A cloud-computing-based cross-platform advertisement pushing and feedback integration system and method

Abstract

This invention discloses a cloud-based cross-platform advertising push and feedback integration system and method, relating to the field of advertising data integration. The system includes: regional aggregation of encrypted behavioral data to generate encrypted regional user profile fragments; verification of the validity of interest tags through cross-platform behavior backtracking to obtain negative feedback markers; processing of the encrypted regional user profile fragments using a pre-trained DNN model to obtain an encrypted global user profile; generation of a dynamic advertising delivery strategy through a reinforcement learning algorithm; triggering a self-healing process for the profile through negative feedback markers to update the encrypted global user profile; decrypting the updated encrypted global user profile; matching the dynamic advertising delivery strategy, the decrypted global user profile, and the local advertising inventory to obtain and execute advertising push instructions. This invention improves advertising conversion rates and real-time performance through optimized strategies, enhancing the accuracy and long-term effectiveness of cross-platform advertising.

Description

Technical Field

[0001] This invention relates to the field of advertising data integration, and in particular to a cloud-based cross-platform advertising push and feedback integration system and method. Background Technology

[0002] In recent years, with the rapid development of cloud computing, edge computing, and artificial intelligence technologies, cross-platform advertising push systems have made significant progress in the field of precision marketing. Traditional advertising push systems mainly rely on centralized data processing, building user profiles through user behavior data and generating personalized recommendations using machine learning algorithms. In recent years, the rise of federated learning technology has supported distributed data processing, allowing model training without sharing raw data, partially solving privacy protection issues. Homomorphic encryption technology has further promoted encrypted domain data analysis, ensuring the security of user data and significantly improving real-time performance and processing efficiency.

[0003] Despite the significant progress made in ad delivery technology, there are still areas for improvement. First, there are limitations in dynamic adaptability, making it difficult to quickly adjust user profiles and delivery strategies based on real-time user feedback, especially in negative feedback scenarios, where there is a lack of self-healing mechanisms to dynamically remove invalid interest tags. In addition, existing feedback loop mechanisms are mostly offline updates, making it difficult to achieve a closed loop of real-time data collection and profile optimization, which limits the long-term improvement of user satisfaction and advertising effectiveness. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a cloud computing-based cross-platform advertising push and feedback integration method to solve the problems of lacking a self-healing mechanism to dynamically remove invalid interest tags and difficulty in achieving a closed loop of real-time data collection and profile optimization.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a cross-platform advertising push and feedback integration method based on cloud computing, comprising,

[0008] Real-time collection of user interaction data to generate encrypted behavioral data;

[0009] The encrypted behavioral data is aggregated by region to generate encrypted regional user profile fragments, and the validity of interest tags is verified through cross-platform behavior backtracking to obtain negative feedback markers;

[0010] The encrypted regional user profile fragments are processed using a pre-trained DNN model to obtain an encrypted global user profile. A dynamic advertising strategy is generated through a reinforcement learning algorithm, and a self-healing process is triggered by negative feedback marking to update the encrypted global user profile.

[0011] The updated encrypted global user profile is decrypted, and the dynamic ad delivery strategy, the decrypted global user profile, and the local ad inventory are matched to obtain and execute ad push instructions. At the same time, user feedback data is collected, and the decrypted regional user profile fragments are updated based on the user feedback data and then re-encrypted.

[0012] Based on the updated encrypted regional user profile fragments, the encrypted global user profile is adjusted to generate an optimized dynamic advertising delivery strategy.

[0013] As a preferred embodiment of the cloud computing-based cross-platform advertising push and feedback integration method described in this invention, the generation of encrypted behavioral data refers to capturing user clicks, page dwell time, swiping behavior and device metadata, filtering invalid data to generate valid interaction data, and using the Paillier homomorphic encryption algorithm to encrypt and generate encrypted behavioral data.

[0014] As a preferred embodiment of the cloud computing-based cross-platform advertising push and feedback integration method of the present invention, the step of performing regional aggregation on encrypted behavioral data to generate encrypted regional user profile fragments specifically includes the following steps.

[0015] The fog computing layer assigns weights to encrypted behavior data by calculating time weights, and performs ciphertext domain weighted aggregation on the encrypted behavior data through federated learning client algorithms to generate encrypted regional interest vectors.

[0016] The encrypted regional interest vector is combined with the corresponding regional identifier and timestamp to form an encrypted regional user profile fragment.

[0017] As a preferred embodiment of the cloud-based cross-platform advertising push and feedback integration method of the present invention, the step of verifying the validity of interest tags and obtaining negative feedback markers through cross-platform behavior backtracking specifically includes the following steps.

[0018] Interest tags are extracted from encrypted user profile fragments in the region. External data is queried through social media APIs. VADER sentiment analysis is used to process the external data, calculate the composite sentiment score for each external data point, and count the percentage of negative sentiment.

[0019] When negative emotions account for more than half, a negative feedback tag is generated for the interest tag.

[0020] As a preferred embodiment of the cloud computing-based cross-platform advertising push and feedback integration method of the present invention, the step of obtaining the encrypted global user profile specifically includes the following steps.

[0021] Initialize the DNN model, use the FedAvg framework, process historically encrypted regional user profile fragments in the ciphertext domain, and read historically encrypted regional user profile fragments in each training round.

[0022] For each training round, the encrypted gradient is calculated in the encrypted domain, and the accuracy of the DNN model is measured based on the mean squared error loss function.

[0023] The federated learning server algorithm is used to perform a weighted average of the cryptographic gradients, update the cryptographic weights of the DNN model, and generate the trained DNN model.

[0024] The encrypted regional user profile fragments are input into the trained DNN model to output an encrypted global interest vector. The encrypted global interest vector is then integrated with the corresponding regional identifier and timestamp to form an encrypted global user profile.

[0025] As a preferred embodiment of the cloud-based cross-platform advertising push and feedback integration method of the present invention, the step of generating a dynamic advertising delivery strategy through a reinforcement learning algorithm and simultaneously triggering a profile self-healing process through negative feedback labeling to update the encrypted global user profile specifically includes the following steps.

[0026] The encrypted global user profile is input into Deep-Q-Learning, the Q value of the ad delivery action is calculated, a Q value vector is formed, and N optimal actions are selected from the Q value vector;

[0027] Integrate the selected actions into a dynamic ad delivery strategy;

[0028] Analyze negative feedback markers, identify invalid interest labels, calculate the encryption weight adjustment amount, update the encryption weights, and generate an updated encryption global interest vector.

[0029] As a preferred embodiment of the cloud-based cross-platform advertising push and feedback integration method described in this invention, the method includes: decrypting the updated encrypted global user profile, matching the dynamic advertising delivery strategy, the decrypted global user profile, and the local advertising inventory, obtaining and executing advertising push instructions, collecting user feedback data during execution, updating the decrypted regional user profile fragments based on the user feedback data, and re-encrypting them. Specifically, this includes the following steps.

[0030] Receive the updated encrypted global user profile and decrypt it to obtain the plaintext global interest vector;

[0031] The plaintext global interest vector is mapped to the regional user segment to generate a decrypted regional user profile fragment.

[0032] A candidate ad list is generated based on the ad records in the local ad inventory. The cosine similarity between the decrypted regional user profile fragments is calculated. The ad with the highest cosine similarity is selected and verified using a dynamic ad delivery strategy. If there are no errors, an ad push instruction is obtained.

[0033] Execute advertising push instructions and collect user click data and dwell time, filter them to form user feedback data, and perform negative feedback detection. If negative feedback is detected, check the authenticity of the negative feedback. If it is true, update the decrypted regional user profile fragment.

[0034] The updated and decrypted regional user profile fragments and user feedback data will be encrypted and then transmitted to the cloud.

[0035] Secondly, this invention provides a cloud-based cross-platform advertising push and feedback integration system, including:

[0036] The data acquisition module collects user interaction data in real time and generates encrypted behavioral data.

[0037] The aggregation module performs regional aggregation on encrypted behavioral data, generates encrypted regional user profile fragments, and verifies the validity of interest tags through cross-platform behavior backtracking to obtain negative feedback markers.

[0038] The generation module uses a pre-trained DNN model to process encrypted regional user profile fragments to obtain encrypted global user profiles. It also generates dynamic advertising strategies through reinforcement learning algorithms and triggers a self-healing process for the profiles by using negative feedback tags to update the encrypted global user profiles.

[0039] The delivery module decrypts the updated encrypted global user profile, matches the dynamic ad delivery strategy, the decrypted global user profile, and the local ad inventory, obtains and executes ad push instructions, collects user feedback data during execution, updates the decrypted regional user profile fragments based on the user feedback data, and re-encrypts them.

[0040] The feedback module, based on the updated encrypted regional user profile fragments, uses negative feedback to drive adjustments to the encrypted global user profile, generating an optimized dynamic advertising delivery strategy.

[0041] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the cloud computing-based cross-platform advertising push and feedback integration method described in the first aspect of the present invention.

[0042] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the cloud computing-based cross-platform advertising push and feedback integration method described in the first aspect of the present invention.

[0043] The beneficial effects of this invention are as follows: It generates an encrypted global user profile by fusing encrypted regional user profile fragments with a pre-trained encrypted deep neural network model. Combined with a gradient descent-based self-healing algorithm to analyze negative feedback markers and dynamically adjust encryption weights, it reduces the proportion of invalid interest tags. Furthermore, it generates dynamic advertising delivery strategies through reinforcement learning algorithms, improving ad click-through rates and achieving dynamic adaptation between user profiles and strategies, while reducing invalid push notifications. Simultaneously, through real-time user feedback data collection and cloud-based closed-loop optimization, it utilizes the optimized encrypted deep neural network model and Deep-Q-Learning to generate optimization strategies, improving ad conversion rates and real-time performance, and significantly enhancing the accuracy and long-term effectiveness of cross-platform advertising. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of a cloud-based cross-platform advertising push and feedback integration method.

[0046] Figure 2 This is a schematic diagram of a cloud-based cross-platform advertising push and feedback integration system.

[0047] Figure 3 A flowchart illustrating the self-healing mechanism of the image.

[0048] Figure 4 A schematic diagram for closed-loop optimization. Detailed Implementation

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0050] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0051] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0052] Reference Figures 1-4 This is one embodiment of the present invention, which provides a cloud computing-based cross-platform advertising push and feedback integration method, including the following steps:

[0053] S1: Collect user interaction data in real time and generate encrypted behavioral data;

[0054] Specifically, the steps include the following:

[0055] At the edge layer, user interaction data is collected in real time using lightweight software development kits (SDKs). This data includes clicks, page dwell time, swipe actions, and device metadata such as operating system type and screen resolution. The SSDs are implemented in JavaScript to support web platforms or in native code to support iOS and Android platforms, generating JSON objects containing anonymous user IDs, timestamps, behavior types, and contextual information such as geolocation.

[0056] User interaction data is filtered through a lightweight rules engine, and invalid interaction data is removed. For example, records with a page dwell time of less than 1 second are marked as invalid and discarded, generating valid interaction data. Valid interaction data includes records of ad clicks or video ad views exceeding 5 seconds, which are stored as JSON objects containing anonymous user ID, timestamp, behavior type, and context information.

[0057] To further explain, filtering invalid data reduces interference from noisy data, improves data quality, ensures that the fog computing layer receives highly relevant and effective interactive data, reduces computing and transmission costs, provides accurate input for generating regional user profile fragments, and enhances the targeting of advertising pushes.

[0058] The Paillier homomorphic encryption algorithm is applied to encrypt valid interaction data to generate encrypted behavior data. This algorithm, based on the PySEAL library, is executed on the user terminal. The specific process is as follows: the JSON object of the valid interaction data is converted into a numerical vector, for example, mapping behavior type and context information to a 64-bit floating-point vector; the numerical vector is then encrypted element-by-element using the Paillier homomorphic encryption algorithm to generate encrypted behavior vectors, such as a 64-dimensional ciphertext array; the anonymous user ID remains unchanged after SHA-256 hashing, and the timestamp is preserved in plaintext to support subsequent time series analysis; the encrypted behavior data is stored in JSON format, containing fields such as the encrypted behavior vector, the anonymous user ID hash, and the timestamp; and the encrypted behavior data is transmitted to the fog computing layer via a low-latency network using the HTTPS protocol.

[0059] To further explain, the Paillier homomorphic encryption algorithm supports ciphertext addition operations, allowing data analysis while protecting user privacy, preventing data leakage, and providing a trust foundation for cross-platform data transmission.

[0060] S2: Perform regional aggregation on encrypted behavioral data to generate encrypted regional user profile fragments, and verify the validity of interest tags through cross-platform behavior backtracking to obtain negative feedback markers;

[0061] Specifically, the steps include the following:

[0062] The fog computing layer is deployed on regional nodes, such as edge servers in cities or business districts. Regional nodes receive encrypted behavioral data from the edge layer. The encrypted behavioral data is transmitted in JSON format via HTTPS protocol and includes fields such as encrypted behavioral vectors, anonymous user ID hashes, and timestamps. The receiving process distributes traffic through a load balancer. Regional nodes use a Redis database to cache the encrypted behavioral data to support low-latency access.

[0063] S2.1 The cached encrypted behavior data is aggregated using a federated learning client algorithm to generate encrypted regional user profile fragments. The federated learning client algorithm is based on the FedAvg framework and implemented using TensorFlow Federated. The aggregation process performs a weighted average calculation on the encrypted behavior data of multiple anonymous users within a region. It utilizes the addition property of the Paillier homomorphic encryption algorithm to perform operations in the ciphertext domain, eliminating the need to decrypt the original data. Specifically, the cached encrypted behavior data is read from Redis, for example, 1000 encrypted behavior data entries from a business district in the last hour. Each entry contains a 64-dimensional encrypted behavior vector, corresponding to predefined interest tags, such as the first dimension representing "coffee" interest and the second dimension representing "sports" interest. Based on time weights, a weight is assigned to each encrypted behavior data entry. For example, the weight of data from the last 10 minutes is 1.0, the weight from 10-30 minutes is 0.9, and the weight from 30-60 minutes is 0.8. The weights are calculated using a linear decay formula:

[0064] ;

[0065] in, This represents the weight assigned to the encrypted behavioral data, determining its contribution to region aggregation in the joint learning client algorithm. This indicates that a larger value is selected. This indicates the current time of the fog computing layer when aggregation is performed. This represents the timestamp of the encrypted action data stored in the JSON object. The time difference between the current time and the data timestamp indicates freshness. It is the number of seconds within one hour.

[0066] By using the addition and scalar multiplication operations of the Paillier homomorphic encryption algorithm, the corresponding dimensions of the encryption behavior vector are weighted and summed one by one to obtain the weighted sum vector of the encryption, the expression of which is:

[0067] ;

[0068] in, The first element represents the weighted sum vector of the encryption. Victoria's Secret value, Indicates the first The weight of each encrypted action data, Indicates the first The first encrypted action vector of the encrypted action data Victoria's Secret value, An index representing encrypted behavioral data. The dimension index of the encrypted behavior vector.

[0069] Divide the encrypted weighted sum vector by the total weights to generate a 64-dimensional encrypted regional interest vector. For example, the encrypted value representing the interest in "coffee" corresponds to 0.7, but it is actually stored as ciphertext. After obtaining the 64-dimensional encrypted regional interest vector, combine the encrypted regional interest vector with the regional identifier and timestamp into a JSON object to form an encrypted regional user profile fragment. The encrypted regional interest vector is a 64-dimensional ciphertext array, and the encrypted regional user profile fragment is cached in Redis.

[0070] To further explain, the distributed aggregation of the federated learning client algorithm reduces computational overhead, generates accurate encrypted regional user profile fragments, provides high-quality data support for regionalized advertising push, and improves the computational efficiency of the fog computing layer.

[0071] S2.2 Perform cross-platform behavior backtracking on interest tags in encrypted regional user profile fragments to verify the validity of interest tags and generate negative feedback markers. Cross-platform behavior backtracking obtains anonymized user data within the region through social media APIs, such as using the Meituan API to query public posts related to interest tags. Specifically, interest tag identifiers, such as "coffee," "sports," and "travel," are extracted from encrypted regional user profile fragments. These identifiers are stored in plaintext in the metadata field of a JSON object. Relevant posts within a certain region are queried through the search endpoint of the Meituan API, for example, retrieving 1000 public posts related to "coffee" in the last 24 hours.

[0072] The VADER sentiment analysis is applied to the post text to calculate the sentiment score for each post. Specifically, the post text is extracted from the post data obtained from the Meituan API, and a text list containing the text content is generated. For example, 1000 post texts related to "coffee" are extracted, such as "coffee is too bitter" and "this coffee shop is great". The Python-NLTK library's preprocessing function is applied to each post text in the text list to generate standardized text. VADER sentiment analysis is applied to the standardized text to generate a sentiment score dictionary for each post. Composite sentiment scores are extracted from the sentiment score dictionary to generate a composite sentiment score list. Each post is classified according to the composite sentiment score list to generate sentiment classification results. For example, a positive score > 0.05 is positive, a negative score < -0.05 is negative, and a neutral score is between [-0.05, 0.05].

[0073] The process involves calculating the percentage of negative posts in the sentiment category results and generating a negative percentage value. For example, if 600 negative posts represent 60% of 1000 total posts, a negative percentage value (e.g., 0.6) is generated. Based on this negative percentage value, negative feedback tags are generated and stored in JSON format. The negative percentage value is compared for each interest tag; if it exceeds half, a negative feedback tag is generated. For instance, if 600 negative posts represent 60% of 1000 total posts, a negative feedback tag is generated. These tags are stored as a hash table using the Redis HSET command. The Redis HSET command is then used to associate the region identifier of the negative feedback tag with the region identifier of an encrypted user profile fragment. Finally, the encrypted user profile fragment and the negative feedback tags are uploaded to the cloud.

[0074] To further explain, cross-platform behavior retrospective analysis improves the accuracy of interest tags through Meituan API and VADER sentiment analysis, optimizing the precision of subsequent advertising strategies. Moreover, negative feedback marking provides key feedback for cloud-based adjustments to encrypted global user profiles and optimization of dynamic advertising strategies, enhancing the targeting of advertising and user satisfaction, and supports cross-platform data integration.

[0075] S3: Use a pre-trained DNN model to process encrypted regional user profile fragments to obtain encrypted global user profiles, and generate dynamic advertising strategies through reinforcement learning algorithms. At the same time, trigger the profile self-healing process through negative feedback marking to update the encrypted global user profiles.

[0076] Specifically, the steps include the following:

[0077] S3.1: Deployed on a scalable cloud computing platform, such as AWS or Alibaba Cloud, it receives encrypted regional user profile fragments and negative feedback tags from multiple fog computing nodes. For the encrypted regional user profile fragments, a pre-trained encrypted deep neural network model is applied for processing to obtain an encrypted global user profile. The specific process is as follows:

[0078] (1) Define the architecture of the encrypted deep neural network model: The encrypted deep neural network model consists of 3 fully connected layers with 256 neurons in each layer. The ReLU activation function is used. The input layer receives a 64-dimensional encrypted regional interest vector, and the output layer generates a 64-dimensional encrypted global interest vector.

[0079] (2) Training the encrypted deep neural network model: The model is trained on historically encrypted regional user profile fragments using the federated learning server algorithm, and the FedAvg framework is used to complete the training in the ciphertext domain. The training data consists of historically encrypted regional user profile fragments, for example, 10,000 fragments covering 100 regions over 30 days. The training process is as follows: Initialize the model weights, read 1,000 encrypted regional user profile fragments in each training round, divide them into 10 batches, calculate the gradients using the Paillier homomorphic encryption algorithm with addition and scalar multiplication, compare the model output with the historically encrypted global interest vector based on the mean squared error loss function, aggregate the gradients using the FedAvg algorithm, for example, by weighted averaging the encryption gradients of 10 regions, with the weights based on the number of users in each region, update the weights of the encrypted deep neural network model, optimize through gradient descent, train for 100 rounds, and generate the trained encrypted deep neural network model.

[0080] (3) Encrypted deep neural network model processes encrypted regional user profile fragments to generate encrypted global user profiles: Encrypted regional user profile fragments are extracted, each fragment containing a 64-dimensional encrypted regional interest vector. The encrypted regional interest vectors are extracted through MongoDB index and used as input to the encrypted deep neural network model. The trained encrypted deep neural network model is weighted and summed to fuse multi-region data to generate a 64-dimensional encrypted global interest vector. After obtaining the 64-dimensional encrypted global interest vector, the global identifier and timestamp are integrated to form an encrypted global user profile. The encrypted global user profile is stored as a JSON object, containing fields such as the global identifier, encrypted global interest vector, and timestamp.

[0081] To further explain, the encrypted global user profile formed by the pre-trained encrypted deep neural network model captures the cross-platform interest distribution, improves the accuracy of user profiles, and provides a high-quality foundation for dynamic advertising.

[0082] S3.2 Apply reinforcement learning algorithm to the encrypted global user profile to generate dynamic advertising delivery strategy. The reinforcement learning algorithm is based on Deep-Q-Learning, runs in the TensorFlow environment, uses a 4-layer fully connected network with 128 neurons in each layer, and the activation function is ReLU. The input is the encrypted global interest vector, and the output layer generates the Q-value vector. The network parameters are initialized by pre-training with historical advertising delivery data to obtain the pre-trained Deep-Q-Learning network.

[0083] The encrypted global interest vector is input into a pre-trained Deep-Q-Learning network to calculate the Q-value of ad delivery actions and optimize long-term metrics such as click-through rate and conversion rate. Ad delivery actions are defined as selecting ad identifiers, target user segments, and platforms, such as the action "push coffee_ad_001 to mobile users". The Q-value function calculates the expected return for each action, and its expression is:

[0084] ;

[0085] in, This represents the Q-value function calculated by the Deep-Q-Learning network in the reinforcement learning algorithm, which measures the action. Next, execute the advertising placement action. Expected returns This indicates an action, specifically encrypting the global interest vector. This indicates the act of placing an advertisement. This indicates the click-through rate. Weighting that represents click-through rate. Indicates conversion rate. The weight representing the conversion rate.

[0086] Deep-Q-Learning networks process the input through forward propagation, calculate the Q-value vector of actions, select the action with the highest Q-value, generate a Q-value vector, and select the 10 actions with the highest Q-values ​​based on the Q-value vector to form a dynamic advertising delivery strategy, which includes fields such as ad identifier, target user segment, platform, and delivery time.

[0087] To further explain, based on Deep-Q-Learning, the dynamic ad delivery strategy obtained by using encrypted global interest vectors as input can improve ad click-through rate and conversion rate, dynamically adapt to regional interests, and improve the accuracy of ad delivery.

[0088] S3.3 Execute the profile self-healing process to check for invalid conditions for negative feedback tags and determine the interest tags that need adjustment. The profile self-healing process analyzes the tag status of negative feedback tags and related historical advertising data, such as 10,000 ad placement records within 30 days, including click-through rate and conversion rate. An invalid condition is that an interest tag has no clicks or conversions for three consecutive times. For example, the click-through rate of the "coffee" tag is 0.08%, 0.06%, and 0.0% in three ad placements. The check process counts the interest tags that meet the invalid condition and generates a list of invalid interest tags.

[0089] The algorithm extracts interest tags that need adjustment from the list of invalid interest tags, generates a target interest tag index, extracts the encryption weights corresponding to the target interest tag index from the pre-trained encrypted deep neural network model, and generates a subset of encryption weights. For example, it extracts the first-dimensional encryption weight matrix of "coffee". It uses a gradient descent-based self-healing algorithm to calculate the encryption weight adjustment amount and generates an encryption gradient vector. It uses the Paillier homomorphic encryption algorithm to perform addition operations and updates the encryption weight subset with the encryption gradient vector to obtain the updated encryption weight subset. It integrates the updated encryption weight subset back into the pre-trained encrypted deep neural network model to generate an updated encryption weight matrix. It replaces the encryption weights of the corresponding dimensions in the original weight matrix. It re-runs the pre-trained encrypted deep neural network model with the updated encryption weight matrix to generate an updated encrypted global user profile. The updated encrypted global user profile is stored as a JSON object and contains fields such as global identifier, encrypted global interest vector, and timestamp. The updated encrypted global user profile and dynamic advertising delivery strategy are sent to the fog computing layer via a RabbitMQ message queue.

[0090] To further explain, the user profile self-healing process can reduce invalid interest tags, improve the dynamic adaptability of user profiles, reduce invalid ad pushes, and support long-term user satisfaction improvement.

[0091] S4: Decrypt the updated encrypted global user profile, match the dynamic ad delivery strategy, the decrypted global user profile, and the local ad inventory, obtain and execute ad push instructions, collect user feedback data during execution, update the decrypted regional user profile fragments based on user feedback data, and re-encrypt them;

[0092] Specifically, the steps include the following:

[0093] S4.1: Receives the updated encrypted global user profile and dynamic advertising strategy from the cloud. The updated encrypted global user profile contains a 64-dimensional encrypted global interest vector. It decrypts the encrypted global interest vector using locally stored Paillier homomorphic encryption, generating a plaintext global interest vector. Based on the plaintext global interest vector, combined with the region identifier and local user segmentation information, it generates decrypted regional user profile fragments. The region identifier and user segmentation information (e.g., hashes of 1000 anonymous user IDs) are stored in a Redis database. The generation process maps the plaintext global interest vector to regional user segments, generating decrypted regional user profile fragments. Each fragment contains a 64-dimensional interest vector, a region identifier, and a timestamp. For example, the first dimension of the interest vector represents an interest of 0.1 for "coffee," and the second dimension represents an interest of 0.3 for "sports."

[0094] The system queries local ad inventory for ad records. Each ad record contains an ad identifier, ad content, interest tags, and platform information. Matching ad records are retrieved using the interest tags as keys, generating a candidate ad list. The system compares the interest vectors of the decrypted regional user profile fragment with the interest tags of the candidate ad list, calculating a matching score using cosine similarity. The expression for this score is:

[0095] ;

[0096] in, Represents cosine similarity, with a value range of [0,1]. Index representing the dimension of the vector. This represents the total number of vector dimensions. The first element representing the interest vector of the decrypted regional user profile fragment. Dimensional value, The first element representing the interest tag vector of the candidate advertisement Dimensional value.

[0097] Interest thresholds are set based on the interest distribution of historical ad delivery data, for example, 0.05. Cosine similarity thresholds are set based on the cosine similarity distribution of historical ad delivery data, for example, 0.1. If the value of the "coffee" dimension in the interest vector is >0.05 and the cosine similarity is >0.1, then the ad is considered a matching result. Each matching result includes the ad identifier, score, and platform metadata.

[0098] For each decrypted regional user profile segment, the advertisement with the highest cosine similarity is selected, and platform consistency is verified in conjunction with the dynamic ad delivery strategy. The verification process is as follows: Records matching the ad identifier are searched from the dynamic ad delivery strategy; platform fields are extracted; and the platform metadata of the ad inventory is compared with the platform fields of the dynamic ad delivery strategy. If they match, the verification passes; if they do not match, the ad with the second highest similarity is selected, and the verification is repeated until the platform matches or no ads are available. The ad push instruction includes ad content, anonymous user ID hash, and platform information.

[0099] To further explain, based on dynamic ad delivery strategies, decrypted regional user profile fragments, and local ad inventory matching ads, using cosine similarity to calculate matching scores and generate ad push instructions can improve click-through rates and conversion rates, reduce mismatches, and enhance ad accuracy.

[0100] S4.2: Ad push instructions are sent to edge devices in real time via the WebSocket protocol. The edge devices use the SDK to adaptively adjust the ad format according to device characteristics based on the ad push instructions, such as displaying banner ads in mobile apps or pop-ups on web pages. The edge devices track user interactions through the SDK and generate user feedback data, including clicks, dwell time, conversions (such as purchases) or skips. The SDK applies filtering rules to filter user feedback data, such as discarding invalid interactions with dwell time of less than 1 second. Valid user feedback data is encrypted using the Paillier homomorphic encryption algorithm. If negative feedback is detected, the SDK marks an interest tag and includes it in the user feedback data. For example, "coffee_ad_001" has no clicks for three consecutive times, with a click-through rate of less than 0.1%. Encrypted user feedback data is transmitted to the fog computing layer via HTTPS. After receiving the encrypted user feedback data, the fog computing layer decrypts it into plaintext user feedback data, detects negative feedback in the plaintext user feedback data, and if the negative feedback in the plaintext user feedback data is confirmed to be real, it updates the decrypted regional user profile fragment and freezes the interest tags that have been confirmed to be real negative feedback. For example, if a query through the Meituan API shows that 60% of the posts are negative, such as "coffee is too bitter", the "coffee" tag is frozen, the weight of the tag is set to 0, and an updated decrypted regional user profile fragment is generated.

[0101] The updated and decrypted regional user profile fragments and user feedback data are re-encrypted to obtain encrypted regional user profile fragments and encrypted user feedback data, which are then transmitted to the cloud together.

[0102] To further explain, real-time feedback collection and negative feedback detection can reduce invalid ad pushes, API verification improves the accuracy of negative feedback, and updating regional user profile fragments, dynamic updates reduce invalid interest tags, improve user profile adaptability, and enhance the accuracy of long-term advertising.

[0103] S5: Adjust the encrypted global user profile based on the updated encrypted regional user profile fragments to generate an optimized dynamic ad delivery strategy.

[0104] Specifically, the steps include the following:

[0105] Cloud-based analysis of encrypted user feedback data checks for invalid conditions based on negative feedback markers to determine interest tags that need adjustment. The analysis process involves: extracting encrypted user feedback data from the Cassandra distributed database, decrypting it using the Paillier homomorphic encryption algorithm with a private key to generate plaintext user feedback data, querying historical advertising data (10,000 records within 30 days, including click-through rate and conversion rate), statistically analyzing negative feedback patterns, and combining this with cross-platform backtracking results from the fog computing layer (using the Meituan API to query negative sentiment). An invalid condition is an interest tag with three consecutive days without clicks or conversions. The check process then compiles a list of interest tags that meet the invalid condition, generating a comprehensive list of invalid interest tags. The algorithm extracts interest tags that need adjustment from the comprehensive list of invalid interest tags, generates a target interest tag index, extracts 64-dimensional encrypted regional interest vectors corresponding to the target interest tag index from the updated encrypted regional user profile fragments, generates an encrypted interest vector subset, calculates the adjustment amount of the encrypted interest vector using a gradient descent-based self-healing algorithm, generates an encrypted gradient vector, updates the encrypted interest vector subset using the encrypted gradient vector through the Paillier homomorphic encryption algorithm, obtains the updated encrypted interest vector subset, integrates the updated encrypted interest vector subset back into the encrypted global user profile, generates the updated encrypted global interest vector. The federated learning deep neural network model is retrained based on the FedAvg framework, running in the encrypted domain, with 1000 encrypted global user profiles and 1000 updated encrypted regional user profile fragments as input. Each round reads 1000 encrypted fragments in 10 batches, calculates the encryption gradient using the Paillier homomorphic encryption algorithm's addition and scalar multiplication, updates the encryption weights using a weighted average, trains for 100 rounds, and generates the optimized federated learning deep neural network model. An optimized federated learning deep neural network model is used to generate an updated encrypted global user profile. The updated encrypted global user profile and the optimized dynamic advertising delivery strategy are then sent to the fog computing layer via a RabbitMQ message queue.

[0106] The optimized federated learning deep neural network model processes 1000 updated encrypted global user profiles to generate 64-dimensional encrypted global interest vectors. These vectors are then input into the Deep-Q-Learning network, which uses the Paillier homomorphic encryption algorithm with a private key to decrypt the encrypted global interest vectors, generating plaintext global interest vectors. A 100-dimensional Q-value vector (corresponding to 100 actions) is calculated, and the expected return for each action is calculated. The action with the highest Q-value is selected to generate 1000 optimized dynamic advertising strategies, completing the closed loop.

[0107] This embodiment also provides a cloud computing-based cross-platform advertising push and feedback integration system, including:

[0108] The data acquisition module collects user interaction data in real time and generates encrypted behavioral data.

[0109] The aggregation module performs regional aggregation on encrypted behavioral data, generates encrypted regional user profile fragments, and verifies the validity of interest tags through cross-platform behavior backtracking to obtain negative feedback markers.

[0110] The generation module uses a pre-trained DNN model to process encrypted regional user profile fragments to obtain encrypted global user profiles. It also generates dynamic advertising strategies through reinforcement learning algorithms and triggers a self-healing process for the profiles by using negative feedback tags to update the encrypted global user profiles.

[0111] The delivery module decrypts the updated encrypted global user profile, matches the dynamic ad delivery strategy, the decrypted global user profile, and the local ad inventory, obtains and executes ad push instructions, collects user feedback data during execution, updates the decrypted regional user profile fragments based on the user feedback data, and re-encrypts them.

[0112] The feedback module, based on the updated encrypted regional user profile fragments, uses negative feedback to drive adjustments to the encrypted global user profile, generating an optimized dynamic advertising delivery strategy.

[0113] This embodiment also provides a computer device applicable to the cross-platform advertising push and feedback integration method based on cloud computing, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the cross-platform advertising push and feedback integration method based on cloud computing as proposed in the above embodiment.

[0114] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0115] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the cross-platform advertising push and feedback integration method based on cloud computing as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0116] In summary, this invention generates an encrypted global user profile by fusing encrypted regional user profile fragments with a pre-trained encrypted deep neural network model. It then analyzes negative feedback markers using a gradient descent-based self-healing algorithm, dynamically adjusting encryption weights to reduce the proportion of invalid interest tags. Furthermore, it generates dynamic advertising delivery strategies through reinforcement learning algorithms, improving ad click-through rates and achieving dynamic adaptation between user profiles and strategies while reducing invalid push notifications. Simultaneously, through real-time user feedback data collection and cloud-based closed-loop optimization, it utilizes the optimized encrypted deep neural network model and Deep-Q-Learning to generate optimization strategies, thereby improving ad conversion rates and real-time performance, and significantly enhancing cross-platform ad accuracy and long-term effectiveness.

[0117] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A cloud computing-based cross-platform advertisement pushing and feedback integration method, characterized in that: Comprising, Real-time collection of user interaction data to generate encrypted behavior data; Region aggregation of the encrypted behavior data to generate encrypted regional user portrait segments, and verification of interest tag validity through cross-platform behavior backtracking to obtain negative feedback markers; The region aggregation of the encrypted behavior data to generate encrypted regional user portrait segments specifically includes the following steps, The fog computing layer assigns weights to the encrypted behavior data by calculating time weights, and performs ciphertext domain weighted aggregation on the encrypted behavior data to generate an encrypted regional interest vector; The encrypted regional interest vector is combined with the corresponding regional identifier and timestamp to form an encrypted regional user portrait segment; The encrypted global user portrait is obtained by processing the encrypted regional user portrait segment using a pre-trained DNN model, and a dynamic ad placement strategy is generated through a reinforcement learning algorithm, while the portrait self-healing process is triggered by the negative feedback markers to update the encrypted global user portrait; The encrypted global user portrait specifically includes the following steps, Initialize the DNN model, use the FedAvg framework to process historical encrypted regional user portrait segments in the ciphertext domain, and read historical encrypted regional user portrait segments in each training round; For each training round, calculate the encrypted gradient in the ciphertext domain, and measure the accuracy of the DNN model based on the mean square error loss function; Weighted average of the encrypted gradient, update the encrypted weight of the DNN model, generate a trained DNN model; The encrypted regional user portrait segment is input into the trained DNN model to output an encrypted global interest vector, and the encrypted global interest vector is integrated with the corresponding regional identifier and timestamp to form an encrypted global user portrait; The reinforcement learning algorithm is used to generate a dynamic ad placement strategy, and the portrait self-healing process is triggered by the negative feedback markers to update the encrypted global user portrait, specifically including the following steps, Input the encrypted global user portrait into Deep-Q-Learning to calculate the Q value of the ad placement action, form a Q value vector, and select N optimal actions from the Q value vector; The selected actions are integrated into a dynamic ad placement strategy; Analyze the negative feedback markers, identify invalid interest tags, calculate the encrypted weight adjustment amount, and update the encrypted weight to generate an updated encrypted global interest vector; Decrypt the updated encrypted global user portrait, match the dynamic ad placement strategy, the decrypted global user portrait, and the local ad inventory to obtain an ad push instruction and execute it, collect user feedback data during execution, update the decrypted regional user portrait segment through the user feedback data, and re-encrypt it; Decrypting the updated encrypted global user portrait, matching the dynamic ad placement strategy, the decrypted global user portrait, and the local ad inventory to obtain an ad push instruction specifically includes the following steps, Receive the updated encrypted global user portrait and decrypt it to obtain a plaintext global interest vector; Map the plaintext global interest vector to the regional user segment to generate a decrypted regional user portrait segment; The advertisement record based on the local advertisement inventory generates a candidate advertisement list, and calculates the cosine similarity between the decrypted regional user portrait fragments, selects the advertisement with the highest cosine similarity, and verifies it using a dynamic advertisement delivery strategy, and obtains an advertisement pushing instruction after no error is found; The encrypted global user portrait is adjusted based on the updated encrypted regional user portrait fragment, and an optimized dynamic advertisement delivery strategy is generated.

2. The cloud-computing-based cross-platform advertisement pushing and feedback integration method of claim 1, wherein: The encrypted behavior data is generated by capturing the user's clicks, page dwell time, sliding behavior and device metadata, filtering invalid data to generate valid interaction data, and using the Paillier homomorphic encryption algorithm to encrypt the encrypted behavior data. 3.The cloud-computing-based cross-platform advertisement pushing and feedback integration method of claim 2, wherein: The steps of verifying the validity of the interest label by cross-platform behavior backtracking and obtaining a negative feedback mark include the following steps, Extract the interest label identifier from the encrypted regional user portrait fragment, query external data through a social media API, process the external data using VADER sentiment analysis, calculate the composite sentiment score of each external data, and count the proportion of negative emotions; When the proportion of negative emotions exceeds half, generate a negative feedback mark for the interest label identifier. 4.The cloud-computing-based cross-platform advertisement pushing and feedback integration method of claim 3, wherein: Get the advertisement pushing instruction and execute it, collect user feedback data at the same time, update the decrypted regional user portrait fragment through the user feedback data, and re-encrypt, which includes the following steps, Execute the advertisement pushing instruction and collect the user's click data and dwell time, filter them to form user feedback data, and perform negative feedback detection, if negative feedback is detected, detect the authenticity of the negative feedback, and update the decrypted regional user portrait fragment in the real case; The updated decrypted regional user portrait fragment and user feedback data are encrypted and transmitted to the cloud.

5. A cloud computing-based cross-platform advertisement pushing and feedback integration system based on the cloud computing-based cross-platform advertisement pushing and feedback integration method of any one of claims 1-4. It includes, The collection module collects user interaction data in real time to generate encrypted behavior data; The aggregation module aggregates the encrypted behavior data by region to generate encrypted regional user portrait fragments, and verifies the validity of the interest label by cross-platform behavior backtracking to obtain a negative feedback mark; The generation module uses a pre-trained DNN model to process the encrypted regional user portrait fragments to obtain an encrypted global user portrait, and generates a dynamic advertisement delivery strategy through a reinforcement learning algorithm, and triggers the portrait self-healing process through the negative feedback mark to update the encrypted global user portrait; The delivery module decrypts the updated encrypted global user portrait, matches the dynamic advertisement delivery strategy, the decrypted global user portrait and the local advertisement inventory, obtains the advertisement pushing instruction and executes it, collects user feedback data at the same time, updates the decrypted regional user portrait fragment through the user feedback data, and re-encrypts it. The feedback module adjusts the encrypted global user portrait based on the updated encrypted regional user portrait fragment, generates an optimized dynamic advertisement delivery strategy.

6. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that: The processor executes the computer program to realize the steps of the cloud computing-based cross-platform advertisement pushing and feedback integration method of any one of claims 1-4.

7. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the cloud computing-based cross-platform advertisement pushing and feedback integration method of any one of claims 1-4.

Images