporarily storing it in memory. In this embodiment, the graph topology refers to the connection relationships defined by the set of nodes and edges of the graph, typically represented by an adjacency matrix or adjacency list. The causal graph signature is a composite data object, containing at least the graph's adjacency matrix and probability matrix.

[0040] For example, calling the time-series causal graph builder processes the user The generated dynamic temporal graph contains nodes {E1, E2, E3, E4} and directed edges {(E1, E2), (E2, E3)}. First, an initial feature vector is generated for each node. Assuming the event type is {exposure, click, conversion}, then the vector for E1 is [1, 0, 0], for E2 it is [0, 1, 0], for E3 it is [0, 0, 1], and for E4 it is [1, 0, 0]. The graph neural network module receives these initial vectors and the graph's adjacency matrix. After two layers of graph convolution, it outputs updated node feature representations. Assuming the output two-dimensional features are column vectors: E1's... for E2 for E3 for Subsequently, the probability calculation unit calculates the directed transit probabilities between nodes. For example, when calculating the probability from E1 to E2, it is assumed that a learnable weight matrix is ​​used. If it is an identity matrix, then by The calculation process is as follows ,Right now The resulting probability value is approximately 0.61. The probabilities between all node pairs are calculated in the same way, generating a 4×4 probability matrix, where the probability values ​​between most non-directly connected node pairs will be close to 0. Finally, the adjacency matrix of the graph and this generated 4×4 probability matrix are packaged to generate a causal graph signature specific to this user behavior segment. This signature fully describes the structure of this behavior graph and the strength of its inherent causal flow.

[0041] In one embodiment of the present invention, the operator activation module is configured to perform the following steps: The graph topology in the causal graph signature is analyzed, and the temporal dependency strength and feature cross frequency corresponding to the graph topology are calculated. The temporal dependency strength and feature cross frequency are linearly weighted, and the matching weight of each operator in the meta-algorithm library is calculated. The matching weight is compared with a preset activation threshold, and target operators with matching weights exceeding the activation threshold are selected. The target operators are activated as basic signal processing operators, and the learnable meta-parameters built into the basic signal processing operators are extracted.

[0042] Specifically, after generating the causal graph signature, it is sent to a pre-defined meta-algorithm configurator. This configurator is responsible for dynamically selecting and preparing the basic components needed to build the model from a pre-defined meta-algorithm library. The meta-algorithm configurator is a rule-based and computational decision engine used to map data features to model architecture. The meta-algorithm library stores a collection of various parameterized and differentiable basic computational modules. The parsing module of the meta-algorithm configurator receives and parses the causal graph signature, extracting graph topology information such as the graph's adjacency matrix and node type list.

[0043] Based on the graph topology, the configurator's analysis module calculates two core metrics. The first is the temporal dependency strength of user behavior in the target heterogeneous channel data. This strength is quantified by calculating the ratio of the number of edges to the number of nodes in the graph, reflecting the tightness of the user behavior sequence. The second is the feature cross frequency, which is measured by statistically analyzing the proportion of edges connecting nodes of different event types to the total number of edges, reflecting the frequency of interactions between different user behaviors. It should be understood that temporal dependency strength... and feature cross frequency The calculation method is as follows:

[0044]

[0045] In the formula, It is the total number of nodes in the dynamic sequence graph. It is the total number of edges. It is the number of edges connecting nodes of different event types. and All are dimensionless normalized values.

[0046] Subsequently, the weight calculation module calculates the matching weight for each operator in the meta-algorithm library based on the previously calculated temporal dependency strength and feature cross-frequency. The meta-algorithm library pre-constructs various basic signal processing operators, such as the time decay operator, feature cross-attention operator, and information flow gating operator. The time decay operator is a module used to apply a decay effect to historical information according to time intervals, commonly used for processing time-series data. The feature cross-attention operator is a module that uses an attention mechanism to calculate the interaction strength between different features, effectively capturing feature combination patterns. The information flow gating operator is a module that controls the flow of information through gating units, similar to the update and reset gates in a gated loop unit. The matching weight calculation rule is designed to map high temporal dependency strength to high weights in the time decay operator and high feature cross-frequency to high weights in the feature cross-attention operator. Specifically, the matching weights of each operator... Determined by the following rules:

[0047]

[0048]

[0049] in, The first matching weight for the time decay operator. The second matching weight of the feature cross-attention operator. This is the third matching weight for the information flow gating operator. These are the preset first hyperparameter and second hyperparameter. These are the preset third and fourth hyperparameters. The product of the third hyperparameter and the temporal dependency strength is added to the product of the fourth hyperparameter and the feature cross frequency to generate the aforementioned third matching weight of the information flow gating operator. Used to adjust the influence of various metrics on operator weights, the matching weight is a scalar between 0 and 1, which quantifies the degree of fit between a specific operator and the current data pattern.

[0050] Finally, the filtering and activation module compares the matching weights calculated for each operator with a preset activation threshold. The preset activation threshold is a hyperparameter used to decide whether to enable a particular operator, typically set between 0.5 and 0.8, based on a trade-off between ensuring model expressiveness and controlling model complexity. Any operator whose matching weights exceed the activation threshold is marked as activated. The system activates a corresponding number of basic signal processing operators and extracts their built-in learnable meta-parameters from their internal definitions. These parameters form the basis for subsequent model instantiation. It should be noted that learnable meta-parameters are weight matrices or bias vectors stored within each operator in the meta-algorithm library, learned during the meta-training phase. They represent universal knowledge extracted from a large number of historical tasks.

[0051] For example, the meta-algorithm configurator receives information about the user. The causal graph signature. The allocator resolves its graph topology to contain 4 nodes and 2 edges, i.e. , The analysis module begins calculating metrics. First, it calculates the strength of time-series dependencies. Next, the characteristic crossover frequency is calculated. Since the two edges (E1,E2) and (E2,E3) connect event nodes of different types, the number of crossover edges is... Therefore Assume the system's preset hyperparameters are... , , , The preset activation threshold is 0.5. The weight calculation module calculates the matching weights of each operator as follows: weight of the time decay operator. Weights of the feature cross-attention operator Weights of information flow gating operators Finally, the filtering and activation module makes a judgment: The time decay operator was not activated. The feature cross-attention operator is activated. The information flow gating operator is activated. Therefore, the system activates the feature cross attention operator and the information flow gating operator, and extracts the learnable meta-parameters inherent in these two operators from the meta-algorithm library, such as the query, key, and value weight matrices of the attention operator and the update and reset gate weight matrices of the gating operator, for use in the next step.

[0052] In one embodiment of the present invention, the model building module is configured to perform the following steps: The graph topology is mapped to a computation graph node connection rule; the basic signal processing operators are connected in series and parallel according to the computation graph node connection rule to generate a dynamic computation graph with a specific information flow path; the Hadamard product operation between the probability matrix and the learnable meta-parameters is performed, and the result of the Hadamard product operation is used as the specific instance parameter; the specific instance parameter is injected into the corresponding operator node in the dynamic computation graph to generate an advertising effectiveness prediction model.

[0053] Specifically, after activating the required basic signal processing operators and extracting their meta-parameters, the model building server executes this step to dynamically assemble these discrete operators into a fully functional predictive model. First, the topology reorganization module extracts the graph topology from the causal graph signature, which clarifies the sequence and dependencies between user behavior events. This module directly maps this event flow graph to computation graph node connection rules. These rules are a set of instructions that specify the data flow between computation operators based on the overall adjacency of the user behavior graph and the data processing attributes of the activated operators. According to these rules, the system creates an empty computation graph in memory and instantiates each activated basic signal processing operator as a computation node within the graph. Specifically, since the number of activated basic signal processing operators does not correspond one-to-one with the number of specific event nodes, the mapping rule is as follows: the system establishes a pipeline based on the functional hierarchy of the operators. For example, the feature cross-attention operator responsible for extracting spatial / interactive features is placed at the bottom layer, and the information flow gating operator responsible for information temporal filtering is placed at the top layer. Subsequently, the input feature vector flows sequentially through the operator nodes arranged in a directed manner according to the preset hierarchy, thereby establishing data flow connections between each operator node, and finally generating a complete dynamic computation graph model composed of activated operators connected in series, with a specific information flow path. It should be understood that a dynamic computation graph is a directed acyclic graph data structure, where the nodes are activated basic signal processing operators, the edges are data flows between nodes, and its topology is dynamically generated at runtime based on the input causal graph signature.

[0054] In parallel, the parameter instantiation module performs parameter generation. This module reads the probability matrix from the causal graph signature and obtains the extracted learnable meta-parameters from the previous step. The Hadamard product operation is performed on these two sets of numerical tensors, an element-wise multiplication operation. It should be noted that the Hadamard product operation is a binary operation on matrices or tensors, the result of which is a new matrix or tensor obtained by multiplying corresponding elements of the two operands. The specific process of generating instance parameters can be represented by the following formula:

[0055] In the formula, This represents the learnable meta-parameters extracted from the meta-algorithm library. This represents the probability matrix extracted from the causal graph signature. The symbol "" represents the Hadamard product operation. It is a dimension alignment function, which adjusts the input... Transformations can be performed, such as dimensional projection, tensor tiling, or slicing operations using a multilayer perceptron (MLP), to ensure that the tensor shape matches the extracted learnable meta-parameters. The hidden layer dimensions are strictly aligned and compatible, allowing safe element-wise multiplication to be performed without causing shape errors.

[0056] If the dimensions of two tensors do not match, they are aligned using methods such as broadcasting or slicing. The final result of the operation forms a new set of parameters, which are the specific instance parameters of the basic signal processing operator under the current target heterogeneous channel. These specific instance parameters are generated after the Hadamard product operation, and are the final model parameters that integrate meta-knowledge and channel characteristics. They will be directly used for subsequent prediction calculations. Finally, the model assembly module injects the generated specific instance parameters into the corresponding operator nodes in the dynamic computation graph, completing the parameter instantiation. At this point, an advertising performance prediction model with an architecture and parameters customized for the characteristics of the current channel data is completed. It can be said that the advertising performance prediction model is the final product of this step. It is a lightweight and highly specialized neural network model whose architecture and parameters are adapted to the specific target heterogeneous channel.

[0057] For example, following the previous step, the model building server knows that the activated feature cross-attention operator and information flow gating operator need to be used. The server first reads the graph topology in the causal graph signature, i.e., the existence of directed edges (E1, E2) and (E2, E3). Based on this, the connection rules for generating the dynamic computation graph are: the feature cross-attention operator... and information flow gating operators As a computing node, and to establish from arrive The data stream connection. Meanwhile, the parameter instantiation module reads the probability matrix, assuming its key numerical parts are matrix... The module simultaneously extracts Learnable meta-parameters, assumed to be matrices The system performs the Hadamard product operation to calculate... Specific instance parameters: Finally, the model assembly module will calculate the matrix. Loaded into the dynamic computation graph Within the node. In a similar manner... After generating and injecting its specific instance parameters, the entire advertising effectiveness prediction model adapted to this short video channel is completed.

[0058] In one embodiment of the present invention, the performance prediction module is configured to perform the following steps: The system collects the current context features and real-time interaction history sequence of the user to be predicted, and performs standardization processing on the current context features and real-time interaction history sequence to generate a standard input feature vector. The standard input feature vector is then input into the advertising effectiveness prediction model, and the model performs feature mapping and attenuation operations sequentially through basic signal processing operators along the information flow path of the dynamic computation graph to generate a target feature vector. The target feature vector is then subjected to normalization mapping calculation to generate a two-dimensional probability vector representing the predicted click-through rate and predicted conversion rate. The two-dimensional probability vector is then used as the advertising effectiveness prediction result.

[0059] Once the advertising effectiveness prediction model is built, the real-time prediction server uses this model to estimate the real-time advertising effectiveness of online traffic. First, the data acquisition module captures advertising requests from users to be predicted in real time through the server interface. These users are online users who have triggered an ad display request and require the system to estimate the click-through and conversion probabilities of the recommended ads. The request includes the user's current contextual characteristics, such as device model, operating system version, network environment, and geographical location, and is associated with the user's real-time interaction history sequence, such as records of recent clicks or browsing behavior. The current contextual characteristics describe the instantaneous information of the user's current environment. The real-time interaction history sequence records the user's behavior in the current session or in the recent past, used to capture the user's short-term interests.

[0060] The feature processing module receives this raw data and standardizes it. For categorical features, such as device model, one-hot encoding or embedding lookup is used to convert them into numerical vectors; for numerical features, such as the time since the last active event, min-max normalization is used to scale them to a preset range. All processed feature vectors are concatenated into a single-dimensional standard input feature vector. It should be noted that the standard input feature vector is a standardized numerical vector that can be directly accepted by the model.

[0061] Subsequently, the prediction execution engine takes the standard input feature vector as input and feeds it into the loaded advertising effectiveness prediction model, initiating a complete forward propagation operation. Forward propagation refers to the process of calculating data layer by layer from the model's input layer to the output layer until the final result is produced. The data flow strictly follows the pre-defined information flow path of the dynamic computation graph, starting from the input layer and flowing sequentially through each instantiated basic signal processing operator in the graph. At each operator node, the input data undergoes pre-defined feature mapping and attenuation operations with the specific instance parameters of that node, and the output result is then used as the input to the next connected operator. In this embodiment, feature mapping and attenuation operations are the core functions executed by the basic signal processing operators; the former aims to discover complex relationships between features, while the latter is used to reduce the weight of information based on time or correlation.

[0062] After the data stream is passed to the terminal operator of the advertising performance prediction model and the calculation is completed, a high-dimensional target feature vector is obtained. Finally, the output mapping layer performs a normalization mapping calculation on this target feature vector, linearly transforming it to two dimensions through a fully connected layer, and applying the Sigmoid activation function to map each dimension of the transformation result to the interval between 0 and 1. The final outputs advertising performance prediction results representing the predicted click-through rate and predicted conversion rate, respectively. The normalization mapping calculation is the process of converting the abstract feature representation within the model into probability values ​​with clear business meaning. It generates the advertising performance prediction results. The generation of is defined by the following formula:

[0063] In the above formula, It is the target feature vector output by the terminal operator of the advertising effectiveness prediction model. and These represent the weight matrix and bias vector of the fully connected layer in the output mapping layer, respectively. It is the Sigmoid activation function. The result of this formula is a two-dimensional vector with two elements. and These are the predicted click-through rate and conversion rate, both dimensionless probability values. The ad performance prediction result is a tuple or vector containing two floating-point numbers, representing the probability that the user will click on the ad and the probability that the user will complete the conversion, as predicted by the model.

[0064] For example, the real-time prediction server loads an advertising effectiveness prediction model built for short video channels. A user to be predicted... A request was initiated, with its current context features being {Device: 'iOS', Network: '5G'} and its real-time interaction history sequence being {Click: 'vid_A', Browse: 'vid_B'}. The feature processing module standardizes this information to generate a standard input feature vector, assumed to be [1.0, 0.0, 0.8, 0.2, 0.9, 0.1]. This vector is then input to a... and A series-connected advertising effectiveness prediction model. First, the vector flows through... The node, and the parameters that have been instantiated within that node. Attention is calculated, assuming the intermediate feature of the output is [0.45, 0.62]. This intermediate feature is then used as input. After the node is processed by the gating unit, the target feature vector is obtained. Finally, the output mapping layer... Perform the calculation, assuming the weights of this layer. For identity matrix, bias Given the range [-0.5, -1.0], the prediction result is calculated as Sigmoid([-2.5, -5.0] + [-0.5, -1.0]), which is equivalent to Sigmoid([-3.0, -6.0]). The calculated [Sigmoid(-3.0), Sigmoid(-6.0)] is approximately equal to [0.047, 0.0025]. Therefore, the final output of the advertising effectiveness prediction is: predicted click-through rate of 0.047 and predicted conversion rate of 0.0025.

[0065] In one embodiment of the present invention, the sampling triggering module is configured to perform the following steps: The advertising performance prediction results are decomposed into two independent Bernoulli distributions corresponding to click events and conversion events. The Shannon entropy of the two independent Bernoulli distributions is calculated independently, and the two Shannon entropies are arithmetically averaged to generate the prediction uncertainty. It is determined whether the prediction uncertainty is higher than a preset uncertainty threshold. If the prediction uncertainty is higher than the uncertainty threshold, the corresponding user sample to be predicted is marked as an abnormal prediction sample. The abnormal prediction sample is sent to the data acquisition system for automated alignment and verification of offline attribution logs to generate real performance label data.

[0066] Specifically, after the real-time prediction server outputs the advertising performance prediction results, the uncertainty assessment system immediately quantifies the reliability of these results and determines whether to initiate an active sampling process based on this assessment. First, the uncertainty calculation module receives the advertising performance prediction results, which are two-dimensional probability vectors containing predicted click-through rate and predicted conversion rate. This two-dimensional probability vector is decomposed into two independent Bernoulli distributions, corresponding to the probability distributions of click events and conversion events, respectively. Then, using information entropy as a measure of uncertainty, the Shannon entropy of these two Bernoulli distributions is calculated independently. The two calculated entropy values ​​are arithmetically averaged to obtain a comprehensive entropy value, which is defined as the prediction uncertainty, measuring the uncertainty of the current prediction result. A higher entropy value indicates lower prediction reliability. Prediction Uncertainty The calculation formula is as follows:

[0067] in The information entropy function for a Bernoulli distribution is defined as:

[0068] In the formula, and These represent the predicted click-through rate and predicted conversion rate in the advertising effectiveness prediction results, respectively. and These are the prediction uncertainties for click events and conversion events, respectively. This is the average of the two values, serving as a comprehensive measure of uncertainty. This value is a dimensionless scalar, ranging from 0 to 1. In this embodiment, prediction uncertainty specifically refers to the uncertainty score of the prediction result; the higher the value, the more uncertain the model is about the prediction result, i.e., the higher the uncertainty. It should be noted that information entropy is an indicator used in probability theory to measure the degree of uncertainty of a random variable. When the probability distribution is (0.5, 0.5), the entropy reaches its maximum value of 1, indicating the highest uncertainty.

[0069] Next, the sample screening module compares the calculated prediction uncertainty with a preset uncertainty threshold. The preset uncertainty threshold is a floating-point number between 0 and 1, typically set in the higher range of 0.85 to 0.98 based on business needs. This threshold is set to balance log verification load with model improvement benefits; a higher threshold means the system only needs to perform attribution point alignment confirmation on a very small number of samples with high prediction uncertainty. The system determines whether the prediction uncertainty exceeds the preset uncertainty threshold. If the result is yes, the current sample is identified as a high-uncertainty sample and screened as an anomalous prediction sample. An anomalous prediction sample is one whose prediction probability is close to random guessing, thus possessing high informational value. The data packet for this anomalous prediction sample not only contains the original input feature vector and the model's prediction results but also includes its prediction uncertainty score.

[0070] Finally, the active sampling module asynchronously sends the selected anomaly prediction samples to the data acquisition system via message queue or remote procedure call. The data acquisition system is an independent background service responsible for managing and scheduling offline data verification. Upon receiving the samples, the data acquisition system places them in a high-priority labeling queue to trigger the offline delayed attribution verification process, ultimately generating true performance label data for subsequent model fine-tuning. The true performance label data is determined by the system based on the attribution tracking logs from the user's terminal device, automatically comparing and verifying the user's subsequent actual operational behavior; for example, it may contain structured data such as {click:1, conversion:0}.

[0071] For example, the uncertainty assessment system receives a request from the user. The predicted advertising effectiveness is [0.047, 0.0025]. The uncertainty calculation module first calculates the information entropy of click-through rate and conversion rate respectively. For click-through rate... Its information entropy Regarding conversion rate Its information entropy Then, the prediction uncertainty was calculated. Assuming a preset uncertainty threshold for 0.9. The sample selection module determined that 0.149 was not higher than 0.9, therefore this sample was considered a low-uncertainty sample and did not trigger active sampling. Assume that in another prediction instance, the model... The prediction result is [0.52, 0.60]. The calculated prediction uncertainty is... Because the uncertainty value of 0.985 is higher than the preset uncertainty threshold of 0.9, this sample is filtered out as an abnormal prediction sample. The active sampling module will include user... The data packet containing the feature data and prediction results [0.52, 0.60] is sent to the data acquisition system. The system backend pushes this task to the underlying log attribution verification pipeline. Based on the real terminal playback data collected by the system backend, it is verified that the user did not actually click but completed the conversion, thereby generating the real performance label data {features:[...],label:[0,1]}.

[0072] In one embodiment of the present invention, the meta-knowledge update module is configured to perform the following steps: The advertising effectiveness prediction results of the abnormal prediction samples and the true effectiveness label data are input into the binary cross-entropy loss function to calculate the cross-entropy loss gradient of the current prediction task. The cross-entropy loss gradient is used to perform an internal loop update on the specific instance parameters in the advertising effectiveness prediction model to obtain temporary adaptation parameters. The reserved validation set is extracted, and the advertising effectiveness prediction model with temporary adaptation parameters is applied to the validation set for prediction calculation to generate meta-loss values. The meta-loss values ​​are backpropagated along the dynamic computation graph to the meta-algorithm library and the temporal causal graph builder to update the learnable meta-parameters and the network weights of the temporal causal graph builder.

[0073] Specifically, after obtaining the true performance label data, the meta-update server executes the final step of the method of this invention to achieve iterative optimization of the meta-knowledge of the entire prediction system. First, the loss calculation module compares the advertising performance prediction results of the abnormal prediction samples with the true performance label data. Specifically, it uses the model's two-dimensional prediction probability vector and the two-dimensional true label vector as input, applies the binary cross-entropy loss function, and calculates the cross-entropy loss value for the current prediction task. It should be understood that the core update logic of this step is defined by the following set of formulas. First, calculate the task loss. :

[0074] in, For the specific instance parameters, The aforementioned actual performance tag data corresponds to the actual tags for clicks and conversions mentioned earlier. This represents the advertising effectiveness prediction result. The numerical elements are... This refers to the predicted click-through rate mentioned earlier. Numerical elements This refers to the predicted conversion rate mentioned above. .

[0075] Based on this loss value, the gradient of the loss relative to each specific instance parameter in the advertising effectiveness prediction model is calculated using an automatic differentiation mechanism; this is the cross-entropy loss gradient. The cross-entropy loss gradient is a vector with the same dimension as the model parameters, indicating the direction of parameter adjustment that can most quickly reduce the prediction error of the current sample. Subsequently, the internal update module performs a simulated update, also known as the internal loop update. The internal loop update is a simulated, in-task learning process designed to evaluate the plasticity of the current meta-parameters. This module uses the just-calculated cross-entropy loss gradient to perform a gradient descent step on the specific instance parameters, thereby obtaining a set of temporary adaptive parameters. This step can be expressed by the following formula to obtain the temporary adaptive parameters. :

[0076] in The learning rate is the internal loop rate. The temporary fitting parameters are the parameter states obtained after the model makes a one-step adjustment to fit the current outlier sample.

[0077] Next, the reserved validation set is extracted from the data cache of the target heterogeneous channel. The model, equipped with temporary adaptive parameters, is applied to the validation set for prediction, and its average loss on that set is calculated; this loss is the meta-loss value. The validation set is a small batch of labeled data drawn from the same source that has not participated in the internal loop update. The meta-loss value measures the performance of the model after one adaptive adjustment step on unseen data and is the true optimization target of meta-learning. The calculation method is as follows:

[0078] In the formula, The verification set One of the sample instances, where For the input features of this sample, This is the true label corresponding to the sample. Indicates the use of parameters The cross-entropy task loss generated by the model predicting the sample.

[0079] Finally, the meta-update module performs backpropagation meta-update. Backpropagation meta-update is a second-order gradient optimization process aimed at optimizing the initial meta-parameters to produce better temporary fit parameters. The meta-loss value is used as the final optimization objective, and its gradient is propagated back along the entire model generation path. This gradient flow first passes through the inner loop update step, then through the Hadamard product operation during parameter instantiation, ultimately tracing back to its source: the learnable meta-parameters in the meta-algorithm library and the network weights of the graph neural network in the temporal causal graph builder. Its update process can be represented as:

[0080]

[0081] In the formula, Represents learnable meta-parameters. Represents the weights of a graph neural network. The learning rate is denoted as . All loss values ​​in the formulas are dimensionless scalars, and the parameters and gradients are tensors. Using this backtracked gradient, the learnable meta-parameters and network weights are updated through the meta-optimizer, thereby generalizing the experience learned from the anomalous sample and precipitating it as the system's meta-knowledge, completing the meta-knowledge deposition loop of the entire prediction system. This is the meta-knowledge deposition loop, which means that by handling prediction failures, the system not only corrects the current task but also improves its ability to handle all similar tasks in the future.

[0082] For example, following the previous step, the meta update server receives the user's... The predicted results [0.52, 0.60] and their corresponding true performance label data [0, 1] are used. The loss calculation module first calculates the cross-entropy loss. The system calculates the cross-entropy loss gradient based on this. Assuming the specific instance parameter of the current attention operator is 0.375, the corresponding gradient is calculated to be 0.9. The internal update module uses the learning rate... The update was performed, resulting in a temporary adaptation parameter of 0.375 - 0.01 × 0.9 = 0.366. Subsequently, the model equipped with this set of temporary adaptation parameters was used to analyze user data. The validation set is used for prediction, and the meta-loss is calculated to be 0.8. Finally, the meta-update module backpropagates the meta-loss of 0.8. Assuming this gradient is ultimately backpropagated to the corresponding learnable meta-parameter in the meta-algorithm library, whose original value is 0.5, the calculated meta-gradient is 0.2. The meta-learning rate is used... The learnable meta-parameter is updated, with a new value of 0.5 - 0.001 × 0.2 = 0.4998. Simultaneously, the weights of the graph neural network in the temporal causal graph builder are fine-tuned in the same way. At this point, the system has completed the meta-knowledge learning and accumulation for this anomaly prediction.

[0083] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0084] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A multi-channel advertising performance prediction system integrating meta-learning, characterized in that, include: The timing diagram construction module is used to acquire the raw event sequence data stream and construct a dynamic timing diagram using the raw event sequence data stream. The signature generation module is used to input dynamic time series graphs into a preset time series causal graph builder to generate causal graph signatures that include graph topology and probability matrices. The operator activation module is used to input the causal graph signature into the preset meta-algorithm configurator, activate the basic signal processing operators in the preset meta-algorithm library, and extract the learnable meta-parameters within the basic signal processing operators. The model building module connects basic signal processing operators based on the graph topology to generate a dynamic computation graph. Specific instance parameters are then injected into the dynamic computation graph to generate an advertising effectiveness prediction model. The performance prediction module is used to obtain standard input feature vectors and input them into the advertising performance prediction model to generate advertising performance prediction results; The sampling trigger module is used to calculate the prediction uncertainty of the advertising effectiveness prediction results, filter out abnormal prediction samples based on the prediction uncertainty, label them, and generate real effectiveness label data. The meta-knowledge update module calculates the meta-loss value based on the advertising effectiveness prediction results of the abnormal prediction samples and the actual effectiveness label data. It then uses the meta-loss value to update the learnable meta-parameters and the network weights of the time-series causal graph builder.

2. The multi-channel advertising performance prediction system integrating meta-learning according to claim 1, characterized in that, Constructing a dynamic time series graph using the original event sequence data stream includes the following steps: Parse the original event sequence data stream, extract user exposure event data, click event data, and conversion event data, and extract the timestamp and user identification corresponding to each event; Cluster user exposure event data, click event data, and conversion event data using user identity identifiers to generate a time-series behavior trajectory sequence; Abstract each event in the temporal behavior trajectory sequence into a graph node, and calculate the difference in timestamps between adjacent event nodes; Determine if the difference is less than the preset maximum time interval threshold. If the difference is less than the maximum time interval threshold, establish a directed edge between adjacent event nodes and generate a dynamic time sequence graph.

3. The multi-channel advertising performance prediction system integrating meta-learning according to claim 1, characterized in that, Generating a causal graph that includes the graph topology and probability matrix involves the following steps: The graph neural network in the temporal causal graph builder is invoked to aggregate node features of the dynamic temporal graph and extract the time delay distribution features and path propagation features represented by directed edges; Input the time delay distribution characteristics and path propagation characteristics into the probability calculation unit in the temporal causal graph builder to calculate the directed propagation probability between each graph node; A probability matrix is ​​generated using the directed transitive probabilities between nodes in the graph. Extract the graph topology of the dynamic time series graph, concatenate the graph topology with the probability matrix, and generate a causal graph signature.

4. The multi-channel advertising performance prediction system based on integrated meta-learning according to claim 1, characterized in that, Extracting learnable meta-parameters within basic signal processing operators includes the following steps: Analyze the graph topology in the causal graph signature and calculate the temporal dependency strength and feature cross frequency corresponding to the graph topology. The matching weights of each operator in the meta-algorithm library are calculated by linearly weighting the temporal dependency strength and the feature cross frequency. Compare the matching weights with the preset activation thresholds, and filter out target operators whose matching weights exceed the activation thresholds; The target operator is activated as a basic signal processing operator, and the learnable meta-parameters inherent in the basic signal processing operator are extracted.

5. The multi-channel advertising performance prediction system integrating meta-learning according to claim 1, characterized in that, Generating an advertising effectiveness prediction model includes the following steps: Map the graph topology to computation graph node connection rules; The basic signal processing operators are connected in series and parallel according to the node connection rules of the computation graph to generate a dynamic computation graph with a specific information flow path; Perform the Hadamard product operation between the probability matrix and the learnable meta-parameters, and use the result of the Hadamard product operation as the specific instance parameter; Specific instance parameters are injected into the corresponding operator nodes in the dynamic computation graph to generate an advertising effectiveness prediction model.

6. The multi-channel advertising performance prediction system based on integrated meta-learning according to claim 1, characterized in that, Generating advertising effectiveness prediction results includes the following steps: Collect the current context features and real-time interaction history sequence of the user to be predicted, and perform standardization processing on the current context features and real-time interaction history sequence to generate a standard input feature vector; The standard input feature vector is input into the advertising effectiveness prediction model, and the feature mapping and attenuation operations are performed sequentially through the basic signal processing operators along the information flow path of the dynamic computation graph to generate the target feature vector. Normalization mapping is performed on the target feature vector to generate a two-dimensional probability vector representing the predicted click-through rate and the predicted conversion rate; the two-dimensional probability vector is used as the result of advertising effectiveness prediction.

7. The multi-channel advertising performance prediction system integrating meta-learning according to claim 1, characterized in that, Generating realistic performance label data includes the following steps: The advertising effectiveness prediction results are decomposed into two independent Bernoulli distributions corresponding to click events and conversion events; Calculate the Shannon entropy of two independent Bernoulli distributions independently, and then calculate the prediction uncertainty by arithmetically averaging the two Shannon entropies. Determine whether the prediction uncertainty is higher than the preset uncertainty threshold. If the prediction uncertainty is higher than the uncertainty threshold, mark the corresponding user sample to be predicted as an abnormal prediction sample. Anomaly prediction samples are sent to the data acquisition system for automated alignment and verification of offline attribution logs, generating real performance label data.

8. The multi-channel advertising performance prediction system based on integrated meta-learning according to claim 1, characterized in that, Updating the learnable meta-parameters and network weights of the time-series causal graph builder using the meta-loss value includes the following steps: Input the advertising effectiveness prediction results of the abnormal prediction samples and the actual effectiveness label data into the binary cross-entropy loss function to calculate the cross-entropy loss gradient of the current prediction task. The cross-entropy loss gradient is used to perform an internal loop update on the specific instance parameters in the advertising effectiveness prediction model to obtain temporary adaptation parameters. Extract the reserved validation set, apply the advertising effectiveness prediction model with temporary adaptation parameters to the validation set for prediction calculation, and generate meta-loss value; The meta-loss value is backpropagated along the dynamic computation graph to the meta-algorithm library and the temporal causal graph builder, updating the learnable meta-parameters and the network weights of the temporal causal graph builder.

9. The multi-channel advertising performance prediction system based on integrated meta-learning according to claim 4, characterized in that, The matching weights of each operator in the meta-algorithm library are calculated by linearly weighting the temporal dependency strength and the feature cross-frequency, including the following steps: Extract the preset first hyperparameter, second hyperparameter, third hyperparameter, and fourth hyperparameter; Multiply the first hyperparameter by the temporal dependency strength to calculate the first matching weight for generating the time decay operator; The second hyperparameter is multiplied by the feature cross frequency to calculate the second matching weight for generating the feature cross attention operator; The product of the third hyperparameter and the temporal dependency strength is added to the product of the fourth hyperparameter and the feature cross frequency to generate the third matching weight of the information flow gating operator. The first matching weight, the second matching weight, and the third matching weight are used as the matching weights of the corresponding operators.

10. A multi-channel advertising performance prediction system integrating meta-learning according to claim 8, characterized in that, The cross-entropy loss gradient is used to perform an inner loop update on the parameters of specific instances in the advertising effectiveness prediction model to obtain temporary adaptation parameters, including the following steps: Extract the preset internal loop learning rate; multiply the internal loop learning rate with the cross-entropy loss gradient to generate a parameter adjustment step size tensor; The difference between the specific instance parameters and the parameter adjustment step tensor is calculated to generate an updated parameter matrix; the updated parameter matrix is ​​then used as the temporary adaptation parameters.

Images