A Big Data-Based Method for Predicting Internet User Ad Clicks
By refining and multi-level processing user and advertising features, and combining factorization machines, sequence modeling, and deep neural networks, the problems of high-dimensional sparse features and complex interactions are solved, improving the accuracy and generalization ability of ad click prediction and meeting actual business needs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-02-11
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies struggle to effectively handle high-dimensional sparse features, capture complex feature interactions, and adapt to the generalization needs of real-world scenarios, resulting in insufficient accuracy and reliability in predicting ad clicks for internet users.
By refining user and advertising features into coarse and fine granular categories, low-dimensional dense features are extracted using a factorization machine model. Clustering is then performed using sequence modeling and the K-medoids algorithm to construct an independent gradient boosting decision tree model. Finally, feature fusion and prediction are achieved through a deep neural network.
It significantly improves the accuracy and generalization ability of internet user ad click prediction, better adapts to the diverse user interests and complex data needs in real-world scenarios, optimizes traffic allocation, reduces campaign costs, and enhances user experience.
Smart Images
Figure CN122089398A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of internet advertising technology, and in particular relates to a method for predicting internet user ad clicks based on big data. Background Technology
[0002] The rapid development of the internet has propelled the advertising industry into an era of large-scale campaigns. The continuous emergence of massive amounts of user behavior and advertising data, coupled with the diversification of user interests and the enrichment of advertising formats, has presented a severe challenge to the accuracy of ad targeting. Accurately predicting user ad click behavior is crucial for advertising platforms to optimize traffic allocation and improve monetization efficiency. It is also a core need for advertisers to reduce ineffective campaign costs and improve conversion rates. Furthermore, it can reduce user distraction from irrelevant ads, enhancing user experience and possessing irreplaceable research and application value within the digital marketing ecosystem.
[0003] Current mainstream ad click prediction methods primarily rely on statistical models, collaborative filtering, traditional machine learning, and deep learning techniques. User ad click-related data contains rich user features, ad features, and contextual features, such as user age, gender, browsing history, ad type, content, display location, and information about the user's browsing time and device. These features are extremely high-dimensional, and traditional encoding methods easily generate high-dimensional sparse feature vectors. The sparse data structure makes it difficult for models to uncover effective correlations between features, directly affecting the model's generalization ability. Simultaneously, complex nonlinear interactions exist between different features; traditional methods often only capture the influence of a single or a few features, failing to comprehensively cover the deep interactions between multiple features, thus missing crucial prediction information. While deep learning models theoretically possess the potential to handle complex relationships, in practical applications, due to data sparsity and overfitting issues, they are prone to overlearning noise in the training data, resulting in poor prediction performance in real-world scenarios and difficulty in accurately matching users' true interests.
[0004] The existing technologies mentioned above have significant shortcomings in handling high-dimensional sparse features, capturing complex feature interactions, and adapting to the generalization requirements of real-world scenarios, resulting in the accuracy and reliability of user ad click prediction failing to meet actual business requirements. Therefore, how to effectively solve the modeling challenges posed by high-dimensional sparse features, comprehensively capture the complex interaction relationships between multiple features, and improve the model's generalization ability to adapt to real-world application scenarios has become a core technical bottleneck that urgently needs to be overcome in the current field of ad click prediction technology. Summary of the Invention
[0005] To address the problems existing in the background art, one aspect of the present invention provides a method for predicting internet user ad clicks based on big data, comprising:
[0006] S1: Obtain the user ad click dataset, wherein the user ad click data includes: user characteristics, ad characteristics, and the user's historical behavior sequence;
[0007] S2: Divide user features and advertising features into coarse-grained features and fine-grained features according to preset rules; perform one-hot encoding on the coarse-grained features and fine-grained features respectively to obtain the corresponding high-dimensional sparse feature vectors;
[0008] S3: Use the factorization machine model to process the high-dimensional sparse feature vectors of coarse-grained and fine-grained features and extract the corresponding low-dimensional dense features.
[0009] S4: Embed the user's historical behavior sequence and use a sequence modeling model to process the embedded feature representation to obtain the user's dynamic interest vector;
[0010] S5: Concatenate the user's dynamic interest vector and the low-dimensional dense features of the coarse-grained features to obtain the coarse-grained interest vector between user ads; use the K-medoids algorithm to cluster the coarse-grained interest vector between user ads to obtain K dynamic sample clusters;
[0011] S6: Construct an independent gradient boosting decision tree model for each dynamic sample cluster, input the corresponding fine-grained features of each sample cluster into the gradient boosting decision tree model for feature fusion, and output the fusion feature representation between user ads through one-hot encoding of leaf node IDs;
[0012] S7: Input the fusion feature representation between user ads into a deep neural network for prediction to obtain the user's click prediction result for the ads.
[0013] Another aspect of the present invention provides an internet user ad click prediction system based on big data, the system including a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the aforementioned internet user ad click prediction method based on big data.
[0014] Another aspect of the present invention provides a computer storage medium storing a remote monitoring program, which, when executed by a processor, implements the aforementioned method for predicting internet user ad clicks based on big data.
[0015] The present invention has at least the following beneficial effects
[0016] This invention effectively solves the challenge of high-dimensional sparse feature modeling by finely segmenting user and advertising features into coarse and fine granular levels and processing them accordingly. It also fully captures the deep correlations between multi-dimensional features through feature interaction mining technology. Simultaneously, it accurately captures users' dynamic interests through sequence modeling, combines clustering and dedicated models to adapt to the feature distribution of different interest groups, and further enhances prediction robustness with a dual-task deep neural network. Ultimately, this significantly improves the accuracy and generalization ability of internet user ad click prediction, better adapting to the diverse user interests and complex data needs in real-world scenarios. This provides reliable support for advertising platforms to optimize traffic allocation, advertisers to reduce campaign costs, and improve user experience. Attached Figure Description
[0017] Figure 1 This is a schematic diagram of the method flow of the present invention;
[0018] Figure 2 This is a flowchart of the method framework of the present invention. Detailed Implementation
[0019] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0020] Please see Figure 1 and Figure 2 One aspect of the present invention provides a method for predicting internet user ad clicks based on big data, comprising:
[0021] S1: Obtain the user ad click dataset, wherein the user ad click data includes: user characteristics, ad characteristics, and the user's historical behavior sequence;
[0022] Preferably, the user characteristics include coarse-grained characteristics and fine-grained characteristics, wherein coarse-grained characteristics include user gender, user's geographical region, user device type, and user occupation, etc., and fine-grained characteristics include user age, user interest tags, and user's following list, etc.; the advertising characteristics include coarse-grained characteristics and fine-grained characteristics, wherein coarse-grained characteristics include the major industry to which the advertisement belongs, advertisement type, and type of page on which the advertisement is located, etc.; fine-grained characteristics include advertising keywords, advertising value words, and thematic keywords of the page on which the advertisement is located, etc.
[0023] In this embodiment, data is collected synchronously through multiple channels, including the advertising platform's log collection system, user behavior tracking module, and advertising delivery management system. After collection, preliminary data cleaning is required, including removing invalid records with format errors, duplication, redundancy, and logical contradictions (such as empty user IDs or data with logical conflicts between ad display time and click time). Missing values in key fields are appropriately filled (e.g., if a user's occupation is missing, it can be filled based on other user characteristics). Finally, a user ad click dataset with a standardized structure and complete information is formed. This dataset specifically includes the following three core data categories: User characteristics: These refer to the user's long-term stable basic attribute information that does not change with short-term behavior. They are key data for characterizing the user's inherent labels, specifically including: Coarse-grained features: User gender (e.g., male, female, unknown), user's geographical region (e.g., North China, East China, South China, etc., first-level administrative divisions), user device type (e.g., smartphone, tablet, desktop computer, smart TV, etc.), user occupation (e.g., corporate employee, student, freelancer, public official, etc.); Fine-grained features: User age (accurate to a specific age range or specific value, such as 25-30 years old, 35 years old, etc.). User characteristics include: age, user interest tags (such as interest categories set by users or marked by the platform based on long-term behavior, such as digital, beauty, fitness, maternal and infant, finance, etc.), user following list (such as specific information such as accounts, topics, columns, etc. that users follow within the platform); advertising features: refer to the inherent attribute information of the advertisement itself, which directly reflects the content, type, and placement scenario of the advertisement, specifically including: coarse-grained features: the major industry to which the advertisement belongs (such as e-commerce retail, financial services, education and training, automotive transportation, etc.), advertisement type (such as information flow advertisement, splash screen advertisement, pop-up advertisement, video ad, etc.), and the type of page on which the advertisement is located (such as homepage, content details page, personal center page, search results page, etc.); fine-grained features: advertising keywords (such as core promotional terms in the advertising copy, such as "wireless Bluetooth headphones", "low-fat meal replacement", etc.), advertising value words (such as those reflecting the core selling point of the advertisement). The keywords for each ad include: keywords such as "limited-time discount," "official authentic product," and "free trial"; the theme keywords of the page where the ad is displayed (such as the core content keywords of the ad display page, such as "home-style dishes" and "cooking skills" on a food tutorial page); and the user's historical behavior sequence, which refers to the user's interaction behavior with various ads or related content on the platform within a certain period of time before the current ad is displayed. This is the core data for capturing the user's dynamic interests. Specifically, it includes the sequence of ad IDs that the user has recently interacted with (such as the unique identifier of ads that the user has clicked, viewed, or favorited in the past 30 days), the behavior type of each interaction (click, view, favorite, share, skip, etc.), the interaction timestamp (accurate to the second, used to reflect the temporal relationship of user behavior), and the interaction duration (the time the user views or stays on the ad). The user's interest change trends and preferences will be mined based on this sequence in the future.
[0024] S2: Divide the features and advertising features into coarse-grained features and fine-grained features according to preset rules; perform one-hot encoding on the coarse-grained features and fine-grained features respectively to obtain the corresponding high-dimensional sparse feature vectors;
[0025] In this embodiment, refined feature segmentation and standardized coding provide standardized input for subsequent feature processing. First, based on the abstraction level, coverage, and information granularity of the features, a predefined segmentation rule is established to divide user static features and advertising features into coarse-grained features and fine-grained features, respectively. Coarse-grained features focus on macro-level classification information (such as user gender, geographical region, and the major industry of the advertisement), while fine-grained features focus on specific details (such as user age, interest tags, and advertising keywords). This segmentation enables targeted processing of features at different levels. Subsequently, one-hot coding is used to transform the two types of features. Specifically, for each discrete coarse-grained and fine-grained feature, a complete set of all possible values is first constructed, and then a unique vector dimension is assigned to each value. When the feature takes a specific value in the sample, the element of the corresponding dimension is set to 1, and the elements of all other dimensions are set to 0. Finally, each category of features is transformed into a corresponding high-dimensional sparse feature vector, which not only fully preserves the category attribute information of the features but also meets the adaptation requirements of the subsequent model for numerical inputs.
[0026] S3: Use the factorization machine model to process the high-dimensional sparse feature vectors of coarse-grained and fine-grained features and extract the corresponding low-dimensional dense features.
[0027] Preferably, the process of using a factorization machine model to process the high-dimensional sparse feature vectors of coarse-grained and fine-grained features includes: training the factorization machine model using a user ad click dataset, and extracting the latent vectors of each coarse-grained and fine-grained feature as the corresponding low-dimensional dense features, wherein the factorization machine model includes:
[0028]
[0029] in, This indicates the probability that a user will click on the advertisement. This represents the total number of features, including both coarse-grained and fine-grained features. Indicates the global bias term; Indicates the first Linear weights of each feature; Indicates the first The latent vector of each feature; Indicates the first The latent vector of each feature; Indicates the first One feature; Indicates the first One characteristic.
[0030] In this embodiment, the high-dimensional sparse feature vectors of coarse-grained and fine-grained features are subjected to dimensionality reduction and feature interaction mining. Specifically, the factorization machine model is fully trained using the constructed user ad click dataset. During the training process, the model not only learns the global bias term and the linear weights of each feature to capture the independent influence of a single feature, but also learns the latent vectors corresponding to each coarse-grained and fine-grained feature, transforming the complex interaction relationship between features into the inner product operation of latent vectors. Finally, the latent vectors corresponding to each feature output by the model are directly extracted as low-dimensional dense features. This effectively solves the problems of high computational complexity and weak generalization ability caused by high-dimensional sparse data, while fully preserving the feature information itself and the deep correlation between features.
[0031] S4: Embed the user's historical behavior sequence and use a sequence modeling model to process the embedded feature representation to obtain the user's dynamic interest vector;
[0032] Preferably, step S4 includes: setting a maximum sequence length T, extracting the T most recently interacted ad IDs, embedding them to obtain the embedded feature representation, and processing the embedded feature representation using a GRU model or a Transformer model to obtain the user's dynamic interest vector.
[0033] In this embodiment, a reasonable maximum sequence length T (e.g., 30, 50, etc., to ensure coverage of recent key user interactions) is first set according to the business scenario and data distribution characteristics. The T most recently interacted ad IDs are extracted from the user's historical behavior sequence (if the number of interactions is less than T, zero vectors are used for padding). Each discrete ad ID is mapped to a fixed-dimensional low-dimensional dense vector through an embedding layer, which is the feature representation after embedding. This vector does not directly encode the explicit attributes of the ad, but rather contains a numerical representation of potential information such as semantic and type associations between ads, transforming symbolic ad identifiers into continuous features that the model can process. Subsequently, a GRU model or a Transformer model is used to process this embedded feature sequence. Temporal modeling is performed. The GRU model adaptively forgets irrelevant information from the user's historical behavior and retains key temporal features strongly correlated with current interests through the synergistic effect of reset and update gates. The Transformer model captures the long- and short-term dependencies between different interactions through a self-attention mechanism. Both models ultimately output a dynamic user interest vector that integrates the user's recent behavioral temporal patterns and interest preferences, intuitively reflecting the user's interest tendency at the current moment. If the Transformer model is used, a special [CLS] token (ClassificationToken) is added to the beginning of the input sequence. The model uses a self-attention mechanism to capture long-distance dependencies between different behaviors in the sequence in parallel. Since the [CLS] token is designed to aggregate global semantic information of the entire sequence during training, the vector corresponding to the [CLS] token position in the Transformer model's output layer is extracted as the user's dynamic interest vector. If the GRU model is used, the hidden state of the model at the last time step is extracted as the user's dynamic interest vector.
[0034] S5: Concatenate the user's dynamic interest vector and the low-dimensional dense features of the coarse-grained features to obtain the coarse-grained interest vector between user ads; use the K-medoids algorithm to cluster the coarse-grained interest vector between user ads to obtain K dynamic sample clusters;
[0035] Preferably, the clustering of coarse-grained interest vectors among user advertisements using the K-medoids algorithm includes:
[0036] S51: Randomly select K distinct samples as the initial cluster centers;
[0037] S52: For each sample, calculate its Euclidean distance to all K cluster centers, assign the sample to the cluster corresponding to the nearest cluster center, and after all samples are assigned, obtain K disjoint clusters;
[0038] S53: For each cluster Iterate through each sample within the cluster. Calculate if the cluster center is moved from Replace with Cost changes over time :
[0039]
[0040] in, Represents the Euclidean distance function;
[0041] S54: If a candidate sample exists , making Then the center of the cluster is updated to Otherwise, keep the original center. constant;
[0042] S55: Repeat steps S53~S54 to obtain a new set of cluster centers. If the set of cluster centers before and after the update has not changed, the algorithm converges and outputs the final K clusters; otherwise, return to step S52.
[0043] In this embodiment, feature fusion and clustering are used to classify user interests into groups, laying the foundation for subsequent personalized feature fusion. Specifically, the user dynamic interest vector obtained in step S4 (containing the time sequence pattern of recent user behavior and real-time interest tendency) is first concatenated with the coarse-grained low-dimensional dense features extracted in step S3 (covering the macro attributes and interaction associations between users and advertisements) to form a user-advertisement coarse-grained interest vector that simultaneously contains user dynamic preferences and macro feature associations. This vector can comprehensively and concisely characterize the basic interest matching degree of users to advertisements. Then, the K-medoids algorithm is used to cluster all coarse-grained interest vectors. First, K different samples are randomly selected as initial cluster centers. By calculating the Euclidean distance between each sample and each cluster center, the sample is assigned to the nearest cluster. Then, the cost change of replacing the cluster center is calculated by traversing the samples in each cluster. The cluster centers are iteratively updated until the cluster center set is stably converged, and finally K dynamic sample clusters are obtained. The samples in each cluster have similar coarse-grained interest features, thereby achieving effective division of heterogeneous user groups and providing accurate sample grouping support for building exclusive models for different interest groups in the future.
[0044] S6: Construct an independent gradient boosting decision tree model for each dynamic sample cluster, input the corresponding fine-grained features of each sample cluster into the gradient boosting decision tree model for feature fusion, and output the fusion feature representation between user ads through one-hot encoding of leaf node IDs;
[0045] Preferably, the step of constructing an independent gradient boosting decision tree model for each sample cluster and using the gradient boosting decision tree model to perform feature fusion and output the fused feature representation of each sample includes:
[0046] S61: For the K dynamic sample clusters obtained by partitioning { , ,…, ,…, For each dynamic sample cluster, an independent gradient boosting decision tree model is constructed, denoted as}. This forms a model group containing K gradient boosting decision tree models. , ,…, ,…, During the training phase, for the first... Gradient boosting decision tree model Only use samples belonging to the dynamic sample cluster. The samples were used for training;
[0047] S62: For any sample to be predicted Identify the ID of the dynamic sample cluster to which it belongs. And call the corresponding gradient boosting decision tree model from the model group. ; Extracting samples Fine-grained features As a gradient boosting decision tree model Input;
[0048] S63: Gradient Boosting Decision Tree Model It consists of M decision trees, denoted as { , ,…, ,…, }; will sample Fine-grained features Input gradient boosting decision tree model ,sample In each decision tree The process involves traversing from the root node to the leaf node according to the feature splitting rules; recording the samples. In the Leaf node index in a decision tree Then the index sequence of the sample on M trees is represented as ={ , ,…, ,…, };
[0049] S64: For the index sequence Each leaf node index in Perform one-hot encoding transformation to obtain the corresponding sparse vector. Concatenate all M sparse vectors end-to-end to generate samples. High-dimensional sparse fusion feature representation , ,…, In this context, Z represents the fusion feature between user ads and is input into the deep neural network for prediction.
[0050] In this embodiment, personalized fusion of fine-grained features is achieved for different interest groups to enhance the relevance and effectiveness of feature representation. Specifically, based on the K dynamic sample clusters obtained in step S5, a dedicated gradient boosting decision tree model is independently constructed for each cluster to form a model group. Since the coarse-grained interest features of samples within the same cluster are highly similar, the fine-grained features exhibit a unique distribution pattern. Independent modeling allows each model to focus on the unique fine-grained feature interaction patterns within the cluster. During training, only the sample data of the corresponding cluster is used for iterative optimization to ensure that the model accurately adapts to the feature patterns within the cluster. In the prediction stage, the model first identifies the target... The dynamic sample cluster number to which the predicted sample belongs is called, and the gradient boosting decision tree model corresponding to that cluster is invoked. The fine-grained features of the sample are input into the model. The sample will be traversed layer by layer from the root node to the unique leaf node in each decision tree contained in the model according to the feature splitting rules. The index of the corresponding leaf node in each tree is recorded and formed into an index sequence. Then, each index in the index sequence is one-hot encoded to obtain a sparse vector. Finally, all sparse vectors are concatenated to generate a fusion feature representation containing deep interaction information of fine-grained features, which provides high-quality and personalized feature input for the accurate prediction of the subsequent deep neural network.
[0051] S7: Input the fusion feature representation between user ads into a deep neural network for prediction to obtain the user's click prediction result for the ads.
[0052] Preferably, the deep neural network includes: a feature extraction backbone network, a click task prediction branch, and a deep interaction task branch; the click task prediction branch is used to predict the probability of a user clicking on an advertisement; the deep interaction task branch is used to predict the probability of a user engaging in deep interaction with an advertisement; wherein, the feature extraction backbone network adopts a fully connected network; both the click task prediction branch and the deep interaction task branch adopt a fully connected network and a sigmoid activation function; the click task prediction branch and the deep interaction task branch share the output features of the feature extraction backbone network; the deep interaction refers to whether the user stays on the advertisement for a time greater than a set threshold, if so, then a deep interaction has occurred between the user and the advertisement, otherwise, no deep interaction has occurred between the user and the advertisement.
[0053] Preferably, the loss function used during the training of the deep neural network includes:
[0054]
[0055] in, Represents the total loss function. This represents the cross-entropy loss function for click tasks. Indicates the weighting parameter; This represents the cross-entropy loss function for deep interactive tasks.
[0056] In this embodiment, accurate click prediction is achieved through a multi-task deep neural network. The fused feature representation is input into the feature extraction backbone of the deep neural network. The backbone performs deep information mining and abstraction transformation on the fused features, outputting a unified feature vector with strong representational capabilities. This deep neural network includes a click task prediction branch and a deep interaction task prediction branch, with both branches sharing the output features of the backbone network. The click task prediction branch is specifically used to predict the probability of a user clicking on an advertisement, while the deep interaction task branch predicts the probability of a deep interaction between a user and an advertisement. Both branches use fully connected networks combined with a sigmoid activation function to output probability results. A total loss function is used during model training. The model is optimized by using dual-task collaborative training to improve its robustness and generalization ability. Finally, the click task prediction branch outputs the probability of the user clicking the advertisement, which serves as the final click prediction result, thus achieving accurate judgment of the user's click behavior.
[0057] Another aspect of the present invention provides an internet user ad click prediction system based on big data, the system including a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the aforementioned internet user ad click prediction method based on big data.
[0058] Another aspect of the present invention provides a computer storage medium storing a remote monitoring program, which, when executed by a processor, implements the aforementioned method for predicting internet user ad clicks based on big data.
[0059] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0060] In summary, this invention effectively solves the challenge of modeling high-dimensional sparse features by finely segmenting user and advertising features into coarse and fine granular levels and processing them accordingly. It also fully captures the deep correlations between multi-dimensional features through feature interaction mining technology. Furthermore, it accurately captures users' dynamic interests through sequence modeling, combines clustering and dedicated models to adapt to the feature distribution of different interest groups, and further enhances prediction robustness with a dual-task deep neural network. Ultimately, this significantly improves the accuracy and generalization ability of predicting ad clicks for internet users, better adapting to the diverse user interests and complex data needs in real-world scenarios. This provides reliable support for advertising platforms to optimize traffic allocation, advertisers to reduce campaign costs, and improve user experience.
[0061] Finally, 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 present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for predicting internet user ad clicks based on big data, characterized in that, include: S1: Obtain the user ad click dataset, wherein the user ad click data includes: user characteristics, ad characteristics, and the user's historical behavior sequence; S2: Divide user features and advertising features into coarse-grained features and fine-grained features according to preset rules; perform one-hot encoding on the coarse-grained features and fine-grained features respectively to obtain the corresponding high-dimensional sparse feature vectors; S3: Use the factorization machine model to process the high-dimensional sparse feature vectors of coarse-grained and fine-grained features and extract the corresponding low-dimensional dense features. S4: Embed the user's historical behavior sequence and use a sequence modeling model to process the embedded feature representation to obtain the user's dynamic interest vector; S5: Concatenate the user's dynamic interest vector and the low-dimensional dense features of the coarse-grained features to obtain the coarse-grained interest vector between user ads; use the K-medoids algorithm to cluster the coarse-grained interest vector between user ads to obtain K dynamic sample clusters; S6: Construct an independent gradient boosting decision tree model for each dynamic sample cluster, input the corresponding fine-grained features of each sample cluster into the gradient boosting decision tree model for feature fusion, and output the fusion feature representation between user ads through one-hot encoding of leaf node IDs; S7: Input the fusion feature representation between user ads into a deep neural network for prediction to obtain the user's click prediction result for the ads.
2. The method for predicting internet user ad clicks based on big data according to claim 1, characterized in that, The user characteristics include coarse-grained characteristics and fine-grained characteristics. Coarse-grained characteristics include user gender, user's geographical region, user device type, and user occupation. Fine-grained characteristics include user age, user interest tags, and user's following list. The advertising characteristics include coarse-grained characteristics and fine-grained characteristics. Coarse-grained characteristics include the major industry to which the advertisement belongs, the type of advertisement, and the type of page on which the advertisement is located. Fine-grained characteristics include advertising keywords, advertising value words, and the topic keywords of the page on which the advertisement is located.
3. The method for predicting internet user ad clicks based on big data according to claim 1, characterized in that, The process of using a factorization machine model to process the high-dimensional sparse feature vectors of coarse-grained and fine-grained features includes: training the factorization machine model using a user ad click dataset, and extracting the latent vectors of each coarse-grained and fine-grained feature as the corresponding low-dimensional dense features. The factorization machine model includes: in, This indicates the probability that a user will click on the advertisement. This represents the total number of features, including both coarse-grained and fine-grained features. Indicates the global bias term; Indicates the first Linear weights of each feature; Indicates the first The latent vector of each feature; Indicates the first The latent vector of each feature; Indicates the first One feature; Indicates the first One characteristic.
4. The method for predicting internet user ad clicks based on big data according to claim 1, characterized in that, Step S4 includes: setting a maximum sequence length T, extracting the T most recently interacted ad IDs, embedding them to obtain the embedded feature representation, and using a GRU model or Transformer model to process the embedded feature representation to obtain the user's dynamic interest vector.
5. The method for predicting internet user ad clicks based on big data according to claim 1, characterized in that, The method of clustering coarse-grained interest vectors among user advertisements using the K-medoids algorithm includes: S51: Randomly select K distinct samples as the initial cluster centers; S52: For each sample, calculate its Euclidean distance to all K cluster centers, assign the sample to the cluster corresponding to the nearest cluster center, and after all samples are assigned, obtain K disjoint clusters; S53: For each cluster Iterate through each sample within the cluster. Calculate if the cluster center is moved from Replace with Cost changes over time : in, Represents the Euclidean distance function; S54: If a candidate sample exists , making Then the center of the cluster is updated to Otherwise, keep the original center. constant; S55: Repeat steps S53~S54 to obtain a new set of cluster centers. If the set of cluster centers before and after the update has not changed, the algorithm converges and outputs the final K clusters; otherwise, return to step S52.
6. The method for predicting internet user ad clicks based on big data according to claim 1, characterized in that, The step of constructing an independent gradient boosting decision tree model for each sample cluster, and using the gradient boosting decision tree model to perform feature fusion and output the fused feature representation of each sample includes: S61: For the K dynamic sample clusters obtained by partitioning { , ,…, ,…, For each dynamic sample cluster, an independent gradient boosting decision tree model is constructed, denoted as}. This forms a model group containing K gradient boosting decision tree models. , ,…, ,…, During the training phase, for the first... Gradient boosting decision tree model Only use samples belonging to the dynamic sample cluster. The samples were used for training; S62: For any sample to be predicted Identify the ID of the dynamic sample cluster to which it belongs. And call the corresponding gradient boosting decision tree model from the model group. ; Extracting samples Fine-grained features As a gradient boosting decision tree model Input; S63: Gradient Boosting Decision Tree Model It consists of M decision trees, denoted as { , ,…, ,…, }; to sample Fine-grained features Input gradient boosting decision tree model ,sample In each decision tree The process involves traversing from the root node to the leaf node according to the feature splitting rules; recording the samples. In the Leaf node index in a decision tree Then the index sequence of the sample on M trees is represented as ={ , ,…, ,…, }; S64: For the index sequence Each leaf node index in Perform one-hot encoding transformation to obtain the corresponding sparse vector. Concatenate all M sparse vectors end-to-end to generate samples. High-dimensional sparse fusion feature representation , , …, In this context, Z represents the fusion feature between user ads and is input into the deep neural network for prediction.
7. The method for predicting internet user ad clicks based on big data according to claim 1, characterized in that, The deep neural network includes: a feature extraction backbone network, a click task prediction branch, and a deep interaction task branch; the click task prediction branch is used to predict the probability of a user clicking on an advertisement; the deep interaction task branch is used to predict the probability of a user engaging in deep interaction with an advertisement; wherein, the feature extraction backbone network adopts a fully connected network; both the click task prediction branch and the deep interaction task branch adopt a fully connected network and a sigmoid activation function; the click task prediction branch and the deep interaction task branch share the output features of the feature extraction backbone network; deep interaction refers to whether the user stays on the advertisement for a time greater than a set threshold, if so, then a deep interaction has occurred between the user and the advertisement, otherwise, no deep interaction has occurred between the user and the advertisement.
8. The method for predicting internet user ad clicks based on big data according to claim 7, characterized in that, The loss function used during the training of the deep neural network includes: in, Represents the total loss function. This represents the cross-entropy loss function for click tasks. Indicates the weighting parameter; This represents the cross-entropy loss function for deep interactive tasks.
9. A big data-based internet user ad click prediction system, characterized in that, The system includes a memory and a processor; the memory is used to store an application; the processor is used to run the application and execute the big data-based Internet user advertising click prediction method as described in any one of claims 1 to 8.
10. A computer storage medium, characterized in that, The computer storage medium stores a remote monitoring program, which, when executed by a processor, implements a big data-based method for predicting internet user advertising clicks as described in any one of claims 1 to 8.