Intelligent recommendation method for optimizing advertisement keyword combination through cross validation

Abstract

The invention discloses an intelligent recommendation method for optimizing advertisement keyword combination through cross validation, and relates to the technical field of advertisement technology and search engine marketing, which comprises the following steps: constructing a heterogeneous data set through multi-modal data fusion, and layering according to data sparseness: training a Transform-XL time sequence model by adopting time cross validation of a dynamic K value in a high resource layer; a graph neural network association graph is introduced into a low resource layer, semantic expression of a long tail word is enhanced, a stratified sampling-transfer learning two-channel mechanism is designed, and the generalization ability is improved in combination with exposure frequency weighting and a parameter freezing strategy; developing a Bayesian fusion engine, and dynamically weighting a high / low resource layer prediction result by using an improved Materon kernel function Gaussian process; and generating a confidence interval based on neural quantile regression, and outputting an optimal keyword combination sequence under ROI-risk-diversity constraint in combination with multi-target Pareto optimization. According to the method, the cold start efficiency and the long-tail resource utilization rate are improved, and high-robustness decision support is provided for advertisement putting.

Description

Technical Field

[0001] The present invention relates to the fields of advertising technology and search engine marketing technology, and in particular to an intelligent recommendation method for optimizing advertising keyword combinations through cross-validation. Background Art

[0002] Traditional manual keyword adjustment methods are inefficient and rely heavily on experience, making them difficult to adapt to the massive volume of keywords and dynamic market changes. For example, when promoting a new product, it's necessary to quickly identify long-tail keywords to break through market bottlenecks, while seasonal fluctuations in search trends require real-time strategy adjustments.

[0003] At the data level, the accumulation of multi-source data, such as user search behavior and click-through conversions, provides a foundation for model training. For example, DeepBI integrates product information, sales data, and consumer behavior data to achieve intelligent keyword mining. At the algorithmic level, the development of deep learning models and generative AI enables semantic understanding and dynamic optimization. For example, Microsoft's pre-trained models improve the accuracy of keyword generation through the Mask Language Model task. Keyword bidding in vertical sectors such as education and e-commerce is fierce, driving long-tail keyword mining and landing page optimization as key to competitive differentiation.

[0004] However, there are still problems in the existing technology that need to be solved urgently: data sparsity and cold start problems. The exposure and conversion data of long-tail keywords and new keyword combinations are extremely scarce, making it difficult to train reliable models for effect prediction and cross-validation evaluation. The model's generalization ability assessment of these keyword combinations may be unreliable, making it difficult to discover potential high-value long-tail opportunities. Summary of the Invention

[0005] In order to solve the above technical problems, an intelligent recommendation method is provided for optimizing advertising keyword combinations through cross-validation. This technical solution solves the above-mentioned problems of data sparsity and cold start.

[0006] In order to achieve the above objects, the technical solution adopted by the present invention is:

[0007] The intelligent recommendation method for optimizing advertising keyword combinations through cross-validation includes:

[0008] S1. Obtain historical interaction data from the advertising platform and construct a multimodal dataset containing user behavior characteristics, keyword semantic vectors, advertising creative attributes, and conversion labels;

[0009] S2. Divide the dataset into high-resource layer and low-resource layer according to the data sparsity of keyword combinations, where the low-resource layer contains long-tail keywords and new combinations;

[0010] S3. K-fold cross-validation is used to train the initial prediction model for the high-resource layer. Meanwhile, the following operations are performed on the low-resource layer: a keyword-user intent association graph is constructed based on a graph neural network, and node embedding technology is used to generate semantic enhancement features for the low-resource combination. The hidden layer parameters of the high-resource layer model are mapped to the low-resource layer through transfer learning, and the underlying semantic encoder is frozen. A stratified sampling strategy is used to distribute the low-resource layer data to the cross-validation folds based on the exposure frequency, ensuring that each fold contains representative samples.

[0011] S4. Dynamically integrate the prediction results of the high / low resource layer models, use Bayesian optimization to adjust the fusion weights, and generate the final effect estimate;

[0012] S5. Based on the confidence interval of the estimated value and the potential ROI, a multi-objective optimization function is constructed to output the recommended keyword combination and delivery priority ranking.

[0013] Preferably, the S1 specifically includes:

[0014] Acquisition of historical interaction data, including user behavior logs, advertising interaction records, and third-party supplementary data;

[0015] User behavior logs, including search queries, click sequences, page dwell time, shopping cart operations, and cross-session behavior links, are collected in real time using the tracking SDK and cleaned of abnormal traffic using the Apache Flink stream processing engine.

[0016] Ad interaction records, including impressions, click-through rates, conversion rates, cost per conversion, and bidding environment parameters, are connected to the advertising platform API to obtain data down to the keyword group level;

[0017] Third-party supplementary data, including user device information, geographic location tags, and time period characteristics, are integrated with DMP's desensitized portrait tags through joint modeling.

[0018] Preferably, the S1 specifically includes:

[0019] We constructed a multimodal dataset and generated keyword semantic vectors. By fine-tuning a pre-trained language model, we used the RoBERTa-large model and continued pre-training on vertical domain corpora to capture industry terminology. We also used context-aware encoding and adversarial training to enhance generalization. We also introduced FGSM adversarial training and added perturbations to the input embeddings.

[0020] User behavior features are constructed using an LSTM-Attention model based on behavior sequences to capture intent strength and differentiate between browsing and decision-making users. The reach time difference and conversion contribution between search ads and social media ads are calculated to achieve cross-channel collaboration and quantify multi-channel attribution effectiveness. Sliding window statistics are used to calculate the distribution of click categories, and JS divergence is used to measure the rate of interest change, enabling real-time interest drift and timely adjustment of keywords based on demand fluctuations.

[0021] Quantifying advertising creative attributes and analyzing textual creativity, using BERT to extract copywriting sentiment polarity and promotional information density; visual creative analysis, deploying the ResNet-50 convolutional network to identify product entities, scene types, and character attributes in images and videos, and outputting structured labels;

[0022] Conversion label definition, multi-attribution model integration, definition of basic labels including last-click conversion and linear attribution, and use of Shapley Value algorithm to assign keyword combination contribution weights.

[0023] Preferably, the S2 specifically includes:

[0024] Data sparsity is quantified by using a three-dimensional weighted scoring model to calculate the sparsity score of keyword combinations, semantic coverage, and node relevance based on the knowledge graph;

[0025] Sparsity score:

[0026]

[0027] Where, is the sparsity score; The weight of the exposure dimension indicates the importance of exposure in calculating the sparsity score. is the weight of the semantic coverage dimension, which is used to measure the influence of semantic coverage on the sparsity score; The weight of the time decay factor reflects the importance of time in the sparsity score. Exposure refers to the cumulative number of exposures within a given period of time, which is used to measure the frequency of attention and display of a keyword combination over a period of time. After the Box-Cox transformation is applied to eliminate the long-tail distribution bias, it is used to calculate the sparsity score. Semantic coverage refers to the extent to which a keyword combination can cover the scope of knowledge or information at the semantic level. 1 minus semantic coverage represents the proportion of the semantically uncovered portion.

[0028] Dynamic threshold setting and high / low resource layer boundaries are automatically divided using a dual clustering algorithm: the first layer uses K-means++ to cluster data into high, medium, and low resource candidates; the second layer uses DBSCAN to detect edge points and merge them into the low resource layer; abnormal scenario processing: if the advertiser specifies core words, they are forced to be retained in the high resource layer even if they are underexposed.

[0029] Preferably, the S2 specifically includes:

[0030] Specialized identification of low-resource layers and long-tail keyword mining, with sparse scores greater than a predetermined threshold and exposure at a predetermined ratio quantile; identification technology stack, extracting low-frequency co-occurring word pairs from search sessions through FP-Growth sequence pattern mining; and reconstructing exposure distributions and identifying statistically significant tail outliers through adversarial autoencoders.

[0031] New combination cold start determination: whether the new keyword combination meets any of the following conditions: zero historical exposure, change in related product categories, or expansion of advertiser categories. If so, it is classified as a low-resource tier. Semantic similarity compensation: if the Jaccard similarity between the new combination and the high-resource word is greater than a predetermined threshold, an initial exposure estimate is assigned.

[0032] Layered data governance strategy: at the high-resource layer, dimensionality reduction accelerates training, feature hashing is used to compress the dimensions of head keywords, and sparse features use Tf-Idf weighting instead of One-Hot encoding; at the low-resource layer, synthetic minority oversampling is used, and virtual samples are created using the CTGAN generative model; a cross-advertiser migration pool is established to establish a desensitized low-resource data sharing pool for non-competing advertisers in the same industry, and knowledge is aggregated through federated learning.

[0033] Preferably, the S3 specifically includes:

[0034] The high-resource layer uses enhanced K-fold cross-validation training. The base network adopts the Transformer-XL time series model to capture the long-term dependencies of user behavior. The feature fusion layer introduces a multi-head cross-attention mechanism to align user behavior sequences, keyword semantic vectors, and creative visual features. The output layer has a dual-channel prediction structure consisting of a CTR estimation channel and a CVR estimation channel, which share underlying parameters but independently optimize the loss function.

[0035] Cross-validation, dynamic K value adjustment, adaptive selection of K value based on high-resource data scale, time series protection, using time series cross-validation to prohibit future data from entering the training set; business-oriented evaluation indicators, in addition to conventional AUC / LogLoss, define advertiser revenue-sensitive indicators.

[0036] Preferably, the S3 specifically includes:

[0037] At the low-resource layer, semantic enhancement driven by graph neural networks builds a keyword-user intent association graph. Heterogeneous nodes include keyword nodes, user intent nodes, and ad creative nodes; multi-relationship edges include search co-occurrence relationships, creative conversion relationships, and semantic adversarial relationships. The graph neural network architecture uses a hybrid model of RGCN and GraphSAGE to output semantically enhanced vectors for low-resource keywords, integrating graph topology and cross-modal associations.

[0038] Transfer learning parameter mapping, dual-channel parameter transfer architecture, based on the high-resource model, establishes a shared semantic space by fixing the underlying encoder, and maps it to the low-resource adaptation layer and low-resource prediction head; the frozen layer includes the BERT text encoder and the ResNet visual encoder; the trainable layer includes the domain adapter, and introduces fine-tuning of the adaptation layer; the dynamic gating mechanism automatically adjusts the transfer intensity according to the sparsity of the low-resource data.

[0039] Preferably, the S3 specifically includes:

[0040] Stratified sampling strategy, low-resource data, exposure frequency weighting rules, weight function calculation, dynamic adjustment mechanism, after each round of cross-validation, reverse update of sample weights based on the prediction error of the validation set;

[0041] Fold allocation algorithm, hierarchical clustering preprocessing, uses the K-prototypes algorithm to cluster low-resource data into K categories based on semantic vectors, exposure time, and user profiles; block random allocation, each cluster is assigned to at least two different folds, and new combinations are forced to be evenly distributed across folds; conflict resolution mechanism, if the proportion of long-tail words in a fold is less than the established proportion, oversampling compensation is triggered.

[0042] The S4 specifically includes:

[0043] In addition to basic CTR / CVR estimates, high-resource model outputs extract hidden layer activation vectors to capture intermediate states of user intent understanding. Low-resource model outputs add uncertainty quantification metrics: Monte Carlo Dropout variance and graph neural network node embedding distance; business features are injected into the competitive environment coefficient and advertiser risk preference.

[0044] Fusion function, using a dual-channel adaptive fusion framework:

[0045]

[0046] Where, For the final fusion output, the prediction information of high-resource channels and low-resource channels are comprehensively considered; is the transformation function of the high resource channel, which is used to adjust the high resource prediction results. is the predicted output of the high resource channel; is the transformation function of the low-resource channel; is the weight coefficient corresponding to the high-resource channel, which is used to measure the importance of the high-resource channel in the fusion process; is the weight coefficient corresponding to the low resource channel, which is equal to the weight coefficient of the high resource channel Complementary, that is, the sum of the two is 1;

[0047] For high-resource channel transformation, a Sigmoid calibration layer is introduced; for low-resource channel transformation, an uncertainty gating mechanism is designed:

[0048] Bayesian optimization weight parameter tuning engine, optimized objective function design, Gaussian process regression modeling, kernel function selection, using the Matérn 5 / 2 kernel; acquisition function optimization, improved Expected Improvement strategy, and added exploration penalty terms.

[0049] Preferably, the S5 specifically includes:

[0050] Confidence interval driven risk-return modeling, neural quantile regression, embedding multi-quantile output heads in the estimation model to form dynamic confidence intervals; dynamic mapping of risk levels, bandwidth calculation;

[0051] Multi-objective optimization function design, expected ROI, calculating comprehensive benefits based on the integral of the probability density function; risk control, setting the minimum acceptable ROI threshold; diversity constraints, ensuring keyword coverage of multiple product categories through Shannon entropy;

[0052] Pareto front solution, NSGA-III algorithm optimization, population initialization, sampling from historical high-quality combinations, and injecting semantically similar long-tail keywords; crossover mutation rules, including semantic hybridization, exchanging subnodes with keywords in the same category, adversarial mutation, and generating counterfactual samples based on GAN; reference point generation, constructing a non-dominated solution set in the three-dimensional space of ROI, risk, and diversity;

[0053] Real-time feedback mechanism, Bandit online learning, regular update of target weights, dynamic adjustment of regret values, and priority reduction of ranking weights for combinations with high regret values.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] This invention proposes to construct a multimodal risk-return dynamic coupling model to effectively solve the problem of strategy rigidity caused by static threshold setting in traditional methods. The introduction of neural quantile regression and confidence bandwidth self-calibration mechanism can dynamically generate asymmetric confidence intervals based on the temporal characteristics of user behavior and changes in competitive situation, thereby increasing the risk warning response speed to minutes while avoiding the loss of business opportunities caused by excessive conservatism. Secondly, at the multi-objective optimization level under complex constraints, traditional solutions often use simplified methods such as weighted summation or hierarchical optimization, resulting in insufficient exploration of the Pareto frontier.

[0056] This invention creatively integrates the NSGA-III quantum acceleration algorithm and the semantically enhanced evolutionary mechanism, which not only breaks through the local optimal trap of traditional genetic algorithms, but also significantly expands the exploration dimension of the understanding space through adversarial sample generation and semantic subtree hybridization technology, achieving a double breakthrough in the diversity coverage of keyword combinations and the stability of returns.

[0057] Through a dual-channel adaptive migration framework and a dynamic gating mechanism, high-resource knowledge distillation and low-resource semantic enhancement are effectively synergized, ensuring model robustness while improving cold-start efficiency for new category keywords. A revolutionary upgrade to the real-time feedback mechanism overcomes the limitations of the traditional Bandit algorithm, which relies on a fixed exploration-exploitation ratio. A dynamic weight adjustment system driven by regret values ​​reconstructs the objective function in real time based on delivery performance. Combined with an online optimization engine accelerated by quantum computing, this reduces the strategy iteration cycle from hours to seconds, significantly improving the system's responsiveness to sudden traffic fluctuations and competitive dynamics. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Flowchart of an intelligent recommendation method for optimizing advertising keyword combinations through cross-validation. DETAILED DESCRIPTION

[0059] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are merely examples, and those skilled in the art may conceive of other obvious variations.

[0060] Reference Figure 1 As shown in FIG, the intelligent recommendation method for optimizing advertising keyword combinations through cross-validation includes:

[0061] S1. Obtain historical interaction data from the advertising platform and construct a multimodal dataset containing user behavior characteristics, keyword semantic vectors, advertising creative attributes, and conversion labels;

[0062] S2. Divide the dataset into high-resource layer and low-resource layer according to the data sparsity of keyword combinations, where the low-resource layer contains long-tail keywords and new combinations;

[0063] S3. K-fold cross-validation is used to train the initial prediction model for the high-resource layer. Meanwhile, the following operations are performed on the low-resource layer: a keyword-user intent association graph is constructed based on a graph neural network, and node embedding technology is used to generate semantic enhancement features for the low-resource combination. The hidden layer parameters of the high-resource layer model are mapped to the low-resource layer through transfer learning, and the underlying semantic encoder is frozen. A stratified sampling strategy is used to distribute the low-resource layer data to the cross-validation folds based on the exposure frequency, ensuring that each fold contains representative samples.

[0064] S4. Dynamically integrate the prediction results of the high / low resource layer models, use Bayesian optimization to adjust the fusion weights, and generate the final effect estimate;

[0065] S5. Based on the confidence interval of the estimated value and the potential ROI, a multi-objective optimization function is constructed to output the recommended keyword combination and delivery priority ranking.

[0066] It's important to note that the multimodal dataset built by S1 serves not only as a feature repository but also as an evolutionary engine, fed back in real time by S5's delivery results. When a new keyword combination generates high conversion rates, the system automatically injects its semantic vector into S3's graph neural network, dynamically expanding the node connections in the user intent graph. This closed-loop learning enables the model to continuously capture evolving consumer trends, significantly accelerating its response to new product categories compared to traditional static models.

[0067] The sparsity threshold in S2 isn't fixed, but dynamically adjusted based on the confidence interval of the S4 fusion prediction. If a low-resource tier keyword's estimated ROI variance falls below 0.03 for five consecutive times, it's automatically promoted to the high-resource tier training queue. Meanwhile, declining high-resource tier keywords (monthly CTR decline >15%) are demoted to the low-resource tier for renewed semantic enhancement, creating a "metabolism" mechanism for data value.

[0068] The real-time environmental perception engine and S4's Bayesian optimizer integrate external environmental variables (such as competitor bid intensity index and social media hot events) to adjust fusion weights through covariates. Environment-sensitive weight updates increase the weight of high-resource models and strengthen innovative discovery of low-resource models during promotional periods.

[0069] Establish a risk control chain from the feature layer to the decision layer:

[0070] Input layer: Filter high variance samples during S3 stratified sampling;

[0071] Fusion layer: S4 constrains the maximum weight of the low-resource model;

[0072] Output layer: S5 starts manual review sandbox for combinations with confidence bandwidth > 0.1;

[0073] When external shocks, such as changes in policies and regulations, cause historical models to become invalid: initiate the topology reconstruction protocol of the graph neural network: delete the offending word nodes and rebuild the associated edges; activate cross-industry migration channels and reuse the new regulations of the fast-moving consumer goods industry to adapt to the model parameters.

[0074] The ability to reversely predict demand uses low-resource semantic graphs to analyze unmet needs and identify isolated but highly central nodes. For example, by associating "camping power supply" with "photography drone," it generates cross-category combination suggestions (recommending the "outdoor power supply + starry sky photography" bundled keyword). This allows an outdoor brand to explore new scenarios and increase average order value.

[0075] Said S1 specifically includes:

[0076] Acquisition of historical interaction data, including user behavior logs, advertising interaction records, and third-party supplementary data;

[0077] User behavior logs, including search queries, click sequences, page dwell time, shopping cart operations, and cross-session behavior links, are collected in real time using the tracking SDK and cleaned of abnormal traffic using the Apache Flink stream processing engine.

[0078] Ad interaction records, including impressions, click-through rates, conversion rates, cost per conversion, and bidding environment parameters, are connected to the advertising platform API to obtain data down to the keyword group level;

[0079] Third-party supplementary data, including user device information, geographic location tags, and time period characteristics, is integrated with the DMP's desensitized portrait tags through joint modeling;

[0080] We constructed a multimodal dataset and generated keyword semantic vectors. By fine-tuning a pre-trained language model, we used the RoBERTa-large model and continued pre-training on vertical domain corpora to capture industry terminology. We also used context-aware encoding and adversarial training to enhance generalization. We also introduced FGSM adversarial training and added perturbations to the input embeddings.

[0081] User behavior features are constructed using an LSTM-Attention model based on behavior sequences to capture intent strength and differentiate between browsing and decision-making users. The reach time difference and conversion contribution between search ads and social media ads are calculated to achieve cross-channel collaboration and quantify multi-channel attribution effectiveness. Sliding window statistics are used to calculate the distribution of click categories, and JS divergence is used to measure the rate of interest change, enabling real-time interest drift and timely adjustment of keywords based on demand fluctuations.

[0082] Quantifying advertising creative attributes and analyzing textual creativity, using BERT to extract copywriting sentiment polarity and promotional information density; visual creative analysis, deploying the ResNet-50 convolutional network to identify product entities, scene types, and character attributes in images and videos, and outputting structured labels;

[0083] Conversion label definition, multi-attribution model integration, definition of basic labels including last-click conversion and linear attribution, and use of Shapley Value algorithm to assign keyword combination contribution weights.

[0084] It should be noted that historical interaction data is obtained:

[0085] Multi-dimensional reconstruction of user behavior logs, cross-platform behavior stitching, and integration of APP / Web / Mini Program behaviors through device fingerprint fuzzy matching:

[0086] Construct a four-tuple consisting of a pseudo user ID, timestamp, behavior type, and content hash; and use the Transformer-XL model to reconstruct the cross-end session path (session interruption tolerance ≤ 4 hours).

[0087] Cognitive load quantification, calculating attention metrics based on page dwell time and scroll depth:

[0088]

[0089] Used to distinguish impulsive clicks (<0.3) from deep decisions (>0.7);

[0090] Dynamic extraction of bidding environment parameters, real-time bidding snapshot technology, captured every 15 seconds via the ad platform API, win rate distribution histogram (by price range), and hash values ​​of top competitor materials (for plagiarism detection);

[0091] Game theory feature construction, calculation of Nash equilibrium offset:

[0092]

[0093] >0.25 triggers a bidding strategy warning;

[0094] Secure integration of third-party data, federated feature engineering, and establishment of an encrypted vector exchange channel with the DMP platform:

[0095] Step 1: Generate user group embeddings locally (e.g., "high-end digital enthusiasts");

[0096] Step 2: Upload to DMP using Paillier homomorphic encryption;

[0097] Step 3: Return the K-anonymized group feature statistics (such as the average monthly search increase for the category).

[0098] Keyword semantic vectors, vertical domain pre-training, and term injection training: Add industry term masks to the MLM task of RoBERTa-large, perform semantic alignment, and use Triplet Loss to narrow the distance between synonymous terms.

[0099] Enhanced adversarial robustness, a gradient penalty mechanism, and the introduction of Wasserstein distance constraints in FGSM perturbation training to prevent excessive perturbations; multi-view consistency, generating three variant interpretations for the same keyword, requiring vector cosine similarity > 0.93;

[0100] Cognitive modeling of user behavior characteristics and LSTM-Attention architecture upgrade;

[0101] Time decay Attention: recent behavior weight = e −λΔt (λ is dynamically adjusted by user activity), behavior type gating, distinguishing between information acquisition (high clicks / low conversions) and purchase intent (low retention / high add-to-cart).

[0102] Cross-channel attribution, using the Spatio-Temporal Convolutional Attribution Network (STCAN):

[0103] Input layer: timestamp sequence of each channel’s access and creative content fingerprint;

[0104] Output layer: quantifies channel synergy effects through temporal convolution kernels (e.g., social media exposure increases search conversion rate by 22%).

[0105] Real-time capture of interest drift, concept drift detection algorithm, calculation of demand mutation index based on sliding window, triggering keyword strategy reorganization when it is 1.5.

[0106] Deep quantification of creative attributes, sentiment calculation of creative text, multi-dimensional indicators of promotional intensity, sense of urgency, calculated by multiplying the frequency of "limited time" and "flash sale" keywords by position weight; and density of interest points, calculated by dividing the number of times the discount amount is mentioned by the total length of the copy.

[0107] Sentiment polarity refinement: the Electra model is used to output a 9-level sentiment score (-4 to +4), with special annotation of sarcastic meanings.

[0108] Structured analysis of visual creativity, multimodal alignment technology, and the use of the CLIP model to construct image-text consistency scores;

[0109] Scene semantic distillation: ResNet-50 followed by a graph neural network decoder identifies scene elements (beach / office / night scene) and constructs element association rules, such as coffee cup + laptop = office scene.

[0110] The S2 specifically includes:

[0111] Data sparsity is quantified by using a three-dimensional weighted scoring model to calculate the sparsity score of keyword combinations, semantic coverage, and node relevance based on the knowledge graph;

[0112] Sparsity score:

[0113]

[0114] Where, is the sparsity score; The weight of the exposure dimension indicates the importance of exposure in calculating the sparsity score. is the weight of the semantic coverage dimension, which is used to measure the influence of semantic coverage on the sparsity score; The weight of the time decay factor reflects the importance of time in the sparsity score. Exposure refers to the cumulative number of exposures within a given period of time, which is used to measure the frequency of attention and display of a keyword combination over a period of time. After the Box-Cox transformation is applied to eliminate the long-tail distribution bias, it is used to calculate the sparsity score. Semantic coverage refers to the extent to which a keyword combination can cover the scope of knowledge or information at the semantic level. 1 minus semantic coverage represents the proportion of the semantically uncovered portion.

[0115] Dynamic threshold setting and high / low resource layer boundaries are automatically divided using a dual clustering algorithm: the first layer uses K-means++ to cluster data into high, medium, and low resource candidates; the second layer uses DBSCAN to detect edge points and merge them into the low resource layer; abnormal scenario processing: if the advertiser specifies a core word, it will be forced to be retained in the high resource layer even if it is underexposed;

[0116] Specialized identification of low-resource layers and long-tail keyword mining, with sparse scores greater than a predetermined threshold and exposure at a predetermined ratio quantile; identification technology stack, extracting low-frequency co-occurring word pairs from search sessions through FP-Growth sequence pattern mining; and reconstructing exposure distributions and identifying statistically significant tail outliers through adversarial autoencoders.

[0117] New combination cold start determination: whether the new keyword combination meets any of the following conditions: zero historical exposure, change in related product categories, or expansion of advertiser categories. If so, it is classified as a low-resource tier. Semantic similarity compensation: if the Jaccard similarity between the new combination and the high-resource word is greater than a predetermined threshold, an initial exposure estimate is assigned.

[0118] Layered data governance strategy: at the high-resource layer, dimensionality reduction accelerates training, feature hashing is used to compress the dimensions of head keywords, and sparse features use Tf-Idf weighting instead of One-Hot encoding; at the low-resource layer, synthetic minority oversampling is used, and virtual samples are created using the CTGAN generative model; a cross-advertiser migration pool is established to establish a desensitized low-resource data sharing pool for non-competing advertisers in the same industry, and knowledge is aggregated through federated learning.

[0119] It should be noted that data sparsity is quantified:

[0120] Dynamic weight adaptive mechanism, weight evolution rules, and weights of each dimension are dynamically adjusted according to the advertiser's industry;

[0121] Nonlinear correction of exposure using improved Box-Cox transformation;

[0122] Implementation of knowledge graph with semantic coverage and construction of industry knowledge graph:

[0123] Nodes: product categories, functional attributes, and scene labels;

[0124] Edge relationships: co-occurrence strength (search session co-occurrence rate), functional complementarity (product pairing purchase probability);

[0125] Intelligent demarcation technology for hierarchical boundaries, industrial-grade optimization of the biclustering algorithm, an enhanced version of K-means++, and initial center selection that prioritizes the 85th percentile of the exposure distribution; the distance metric is Mahalanobis distance (the covariance matrix includes exposure / coverage / time decay correlations).

[0126] Specialized identification of low-resource layers, long-tail keywords, and FP-Growth algorithm upgrades:

[0127] The spatiotemporal constraint model only mines emerging combinations that have appeared in the past 30 days; the semantic filtering layer eliminates combinations without commercial intent, such as "how to install" vs. "purchase installation kit."

[0128] The S3 specifically includes:

[0129] The high-resource layer uses enhanced K-fold cross-validation training. The base network adopts the Transformer-XL time series model to capture the long-term dependencies of user behavior. The feature fusion layer introduces a multi-head cross-attention mechanism to align user behavior sequences, keyword semantic vectors, and creative visual features. The output layer has a dual-channel prediction structure consisting of a CTR estimation channel and a CVR estimation channel, which share underlying parameters but independently optimize the loss function.

[0130] Cross-validation and dynamic K value adjustment: adaptively selecting the K value based on the scale of high-resource data. Time series protection: using time series cross-validation to prohibit future data from entering the training set. Business-oriented evaluation indicators: in addition to conventional AUC / LogLoss, define advertiser revenue-sensitive indicators.

[0131] At the low-resource layer, semantic enhancement driven by graph neural networks builds a keyword-user intent association graph. Heterogeneous nodes include keyword nodes, user intent nodes, and ad creative nodes; multi-relationship edges include search co-occurrence relationships, creative conversion relationships, and semantic adversarial relationships. The graph neural network architecture uses a hybrid model of RGCN and GraphSAGE to output semantically enhanced vectors for low-resource keywords, integrating graph topology and cross-modal associations.

[0132] Transfer learning parameter mapping, a dual-channel parameter transfer architecture, based on a high-resource model, establishes a shared semantic space by fixing the underlying encoder, and maps it to the low-resource adaptation layer and low-resource prediction head; the frozen layer includes the BERT text encoder and the ResNet visual encoder; the trainable layer includes the domain adapter, and fine-tuning of the adaptation layer is introduced; the dynamic gating mechanism automatically adjusts the transfer strength according to the sparsity of the low-resource data;

[0133] Stratified sampling strategy, low-resource data, exposure frequency weighting rules, weight function calculation, dynamic adjustment mechanism, after each round of cross-validation, reverse update of sample weights based on the prediction error of the validation set;

[0134] Fold allocation algorithm, hierarchical clustering preprocessing, uses the K-prototypes algorithm to cluster low-resource data into K categories based on semantic vectors, exposure time, and user profiles; block random allocation, each cluster is assigned to at least two different folds, and new combinations are forced to be evenly distributed across folds; conflict resolution mechanism, if the proportion of long-tail words in a fold is less than the established proportion, oversampling compensation is triggered.

[0135] It should be noted that high-resource layer training:

[0136] Transformer-XL's dynamic enhancement and memory segment distillation technology address the training bottleneck of long sequences (>1,000 behavioral events) by encoding user behavior sequences into segmented memory vectors.

[0137] Multimodal aligned attention, the multi-head cross attention mechanism is upgraded to three-way alignment:

[0138] behavior_attn = Attention(Q=behavior sequence, K=V=semantic vector)

[0139] visual_attn = Attention(Q=behavior sequence, K=V=visual feature)

[0140] fused_vector = LayerNorm(behavior_attn + visual_attn + original behavior vector)

[0141] Definition of business sensitive indicators:

[0142] index Calculation formula Business significance Value Consumption Ratio (VCR) GMV / Consumption × CVR Balancing volume and quality Life Cycle Score (LCS) log(keyword effective days)×CTR stability Measuring long-term value

[0143] Low-resource layer semantic enhancement:

[0144] Heterogeneous graph construction, semantic adversarial edges, and negative sample connections generated by adversarial training; cross-modal conversion edges, and the conversion rate covariance between creative materials and keywords is >0.6;

[0145] Virtual intent nodes create bridge nodes through NLP generation technology, improving the relevance of long-tail words and increasing edge coverage by 58%;

[0146] Graph neural network hybrid architecture, RGCN+GraphSAGE collaborative mechanism, RGCN layer, handles heterogeneous relationships (weighted aggregation of different edge types); GraphSAGE layer samples neighbors to generate target node embedding;

[0147] Adaptive control of transfer learning, quantization compression of parameter mapping, lightweight adaptation layer technology, converting high-resource model FP32 parameters → to INT8 precision → adding a low-rank adaptation matrix.

[0148] Intelligent scheduling of stratified sampling and conflict resolution of fold allocation:

[0149] K-prototypes clustering enhancement:

[0150] Numerical features: exposure, click-through rate → Euclidean distance

[0151] Category features: time period, user profile → use Hamming distance

[0152] Mixed distance formula: d=αd num +βd cat +γcos(semantic vector)

[0153] Where d represents the hybrid distance, which is the final distance metric obtained by comprehensively considering numerical features, categorical features, and semantic features; α represents the weight coefficient of the numerical feature part in the hybrid distance calculation; β represents the weight coefficient of the categorical feature part in the hybrid distance calculation; γ is the weight coefficient of the semantic vector part in the hybrid distance; d num is the distance of the numerical feature part. Here, Euclidean distance is used to measure the difference between numerical features; d catThe distance between the category features is calculated using the Hamming distance;

[0154] Oversampling compensation mechanism, using pre-trained VAE to generate synthetic samples of underrepresented categories; weight amplification, the sampling probability of the target category is increased by min(3, target ratio / current ratio).

[0155] The S4 specifically includes:

[0156] In addition to basic CTR / CVR estimates, high-resource model outputs extract hidden layer activation vectors to capture intermediate states of user intent understanding. Low-resource model outputs add uncertainty quantification metrics: Monte Carlo Dropout variance and graph neural network node embedding distance; business features are injected into the competitive environment coefficient and advertiser risk preference.

[0157] Fusion function, using a dual-channel adaptive fusion framework:

[0158]

[0159] Where, For the final fusion output, the prediction information of high-resource channels and low-resource channels are comprehensively considered; is the transformation function of the high resource channel, which is used to adjust the high resource prediction results. is the predicted output of the high resource channel; is the transformation function of the low-resource channel; is the weight coefficient corresponding to the high-resource channel, which is used to measure the importance of the high-resource channel in the fusion process; is the weight coefficient corresponding to the low resource channel, which is equal to the weight coefficient of the high resource channel Complementary, that is, the sum of the two is 1;

[0160] For high-resource channel transformation, a Sigmoid calibration layer is introduced; for low-resource channel transformation, an uncertainty gating mechanism is designed:

[0161] Bayesian optimization weight parameter tuning engine, optimized objective function design, Gaussian process regression modeling, kernel function selection, using the Matérn 5 / 2 kernel; acquisition function optimization, improved Expected Improvement strategy, and added exploration penalty terms.

[0162] It should be noted that the high resource model has multidimensional outputs;

[0163] Intention understanding vector extraction, generating behavioral motivation encoding through hidden layer clustering;

[0164] Attention heat map application to locate high-impact behavioral events;

[0165] Fusion function:

[0166] Dynamic control of weight coefficients:

[0167] Scenario <![CDATA[ϕ h Adjust rules]]> Mathematical Expression New product launch period Reduce the weight of high resources <![CDATA[max(0.3,ϕ h −0.2)]]> Large-scale promotions of competing products Improve high resource stability <![CDATA[min(0.9,ϕ h +0.15)]]> The variance of the low resource model remains low Gradually increase the weight of low resources <![CDATA[ϕ h =0.5+0.1×log(number of consecutive days meeting the standard)

[0168] Bayesian optimization engine and Gaussian process regression upgrades:

[0169] Input → Kernel function selection → Stationary process → Matérn 5 / 2;

[0170] Kernel function selection → Periodic fluctuation → Cosine periodic kernel;

[0171] Matérn 5 / 2 → Automatic hyperparameter optimization;

[0172] Cosine periodic kernel → automatic hyperparameter optimization;

[0173] Automatic optimization of hyperparameters → output fusion weights;

[0174] Asymmetric likelihood function, which imposes a 2-fold penalty weight on negative errors (estimate < actual).

[0175] The S5 specifically includes:

[0176] Confidence interval driven risk-return modeling, neural quantile regression, embedding multi-quantile output heads in the estimation model to form dynamic confidence intervals; dynamic mapping of risk levels, bandwidth calculation;

[0177] Multi-objective optimization function design, expected ROI, calculating comprehensive benefits based on the integral of the probability density function; risk control, setting the minimum acceptable ROI threshold; diversity constraints, ensuring keyword coverage of multiple product categories through Shannon entropy;

[0178] Pareto front solution, NSGA-III algorithm optimization, population initialization, sampling from historical high-quality combinations, and injecting semantically similar long-tail keywords; crossover mutation rules, including semantic hybridization, exchanging subnodes with keywords in the same category, adversarial mutation, and generating counterfactual samples based on GAN; reference point generation, constructing a non-dominated solution set in the three-dimensional space of ROI, risk, and diversity;

[0179] Real-time feedback mechanism, Bandit online learning, regular update of target weights, dynamic adjustment of regret values, and priority reduction of ranking weights for combinations with high regret values.

[0180] It should be noted that risk-return modeling:

[0181] Neural Quantile Regression deploys three quantile output heads (Q10 / Q50 / Q90) in parallel at the end of the Transformer-XL model, using a quantile weighted loss function;

[0182] Dynamic confidence bandwidth calculation introduces industry volatility factors. When the bandwidth is greater than 1.5, a three-level warning mechanism is triggered: automatic 30% demotion → manual review → competitive product comparison analysis;

[0183] Risk level mapping uses fuzzy logic rules:

[0184] Low risk (bandwidth ≤ 0.2): Allocate an additional 15% budget weight;

[0185] Medium risk (0.2 < bandwidth ≤ 0.4): Enable dynamic bid adjuster;

[0186] High risk (bandwidth > 0.4): Freeze delivery and generate attribution reports.

[0187] Multi-objective optimization: expected ROI calculation by fitting probability density function of truncated Gaussian mixture model

[0188] Pareto frontier solution: NSGA-III algorithm modification, population initialization strategy, historical high-quality portfolio, screening portfolios with ROI stability > 0.85 over the past 90 days; semantic long-tail injection, based on GNN embedding space retrieval, meeting COS (v new word, v head) > 0.7; adversarial sample enhancement, using Wasserstein GAN to generate counterfactual samples and minimize;

[0189] Cross-mutation rules and semantic hybridization are used to exchange categories at the subtree level, such as outdoor equipment and sports accessories. Adversarial mutation is used to apply gradient attacks to low ROI combinations.

[0190] Quantum accelerated solution maps the non-dominated sorting process to the D-Wave quantum processor, which speeds up the solution of 300-dimensional problems and meets the real-time response requirements of 100ms.

[0191] In summary, the advantages of the present invention are:

[0192] Breaking through the limitations of traditional static thresholds, this approach builds adaptive confidence intervals based on neural quantile regression, enabling minute-by-minute risk warning responses through real-time volatility feedback and bandwidth calibration. This effectively addresses the lag inherent in empirical rules, significantly reducing misjudgment of business opportunities while avoiding overly conservative strategies.

[0193] The quantum-accelerated NSGA-III framework, combined with semantic evolution, overcomes the local convergence trap of traditional genetic algorithms through adversarial example generation and subtree hybridization techniques. It achieves efficient exploration of the Pareto frontier in the three-dimensional space of ROI, risk, and diversity, simultaneously achieving the dual goals of maximizing returns and promoting combinatorial innovation.

[0194] A dynamic gating mechanism for dual-channel migration was designed to intelligently balance the synergistic relationship between high-resource knowledge distillation and low-resource semantic enhancement. Through graph topology and cross-modal alignment technology, the cognitive efficiency of new category keywords achieved an order of magnitude leap, completely resolving the dilemma of long-tail resource development.

[0195] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent recommendation method for optimizing advertising keyword combinations through cross-validation, characterized in that: include: S1. Obtain historical interaction data from the advertising platform and construct a multimodal dataset containing user behavior characteristics, keyword semantic vectors, advertising creative attributes, and conversion labels; S2. Divide the dataset into high-resource layer and low-resource layer according to the data sparsity of keyword combinations, where the low-resource layer contains long-tail keywords and new combinations; S3. K-fold cross-validation is used to train the initial prediction model for the high-resource layer. Meanwhile, the following operations are performed on the low-resource layer: a keyword-user intent association graph is constructed based on a graph neural network, and semantic enhancement features of low-resource combinations are generated using node embedding technology. Through transfer learning, the hidden layer parameters of the high-resource layer model are mapped to the low-resource layer, and the underlying semantic encoder is frozen. A stratified sampling strategy is used to distribute the low-resource layer data to cross-validation folds based on the exposure frequency, ensuring that each fold contains representative samples. S4. Dynamically integrate the prediction results of the high / low resource layer models, use Bayesian optimization to adjust the fusion weights, and generate the final effect estimate; S5. Based on the confidence interval of the estimated value and the potential ROI, a multi-objective optimization function is constructed to output the recommended keyword combination and delivery priority ranking.

2. The intelligent recommendation method for optimizing advertising keyword combinations through cross-validation according to claim 1, characterized in that: The S3 specifically includes: The high-resource layer uses enhanced K-fold cross-validation training. The base network adopts the Transformer-XL time series model to capture the long-term dependencies of user behavior. The feature fusion layer introduces a multi-head cross-attention mechanism to align user behavior sequences, keyword semantic vectors, and creative visual features. The output layer has a dual-channel prediction structure consisting of a CTR estimation channel and a CVR estimation channel, which share underlying parameters but independently optimize the loss function. Cross-validation, dynamic K value adjustment, adaptive selection of K value based on high-resource data scale, time series protection, using time series cross-validation to prohibit future data from entering the training set; business-oriented evaluation indicators, in addition to conventional AUC / LogLoss, define advertiser revenue-sensitive indicators.

3. The intelligent recommendation method for optimizing advertising keyword combinations through cross-validation according to claim 2, characterized in that: The S3 specifically includes: At the low-resource layer, semantic enhancement driven by graph neural networks builds a keyword-user intent association graph. Heterogeneous nodes include keyword nodes, user intent nodes, and ad creative nodes; multi-relationship edges include search co-occurrence relationships, creative conversion relationships, and semantic adversarial relationships. The graph neural network architecture uses a hybrid model of RGCN and GraphSAGE to output semantically enhanced vectors for low-resource keywords, integrating graph topology and cross-modal associations. Transfer learning parameter mapping, dual-channel parameter transfer architecture, based on the high-resource model, establishes a shared semantic space by fixing the underlying encoder, and maps it to the low-resource adaptation layer and low-resource prediction head; the frozen layer includes the BERT text encoder and the ResNet visual encoder; the trainable layer includes the domain adapter, and introduces fine-tuning of the adaptation layer; the dynamic gating mechanism automatically adjusts the transfer intensity according to the sparsity of the low-resource data.

4. The intelligent recommendation method for optimizing advertising keyword combinations through cross-validation according to claim 3, characterized in that: The S3 specifically includes: Stratified sampling strategy, low-resource data, exposure frequency weighting rules, weight function calculation, dynamic adjustment mechanism, after each round of cross-validation, reverse update of sample weights based on the prediction error of the validation set; Fold allocation algorithm, hierarchical clustering preprocessing, uses the K-prototypes algorithm to cluster low-resource data into K categories based on semantic vectors, exposure time, and user profiles; block random allocation, each cluster is assigned to at least two different folds, and new combinations are forced to be evenly distributed across folds; conflict resolution mechanism, if the proportion of long-tail words in a fold is less than the established proportion, oversampling compensation is triggered.

5. The intelligent recommendation method for optimizing advertising keyword combinations through cross-validation according to claim 4, characterized in that: The S2 specifically includes: Data sparsity is quantified by using a three-dimensional weighted scoring model to calculate the sparsity score of keyword combinations, semantic coverage, and node relevance based on the knowledge graph; Sparsity score: Where, is the sparsity score; The weight of the exposure dimension indicates the importance of exposure in calculating the sparsity score. is the weight of the semantic coverage dimension, which is used to measure the influence of semantic coverage on the sparsity score; The weight of the time decay factor reflects the importance of time in the sparsity score. Exposure is the cumulative number of exposures within a given period of time, which is used to measure the frequency of attention and display of a keyword combination over a period of time. After the Box-Cox transformation eliminates the long-tail distribution bias, it is used to calculate the sparsity score. Semantic coverage is the extent to which the keyword combination can cover the knowledge or information range at the semantic level. 1 minus semantic coverage is the proportion of the semantically uncovered part. Dynamic threshold setting and high / low resource layer boundaries are automatically divided using a dual clustering algorithm: the first layer uses K-means++ to cluster data into high, medium, and low resource candidates; the second layer uses DBSCAN to detect edge points and merge them into the low resource layer; abnormal scenario processing: if the advertiser specifies core words, they are forced to be retained in the high resource layer even if they are underexposed.

6. The intelligent recommendation method for optimizing advertising keyword combinations through cross-validation according to claim 5, characterized in that: The S2 specifically includes: Specialized identification of low-resource layers and long-tail keyword mining, with sparse scores greater than a predetermined threshold and exposure at a predetermined ratio quantile; identification technology stack, extracting low-frequency co-occurring word pairs from search sessions through FP-Growth sequence pattern mining; and reconstructing exposure distributions and identifying statistically significant tail outliers through adversarial autoencoders. New combination cold start determination: whether the new keyword combination meets any of the following conditions: zero historical exposure, change in related product categories, or expansion of advertiser categories. If so, it is classified as a low-resource tier. Semantic similarity compensation: if the Jaccard similarity between the new combination and the high-resource word is greater than a predetermined threshold, an initial exposure estimate is assigned. Layered data governance strategy: at the high-resource layer, dimensionality reduction accelerates training, feature hashing is used to compress the dimensions of head keywords, and sparse features use Tf-Idf weighting instead of One-Hot encoding; at the low-resource layer, synthetic minority oversampling is used, and virtual samples are created using the CTGAN generative model; a cross-advertiser migration pool is established to establish a desensitized low-resource data sharing pool for non-competing advertisers in the same industry, and knowledge is aggregated through federated learning.

7. The intelligent recommendation method for optimizing advertising keyword combinations through cross-validation according to claim 6, characterized in that: The S5 specifically includes: Confidence interval driven risk-return modeling, neural quantile regression, embedding multi-quantile output heads in the estimation model to form dynamic confidence intervals; dynamic mapping of risk levels, bandwidth calculation; Multi-objective optimization function design, expected ROI, calculating comprehensive benefits based on the integral of the probability density function; risk control, setting the minimum acceptable ROI threshold; diversity constraints, ensuring keyword coverage of multiple product categories through Shannon entropy; Pareto front solution, NSGA-III algorithm optimization, population initialization, sampling from historical high-quality combinations, and injecting semantically similar long-tail keywords; crossover mutation rules, including semantic hybridization, exchanging subnodes with keywords in the same category, adversarial mutation, and generating counterfactual samples based on GAN; reference point generation, constructing a non-dominated solution set in the three-dimensional space of ROI, risk, and diversity; Real-time feedback mechanism, Bandit online learning, regular update of target weights, dynamic adjustment of regret values, and priority reduction of ranking weights for combinations with high regret values.

8. The intelligent recommendation method for optimizing advertising keyword combinations through cross-validation according to claim 7, characterized in that: The S4 specifically includes: In addition to basic CTR / CVR estimates, high-resource model outputs extract hidden layer activation vectors to capture intermediate states of user intent understanding. Low-resource model outputs add uncertainty quantification metrics: Monte Carlo Dropout variance and graph neural network node embedding distance; business features are injected into the competitive environment coefficient and advertiser risk preference. Fusion function, using a dual-channel adaptive fusion framework: Where, For the final fusion output, the prediction information of high-resource channels and low-resource channels are comprehensively considered; is the transformation function of the high resource channel, which is used to adjust the high resource prediction results. is the predicted output of the high resource channel; is the transformation function of the low-resource channel; is the weight coefficient corresponding to the high-resource channel, which is used to measure the importance of the high-resource channel in the fusion process; is the weight coefficient corresponding to the low resource channel, which is equal to the weight coefficient of the high resource channel Complementary, that is, the sum of the two is 1; For high-resource channel transformation, a Sigmoid calibration layer is introduced; for low-resource channel transformation, an uncertainty gating mechanism is designed: Bayesian optimization weight parameter tuning engine, optimized objective function design, Gaussian process regression modeling, kernel function selection, using the Matérn 5 / 2 kernel; acquisition function optimization, improved Expected Improvement strategy, and added exploration penalty terms.

9. The intelligent recommendation method for optimizing advertising keyword combinations through cross-validation according to claim 8, characterized in that: Said S1 specifically includes: Acquisition of historical interaction data, including user behavior logs, advertising interaction records, and third-party supplementary data; User behavior logs, including search queries, click sequences, page dwell time, shopping cart operations, and cross-session behavior links, are collected in real time using the tracking SDK and cleaned of abnormal traffic using the Apache Flink stream processing engine. Ad interaction records, including impressions, click-through rates, conversion rates, cost per conversion, and bidding environment parameters, are connected to the advertising platform API to obtain data down to the keyword group level; Third-party supplementary data, including user device information, geographic location tags, and time period characteristics, are integrated with DMP's desensitized portrait tags through joint modeling.

10. The intelligent recommendation method for optimizing advertising keyword combinations through cross-validation according to claim 9, characterized in that: Said S1 specifically includes: We constructed a multimodal dataset and generated keyword semantic vectors. By fine-tuning a pre-trained language model, we used the RoBERTa-large model and continued pre-training on vertical domain corpora to capture industry terminology. We also used context-aware encoding and adversarial training to enhance generalization. We also introduced FGSM adversarial training and added perturbations to the input embeddings. User behavior features are constructed using an LSTM-Attention model based on behavior sequences to capture intent strength and differentiate between browsing and decision-making users. The reach time difference and conversion contribution between search ads and social media ads are calculated to achieve cross-channel collaboration and quantify multi-channel attribution effectiveness. Sliding window statistics are used to calculate the distribution of click categories, and JS divergence is used to measure the rate of interest change, enabling real-time interest drift and timely adjustment of keywords based on demand fluctuations. Quantifying advertising creative attributes and analyzing textual creativity, using BERT to extract copywriting sentiment polarity and promotional information density; visual creative analysis, deploying the ResNet-50 convolutional network to identify product entities, scene types, and character attributes in images and videos, and outputting structured labels; Conversion label definition, multi-attribution model integration, definition of basic labels including last-click conversion and linear attribution, and use of Shapley Value algorithm to assign keyword combination contribution weights.

Images