Precision marketing global AI video promotion method and device, equipment and medium

Abstract

The application relates to the technical field of video promotion. A precise marketing global AI video promotion method, device, equipment and medium are provided, wherein the method comprises the following steps: performing identity graph fusion processing on acquired multi-platform user behavior logs to generate a global interest ontology library; performing model construction processing on promotion effect data acquired by a preset double-channel feedback unit and the global interest ontology library to generate an adaptive attention prediction model; performing three-level trigger condition judgment processing on the adaptive attention prediction model to generate an evolution instruction; the adaptive attention prediction model is subjected to recombination processing in response to the evolution instruction to generate an evolved prediction model; and the output result of the evolved prediction model is subjected to multi-modal narrative unit recombination processing to generate a promotion video adapted to a target platform, so that the technical effects of improving user portrait integrity, enhancing model dynamic adaptability and optimizing cross-platform content adaptability are achieved.

Description

Technical Field

[0001] This invention relates to the field of video promotion technology, and in particular to methods, devices, equipment and media for precision marketing and AI-powered video promotion across the entire domain. Background Technology

[0002] With the rapid development of artificial intelligence and big data technologies, precision marketing is playing an increasingly crucial role in e-commerce, social media, digital advertising, and other fields. Omnichannel video promotion, as a core tool of modern marketing, directly impacts brand communication efficiency and user conversion rates.

[0003] Traditional techniques employ single-platform behavior analysis for user interest modeling. However, this approach suffers from fragmented user profiles due to the separation of identities across multiple platforms. In the face of dynamic market environments, traditional techniques use static prediction models to predict promotional effects, but these models suffer from degradation due to parameter rigidity and feature inertia. Furthermore, traditional techniques often overlook the adaptive requirements of cross-platform video generation, leading to platform adaptation barriers caused by content homogenization and rigid narrative structures. Summary of the Invention

[0004] Therefore, it is necessary to provide precise marketing AI video promotion methods, devices, equipment and media to address the above-mentioned technical issues, so as to achieve the technical effects of improving the completeness of user profiles, enhancing the dynamic adaptability of models, and optimizing cross-platform content adaptability.

[0005] Firstly, this application provides a precise marketing, full-domain AI video promotion method, which includes:

[0006] The acquired user behavior logs from multiple platforms are processed by identity graph fusion to generate a global interest ontology.

[0007] The promotion effect data and the whole domain interest ontology obtained by the preset dual-channel feedback unit are processed to generate an adaptive attention prediction model.

[0008] The adaptive attention prediction model is processed with three levels of trigger conditions to generate evolutionary instructions.

[0009] The adaptive attention prediction model responds to evolutionary instructions by undergoing recombination to generate an evolved prediction model.

[0010] The output of the evolved prediction model is reorganized into multimodal narrative units to generate promotional videos adapted to the target platform.

[0011] In one embodiment, the adaptive attention prediction model undergoes a three-level trigger condition judgment process to generate evolutionary instructions, including:

[0012] The predictive performance metrics of the adaptive attention prediction model are monitored and processed in real time to generate time series data of the model status.

[0013] Deviation fluctuation analysis is performed on the model state time series data. When the fluctuation amplitude exceeds the preset fluctuation threshold, a parameter adjustment trigger signal is generated.

[0014] Frequency detection processing is performed on emerging interest entities in the global interest ontology. When the frequency of occurrence exceeds a preset emergence threshold per unit time, a feature reconstruction trigger signal is generated.

[0015] The architecture degradation index of the adaptive attention prediction model is subjected to decay analysis. When the decay rate exceeds the preset architecture threshold, an architecture mutation trigger signal is generated.

[0016] The parameter adjustment trigger signal, feature reconstruction trigger signal, or architecture mutation trigger signal are processed by instruction encoding to generate evolution instructions.

[0017] In one embodiment, deviation fluctuation analysis is performed on the model state time series data. When the fluctuation amplitude exceeds a preset fluctuation threshold, a parameter adjustment trigger signal is generated, including:

[0018] The model state time series data is processed by sliding window segmentation to generate dynamic time segments;

[0019] For each dynamic time segment, the fluctuation acceleration is calculated and processed to generate the instantaneous fluctuation intensity;

[0020] The instantaneous fluctuation intensity of continuous dynamic time segments is processed by trend accumulation to generate a comprehensive fluctuation index;

[0021] The comprehensive volatility index is compared with a preset volatility threshold. When the comprehensive volatility index exceeds the preset volatility threshold, a parameter adjustment trigger signal is generated.

[0022] In one embodiment, frequency detection processing is performed on emerging interest entities in the global interest ontology database. When the frequency of occurrence per unit time exceeds a preset emergence threshold, a feature reconstruction trigger signal is generated, including:

[0023] Spatiotemporal density clustering is performed on emerging interest entities in the global interest ontology to generate a candidate entity set;

[0024] The candidate entity set is processed to construct an association network, generating an entity propagation graph.

[0025] Propagation path tracing is performed based on entity propagation graphs to generate entity influence diffusion rates.

[0026] When the diffusion rate of entity influence exceeds a preset emergence threshold, a feature reconstruction trigger signal is generated.

[0027] In one embodiment, the entity influence diffusion rate is obtained using the following formula, including:

[0028]

[0029] Among them, R diff The value represents the diffusion rate of entity influence, T represents the observation time window, K represents the total number of candidate entities, and α represents the weight of the time dimension. V represents the set of time-sensitive active entities. k Denotes the total set of related entities, Δt k E represents the mean propagation delay, β represents the topological dimension weight, and E k Let ω(e) represent the propagation edge set, and deg represent the propagation weight. max Indicates the maximum node degree. This represents the propagation speed factor.

[0030] In one embodiment, the promotion effect data acquired by the preset dual-channel feedback unit and the global interest ontology are processed to generate an adaptive attention prediction model, including:

[0031] Multi-dimensional feature decoupling processing is performed on the promotion effect data to generate explicit feedback feature vectors and implicit feedback feature vectors;

[0032] Spatiotemporal interest encoding is performed on the global interest ontology to generate a dynamic interest feature matrix;

[0033] Based on the attention gating mechanism, cross-domain feature fusion processing is performed on explicit feedback feature vectors, implicit feedback feature vectors, and dynamic interest feature matrices to generate joint feature representations;

[0034] By using a pre-defined neural network architecture search unit, the joint feature representation is subjected to adaptive network topology construction processing to generate an adaptive attention prediction model.

[0035] In one embodiment, identity graph fusion processing is performed on the acquired multi-platform user behavior logs to generate a global interest ontology, including:

[0036] Perform cross-domain behavior time sequence alignment processing on user behavior logs from multiple platforms to generate spatiotemporally related behavior sequences;

[0037] By using a pre-defined identity resolution rule base, device fingerprint clustering is performed on spatiotemporally related behavior sequences to generate cross-platform identity clusters.

[0038] Graph neural network embedding processing is applied to cross-platform identity clusters to generate dynamic identity graphs;

[0039] Based on the trained interest entity extraction model, multi-dimensional interest mining is performed on the dynamic identity graph to generate a global interest ontology library.

[0040] Secondly, this application also provides a precision marketing omni-channel AI video promotion device, which includes:

[0041] The identity fusion module is used to perform identity graph fusion processing on the acquired user behavior logs from multiple platforms to generate a global interest ontology library.

[0042] The adaptive modeling module is used to process the promotion effect data obtained from the preset dual-channel feedback unit and the full-domain interest ontology to generate an adaptive attention prediction model.

[0043] The evolution trigger module is used to perform three-level trigger condition judgment processing on the adaptive attention prediction model and generate evolution instructions.

[0044] The model recombination module is used to reorganize the adaptive attention prediction model in response to evolutionary instructions, generating an evolved prediction model.

[0045] The narrative generator module is used to reorganize the output of the evolved prediction model into multimodal narrative units to generate promotional videos adapted to the target platform.

[0046] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods in the first aspect of this application.

[0047] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods in the first aspect of this application.

[0048] The precise marketing full-domain AI video promotion method, device, equipment and medium provided in this application break down the user identity barriers between different platforms by performing identity graph fusion processing on the acquired user behavior logs from multiple platforms. It links and integrates the browsing, clicking, interaction and other behavioral data scattered across various platforms, thereby generating a full-domain interest ontology library that includes users' multi-dimensional preferences and long-term and short-term interests. This effectively makes up for the fragmentation problem of user profiles caused by traditional single-platform analysis and truly improves the completeness of user profiles.

[0049] Based on the promotion effect data obtained by the preset dual-channel feedback unit, the model is constructed in combination with the full-domain interest ontology to generate an adaptive attention prediction model. Then, a three-level trigger condition judgment is implemented for the model and an evolution command is generated based on the judgment result. This allows the model to respond to the evolution command in a timely manner when the prediction performance fluctuates, new user interests emerge, or the architecture degrades, and complete the reorganization to form an evolved prediction model. This avoids the defects of fixed parameters and rigid features in traditional static models and significantly enhances the dynamic adaptability of the model.

[0050] Based on the output of the post-evolutionary prediction model, the multimodal narrative units are reorganized. This allows for the generation of promotional videos that fit the characteristics of different target platforms, such as content style, user viewing habits, and duration requirements. This solves the problems of content homogenization and rigid narrative structure in traditional cross-platform promotion, and optimizes cross-platform content adaptability. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 A flowchart of a precision marketing full-domain AI video promotion method according to an embodiment of the present invention;

[0053] Figure 2 This is a flowchart of a process for performing deviation fluctuation analysis on model state time series data in one embodiment of the present invention, and generating a parameter adjustment trigger signal when the fluctuation amplitude exceeds a preset fluctuation threshold.

[0054] Figure 3 This is a structural diagram of a precision marketing full-domain AI video promotion device according to one embodiment of the present invention. Detailed Implementation

[0055] To make the above-mentioned objects, features, and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0056] First, the application scenarios of the embodiments of this application are described. In the embodiments of this application, a precise marketing AI video promotion method, apparatus, device, and medium applicable to, but not limited to, e-commerce omni-channel marketing promotion scenarios are provided. For example, for apparel products under the same brand, e-commerce platforms need to simultaneously promote them on multiple platforms such as Douyin, Xiaohongshu, and Taobao Live. This requires combining the browsing habits, interaction preferences, and purchasing behavior of users on different platforms to generate AI promotional videos adapted to each platform, in order to achieve cross-platform traffic conversion and enhance brand awareness.

[0057] The illustrative examples of the precision marketing full-domain AI video promotion method, apparatus, equipment, and medium provided in this application embodiment can also be applied to other application scenarios, such as the new product launch promotion scenario in the fast-moving consumer goods industry, where differentiated AI promotion videos can be customized for different social platforms by integrating user behavior data from multiple platforms; or the course enrollment promotion scenario in the education and training industry, where suitable AI promotion videos can be generated based on the learning needs and content consumption characteristics of users on each platform. These are merely illustrative examples and do not limit the specific application scenarios.

[0058] like Figure 1 As shown, this application provides a precise marketing omni-channel AI video promotion method, which includes:

[0059] S101: Perform identity graph fusion processing on the acquired user behavior logs from multiple platforms to generate a global interest ontology.

[0060] For example, the acquired multi-platform user behavior logs are preprocessed to filter out valid behavior records and remove redundant information. Cross-domain behavior time-series alignment is then performed on the preprocessed multi-platform user behavior logs to associate user behavior data from different platforms within the same time period, generating a spatiotemporally related behavior sequence. A preset identity resolution rule library is then invoked to perform device fingerprint clustering on the spatiotemporally related behavior sequences based on device identifiers, account association information, etc., classifying and aggregating cross-platform behavior data belonging to the same user to generate a cross-platform identity cluster.

[0061] Graph neural network embedding technology is used to extract features and model relationships among cross-platform identity clusters, constructing a dynamic identity graph that reflects the cross-platform relationships of users. A trained interest entity extraction model is then used to mine user interest features from the dynamic identity graph in multiple dimensions, identifying and integrating users' long-term interest preferences, short-term interest tendencies, and potential interest needs to generate a comprehensive interest ontology.

[0062] S102: Process the promotion effect data and the global interest ontology obtained from the preset dual-channel feedback unit to generate an adaptive attention prediction model.

[0063] For example, the promotion effect data acquired by the preset dual-channel feedback unit is decoupled into multi-dimensional features, decomposed into explicit feedback feature vectors and implicit feedback feature vectors. Spatiotemporal interest encoding is performed on the global interest ontology, transforming it into a structured dynamic interest feature matrix. Through an attention gating mechanism, cross-domain feature fusion is performed on the explicit feedback feature vectors, implicit feedback feature vectors, and dynamic interest feature matrix to construct a joint feature representation. A preset neural network architecture search unit is invoked to automatically complete network topology design, parameter initialization, and training optimization based on the joint feature representation. After iterative verification and achievement of standards, an adaptive attention prediction model is generated.

[0064] S103: Perform three-level trigger condition judgment processing on the adaptive attention prediction model to generate evolution instructions.

[0065] For example, the predictive performance metrics of the adaptive attention prediction model are monitored in real time, and model state time-series data is generated. Deviation fluctuation analysis is performed on the model state time-series data; if the fluctuation amplitude exceeds a preset fluctuation threshold, a parameter adjustment trigger signal is generated. Simultaneously, frequency detection is performed on emerging interest entities in the global interest ontology; if the frequency of occurrence per unit time exceeds a preset emergence threshold, a feature reconstruction trigger signal is generated. Attenuation analysis is performed on the architecture degradation metrics of the adaptive attention prediction model; if the attenuation rate exceeds a preset architecture critical value, an architecture mutation trigger signal is generated. The parameter adjustment trigger signal, feature reconstruction trigger signal, or architecture mutation trigger signal are encoded into instructions to generate evolutionary instructions.

[0066] S104: The adaptive attention prediction model responds to evolutionary instructions by undergoing recombination processing to generate an evolved prediction model.

[0067] For example, evolutionary instructions are parsed to identify the adjustment type. If it is a parameter adjustment instruction, the core parameters of the adaptive attention prediction model, such as network layer weights and activation function coefficients, are adaptively optimized. If it is a feature reconstruction instruction, an emerging interest feature dimension is added to the input layer of the adaptive attention prediction model, and the mapping rules of the feature processing module are updated. If it is an architecture mutation instruction, the network reconstruction module is called to redesign the network layer number, node connection method, and other topological structures of the adaptive attention prediction model. The recombined adaptive attention prediction model is then locally trained and validated, and once it meets the requirements, the evolved prediction model is generated.

[0068] S105: Perform multimodal narrative unit reorganization on the output of the evolved prediction model to generate a promotional video adapted to the target platform.

[0069] For example, the output of the evolved prediction model is extracted, and user attention trends, content preferences, and target platform adaptation requirements are analyzed. A pre-defined multimodal narrative unit library is invoked to filter matching video clips, text units, background music units, and visual effects units. Based on the target platform's duration guidelines, aspect ratio, and user viewing habits, the filtered multimodal narrative units are arranged chronologically. Next, the arranged multimodal narrative units are fused to ensure compatibility in content logic, visual rhythm, and core information delivery. The fused content undergoes integrity verification and quality review; once it meets the standards, a promotional video adapted for the target platform is generated.

[0070] The precise marketing full-domain AI video promotion method provided in one embodiment of this application breaks down the user identity barriers between different platforms by performing identity graph fusion processing on the acquired user behavior logs from multiple platforms. It integrates and correlates browsing, clicking, interaction and other behavioral data scattered across various platforms, thereby generating a full-domain interest ontology library that includes users' multi-dimensional preferences and long-term and short-term interests. This effectively makes up for the fragmentation problem of user profiles caused by traditional single-platform analysis and truly improves the completeness of user profiles.

[0071] Based on the promotion effect data obtained by the preset dual-channel feedback unit, the model is constructed in combination with the full-domain interest ontology to generate an adaptive attention prediction model. Then, a three-level trigger condition judgment is implemented for the model and an evolution command is generated based on the judgment result. This allows the model to respond to the evolution command in a timely manner when the prediction performance fluctuates, new user interests emerge, or the architecture degrades, and complete the reorganization to form an evolved prediction model. This avoids the defects of fixed parameters and rigid features in traditional static models and significantly enhances the dynamic adaptability of the model.

[0072] Based on the output of the post-evolutionary prediction model, the multimodal narrative units are reorganized. This allows for the generation of promotional videos that fit the characteristics of different target platforms, such as content style, user viewing habits, and duration requirements. This solves the problems of content homogenization and rigid narrative structure in traditional cross-platform promotion, and optimizes cross-platform content adaptability.

[0073] In one embodiment, the adaptive attention prediction model undergoes a three-level trigger condition judgment process to generate evolutionary instructions, including:

[0074] (1) Real-time monitoring and processing of the prediction performance indicators of the adaptive attention prediction model to generate model state time series data.

[0075] For example, the predictive performance indicators of the adaptive attention prediction model are monitored and processed in real time, including determining the monitoring dimensions and collection frequency of the predictive performance indicators, continuously collecting data such as error rate, accuracy, and recall during the operation of the adaptive attention prediction model, and organizing and storing the collected predictive performance indicator data in chronological order to generate ordered model state time series data.

[0076] Among them, the prediction performance indicators include indicators such as error rate, accuracy, and recall rate, which can reflect the prediction effect of the adaptive attention prediction model. Monitoring and processing need to be carried out continuously according to the set collection frequency to ensure the continuity of the model state time series data.

[0077] (2) Perform deviation fluctuation analysis on the model state time series data. When the fluctuation amplitude exceeds the preset fluctuation threshold, generate a parameter adjustment trigger signal.

[0078] For example, deviation fluctuation analysis is performed on the model state time series data, the analysis period of the model state time series data is divided, the fluctuation range and change trend of the model state time series data in each analysis period are calculated, the calculated fluctuation range is compared with the preset fluctuation threshold, and if the fluctuation amplitude exceeds the preset fluctuation threshold, the signal generation process is started to generate a parameter adjustment trigger signal.

[0079] (3) Frequency detection processing is performed on emerging interest entities in the global interest ontology. When the frequency of occurrence per unit time exceeds the preset emergence threshold, a feature reconstruction trigger signal is generated.

[0080] For example, frequency detection processing is performed on emerging interest entities in the global interest ontology. Recently added interest entities are selected from the global interest ontology as emerging interest entities. The time period for frequency detection is determined, and the number of occurrences of emerging interest entities in each time period is counted. The counted frequency of occurrence per unit time is compared with a preset emergence threshold. If the frequency of occurrence per unit time exceeds the preset emergence threshold, a feature reconstruction trigger signal is generated.

[0081] Among them, the selection of emerging interest entities should be based on the addition time and update records of interest entities in the whole domain interest ontology library, and the time period setting should meet the monitoring needs of user interest changes in order to accurately capture the spread of emerging interests.

[0082] (4) Perform attenuation analysis on the architecture degradation index of the adaptive attention prediction model. When the attenuation rate exceeds the preset architecture critical value, generate an architecture mutation trigger signal.

[0083] For example, the architecture degradation index of the adaptive attention prediction model is subjected to decay analysis processing to clarify the specific content of the architecture degradation index, including feature utilization, network layer response speed, model generalization ability, etc. The decay magnitude and decay rate of the architecture degradation index within a certain period of time are calculated, and the decay rate is compared with the preset architecture critical value. If the decay rate exceeds the preset architecture critical value, the architecture adjustment signal generation mechanism is triggered to generate an architecture mutation trigger signal.

[0084] Among them, the architecture degradation index should be able to reflect the operating status of the adaptive attention prediction model architecture, and the calculation of the decay rate should be based on the difference in the architecture degradation index values ​​and the time interval at different time points.

[0085] (5) Perform instruction encoding processing on parameter adjustment trigger signals, feature reconstruction trigger signals or architecture mutation trigger signals to generate evolution instructions.

[0086] For example, the parameter adjustment trigger signal, feature reconstruction trigger signal, or architecture mutation trigger signal are processed by instruction encoding to determine the mapping rules between the trigger signal and the evolution instruction, clarify the instruction code and instruction parameters corresponding to different types of trigger signals, and convert the parameter adjustment trigger signal, feature reconstruction trigger signal, or architecture mutation trigger signal into encoded information that conforms to the recognition format of the adaptive attention prediction model according to the mapping rules to generate evolution instructions.

[0087] Among these, the mapping rules need to be set in advance to ensure the accurate correspondence between the trigger signal and the evolutionary instruction, and the instruction encoding format needs to be compatible with the signal reception and parsing capabilities of the adaptive attention prediction model.

[0088] like Figure 2 As shown, deviation fluctuation analysis is performed on the model state time series data. When the fluctuation amplitude exceeds the preset fluctuation threshold, a parameter adjustment trigger signal is generated, including:

[0089] S201: Perform sliding window segmentation on the model state time series data to generate dynamic time segments.

[0090] For example, the model state time series data is processed by sliding window segmentation. The time length and sliding step size of the sliding window are determined to ensure that the window length can cover the effective fluctuation period of the model state time series data and the sliding step size can capture the continuous changes of the data. The model state time series data is included into the sliding window segment by segment according to the set sliding step size. The integrity of the data in each window is checked. After removing windows with missing or abnormal data, continuous and non-overlapping dynamic time segments are generated.

[0091] The sliding window duration needs to be set in conjunction with the acquisition frequency and fluctuation characteristics of the model state time series data. The sliding step size needs to be smaller than the window length to ensure the continuity of the dynamic time segment. The dynamic time segment needs to include the complete time interval and the model state data within the corresponding interval.

[0092] S202: Perform wave acceleration calculation on each dynamic time segment to generate instantaneous wave intensity.

[0093] For example, fluctuation acceleration calculation is performed on each dynamic time segment, the maximum, minimum and average values ​​of the model state data in each dynamic time segment are extracted, the change amplitude of the model state data in the dynamic time segment relative to the previous dynamic time segment is calculated, and the rate of increase or decrease of the change amplitude per unit time is calculated based on the change amplitude and the time interval between two dynamic time segments. This rate is used as the instantaneous fluctuation intensity of each dynamic time segment.

[0094] The calculation of the magnitude of change should be based on the average value of the model state data within the dynamic time segment, the time interval should be consistent with the acquisition cycle of the model state time series data, and the instantaneous fluctuation intensity should be able to reflect the degree of change of the model state data within a specific dynamic time segment.

[0095] S203: Perform trend accumulation processing on the instantaneous fluctuation intensity of continuous dynamic time segments to generate a comprehensive fluctuation index.

[0096] For example, the instantaneous fluctuation intensity of continuous dynamic time segments is processed by trend accumulation to determine the cumulative number of continuous dynamic time segments, ensuring that the cumulative number can reflect the stability of the fluctuation trend. The instantaneous fluctuation intensity of continuous dynamic time segments is weighted and summed, and different weights are assigned according to the order of the dynamic time segments. The instantaneous fluctuation intensity of recent dynamic time segments is given higher weights. The comprehensive fluctuation index is obtained after weighted summation.

[0097] The cumulative quantity needs to be set according to the fluctuation frequency of the model state time series data, the weight allocation needs to follow the principle that the closer the time, the greater the impact, and the comprehensive fluctuation index needs to be able to comprehensively reflect the overall fluctuation trend of the model state data within a continuous time period.

[0098] S204: Compare the comprehensive volatility index with the preset volatility threshold. When the comprehensive volatility index exceeds the preset volatility threshold, generate a parameter adjustment trigger signal.

[0099] For example, the comprehensive volatility index is compared with a preset volatility threshold. This includes confirming the validity of the preset volatility threshold, which needs to be set based on the model's historical operating data and expected performance standards. The calculated comprehensive volatility index is compared with the preset volatility threshold. If the comprehensive volatility index exceeds the preset volatility threshold, the signal generation program is started to generate a parameter adjustment trigger signal according to a preset format.

[0100] Among these, the preset fluctuation threshold needs to be updated regularly according to the model's operating environment and performance requirements; the numerical comparison needs to adopt a precise numerical comparison method; and the parameter adjustment trigger signal needs to include basic information on the triggering reason and adjustment direction so that it can be used for parameter optimization of the adaptive attention prediction model in the future.

[0101] In one embodiment, frequency detection processing is performed on emerging interest entities in the global interest ontology database. When the frequency of occurrence per unit time exceeds a preset emergence threshold, a feature reconstruction trigger signal is generated, including:

[0102] (1) Spatiotemporal density clustering is performed on emerging interest entities in the global interest ontology to generate a candidate entity set.

[0103] For example, spatiotemporal density clustering is performed on emerging interest entities in the global interest ontology. The division criteria for spatiotemporal dimensions are determined. The time dimension is divided into time periods according to a set period, and the spatial dimension is associated with the user's region or platform scenario. The spatiotemporal attribute information of emerging interest entities in the global interest ontology is extracted. Based on the distribution density of entities in the spatiotemporal dimension, emerging interest entities that are close in distance and have overlapping time periods are grouped into the same cluster. After removing isolated entities with a density lower than a set threshold, a candidate entity set containing multiple clusters is generated.

[0104] The criteria for dividing the spatiotemporal dimension need to be set in combination with the temporal patterns of changes in user interests and the regional distribution characteristics. The calculation of distribution density needs to take into account the frequency of occurrence of entities and spatiotemporal distance. Each cluster in the candidate entity set needs to contain at least two emerging interest entities with related attributes.

[0105] (2) Perform association network construction on the candidate entity set to generate an entity propagation graph.

[0106] For example, the candidate entity set is processed by constructing an association network. First, the association relationships between emerging interest entities in the candidate entity set are analyzed, including user behavior associations and content semantic associations. Then, each emerging interest entity is used as a network node, and the connection edges between nodes are set according to the strength of the association relationship. The higher the association strength, the greater the weight of the connection edge. After the construction of all nodes and connection edges is completed, an entity propagation graph that can reflect the association structure between entities is generated.

[0107] Among them, the analysis of relationships needs to be based on user behavior logs and entity semantic data in the global interest ontology library, the weight of the connection edge needs to be determined by the association frequency or semantic similarity calculation, and the entity propagation graph needs to clearly show the association path and influence range of each emerging interest entity.

[0108] (3) Based on the entity propagation map, the propagation path is traced and the entity influence diffusion rate is generated.

[0109] For example, propagation path tracing processing based on entity propagation graphs includes identifying the initial propagation source node from the entity propagation graph, tracing the propagation path from the initial propagation source node to other nodes through connecting edges, recording the propagation order and propagation time of nodes on each propagation path, calculating the extension length of the propagation path and the number of newly associated nodes per unit time, and combining these data to generate an entity influence diffusion rate that reflects the diffusion speed of emerging interest entities.

[0110] The identification of the initial propagation source node needs to be based on the timestamp of the emerging interest entity, the tracking of the propagation path needs to record the propagation direction and time interval between nodes, and the calculation of the entity's influence diffusion rate needs to combine the two core dimensions of propagation range and propagation time.

[0111] (4) When the diffusion rate of entity influence exceeds the preset emergence threshold, a feature reconstruction trigger signal is generated.

[0112] For example, when the entity's influence diffusion rate exceeds a preset emergence threshold, a feature reconstruction trigger signal is generated to confirm the rationality of the preset emergence threshold. This threshold needs to be set based on the diffusion data of historical emerging interest entities and marketing needs. The calculated entity influence diffusion rate is compared with the preset emergence threshold. If the entity influence diffusion rate exceeds the preset emergence threshold, it is determined that the emerging interest entity has formed a significant spread trend. A feature reconstruction trigger signal containing entity category and diffusion characteristics is generated according to a preset signal format.

[0113] Among these, the preset emergence threshold needs to be updated regularly according to the changing patterns of market environment and user interests, the comparison process needs to adopt a standardized numerical comparison method, and the feature reconstruction trigger signal needs to contain sufficient information to support the subsequent feature update operation of the full-domain interest ontology.

[0114] In one embodiment, the entity influence diffusion rate is obtained using the following formula, including:

[0115]

[0116] Among them, R diff The value represents the diffusion rate of entity influence, T represents the observation time window, K represents the total number of candidate entities, and α represents the weight of the time dimension. V represents the set of time-sensitive active entities.k Denotes the total set of related entities, Δt k E represents the mean propagation delay, β represents the topological dimension weight, and E k Let ω(e) represent the propagation edge set, and deg represent the propagation weight. max Indicates the maximum node degree. This represents the propagation speed factor.

[0117] For example, an observation time window is determined, and the entity influence diffusion rate is calculated within this time window. The total number of candidate entities is counted, and for each candidate entity, its time-related parameters and topology-related parameters are calculated. For the time dimension, the set of active entities and the total set of associated entities are obtained, their ratio is calculated, and combined with the mean propagation delay, a propagation speed factor is obtained. For the topology dimension, the set of propagation edges is determined, the propagation weight of each edge is calculated, and combined with the maximum node degree, the topology-dimensional influence value is calculated. Based on the time-dimensional and topology-dimensional weights, the time-dimensional influence value and the topology-dimensional influence value are weighted and summed to obtain the local diffusion index for each candidate entity. Then, the local diffusion indices of all candidate entities are averaged to obtain the entity influence diffusion rate.

[0118] In one embodiment, the promotion effect data acquired by the preset dual-channel feedback unit and the global interest ontology are processed to generate an adaptive attention prediction model, including:

[0119] (1) Perform multi-dimensional feature decoupling processing on the promotion effect data to generate explicit feedback feature vectors and implicit feedback feature vectors.

[0120] For example, multi-dimensional feature decoupling processing is performed on the promotion effect data to clarify the dimensions involved in the promotion effect data, such as exposure, clicks, and conversion rate. For each dimension, intuitive data that can directly reflect the promotion effect is extracted as explicit feedback features, and information such as potential user interests and behavioral patterns hidden in the data is mined as implicit feedback features. These features are vectorized respectively to generate explicit feedback feature vectors and implicit feedback feature vectors with clear structures.

[0121] Among them, the dimensions of promotion effect data need to comprehensively cover the core indicators of the promotion activity. Explicit feedback features are data related to promotion results that can be directly observed, while implicit feedback features are potential information that needs to be mined through data analysis.

[0122] (2) Spatiotemporal interest encoding is performed on the global interest ontology to generate a dynamic interest feature matrix.

[0123] For example, spatiotemporal interest encoding processing is performed on the global interest ontology, including dividing it into different time periods in the time dimension and different scenarios or regions in the spatial dimension. Information such as the type and intensity of user interests in each time period and scenario is extracted from the global interest ontology. According to the set encoding rules, the above spatiotemporal interest information is converted into numerical features, and after being arranged and combined, a dynamic interest feature matrix that can reflect the changes of user interests with time and space is generated.

[0124] Among them, the time-division segmentation in the time dimension needs to conform to the pattern of changes in user interests, the scene or region segmentation in the spatial dimension needs to consider the regional differences in user interests, and the coding rules need to ensure that spatiotemporal interest information can be accurately converted into matrix form.

[0125] (3) Based on the attention gating mechanism, cross-domain feature fusion processing is performed on explicit feedback feature vectors, implicit feedback feature vectors and dynamic interest feature matrix to generate joint feature representation.

[0126] For example, based on the attention gating mechanism, cross-domain feature fusion processing is performed on explicit feedback feature vectors, implicit feedback feature vectors, and dynamic interest feature matrices. This includes calculating the correlation weights between each feature through the attention gating mechanism, determining the importance of different features in the fusion process, and weighting and integrating the explicit feedback feature vectors, implicit feedback feature vectors, and dynamic interest feature matrices according to the calculated correlation weights. This integrates features from different sources and of different types into a unified feature set, generating a joint feature representation.

[0127] Among them, the attention gating mechanism is used to measure the degree of correlation between features. The calculation of correlation weights needs to be based on the similarity or correlation between features. The joint feature representation needs to contain the key information of each input feature and reflect the interaction between features.

[0128] (4) Through the preset neural network architecture search unit, the joint feature representation is subjected to adaptive network topology construction processing to generate an adaptive attention prediction model.

[0129] For example, the joint feature representation is adaptively constructed using a preset neural network architecture search unit. This includes inputting the joint feature representation into the preset neural network architecture search unit. The unit will automatically explore suitable neural network layer types, number of layers, and inter-layer connection methods based on the characteristics of the joint feature representation, such as dimension and complexity. According to the explored network topology, the unit will build the corresponding neural network model, initialize the model parameters, and generate an adaptive attention prediction model.

[0130] Among them, the neural network architecture search unit needs to have the ability to automatically explore the network topology. The exploration process needs to consider the characteristics of joint feature representation to ensure that the generated network can effectively process the feature. The adaptive attention prediction model needs to be able to accurately predict user attention based on the input features.

[0131] In one embodiment, identity graph fusion processing is performed on the acquired multi-platform user behavior logs to generate a global interest ontology, including:

[0132] (1) Perform cross-domain behavior time sequence alignment processing on user behavior logs from multiple platforms to generate spatiotemporally related behavior sequences.

[0133] For example, cross-domain behavior time sequence alignment processing is performed on user behavior logs from multiple platforms. This includes extracting the timestamp, behavior type, and associated platform identifier of each behavior record in the user behavior logs from multiple platforms. Using a unified time standard as a benchmark, the behavior records belonging to the same user on different platforms are sorted in chronological order. Records with abnormal timestamps or conflicting behavior logic are removed to ensure that the behavior records of each platform are consistent and correspond in the time dimension, and to generate a spatiotemporally related behavior sequence containing the user's continuous behavior trajectory across platforms.

[0134] Among them, the multi-platform user behavior logs cover user operation data in different scenarios such as e-commerce platforms, social platforms, and video platforms. The time standard must adopt a unified time format and precision, and the spatiotemporal related behavior sequences must fully retain the platform attributes and time attributes of the behavior records.

[0135] (2) The device fingerprint clustering process is performed on the spatiotemporal related behavior sequence through the preset identity resolution rule base to generate cross-platform identity clusters.

[0136] For example, device fingerprint clustering is performed on spatiotemporal related behavior sequences through a preset identity resolution rule base. This includes retrieving resolution rules such as device identifier matching rules and account association rules from the preset identity resolution rule base, extracting device fingerprint information corresponding to each record in the spatiotemporal related behavior sequence, performing similarity matching on the device fingerprint information according to the resolution rules, and grouping spatiotemporal related behavior sequences with device fingerprint similarity reaching a set threshold into the same cluster group to form a cross-platform identity cluster containing the behavior data of the same user across multiple platforms.

[0137] Among them, the preset identity resolution rule base needs to be updated regularly to adapt to new device types and authentication methods, device fingerprint information needs to be de-identified to protect user privacy, and cross-platform identity clusters need to contain indexes of all platform behavior records corresponding to the identity.

[0138] (3) Perform graph neural network embedding processing on cross-platform identity clusters to generate dynamic identity graphs.

[0139] For example, graph neural network embedding processing is performed on cross-platform identity clusters, including constructing a graph structure model, taking each user identity in the cross-platform identity cluster as a graph node, taking the behavioral relationship of users on different platforms as graph edges, inputting the graph structure model into the graph neural network, calculating the feature vectors of nodes and edges through network layers, capturing the potential relationship and behavioral patterns between user identities, integrating the calculated node feature vectors with graph structure information, and generating a dynamic identity graph that can dynamically reflect the relationship between user identities.

[0140] In the graph structure model, the nodes must contain user identity identifiers and core behavioral features, the weights of the graph edges must be set according to the closeness of the behavioral associations, and the dynamic identity graph must have the ability to adjust the node and edge features as user behavior is updated.

[0141] (4) Based on the trained interest entity extraction model, perform multi-dimensional interest mining on the dynamic identity graph to generate a global interest ontology library.

[0142] For example, multi-dimensional interest mining processing is performed on the dynamic identity graph based on a trained interest entity extraction model. This includes loading the trained interest entity extraction model, which has been trained using labeled user interest data and has the ability to identify interest categories and extract interest features. User behavior association information and identity attribute information from the dynamic identity graph are input into the model. The model mines user interest features in dimensions such as consumption preferences, content needs, and interaction tendencies through semantic analysis, behavior pattern recognition, etc. These interest features are organized by category and a structured interest system is constructed to generate a full-domain interest ontology.

[0143] Among these, the trained interest entity extraction model needs to undergo multiple rounds of verification to ensure the accuracy of interest recognition; multi-dimensional interest mining needs to cover users' short-term immediate interests and long-term stable interests; and the full-domain interest ontology needs to contain core information such as interest categories, interest intensity, and interest relationships.

[0144] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0145] In one embodiment, such as Figure 3 As shown, this application also provides a precision marketing full-domain AI video promotion device 300, which includes:

[0146] The identity fusion module 301 is used to perform identity graph fusion processing on the acquired user behavior logs from multiple platforms to generate a global interest ontology library.

[0147] The adaptive modeling module 302 is used to process the promotion effect data and the whole domain interest ontology obtained by the preset dual-channel feedback unit to generate an adaptive attention prediction model.

[0148] The evolution trigger module 303 is used to perform three-level trigger condition judgment processing on the adaptive attention prediction model and generate evolution instructions.

[0149] The model recombination module 304 is used to reorganize the adaptive attention prediction model in response to evolutionary instructions, and generate an evolved prediction model.

[0150] The narrative generator module 305 is used to reorganize the output of the evolved prediction model into multimodal narrative units to generate promotional videos adapted to the target platform.

[0151] Specifically, the identity fusion module 301 performs identity graph fusion processing on the acquired user behavior logs from multiple platforms, including collecting and cleaning user operation logs from different platforms, performing cross-domain behavior time sequence alignment and device fingerprint clustering, constructing a dynamic identity graph, mining user multi-dimensional interests, and generating a full-domain interest ontology library.

[0152] The adaptive modeling module 302 processes the promotion effect data and the global interest ontology obtained by the preset dual-channel feedback unit to build a model, including extracting and decoupling the promotion effect data to generate explicit and implicit feedback feature vectors, encoding the global interest ontology to obtain a dynamic interest feature matrix, fusing features through an attention gating mechanism to form a joint feature representation, building a network using a neural network architecture search unit, and generating an adaptive attention prediction model.

[0153] The evolution trigger module 303 performs three-level trigger condition judgment processing on the adaptive attention prediction model, including monitoring the model's prediction performance indicators to generate time-series data, and generating parameter adjustment trigger signals if the data exceeds a threshold. It also counts the frequency of emerging interest entities in the global interest ontology, and generates feature reconstruction trigger signals if the frequency exceeds a threshold. Simultaneously, it monitors model architecture degradation indicators, and generates architecture mutation trigger signals if the frequency exceeds a threshold. Any trigger signal is encoded to generate an evolution instruction.

[0154] The model recombination module 304 is used to reorganize the adaptive attention prediction model in response to evolutionary instructions. It parses the instruction type, optimizes the model parameters, adds interest feature dimensions or reconstructs the network topology, and then conducts local training and verification on the reorganized model. Once the target is met, the evolved prediction model is generated.

[0155] The narrative generator module 305 performs multimodal narrative unit reorganization processing on the output of the evolved prediction model, including extracting user preferences and platform requirements from the output, screening matching narrative units, arranging and merging units according to platform specifications, and generating promotional videos adapted to the target platform after quality review.

[0156] Evolution trigger module 303 is also used for:

[0157] The predictive performance metrics of the adaptive attention prediction model are monitored and processed in real time to generate time series data of the model status.

[0158] Deviation fluctuation analysis is performed on the model state time series data. When the fluctuation amplitude exceeds the preset fluctuation threshold, a parameter adjustment trigger signal is generated.

[0159] Frequency detection processing is performed on emerging interest entities in the global interest ontology. When the frequency of occurrence exceeds a preset emergence threshold per unit time, a feature reconstruction trigger signal is generated.

[0160] The architecture degradation index of the adaptive attention prediction model is subjected to decay analysis. When the decay rate exceeds the preset architecture threshold, an architecture mutation trigger signal is generated.

[0161] The parameter adjustment trigger signal, feature reconstruction trigger signal, or architecture mutation trigger signal are processed by instruction encoding to generate evolution instructions.

[0162] Evolution trigger module 303 is also used for:

[0163] The model state time series data is processed by sliding window segmentation to generate dynamic time segments;

[0164] For each dynamic time segment, the fluctuation acceleration is calculated and processed to generate the instantaneous fluctuation intensity;

[0165] The instantaneous fluctuation intensity of continuous dynamic time segments is processed by trend accumulation to generate a comprehensive fluctuation index;

[0166] The comprehensive volatility index is compared with a preset volatility threshold. When the comprehensive volatility index exceeds the preset volatility threshold, a parameter adjustment trigger signal is generated.

[0167] Evolution trigger module 303 is also used for:

[0168] Spatiotemporal density clustering is performed on emerging interest entities in the global interest ontology to generate a candidate entity set;

[0169] The candidate entity set is processed to construct an association network, generating an entity propagation graph.

[0170] Propagation path tracing is performed based on entity propagation graphs to generate entity influence diffusion rates.

[0171] When the diffusion rate of entity influence exceeds a preset emergence threshold, a feature reconstruction trigger signal is generated.

[0172] Evolution trigger module 303 is also used for:

[0173]

[0174] Among them, R diff The value represents the diffusion rate of entity influence, T represents the observation time window, K represents the total number of candidate entities, and α represents the weight of the time dimension. V represents the set of time-sensitive active entities. k Denotes the total set of related entities, Δt k E represents the mean propagation delay, β represents the topological dimension weight, and E k Let ω(e) represent the propagation edge set, and deg represent the propagation weight. max Indicates the maximum node degree. This represents the propagation speed factor.

[0175] Adaptive modeling model 302 is also used for:

[0176] Multi-dimensional feature decoupling processing is performed on the promotion effect data to generate explicit feedback feature vectors and implicit feedback feature vectors;

[0177] Spatiotemporal interest encoding is performed on the global interest ontology to generate a dynamic interest feature matrix;

[0178] Based on the attention gating mechanism, cross-domain feature fusion processing is performed on explicit feedback feature vectors, implicit feedback feature vectors, and dynamic interest feature matrices to generate joint feature representations;

[0179] By using a pre-defined neural network architecture search unit, the joint feature representation is subjected to adaptive network topology construction processing to generate an adaptive attention prediction model.

[0180] Identity fusion module 301 is also used for:

[0181] Perform cross-domain behavior time sequence alignment processing on user behavior logs from multiple platforms to generate spatiotemporally related behavior sequences;

[0182] By using a pre-defined identity resolution rule base, device fingerprint clustering is performed on spatiotemporally related behavior sequences to generate cross-platform identity clusters.

[0183] Graph neural network embedding processing is applied to cross-platform identity clusters to generate dynamic identity graphs;

[0184] Based on the trained interest entity extraction model, multi-dimensional interest mining is performed on the dynamic identity graph to generate a global interest ontology library.

[0185] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0186] In one embodiment, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0187] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0188] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A precise marketing, full-domain AI video promotion method, characterized by: The method includes: The acquired user behavior logs from multiple platforms are processed by identity graph fusion to generate a global interest ontology. The promotion effect data obtained by the preset dual-channel feedback unit and the full-domain interest ontology are processed to generate an adaptive attention prediction model. The adaptive attention prediction model is subjected to a three-level trigger condition judgment process to generate evolution instructions; The adaptive attention prediction model responds to the evolutionary instruction by undergoing recombination processing to generate an evolved prediction model. The output of the evolved prediction model is reorganized into multimodal narrative units to generate a promotional video adapted to the target platform.

2. The precision marketing full-domain AI video promotion method according to claim 1, characterized in that, The process of performing a three-level trigger condition judgment on the adaptive attention prediction model to generate evolutionary instructions includes: The predictive performance indicators of the adaptive attention prediction model are monitored and processed in real time to generate model state time series data. The model state time series data is subjected to deviation fluctuation analysis. When the fluctuation amplitude exceeds the preset fluctuation threshold, a parameter adjustment trigger signal is generated. Frequency detection processing is performed on emerging interest entities in the global interest ontology library. When the frequency of occurrence exceeds a preset emergence threshold per unit time, a feature reconstruction trigger signal is generated. The architecture degradation index of the adaptive attention prediction model is subjected to decay analysis. When the decay rate exceeds the preset architecture critical value, an architecture mutation trigger signal is generated. The parameter adjustment trigger signal, the feature reconstruction trigger signal, or the architecture mutation trigger signal are processed by instruction encoding to generate the evolution instruction.

3. The precision marketing omni-channel AI video promotion method according to claim 2, characterized in that, The step of performing deviation fluctuation analysis on the model state time series data, and generating a parameter adjustment trigger signal when the fluctuation amplitude exceeds a preset fluctuation threshold, includes: The model state time series data is processed by sliding window segmentation to generate dynamic time segments; For each of the dynamic time segments, fluctuation acceleration is calculated to generate instantaneous fluctuation intensity; The instantaneous fluctuation intensity of the continuous dynamic time segment is subjected to trend accumulation processing to generate a comprehensive fluctuation index; The comprehensive volatility index is compared with a preset volatility threshold. When the comprehensive volatility index exceeds the preset volatility threshold, the parameter adjustment trigger signal is generated.

4. The precision marketing full-domain AI video promotion method according to claim 2, characterized in that, The frequency detection processing of emerging interest entities in the global interest ontology, when the frequency of occurrence exceeds a preset emergence threshold per unit time, generates a feature reconstruction trigger signal, including: Spatiotemporal density clustering is performed on emerging interest entities in the global interest ontology to generate a candidate entity set; The candidate entity set is subjected to association network construction processing to generate an entity propagation graph; Based on the entity propagation map, propagation path tracing is performed to generate the entity influence diffusion rate. When the diffusion rate of the entity's influence exceeds a preset emergence threshold, the feature reconstruction trigger signal is generated.

5. The precision marketing full-domain AI video promotion method according to claim 4, characterized in that, The diffusion rate of the entity's influence is obtained using the following formula, including: Among them, R diff The value represents the diffusion rate of entity influence, T represents the observation time window, K represents the total number of candidate entities, and α represents the weight of the time dimension. V represents the set of time-sensitive active entities. k Denotes the total set of related entities, Δt k E represents the mean propagation delay, β represents the topological dimension weight, and E k Let ω(e) represent the propagation edge set, and deg represent the propagation weight. max Indicates the maximum node degree. This represents the propagation speed factor.

6. The precision marketing full-domain AI video promotion method according to claim 1, characterized in that, The step of processing the promotion effect data obtained from the preset dual-channel feedback unit and the full-domain interest ontology to generate an adaptive attention prediction model includes: The promotion effect data is subjected to multi-dimensional feature decoupling processing to generate explicit feedback feature vectors and implicit feedback feature vectors; The global interest ontology is subjected to spatiotemporal interest encoding to generate a dynamic interest feature matrix; Based on the attention gating mechanism, cross-domain feature fusion processing is performed on the explicit feedback feature vector, implicit feedback feature vector and dynamic interest feature matrix to generate a joint feature representation; An adaptive network topology construction process is performed on the joint feature representation using a preset neural network architecture search unit to generate an adaptive attention prediction model.

7. The precision marketing full-domain AI video promotion method according to claim 1, characterized in that, The step of performing identity graph fusion processing on the acquired multi-platform user behavior logs to generate a global interest ontology includes: Cross-domain behavior time sequence alignment processing is performed on the user behavior logs of the multiple platforms to generate spatiotemporally related behavior sequences; The spatiotemporal related behavior sequences are clustered using device fingerprints through a preset identity resolution rule base to generate cross-platform identity clusters. The cross-platform identity cluster is subjected to graph neural network embedding processing to generate a dynamic identity graph; The dynamic identity graph is processed by multi-dimensional interest mining based on the trained interest entity extraction model to generate the global interest ontology library.

8. A precision marketing omni-channel AI video promotion device, characterized in that: The device includes: The identity fusion module is used to perform identity graph fusion processing on the acquired user behavior logs from multiple platforms to generate a global interest ontology library. The adaptive modeling module is used to process the promotion effect data obtained by the preset dual-channel feedback unit and the global interest ontology to generate an adaptive attention prediction model. The evolution trigger module is used to perform three-level trigger condition judgment processing on the adaptive attention prediction model and generate evolution instructions. The model recombination module is used to reorganize the adaptive attention prediction model in response to the evolutionary instruction, and generate an evolved prediction model. The narrative generator module is used to perform multimodal narrative unit recombination processing on the output of the evolved prediction model to generate promotional videos adapted to the target platform.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the precision marketing full-domain AI video promotion method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the precision marketing full-domain AI video promotion method as described in any one of claims 1 to 7.

Images