Marketing data real-time processing method and system based on cloud edge collaboration

Abstract

The application provides a marketing data real-time processing method and system based on cloud edge cooperation, which comprises the following steps: obtaining original input data uploaded by each client. Each original input data comprises multiple modal data. The cosine similarity between each modal data and a preset marketing strategy vector is obtained, and the cosine similarity is multiplied by a preset modal density weight factor to obtain the significance score of each modal data. The unit content cost, modal value fluctuation factor and significance score of each modal data are calculated to obtain the upload scheduling score of each modal data. According to the upload scheduling score and a preset upload capacity upper limit, the corresponding modal data is selected as the target modal data, and the target modal data is uploaded to the cloud end, so that the modal data matching the task demand can be quickly and accurately screened out and uploaded to the cloud end.

Description

Technical Field

[0001] This invention belongs to the field of electronics, and in particular relates to a method and system for real-time processing of marketing data based on cloud-edge collaboration. Background Technology

[0002] With the continuous expansion of social media and content e-commerce ecosystems, brands are increasingly relying on multimodal content for marketing campaigns, including short videos, images, and text. This type of content is often characterized by its large volume, rapid updates, and complex structure. In actual marketing campaigns, influencer content typically needs to be screened, analyzed, and predicted before it can be used for decision-making. This process heavily relies on a dynamic understanding of content semantics, user interests, and task objectives. However, traditional marketing systems are primarily built on a centralized cloud computing architecture, requiring all data to be uploaded to the cloud for processing. When the content volume reaches millions or even higher, upload bandwidth, cloud processing latency, and frequent network congestion significantly reduce response speed, making it impossible for the system to meet real-time strategy evaluation and response requirements. Furthermore, due to the significant differences in information density between different modalities of content, the cloud often receives a large amount of redundant data, resulting in resource waste and reduced prediction accuracy. On the other hand, marketing decisions need to be precisely matched based on real-time task requirements (i.e., brief requirements), but traditional platforms cannot determine which content best fits the brief objectives before content is uploaded, causing the cloud to bear an unnecessary data analysis burden. Therefore, how to quickly and accurately select modalities that match task requirements and upload them to the cloud has become an urgent technical problem to be solved. Summary of the Invention

[0003] The purpose of this invention is to design a real-time marketing data processing method and system based on cloud-edge collaboration, which can quickly and accurately filter out modalities that match task requirements and upload them to the cloud.

[0004] To achieve the above objectives, a real-time marketing data processing method based on cloud-edge collaboration is provided in a first aspect of the present invention, the method comprising:

[0005] Obtain the raw input data uploaded by each client; wherein each raw input data includes multiple modal data;

[0006] Obtain the cosine similarity between each modal data and a preset marketing strategy vector, and multiply the cosine similarity by a preset modality density weighting factor to obtain the significance score of each modal data.

[0007] The upload scheduling score for each modal data is calculated based on the unit content cost, modal value fluctuation factor, and significance score of each modal data.

[0008] Based on the upload scheduling score and the preset upload capacity limit, the corresponding modal data is selected as the target modality, and the target modality is uploaded to the cloud.

[0009] Furthermore, after selecting the corresponding modal data as the target modality based on the upload scheduling score and the preset upload capacity limit, and uploading the target modality, the method further includes:

[0010] Feature extraction is performed based on the target modality to obtain modality feature vectors. Each modality feature vector is then concatenated to obtain a joint feature vector.

[0011] The marketing behavior impact score is calculated based on the joint feature vector, the marketing strategy vector, the number of segments actually uploaded for each modality, the preset modality coverage penalty weight, the number of segments that should be uploaded for each modality, and the information entropy within the segments of each modality.

[0012] Furthermore, after calculating the marketing behavior impact score based on the joint feature vector, the marketing strategy vector, the actual number of segments uploaded for each modality, the preset modality coverage penalty weight, the number of segments to be uploaded for each modality, and the intra-segment information entropy for each modality, the method further includes:

[0013] Obtain a strategy dictionary; wherein the strategy dictionary includes multiple keywords and corresponding keyword vectors;

[0014] Obtain the dot product similarity between each modality feature vector and each keyword vector. If the dot product similarity is greater than a preset threshold, the keyword corresponding to the keyword vector is defined as a valid strategy label.

[0015] Based on the marketing behavior impact score, the preset behavior impact score threshold, the preset platform priority factor, the effective strategy tags, and the preset strategy keyword set, the target delivery platform is selected from the preset platform delivery configuration table so that the cloud can deliver the target modality to the target delivery platform.

[0016] Furthermore, after delivering the target modality to the target delivery platform, the method further includes:

[0017] Obtain the platform budget conversion ratio and expected click target for the target advertising platform;

[0018] The expected budget value is obtained by calculating the budget based on the marketing activity impact score, the platform budget conversion ratio, and the expected click target.

[0019] Further, the calculation of the upload scheduling score for each modal data based on the unit content cost, modal value fluctuation factor, and significance score includes:

[0020] Multiply the preset weighting factor by the modal value fluctuation factor, and add the unit content cost to obtain the first data;

[0021] The saliency score is divided by the first data to obtain the upload scheduling score.

[0022] Further, the step of selecting the corresponding modal data as the target modality based on the upload scheduling score and the preset upload capacity limit includes:

[0023] The upload scheduling scores are sorted from high to low, and the capacities of the corresponding modal data are added together until the accumulated capacity reaches the upload capacity limit. The selected modal data is then the target modal.

[0024] Further, the original input data includes image modality, video modality, and text modality. The step of obtaining the cosine similarity between each modality data and a preset marketing strategy vector, and multiplying the cosine similarity by a preset modality density weighting factor to obtain a significance score for each modality data, includes:

[0025] Feature extraction is performed on the image modality, the video modality, and the text modality respectively to obtain image feature vector, video feature vector, and text feature vector;

[0026] The cosine similarity between the image feature vector, the video feature vector, and the text feature vector and the preset marketing strategy vector are calculated respectively to obtain the image cosine similarity, video cosine similarity, and text cosine similarity.

[0027] The image cosine similarity, video cosine similarity, and text cosine similarity are multiplied by a preset modality density weighting factor to obtain the image saliency score, video saliency score, and text saliency score, respectively.

[0028] A second aspect of the invention provides a real-time marketing data processing system based on cloud-edge collaboration, the system comprising:

[0029] The acquisition unit is used to acquire the raw input data uploaded by each client; wherein each raw input data includes multiple modal data;

[0030] The scoring unit is used to obtain the cosine similarity between each modal data and the preset marketing strategy vector, and multiply the cosine similarity with the preset modality density weighting factor to obtain the significance score of each modal data.

[0031] The calculation unit is used to calculate the upload scheduling score of each modal data based on the unit content cost, modal value fluctuation factor and significance score of each modal data.

[0032] The selection unit is used to select the corresponding modal data as the target modality based on the upload scheduling score and the preset upload capacity limit, and upload the target modality to the cloud.

[0033] In a third aspect of the invention, an electronic device is provided, the electronic device including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the method described in the first aspect above.

[0034] In a fourth aspect of the invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect above.

[0035] The beneficial technical effects of the present invention are at least as follows:

[0036] To address the aforementioned issues, this invention provides a real-time marketing data processing method and system based on cloud-edge collaboration. Its core lies in extracting features from influencer content across image, video, and text modalities via edge nodes. It then uses Brief vectors to calculate semantic relevance of modalities, accurately determining which modalities are most likely to match marketing objectives through saliency scoring. This is further combined with modal information density weights to construct a content importance ranking, enabling target-oriented filtering of uploaded content. Based on this, an asynchronous upload scheduling strategy is constructed according to upload scheduling scores, upload costs, and modal stability. This ensures that, under limited bandwidth, higher-value and semantically stronger modalities are prioritized for transmission, allowing the cloud to obtain the most predictive data even with incomplete information. A multimodal feature fusion model and a dual-tower structure are used to score the impact of behavioral marketing behavior. Robust modeling of some uploaded modalities is achieved by combining modal coverage penalty terms and modal information entropy regularization, thereby accurately estimating the potential effect of target modalities in real-world deployment scenarios. Simultaneously, a set of effective strategy tags is automatically generated based on a cross-modal semantic matching mechanism, thus constructing a structured semantic bridge between modalities and strategy requirements. Finally, based on the platform's selection function and budget estimation model, structured delivery suggestions that can be directly used for campaign execution are output, achieving real-time processing across the entire chain from content selection, content uploading, content understanding to delivery instruction generation. This invention can significantly reduce redundant uploads, improve the accuracy of content value assessment, and achieve direct mapping from marketing content to delivery strategies, effectively solving the technical pain points of traditional systems that cannot process multimodal marketing content in a real-time, efficient, and consistent manner. Attached Figure Description

[0037] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0038] Figure 1 This is a flowchart of a real-time marketing data processing method based on cloud-edge collaboration provided in an embodiment of this application.

[0039] Figure 2 This is a schematic diagram of the structure of a real-time marketing data processing system based on cloud-edge collaboration provided in an embodiment of this application. Detailed Implementation

[0040] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0041] Please refer to Figure 1, Figure 1 This is a flowchart of a real-time marketing data processing method based on cloud-edge collaboration provided in an embodiment of this application. Figure 1 The method may include, but is not limited to, steps S101 to S104.

[0042] Step S101: Obtain the raw input data uploaded by each client; wherein each raw input data includes multiple modal data;

[0043] Step S102: Obtain the cosine similarity between each modal data and the preset marketing strategy vector, and multiply the cosine similarity with the preset modality density weight factor to obtain the significance score of each modal data.

[0044] Step S103: Calculate the upload scheduling score for each modal data based on the unit content cost, modal value fluctuation factor, and significance score of each modal data.

[0045] Step S104: Select the corresponding modal data as the target modal based on the upload scheduling score and the preset upload capacity limit, and upload the target modal to the cloud.

[0046] In steps S101 to S102 of some embodiments, saliency scoring is performed on multimodal content created by influencers at edge nodes. Specifically, the task is to determine which information from three modalities—image frames, video clips, and text copy—is more likely to be associated with the current marketing strategy objectives. The saliency score serves as the basis for subsequent upload scheduling, controlling the scope of uploaded data, thereby reducing redundant transmission pressure and improving the overall system's real-time performance and decision-making quality. Since this invention focuses on real-time marketing content selection "based on brief objectives," the saliency score considers not only the semantic density of the content itself but also the direction of the current marketing strategy.

[0047] Specifically, edge nodes receive raw input data from local creator clients. The structure is fixed and consists of the following three categories: image modalities. The system automatically extracts representative frames from the videos uploaded by creators as the image modality. For example, it selects one frame from the first 5 seconds and one frame from the last 2 seconds, for a total of 3 frames. This is achieved through a frame sampling module (such as the FFmpeg frame capture tool), in JPEG format, with a fixed size of 224×224×3. (Video modality) Short video clips, no more than 30 seconds long, are collected from the video upload component of short video apps. After the video is compressed, the system extracts the middle 10 seconds as the analysis input, with the encoding format H.264; text modality. The title is filled in by the creator on the publishing interface, either as a product description or video description, or as a title generated by the system. It is in UTF-8 encoding and has a maximum length of 300 characters.

[0048] Feature extraction was performed on the three modalities of data. For the image modality, the MobileNetV2 network structure was used, loaded as a TensorRT model, and deployed lightweight at the edge. After unifying the input image size, 128-dimensional image feature vectors were extracted through convolutional and pooling layers. The video modality is first uniformly sampled, from which 5 frames are extracted (1 frame every 2 seconds). Each frame is processed by MobileNetV2, and the average feature value is then used to construct the video feature vector. The text modality uses the first four layers of the BERT-Base model to form a lightweight network, retaining only the Embedding layer and the first Transformer block. The [CLS] bits are used as the sentence vector, and then average pooling is performed to obtain the text feature vector. .

[0049] Marketing strategy objectives are represented by a marketing strategy vector (Brief vector). The process of constructing this strategy is as follows: When configuring campaigns, operators specify a set of keywords as the strategy keyword set. (e.g., "wedding", "domestic products", "couples"), the system uses a BERT encoder to generate embeddings for this set of keywords, and then averages them to obtain the marketing strategy vector. This vector is pre-generated in the cloud and periodically distributed to the edge node cache.

[0050] Next, the cosine similarity between each modality data point and the marketing strategy vector is calculated, as shown in the following formula:

[0051] ;

[0052] in, Representing modal data The cosine similarity between the marketing strategy vector and the marketing strategy vector has a value range of [-1, 1], and the closer it is to 1, the higher the semantic relevance. This represents the feature vector extracted from an image, video, or text modality. This represents the marketing strategy vector generated based on the marketing brief objective.

[0053] To account for differences in modal information density, a modal density weighting factor is introduced. This parameter, based on the conversion performance of historical marketing samples, is distributed to edge nodes after global statistical calculation, representing the average information contribution of different modalities. The final modal significance score is calculated as follows:

[0054] ;

[0055] in, Modal data The significance score is used for the next step of scheduling and uploading decisions. Image saliency scoring Video saliency score Or text saliency score For example, if an image shows people wearing traditional Chinese clothing, a video shows a wedding scene, and the caption says "Made in China," then the three modalities... Both are relatively high, and based on historical experience... Larger (e.g., videos are more likely to attract users to stay), therefore ultimately The highest scorer will be uploaded first.

[0056] The output of this step is a set of significance scores, S={S_I,S_V,S_T}, which is passed as an ordered key-value table to the next step, the upload scheduling module. All variables are dimensionless scores, and all features have been normalized.

[0057] In step S103 of some embodiments, the objective is to achieve asynchronous upload scheduling of multimodal marketing content in edge nodes, based on a saliency score set. By combining device status and real-time strategy requirements under marketing scenarios, the system dynamically determines which modalities of content should be prioritized for upload to the cloud. In digital marketing, especially in brief-driven influencer content selection systems, the value of different content modalities (images, videos, text) to marketing effectiveness is not balanced and is greatly affected by changes in target strategies. Therefore, the scheduling strategy should not only be based on saliency scores but also incorporate factors such as upload costs, modal stability, and the effectiveness of content in recent campaign strategies to ensure that the cloud model receives input data with high information density and strong strategy relevance.

[0058] Specifically, a modal upload priority function is introduced, which integrates saliency score, upload cost, and recent content target value volatility to form the final upload scheduling score. First, the unit content cost for each modality is defined. (Unit: KB), the average fragment size is obtained through modal slicing statistics; then the modal value fluctuation factor is defined. This indicates the uncertainty of the current modality's performance in recent marketing tasks. A higher value indicates greater fluctuation in the modality's performance over recent task rounds, requiring close monitoring. The modality upload priority function calculates the upload scheduling score for each modality based on its unit content cost, modality value fluctuation factor, and significance score. Specifically, it multiplies the preset weight factor by the modality value fluctuation factor, adds the unit content cost to obtain the first data point, and divides the significance score by the first data point to obtain the upload scheduling score. See the formula below:

[0059] ;

[0060] in, Modal data Upload scheduling score; Modal data The significance score; Modal data The unit content cost was obtained by analyzing the average size of nearly 10 similar content slices. The modal value fluctuation factor is calculated based on the differences in campaign results. It comes from the cloud-based performance tracking module, which calculates the normalized variance of the click-through rate for the last three modalities. The weighting factor (default value is 0.5) controls the sensitivity to fluctuations. This function introduces "modal uncertainty" as a penalty in the scheduling process, prioritizing the uploading of modes that are both important and stable, thus avoiding the system processing too much redundant data with large fluctuations.

[0061] In step S104 of some embodiments, after obtaining the upload scheduling score for each modality of data, the upload scheduling score is further determined based on the preset upload capacity limit. Select the corresponding modal data as the target modality. Sort the upload scheduling scores from highest to lowest, and then add the capacities of the corresponding modal data for each sorted modality until the accumulated capacity reaches the upload capacity limit. The selected modal data is then the target mode. The specific process is as follows:

[0062] ;

[0063] in, This represents the set of target modes; target modes are those with high upload scheduling scores, low costs, and stable value. This indicates the th modality in the modality sequence sorted from highest to lowest by upload scheduling score. One location, This indicates the number of modes currently included in the calculation during the accumulation of modal capacity. Indicates ranking in bit mode The capacity of modal data is calculated from the average data segment size corresponding to that mode during actual transmission. This indicates that the modality set is arranged in order of upload scheduling score. The data is sorted from highest to lowest, meaning modalities with higher scheduling scores are processed first, followed by modalities with lower scores. For example, Creator 1's original input data includes image modality 1, video modality 1, and text modality 1, while Creator 2's original input data includes image modality 2, video modality 2, and text modality 2. The corresponding upload scheduling scores, sorted from highest to lowest, are: image modality 1, video modality 1, image modality 2, text modality 2, video modality 2, and text modality 1. The sum of the data volumes for image modality 1, video modality 1, image modality 2, and text modality 2 reaches the upload capacity limit. The target modalities are image modal 1, video modal 1, image modal 2, and text modal 2.

[0064] Each modality maintains its corresponding modality index number in the local cache, from Select the target mode from the filter. The index numbers form an index number set. Output structured upload instruction set Each key represents the modality name of the target modality, and the value is the index number of the target modality in the local data, which is used by the next module to retrieve from the cache and upload to the cloud. This data structure is packaged into JSON format through edge middleware and pushed to the cloud to receive the API. This step achieves the key link of "real-time and efficient uploading of marketing data" in the goal of this invention by constructing an upload scheduling function that considers saliency scoring, bandwidth cost, and strategy volatility. In this scenario, due to the rapid changes in the Brief strategy and the diversity of influencer content, the system must dynamically select and filter the uploaded content. This method introduces a non-linear penalty term. This makes upload scheduling more policy-sensitive and robust, a key innovation that conventional scoring and sorting strategies cannot cover.

[0065] Steps S101 to S104, as illustrated in this embodiment, involve acquiring the original input data uploaded by each client. Each original input data includes multiple modal data. The cosine similarity between each modal data and a preset marketing strategy vector is obtained. This cosine similarity is multiplied by a preset modality density weighting factor to obtain a saliency score for each modal data. An upload scheduling score for each modal data is calculated based on its unit content cost, modality value fluctuation factor, and saliency score. The corresponding modal data is selected as the target modality based on the upload scheduling score and a preset upload capacity limit, and the target modality is uploaded to the cloud. This enables rapid and accurate filtering of modal data matching task requirements for upload to the cloud.

[0066] In some embodiments, after step S104, a marketing behavior impact score is calculated to assess the conversion potential (behavioral influence) of the uploaded target modality under the current marketing strategy, and to generate structured, effective strategy tags for subsequent campaign decisions. The uploaded target modality originates from the scoring and scheduling system and is selected local data, which may be incomplete or modally uneven. This uncertainty of "incomplete input" is a problem that traditional marketing scoring systems have not addressed. This step specifically models the structural characteristics of this marginally uploaded data in the scoring model and proposes an innovative scoring formula with structural penalties and target guidance capabilities.

[0067] Specifically, the input is ,in Representing the target modality (including image modality) Video modal or text modality ), This is the set of indexes for all target modalities. The system retrieves the corresponding data from the cloud content object storage based on the index and sends it to the respective modality feature extractor: image and video modalities uniformly use ResNet-18 (retaining only the first 3 residual blocks), while text modalities use the first 4 layers of BERT-Base, with [CLS] position vector pooling followed by 128-dimensional mapping. The feature representation output for each modality is as follows: , and The dimensions are all uniformly set to 128, and they are concatenated to form a joint feature vector. Marketing strategy vector The Brief target content is obtained by BERT encoding.

[0068] Furthermore, a dual-tower structure model is used to calculate the score, with the content tower processing the joint feature vector. Strategy Tower handles marketing strategy vectors The tower structures are all two-layer MLPs, with outputs of respectively and After cosine similarity measurement, the final score is calculated. Considering the possibility of incomplete modalities and missing segments in uploaded content, a modality coverage penalty term and a segment entropy regularization term are introduced to avoid score bias. Specifically, the marketing behavior impact score is calculated based on the joint feature vector, marketing strategy vector, the actual number of segments uploaded for each modality, the preset modality coverage penalty weight, the number of segments to be uploaded for each modality, and the intra-segment information entropy for each modality. The formula is shown below:

[0069] ;

[0070] in, Score the impact of marketing activities. This is the Sigmoid function, with an output range of (0,1). This represents the content side vector obtained after the joint feature vector passes through two fully connected layers. This represents the strategy side vector obtained after passing the marketing strategy vector through two layers of fully connected networks. Represents the target mode The actual number of segments transmitted For target mode The required number of segments to be uploaded (e.g., 3 frames of images, 5 frames of video, 3 segments of text). The modality coverage penalty weight. Represents the target mode The intra-fragment information entropy is defined as the variance of the distance distribution of different fragments in the semantic space within that modality, and is used to measure the diversity of uploaded content; The information entropy regularization weight is used to reward groups of uploaded content with a wider coverage. This scoring mechanism improves the consistency of content selection and scoring by penalizing incomplete information and single uploads, and is suitable for the "asynchronous partial upload" data structure in this invention. It should be noted that "different segments within a modality" refers to different information units belonging to the same modality but independent of each other in content, temporal position, or semantics. For example, different segments in an image modality are the first representative frame, the second representative frame, and the third representative frame; different segments in a video modality are the second-second sampling frame, the fourth-second sampling frame, and the sixth-second sampling frame; and different segments in a text modality are the first, second, and third paragraphs of the text.

[0071] In one embodiment, after calculating the impact score of marketing activities, effective strategy tags need to be generated to obtain a set of effective strategy tags. Its function is to map unstructured content into structured tags that can be understood by downstream decision-making modules.

[0072] Specifically, acquire the strategy dictionary (set by operations, containing approximately 100–300 keywords), and for each keyword... Embedded as keyword vectors by BERT For each modal feature vector and all Perform dot product similarity calculation. If a modality has a similarity to a keyword vector that exceeds a preset threshold, the similarity is calculated. If the value is typically set to 0.75, then the keyword corresponding to the keyword vector is considered activated, defined as a valid strategy tag, and added to the valid strategy tag set. middle.

[0073] In one example, to reduce false activations, a cross-modal label consensus mechanism can be used, which requires a keyword to be activated in at least two modalities before it is considered a valid policy label.

[0074] The marketing behavior impact score and effective strategy tag set are transformed into specific campaign recommendations to support automated campaign decision-making and strategy execution. It is important to emphasize that multiple dimensions must be considered, including configuration differences across various campaign platforms, brief objective requirements, and content tag alignment, to ensure that the generated campaign recommendations are highly adaptable and actionable. Required data includes: Marketing Behavior Impact Score. This represents the behavioral driving force of the target modality under the target brief; the set of effective policy labels. Marketing strategy vector The Brief keywords set by operations are encoded using BERT and used to reflect the semantic direction of this round of campaigning; Platform campaign configuration table , of which each This includes the following fields: platform identifier, modal support type (e.g., a platform supports video but not text / image), and behavioral impact scoring threshold. Strategy Keyword Set (Set by operations) and unit budget conversion ratio (Estimated using historical campaign data, for example, how many clicks correspond to a cost of one yuan).

[0075] For example, in practical use, the operations team issues a brief in the system: "Promoting domestic wedding candies, featuring a modern style." This brief is a set of strategy keywords automatically generated by the operations team's configuration system. {“Wedding”, “Domestic Products”, “Modern”}, where the effective strategy tag set is... The system needs to determine if the tags {“wedding,” “couple,” and “red tones” are aligned, and how marketing activities affect the score. Is it enough? Is it worthwhile to advertise? Which platform is the most suitable for advertising?

[0076] The platform selection logic is controlled by a platform selection function. Its main structure includes filtering for the impact of marketing activities on the score, calculating tag alignment, and prioritizing platforms. First, the impact of marketing activities on the score is excluded. Failure to pass affects the scoring threshold The platform is then used to calculate the semantic tag hit rate for each platform. ,Right now Platform priority factor These are normalized values ​​for historical performance metrics (such as conversion rate or ROI), derived from the platform-level campaign statistics and analysis module. The platform selection function is shown below:

[0077] ;

[0078] in, Targeted delivery platforms. This represents the platform index in the candidate platform set P. This is a logical judgment function that ensures that only the impact of marketing activities on the score is considered. A qualified platform; As a platform priority factor; express and The number of tag intersections, express The number of tags, This reflects the matching degree between the content's semantics and the platform's tasks. It should be noted that satisfying... When the logical judgment function is set to 1, the value is 1; otherwise, it is 0. When the logical judgment function is set to 1, the corresponding platform is eligible to participate in the platform selection calculation. In this case, the corresponding platform will not participate in the platform selection calculation, meaning it will not be selected as the target delivery platform. Therefore, if all platforms fail to meet the requirements... If no platform is selected, the target platform will not be chosen; if at least one platform satisfies the condition... The platform with the highest overall score will ultimately be selected from these platforms. .

[0079] In one embodiment, after selecting the target platform, the impact of the marketing activity needs to be scored. Platform budget conversion ratio Derivation of the recommended budget value. Platform budget conversion rates are typically calculated from historical campaign statistics, such as "on average, 500 clicks are generated for 500 yuan worth of wedding-related content on Xiaohongshu." =1.0. Recommended budget value. The derivation formula is as follows:

[0080] ;

[0081] in, It is the target click target set by the target advertising platform (e.g., 1000). To prevent division by zero for extremely small constants, the default value is 0.01. Recommended budget value is based on the impact score of marketing activities. Automatic scaling; higher scores will generate higher budget recommendations.

[0082] The output is a structure This represents the target platform ID and the referral budget value; if none of the platforms meet the requirements... If the output is empty, it indicates that delivery is not recommended. The final result is encapsulated as a standardized JSON object and pushed to the delivery API module for execution by the actual delivery system.

[0083] Please see Figure 2 This application also provides a cloud-edge collaborative real-time marketing data processing system, which can implement the above-mentioned cloud-edge collaborative real-time marketing data processing method. The system includes:

[0084] The acquisition unit 201 is used to acquire the raw input data uploaded by each client; wherein each raw input data includes multiple modal data;

[0085] The scoring unit 202 is used to obtain the cosine similarity between each modal data and the preset marketing strategy vector, and multiply the cosine similarity with the preset modality density weight factor to obtain the significance score of each modal data.

[0086] The calculation unit 203 is used to calculate the upload scheduling score of each modal data based on the unit content cost, modal value fluctuation factor and significance score of each modal data.

[0087] Unit 204 is used to select the corresponding modal data as the target modality based on the upload scheduling score and the preset upload capacity limit, and upload the target modality to the cloud.

[0088] The specific implementation of this cloud-edge collaborative marketing data real-time processing system is basically the same as the specific implementation of the cloud-edge collaborative marketing data real-time processing method described above, and will not be repeated here.

[0089] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A real-time marketing data processing method based on cloud-edge collaboration, characterized in that, The method includes: Obtain the raw input data uploaded by each client; wherein each raw input data includes multiple modal data; The process involves obtaining the cosine similarity between each modal data point and a preset marketing strategy vector, multiplying the cosine similarity by a preset modality density weighting factor to obtain a saliency score for each modal data point. The original input data includes image, video, and text modal data. Obtaining the cosine similarity between each modal data point and the preset marketing strategy vector, and multiplying the cosine similarity by a preset modality density weighting factor to obtain a saliency score for each modal data point, includes: performing feature extraction based on the image, video, and text modal data to obtain image feature vectors, video feature vectors, and text feature vectors; calculating the cosine similarity between each image feature vector, video feature vector, and text feature vector and the preset marketing strategy vector to obtain image cosine similarity, video cosine similarity, and text cosine similarity; and multiplying the image cosine similarity, video cosine similarity, and text cosine similarity by a preset modality density weighting factor to obtain image saliency score, video saliency score, and text saliency score. The upload scheduling score for each modal data is calculated based on the unit content cost, modal value fluctuation factor, and significance score of each modal data; wherein, the modal value fluctuation factor represents the uncertainty of the effect of the current modal content in recent marketing tasks; Based on the upload scheduling score and the preset upload capacity limit, the corresponding modal data is selected as the target modality, and the target modality is uploaded to the cloud; Feature extraction is performed based on the target modality to obtain modality feature vectors. Each modality feature vector is then concatenated to obtain a joint feature vector. The marketing behavior impact score is calculated based on the joint feature vector, the marketing strategy vector, the number of segments actually uploaded for each modality, the preset modality coverage penalty weight, the number of segments that should be uploaded for each modality, and the information entropy within the segments of each modality.

2. The real-time marketing data processing method based on cloud-edge collaboration according to claim 1, characterized in that, After calculating the marketing behavior impact score based on the joint feature vector, the marketing strategy vector, the number of segments actually uploaded for each modality, the preset modality coverage penalty weight, the number of segments to be uploaded for each modality, and the intra-segment information entropy for each modality, the method further includes: Obtain a strategy dictionary; wherein the strategy dictionary includes multiple keywords and corresponding keyword vectors; Obtain the dot product similarity between each modality feature vector and each keyword vector. If the dot product similarity is greater than a preset threshold, the keyword corresponding to the keyword vector is defined as a valid strategy label. Based on the marketing behavior impact score, the preset behavior impact score threshold, the preset platform priority factor, the effective strategy tags, and the preset strategy keyword set, the target delivery platform is selected from the preset platform delivery configuration table so that the cloud can deliver the target modality to the target delivery platform.

3. The real-time marketing data processing method based on cloud-edge collaboration according to claim 2, characterized in that, After the target modality is delivered to the target delivery platform, the method further includes: Obtain the platform budget conversion ratio and expected click target for the target advertising platform; The expected budget value is obtained by calculating the budget based on the marketing activity impact score, the platform budget conversion ratio, and the expected click target.

4. The real-time marketing data processing method based on cloud-edge collaboration according to claim 1, characterized in that, The upload scheduling score for each modal data is calculated based on the unit content cost, modal value fluctuation factor, and significance score of each modal data, including: Multiply the preset weighting factor by the modal value fluctuation factor, and add the unit content cost to obtain the first data; The saliency score is divided by the first data to obtain the upload scheduling score.

5. The real-time marketing data processing method based on cloud-edge collaboration according to claim 1, characterized in that, The step of selecting the corresponding modal data as the target modality based on the upload scheduling score and the preset upload capacity limit includes: The upload scheduling scores are sorted from high to low, and the capacities of the corresponding modal data are added together until the accumulated capacity reaches the upload capacity limit. The selected modal data is then the target modal.

6. A real-time marketing data processing system based on cloud-edge collaboration, characterized in that: The system includes: The acquisition unit is used to acquire the raw input data uploaded by each client; wherein each raw input data includes multiple modal data; A scoring unit is used to obtain the cosine similarity between each modal data and a preset marketing strategy vector, and multiply the cosine similarity by a preset modality density weighting factor to obtain a saliency score for each modal data. The original input data includes image modality, video modality, and text modality. Obtaining the cosine similarity between each modal data and the preset marketing strategy vector, and multiplying the cosine similarity by the preset modality density weighting factor to obtain a saliency score for each modal data, includes: performing feature extraction based on the image modality, video modality, and text modality respectively to obtain image feature vectors, video feature vectors, and text feature vectors; calculating the cosine similarity between the image feature vector, video feature vector, and text feature vector and the preset marketing strategy vector respectively to obtain image cosine similarity, video cosine similarity, and text cosine similarity; and multiplying the image cosine similarity, video cosine similarity, and text cosine similarity by the preset modality density weighting factor to obtain image saliency score, video saliency score, and text saliency score. The calculation unit is used to calculate the upload scheduling score of each modal data based on the unit content cost, modal value fluctuation factor and significance score of each modal data; wherein, the modal value fluctuation factor represents the uncertainty of the effect of the current modal content in the recent marketing task; The selection unit is used to select the corresponding modal data as the target modality based on the upload scheduling score and the preset upload capacity limit, and upload the target modality to the cloud; Feature extraction is performed based on the target modality to obtain modality feature vectors. Each modality feature vector is then concatenated to obtain a joint feature vector. The marketing behavior impact score is calculated based on the joint feature vector, the marketing strategy vector, the number of segments actually uploaded for each modality, the preset modality coverage penalty weight, the number of segments that should be uploaded for each modality, and the information entropy within the segments of each modality.

7. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the real-time marketing data processing method based on cloud-edge collaboration as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the real-time marketing data processing method based on cloud-edge collaboration as described in any one of claims 1 to 5.

Images