Artificial intelligence-based advertisement planning system and method

Abstract

The present application relates to the technical field of advertising planning, in particular to an advertising planning system and method based on artificial intelligence, in which, sudden and regular events are captured in real time through a distributed message middleware, the micro-window length is dynamically adjusted based on the event density of adjacent time units to extract the burst event propagation rate index, and the macro-window threshold is controlled in linkage to determine the batch processing time of regular events, so as to realize efficient coordination of event flow; further, the burst event propagation rate index is used to dynamically adjust the text modal attention weight, a time-space attention mask is constructed by combining the time-effect label to suppress the interference of non-associated areas, and a time-space constrained fusion feature vector is generated; finally, the vector is mapped to a strategy life cycle decay parameter, and the diffusion potential is predicted to dynamically optimize the advertising strategy effective length, so as to improve the timeliness, accuracy and response sensitivity to sudden events of advertising delivery.

Description

Technical Field

[0001] This invention relates to the field of advertising planning technology, specifically to an advertising planning system and method based on artificial intelligence. Background Technology

[0002] Current advertising planning systems generally use static rules or offline models to generate advertising strategies, making it difficult to capture the propagation characteristics of sudden events in real time, resulting in delayed advertising response or wasted resources. Traditional methods lack dynamic coordination mechanisms for handling routine and sudden events, and often cannot adapt to the fluctuations in event density due to fixed window divisions.

[0003] Furthermore, the lack of consideration for spatiotemporal correlation constraints during multimodal feature fusion leads to an imbalance in feature weight allocation and insufficient accuracy in predicting the diffusion potential of advertising strategies. Existing technologies typically employ fixed lifecycle parameters in their strategy decision-making modules, failing to dynamically adjust the strategy's effective duration based on event propagation rates, thus impacting the timeliness and accuracy of ad delivery. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based advertising planning system and method to address the problems mentioned in the background section. Specific technical problems include: how to dynamically capture the propagation rate of sudden events and adaptively adjust the event processing window to solve the timing control problem of coordinating the processing of regular and sudden events; how to dynamically fuse multimodal data based on event propagation characteristics to suppress interference from unrelated areas and improve the spatiotemporal relevance of advertising strategies; and how to predict the strategy lifecycle based on the event propagation potential to achieve dynamic optimization of the effective duration of advertising strategies.

[0005] To achieve the above objectives, a technical solution discloses an artificial intelligence-based advertising planning system, including a real-time event stream dynamic processing module, a multi-modal feature adaptive fusion module, and a deep reinforcement strategy decision-making module, wherein: The real-time event stream dynamic processing module creates independent topic channels through a distributed message middleware to capture sudden events and regular events respectively, realizing data splitting and parallel processing. The dynamic micro-window processing unit in the real-time event stream dynamic processing module slices sudden events into fixed time units, generates histograms by statistical density, determines the baseline density based on a specified quantile value (such as 75%), and calculates the scaling factor according to the density ratio of adjacent time units: when the ratio is ≥1, an exponential function is used to calculate the upper limit value to shorten the micro-window, and when the ratio is <1, a linear decay is used to calculate the lower limit value to extend the micro-window, thereby dynamically adjusting the micro-window duration and generating a propagation rate index that is positively correlated with the current density / baseline density.

[0006] The macro window batch triggering unit in the real-time event stream dynamic processing module stores the propagation rate index, extracts historical quantile values ​​as benchmark values ​​and peak values, and dynamically adjusts the macro window duration through linear interpolation. If the propagation rate exceeds the benchmark value, the macro window is shortened to control the batch processing frequency of regular events. The output unit in the real-time event stream dynamic processing module integrates micro / macro window duration, real-time effect markers, and propagation rate indicators to form a dynamic feature vector; The real-time event stream dynamic processing module enables microsecond-level response to sudden events, and the processing frequency of regular events adaptively adjusts with the sudden pressure, precisely controlling the timing of regular event processing and avoiding resource conflicts. The multimodal feature adaptive fusion module dynamically adjusts the attention weights of text modalities based on the propagation rate index and parses the timeliness markers in the dynamic feature vector to construct a spatiotemporal attention mask: mapping the micro-window duration to a time decay coefficient to suppress historical events exceeding the threshold; prioritizing spatial regions based on the macro-window duration and applying mask decay to non-hotspot regions; finally, fusing the text attention weights and the spatiotemporal mask to output a fused feature vector containing spatiotemporal constraints; the multimodal feature adaptive fusion module reduces interference from features in unrelated regions, improves the matching accuracy between advertising strategies and event spatiotemporal hotspots, and increases the efficiency of multimodal data fusion to the millisecond level.

[0007] The deep reinforcement strategy decision-making module maps the fusion features to parameters of the strategy lifecycle decay function, combines the propagation rate index to predict the diffusion potential value, and derives the dynamic effective duration. For example, it calculates the diffusion potential by the peak ratio of sudden events and generates priority labels by combining the baseline threshold of the decay function to ensure that the strategy matches the event propagation trend.

[0008] The deep enhancement strategy decision-making module maps the fused feature vectors to input parameters of the strategy lifecycle decay function, including the initial influence coefficient (positively correlated with the propagation rate), decay rate (negatively correlated with the time mask), and effective baseline threshold (related to the spatial mask). At the same time, it calculates the diffusion potential value based on the ratio of the propagation rate index to the historical peak value (the potential value doubles when the ratio is >1), and derives the effective duration of the strategy by combining the decay function (such as the exponential model). Finally, it outputs the advertising strategy instruction containing dynamic priority tags and effective duration.

[0009] The enhanced strategy decision-making module reduces the prediction error rate of advertising strategy effectiveness duration and increases the intensity of resource allocation for high-potential events.

[0010] Another technical solution discloses an AI-based advertising planning method, specifically a method using an AI-based advertising planning system, with the following steps: S1. Step: Utilize distributed message middleware to separate and capture sudden events and regular events. Dynamically scale the micro-window duration based on the density of sudden events in adjacent time units (e.g., shorten the window to the second level when the density increases sharply). Calculate the propagation rate index of sudden events (positively correlated with the current density / baseline density). Combine this index with a macro-window trigger threshold to control the timing of batch processing of regular events (e.g., suspend regular processing when the propagation rate is high). Output a dynamic feature vector carrying a time stamp (micro / macro window duration) and the propagation rate index to achieve dynamic allocation of event processing resources and real-time feature extraction. S2. Step 1: Dynamically enhance the keyword attention weight of the text modality based on the propagation rate index. At the same time, analyze the timeliness mark to construct a spatiotemporal attention mask (such as mapping the micro-window duration to a time decay coefficient to suppress old events, and dividing the macro-window duration into spatial priority to decay non-hotspot area features). Integrate text and spatiotemporal constraints to generate highly correlated feature vectors, so that the advertising strategy is strongly bound to the spatiotemporal hotspots of the event. Step S3 maps the fused features to the initial influence coefficient, decay rate, and effective baseline threshold of the strategy lifecycle decay function. It combines the ratio of the propagation rate index to the historical peak value to predict the diffusion potential value and finally outputs the dynamically optimized advertising strategy instruction. This achieves millisecond-level response to sudden events, precise matching of advertising strategies with spatiotemporal scenarios, and adaptive adjustment of the strategy's effective duration, forming a closed-loop dynamic optimization from event perception to strategy delivery.

[0011] Compared with the prior art, the beneficial effects of the present invention are: By leveraging a dynamic micro-window and macro-window collaboration mechanism, the real-time processing of sudden events and the batch processing efficiency of routine events are improved; adaptive feature fusion based on spatiotemporal attention masks enhances the relevance of advertising strategies to hot events and regions; and the combination of strategy lifecycle decay function and diffusion potential prediction enables dynamic optimization of ad effectiveness duration, thereby improving targeting accuracy. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the overall modules of the present invention. Figure 2 This is a schematic diagram of the overall method flow of the present invention.

[0013] In the diagram: 100, Real-time event stream dynamic processing module; 101, Data stream capture unit; 102, Dynamic micro-window processing unit; 103, Macro window batch triggering unit; 104, Output unit; 200, Multimodal feature adaptive fusion module; 300, Deep reinforcement strategy decision-making module. Detailed Implementation

[0014] The technical solutions in 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.

[0015] Next, please refer to Figure 1 One of the objectives of this embodiment is to provide an artificial intelligence-based advertising planning system, including a real-time event flow dynamic processing module 100, a multimodal feature adaptive fusion module 200, and a deep reinforcement strategy decision-making module 300.

[0016] The data stream capture unit 101 in the real-time event stream dynamic processing module 100 deploys Apache Kafka distributed message middleware to create independent topic channels to capture sudden events (such as hot news) and regular events (such as user behavior logs). Apache Kafka distributed message middleware is an open-source stream processing platform for building real-time data pipelines and service applications.

[0017] The dynamic micro-window processing unit 102 in the real-time event stream dynamic processing module 100 extracts the propagation rate index of sudden events according to dynamically adjusted micro-windows. The duration of the micro-window is automatically scaled based on the time unit density of adjacent sudden events, specifically including: The sudden event is sliced ​​into fixed time units (e.g., 1 minute), and the arrival volume of the sudden event in each time unit is counted. The sudden event density is obtained by dividing by the fixed time unit duration, and a sudden event density distribution histogram is generated. The frequency of each density value is recorded and used to calculate and determine the baseline density of the sudden event (take the 30th percentile value of the sudden event density distribution histogram).

[0018] The duration of the micro-window is dynamically adjusted based on the burst event density of adjacent time units. The ratio R of the burst event density of the current time unit to that of the previous time unit is calculated. The scaling factor α is then dynamically calculated based on this ratio R, where: When the ratio R≥1 (density increases), α=min(e^{k*(R-1)},α_{max}), where k is the sensitivity coefficient (recommended 0.5-2), and α_max is the maximum scaling factor (e.g. 10). When the ratio R < 1 (density decreases), α = max(β*R, α_{min}), where β is the attenuation coefficient (recommended 0.3-0.8), and α_min is the minimum window ratio (e.g., 0.1). Dynamically adjust the duration of the micro-window according to the scaling factor α. Multiply the duration of the current micro-window by the scaling factor α to obtain the adjusted duration T_new of the micro-window. Extract the propagation rate index of the sudden event according to the adjusted duration of the micro-window. The extraction formula is as follows: V = N / T_new × D_c / D_b, where V represents the propagation rate index of the sudden event, N represents the total number of sudden events within the micro-window; T_new represents the adjusted duration of the micro-window; D_c represents the density of sudden events in the current time unit, and D_b represents the baseline density of sudden events.

[0019] The micro-window in the dynamic micro-window processing unit 102 is used to monitor and dynamically adjust the processing rhythm of sudden events in real time. Its window duration is automatically scaled according to the event density to accurately calculate the propagation rate index V of sudden events; dynamically adjust the micro-window size through an exponential or linear function to quickly respond to changes in burst traffic and ensure that key events are processed first.

[0020] The macro-window batch trigger unit 103 in the real-time event flow dynamic processing module 100 inputs the propagation rate index of the sudden event into the macro-window trigger threshold calculation to control the batch processing timing of regular events, specifically including: [[ID=eleven]] Whenever the dynamic micro-window processing unit 102 calculates a propagation rate index V of a sudden event, store this value together with the time stamp into the time series database; and retrieve all the stored V values within the past 24 hours from the time series database to form a historical data set; Sort the extracted V values from smallest to largest, and take the value at the 30% position as the baseline value V_b; it means that within 30% of the time, the propagation rate of sudden events is lower than this value, which is used to judge the "low activity period"; take the value at the 99% position as the peak value V_p, indicating that within 99% of the time, the V value is lower than this threshold, and exceeding this value is regarded as an "extreme peak" and needs to be urgently processed.

[0021] Dynamically adjust the macro-window duration T_macro according to the baseline value V_b and the peak value V_p to control the batch processing frequency of regular events, where: When V ≥ V_p (extreme peak): Set the macro-window duration T_macro to the minimum value (such as 10 seconds) to process regular events as soon as possible and release resources; When V ≤ V_b (low activity period), set the macro-window duration T_macro to the maximum value (such as 5 minutes) to reduce the processing frequency to save resources; When V_b < V < V_p (normal fluctuation), linearly interpolate and dynamically adjust the macro-window duration T_macro between the minimum and maximum values.

[0022] The macro window in the macro window batch trigger unit 103 dynamically controls the batch processing frequency of regular events (such as user logs) based on the burst event propagation rate V output by the micro window and its quantile values (V_b / V_p), balancing resource allocation; during peak bursts, the macro window is shortened to quickly clear regular events, and during low activity periods, the macro window is extended to save resources, achieving global load adaptability.

[0023] The output unit 104 in the real-time event flow dynamic processing module 100 outputs a dynamic feature vector containing a timeliness mark and a burst event propagation rate index, where the timeliness mark includes the duration T_new of the adjusted micro window and the macro window duration T_macro.

[0024] The multi-modal feature adaptive fusion module 200 dynamically adjusts the attention weight of the text modality according to the burst event propagation rate index, specifically including: According to the numerical range of the burst event propagation rate index V (low activity period V≤V_b, normal fluctuation V_b<V<V_p, extreme peak V≥V_p), a piecewise function is used to dynamically adjust the attention weight of the text modality; When V≥V_p, the system determines it as an extreme peak. At this time, the attention weight of the text modality is reduced to 30%-50% of the base value to reduce the interference of non-critical text features; When V≤V_b, the attention weight is increased to 120%-150% to strengthen text semantic analysis to compensate for event sparsity; When V_b<V<V_p, the attention weight is smoothly adjusted near the reference value according to the linear ratio of V to V_b / V_p. The weight calculation introduces logarithmic normalization to ensure the consistency of sensitivity for different magnitudes of V values, and finally outputs a text feature weight vector adapted to the current event density.

[0025] The multi-modal feature adaptive fusion module 200 analyzes the timeliness mark in the dynamic feature vector and constructs a spatio-temporal attention mask for suppressing features in non-associated regions, specifically including: By analyzing the timeliness mark (adjusted micro window duration T_new and macro window duration T_macro) in the dynamic feature vector, a spatio-temporal attention mask is constructed; First, map T_new to a time decay coefficient β_t = 1 / (1 + log(T_new)) to suppress the stale event features in historical time units that exceed 3 times T_new; At the same time, divide the spatial region priority according to T_macro. If T_macro is at the minimum value (10 seconds), apply a mask attenuation of 0.2-0.5 to non-hot geographical regions (such as low-frequency IP segments in user behavior logs).

[0026] If T_macro is the maximum value (5 minutes), then only the core region (such as the high-frequency location of breaking news) is retained. The mask achieves a smooth spatial transition through the Gaussian kernel function, and finally generates a binary matrix with spatiotemporal non-uniform suppression characteristics, which serves as a spatiotemporal attention mask.

[0027] The multimodal feature adaptive fusion module 200, based on the attention weights and spatiotemporal attention mask of the text modality, outputs a fused feature vector containing spatiotemporal constraints, specifically including: The attention weights of the text modality are subjected to tensor Hadamard product operation with the spatiotemporal attention mask to achieve feature fusion under dual constraints. Specifically, the text feature vector is first multiplied element-wise with the attention weight vector of the text modality, and then the suppressed spatiotemporal regions are filtered out by the spatiotemporal attention mask (features with a product result less than the threshold of 0.05 are set to zero). After normalizing the retained features, the burst event propagation rate index V is concatenated as a temporal enhancement factor to finally generate a fused feature vector containing spatiotemporal constraints.

[0028] The deep reinforcement strategy decision module 300 transforms the spatiotemporal constraint fusion feature vector into input parameters for the strategy lifecycle decay function, predicts the diffusion potential value of the advertising strategy based on the propagation rate index of sudden events, and outputs advertising strategy instructions containing the effective duration of the strategy, specifically including: The fused feature vector containing spatiotemporal constraints is mapped to three key parameters of the policy lifecycle decay function: the initial influence coefficient (based on L2 norm normalization of the feature vector), the decay rate (inversely proportional to the propagation rate index V of the sudden event, the higher V is, the slower the decay), and the effective baseline threshold (dynamically matching the V_b / V_p quantile values).

[0029] The diffusion potential value is calculated by weighting the ratio of the propagation rate V of the sudden event to the peak value V_p. If V≥V_p, the diffusion potential value is set to 1.0 (highest priority); otherwise, it is smoothly mapped to the 0.2-0.8 range by the Sigmoid function.

[0030] The final policy duration T_policy is derived by combining the decay function and the diffusion potential value. When the diffusion potential value > 0.7, T_policy = min(T_macro×2,1) to ensure continuous exposure of the policy during peak periods; when the diffusion potential value < 0.3, T_policy = max(T_new,30) to achieve rapid elimination.

[0031] The output instructions include the strategy ID, the strategy duration, and a dynamic priority tag, which are used by the ad engine for real-time scheduling.

[0032] Please see Figure 2The second objective of this embodiment is to provide a method for an artificial intelligence-based advertising planning system, including the following method steps: S1. Capture sudden events and regular events through a distributed message middleware; extract the sudden event propagation rate index for sudden events using dynamically adjusted micro-windows, the duration of which is automatically scaled based on the density of sudden events in adjacent time units; input the sudden event propagation rate index into a macro-window to trigger threshold calculation and control the timing of batch processing of regular events; output a dynamic feature vector containing time-sensitive markers and the sudden event propagation rate index. S2. Dynamically adjust the attention weight of the text modality according to the propagation rate index of the sudden event; parse the timeliness marker in the dynamic feature vector and construct a spatiotemporal attention mask for suppressing features of non-related regions; based on the attention weight of the text modality and the spatiotemporal attention mask, output a fusion feature vector containing spatiotemporal constraints. S3. Transform the spatiotemporal constraint fusion feature vector into the input parameters of the strategy lifecycle decay function, predict the diffusion potential value of the advertising strategy based on the sudden event propagation rate index, and output the advertising strategy instruction containing the effective duration of the strategy.

[0033] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An advertising planning system based on artificial intelligence, characterized in that, It includes a real-time event stream dynamic processing module (100), a multimodal feature adaptive fusion module (200), and a deep reinforcement strategy decision module (300), wherein: The real-time event stream dynamic processing module (100) captures sudden events and regular events through a distributed message middleware; extracts the sudden event propagation rate index for sudden events according to a dynamically adjusted micro-window, the duration of which is automatically scaled based on the density of sudden events in adjacent time units; inputs the sudden event propagation rate index into a macro-window to trigger threshold calculation, controlling the timing of batch processing of regular events; and outputs a dynamic feature vector containing time-sensitive markers and the sudden event propagation rate index. The multimodal feature adaptive fusion module (200) dynamically adjusts the attention weight of the text modality according to the sudden event propagation rate index; parses the timeliness marker in the dynamic feature vector, and constructs a spatiotemporal attention mask for suppressing features in non-related regions; based on the attention weight of the text modality and the spatiotemporal attention mask, it outputs a fusion feature vector containing spatiotemporal constraints. The deep enhancement strategy decision module (300) transforms the spatiotemporal constraint fusion feature vector into the input parameter of the strategy life cycle decay function, predicts the diffusion potential value of the advertising strategy based on the sudden event propagation rate index, and outputs the advertising strategy instruction containing the effective duration of the strategy.

2. The artificial intelligence-based advertising planning system according to claim 1, characterized in that, The real-time event stream dynamic processing module (100) includes a data stream capture unit (101), which deploys a distributed message middleware and creates independent topic channels to capture sudden events and regular events respectively.

3. The artificial intelligence-based advertising planning system according to claim 1, characterized in that, The real-time event stream dynamic processing module (100) includes a dynamic micro-window processing unit (102). The dynamic micro-window processing unit (102) dynamically adjusts the micro-window duration based on the density of sudden events in adjacent time units and calculates the sudden event propagation rate index, specifically including: The sudden events are sliced ​​into fixed time units, the density of sudden events within the time units is statistically analyzed, a histogram of sudden event density distribution is generated, and the baseline density of sudden events is calculated. The scaling factor is dynamically calculated based on the ratio of the density of sudden events in adjacent time units, and the duration of the micro-window is adjusted based on the scaling factor. The propagation rate index of sudden events is calculated using the adjusted micro-window duration, and its value is positively correlated with the ratio of the current density to the baseline density.

4. The artificial intelligence-based advertising planning system according to claim 3, characterized in that, The scaling factor calculation includes: When the ratio of the density of burst events in adjacent time units is greater than or equal to 1, the upper limit of the scaling factor is calculated using an exponential function. When the ratio of the density of sudden events in adjacent time units is less than 1, the lower limit of the scaling factor is calculated using a linear function combined with the attenuation coefficient.

5. The artificial intelligence-based advertising planning system according to claim 3, characterized in that, The baseline density of sudden events is determined by the specified quantile value of the sudden event density distribution histogram.

6. The artificial intelligence-based advertising planning system according to claim 1, characterized in that, The real-time event stream dynamic processing module (100) includes a macro window batch triggering unit (103). The macro window batch triggering unit (103) dynamically adjusts the macro window duration according to the quantile value of the propagation rate index of sudden events, and controls the batch processing frequency of regular events. Specifically, it includes: Store the propagation rate index of sudden events and extract specified quantile values ​​from historical datasets as baseline and peak values; Based on the comparison results of the propagation rate index of sudden events with the baseline and peak values, the duration of the macro window is dynamically adjusted using linear interpolation.

7. The artificial intelligence-based advertising planning system according to claim 1, characterized in that, The real-time event stream dynamic processing module (100) includes an output unit (104), which is used to output a dynamic feature vector containing a time-sensitive marker and a sudden event propagation rate index, wherein the time-sensitive marker includes the adjusted micro-window duration and the macro-window duration.

8. The artificial intelligence-based advertising planning system according to claim 1, characterized in that, The construction of the spatiotemporal attention mask includes: The adjusted micro-window duration is mapped to a time decay coefficient to suppress stale event features that exceed a set threshold. Spatial regions are prioritized based on the duration of the macro window, and non-hotspot regions are either masked for attenuation or their core features are preserved.

9. The artificial intelligence-based advertising planning system according to claim 1, characterized in that, The deep reinforcement strategy decision module (300) specifically includes: The fused feature vectors are mapped to the initial influence coefficient, decay rate, and effective baseline threshold of the policy lifecycle decay function. The diffusion potential value is calculated based on the ratio of the propagation rate index of the sudden event to the peak value, and the effective duration of the strategy is derived by combining the decay function. The output includes advertising strategy instructions with dynamic priority tags and effective duration.

10. A method using an artificial intelligence-based advertising planning system comprising any one of claims 1-9, characterized in that, The methods and steps include the following: S1. Capture sudden events and regular events through a distributed message middleware; extract the sudden event propagation rate index for sudden events according to a dynamically adjusted micro-window, the duration of which is automatically scaled based on the density of sudden events in adjacent time units; input the sudden event propagation rate index into a macro-window to trigger threshold calculation and control the timing of batch processing of regular events. The output is a dynamic feature vector containing timeliness markers and indicators of the propagation rate of sudden events; S2. Dynamically adjust the attention weight of the text modality according to the propagation rate index of the sudden event; parse the timeliness marker in the dynamic feature vector and construct a spatiotemporal attention mask to suppress features of non-related regions; Based on text modality attention weights and spatiotemporal attention masks, the output is a fused feature vector containing spatiotemporal constraints; S3. Transform the spatiotemporal constraint fusion feature vector into the input parameters of the strategy lifecycle decay function, predict the diffusion potential value of the advertising strategy based on the sudden event propagation rate index, and output the advertising strategy instruction containing the effective duration of the strategy.

Images