Intelligent advertisement content recommendation method and system applying prediction model

Abstract

The invention provides an intelligent advertisement content recommendation method and system applying a prediction model, and the method comprises the steps: firstly obtaining a historical interaction data set of a target user in an advertisement interaction scene, the historical interaction data set comprises a plurality of advertisement interaction sequences composed of user identifiers and the like, and then carrying out the cross-dimension feature fusion processing of the historical interaction data set, generating a combined behavior feature set and an advertisement content association feature set, calling a preset depth prediction model to carry out time sequence dependence modeling on the combined behavior feature set, and obtaining an advertisement content preference prediction result containing click probability distribution and an interaction intention classification label, the method comprises the following steps: selecting an advertisement content preference prediction result, performing dynamic weight matching according to the advertisement content preference prediction result and an advertisement content association feature set, determining a candidate advertisement content priority ranking list, finally pushing an adaptive advertisement content set based on the candidate advertisement content priority ranking list, and updating parameter configuration of a depth prediction model according to real-time feedback data. And intelligent and accurate advertisement recommendation is realized.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to an intelligent advertising content recommendation method and system using a prediction model. Background Art

[0002] In the current digital advertising field, with the explosive growth of information and the increasing prominence of user personalized needs, traditional advertising content recommendation methods have been unable to meet the high standards of the market. Traditional methods often rely on simple user behavior statistics or rule-based matching strategies, which are unable to cope with complex and changing user behavior patterns.

[0003] Specifically, existing advertising recommendation systems can usually only process features of a single dimension or limited dimensions, and are unable to effectively integrate user behavior characteristics, advertising content characteristics, and the complex relationships between them. This limits the accuracy and personalization of the recommendation system, making the advertising content often out of touch with the actual needs of users, and reducing the effective reach of advertising and user satisfaction.

[0004] In addition, traditional advertising recommendation methods lack dynamic adjustment and optimization mechanisms. For example, they often adopt fixed recommendation strategies and cannot be flexibly adjusted based on real-time feedback from users and dynamic changes in the market. In the rapidly changing digital marketing environment, this static recommendation model is obviously unable to adapt to the competitive needs of the market. Summary of the invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present application, an embodiment of the present application provides an intelligent advertising content recommendation method using a prediction model, the method comprising: Acquire a historical interaction data set of a target user in an advertisement interaction scenario, wherein the historical interaction data set includes a plurality of advertisement interaction sequences, each advertisement interaction sequence consisting of a user identifier, an advertisement content identifier, an interaction behavior type, and an interaction timestamp; Performing cross-dimensional feature fusion processing on the historical interaction data set to generate a combined behavior feature set and an advertisement content-related feature set for each advertisement interaction sequence; Calling a preset deep prediction model to perform time-series dependency modeling on the combined behavior feature set to generate an advertising content preference prediction result, wherein the advertising content preference prediction result includes a user's click probability distribution for the candidate advertising content and an interaction intention classification label; Performing dynamic weight matching based on the advertising content preference prediction result and the advertising content associated feature set to determine a priority ranking list of candidate advertising contents; Based on the priority sorting list, a set of adapted advertising content is pushed to target users, and parameter configuration of the deep prediction model is updated according to real-time feedback data.

[0006] On the other hand, an embodiment of the present application also provides an advertising service system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to run the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0007] Based on the above aspects, the embodiment of the present application not only captures the temporal characteristics of user behavior by integrating the multi-dimensional historical interaction data of the target user in the advertising interaction scenario, but also deeply associates the user behavior pattern with the advertising content characteristics through cross-dimensional feature fusion processing, forming a comprehensive and sophisticated combination of behavioral feature sets and advertising content related feature sets. The introduction of the deep prediction model further explores the temporal dependencies in the user behavior sequence, effectively predicts the user's click probability distribution and interaction intention classification labels for candidate advertising content, and greatly improves the accuracy and personalization level of advertising content recommendation. On this basis, the dynamic weight matching mechanism intelligently determines the priority ranking list of candidate advertising content based on the deep fusion of the advertising content preference prediction results and the advertising content related feature set, ensuring the timeliness and relevance of advertising push. Finally, based on the priority ranking list, the adapted advertising content set is pushed to the target user, and the deep prediction model is continuously optimized based on real-time feedback data, forming a self-iterative and evolving intelligent recommendation system, which not only significantly improves the click-through rate and conversion rate of advertising recommendations, but also enhances the satisfaction and loyalty of user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a schematic diagram of the execution flow of the intelligent advertising content recommendation method using the prediction model provided in the embodiment of the present application.

[0009] Figure 2 It is a schematic diagram of the hardware architecture of the advertising service system provided in the embodiment of the present application. DETAILED DESCRIPTION

[0010] The present application will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of an intelligent advertising content recommendation method using an application prediction model provided by an embodiment of the present application. The intelligent advertising content recommendation method using the application prediction model is introduced in detail below.

[0011] Example section: Step S110: Acquire a historical interaction data set of a target user in an advertisement interaction scenario, wherein the historical interaction data set includes a plurality of advertisement interaction sequences, each of which is composed of a user identifier, an advertisement content identifier, an interaction behavior type, and an interaction timestamp.

[0012] In this embodiment, in order to realize intelligent advertising content recommendation, taking the advertising interaction scenario of an online shopping platform as an example, the server of the online shopping platform can continuously record a large amount of interaction information between users and advertisements. Assume that the online shopping platform has many users such as user A and user B. Each user will see different advertisements and interact with the above advertisements when browsing product pages, searching for products, etc. For example, user A interacts with multiple advertisements within a week, and the server will record the above interaction information to form an advertising interaction sequence. Each advertising interaction sequence contains a user identifier, such as user A's unique identification number "12345"; an advertising content identifier, such as the number "AD001" of a mobile phone advertisement; an interaction behavior type, such as clicking on an advertisement, adding the product in the advertisement to a shopping cart, purchasing an advertised product, etc.; and an interaction timestamp. For example, user A clicks on an advertisement numbered "AD001" at 10:30 on July 1, 2024, then the time "2024-07-0110:30:00" is the interaction timestamp. The server integrates the above-mentioned advertising interaction sequences collected from all users to form a historical interaction data set. The data in the historical interaction data set is the basis for subsequent advertising recommendations. By analyzing the above data, the user's advertising interaction habits and preferences can be understood.

[0013] Step S120: performing cross-dimensional feature fusion processing on the historical interaction data set to generate a combined behavior feature set and an advertisement content-related feature set for each advertisement interaction sequence.

[0014] In this embodiment, after obtaining the historical interaction data set, it is necessary to perform cross-dimensional feature fusion processing on it to generate a useful feature set. Continuing with the above online shopping platform example, each advertising interaction sequence in the historical interaction data set is deeply analyzed and processed.

[0015] Step S121: performing time window segmentation processing on the interactive behavior types in the advertisement interactive sequence to generate a plurality of behavior segment units, each of which includes continuous interactive behaviors within a preset time length.

[0016] In this embodiment, taking the advertising interaction sequence of user A as an example, in order to better analyze the user's behavior pattern, the interaction behavior type is segmented by time window. Assuming that the preset duration is 1 day, the advertising interaction behavior of user A within a week is divided into daily periods. For example, on July 1, 2024, user A clicked on the mobile phone ad "AD001" and viewed the details page of the computer ad "AD002". The above continuous interaction behaviors within one day constitute a behavior fragment unit. Similarly, the interaction behaviors on July 2, 2024, July 3, and other days constitute different behavior fragment units respectively. In this way, through time window segmentation, the user's advertising interaction behavior is divided into multiple segments with time continuity, which facilitates the subsequent extraction and analysis of the features of each segment.

[0017] Step S122: extracting the behavior density feature and the interaction strength feature of each behavior segment unit, wherein the behavior density feature is determined by the interaction frequency per unit time, and the interaction strength feature is determined by the weighted value of the interaction duration and the operation depth.

[0018] In this embodiment, for each behavior segment unit, the behavior density feature and the interaction intensity feature are further extracted. Taking the behavior segment unit of July 1, 2024 as an example, the behavior density feature can be determined by calculating the interaction frequency within a unit time (here is 1 day). If user A clicks on the advertisement 3 times and views the advertisement detail page 2 times in this day, then the interaction frequency of this day is 5 times, and this 5 times is an important reference value for the behavior density feature of the behavior segment unit. The interaction intensity feature needs to comprehensively consider the duration of interaction and the depth of operation. Assume that after user A clicks on the mobile phone advertisement "AD001", he stays on the advertisement page for 3 minutes and performs operations such as viewing product parameters and comparing different styles. The above operations can be quantified as the depth of operation. Assuming that the weight of the interaction duration is 0.6 and the weight of the operation depth is 0.4, the interaction duration and the operation depth are weighted according to the weight, and the result is the interaction intensity feature of the behavior segment unit. For example, the duration of interaction is quantified as 3 points (out of 10 points), and the operation depth is quantified as 6 points (out of 10 points), then the interaction intensity feature = 3×0.6+6×0.4=4.2 points.

[0019] Step S123: performing semantic analysis on the advertisement material corresponding to the advertisement content identifier, and extracting semantic topic distribution features of the advertisement text and attention weight features of the visual elements.

[0020] In this embodiment, the advertising material corresponding to the advertising content identifier in each advertising interaction sequence is subjected to semantic analysis. Taking the advertisement "AD001" as an example, the advertising material may contain advertising text and visual elements. The advertising text is semantically analyzed. For example, the advertising text is "new smart phone, super performance, clear photo taking, affordable price". The semantic topic distribution characteristics of the advertising text are analyzed through natural language processing technology. The semantic topics can be divided into "performance", "photo taking function", "price", etc., and the proportion of each topic in the text is calculated. Assuming that the proportion of the "performance" topic is 0.3, the proportion of the "photo taking function" topic is 0.4, and the proportion of the "price" topic is 0.3, the above-mentioned proportion values ​​constitute the semantic topic distribution characteristics of the advertising text. For the visual elements of the advertisement, such as pictures, videos, etc., the attention weight characteristics are analyzed by computer vision technology. For example, the camera of the mobile phone is highlighted in the advertising picture. After analysis, the attention weight of the camera element is 0.7, and the sum of the attention weights of other elements is 0.3. The above-mentioned attention weight value is the attention weight feature of the visual element.

[0021] Step S124: The behavior density features, interaction intensity features, semantic topic distribution features and attention weight features corresponding to the same advertisement interaction sequence are concatenated to generate the combined behavior feature set.

[0022] In this embodiment, taking the advertisement interaction sequence between user A and advertisement "AD001" as an example, the previously extracted behavior density features, interaction intensity features, semantic topic distribution features and attention weight features are spliced. Assume that the behavior density feature is 5 times / day, the interaction intensity feature is 4.2 points, the semantic topic distribution feature is (the "performance" theme accounts for 0.3, the "photographing function" theme accounts for 0.4, and the "price" theme accounts for 0.3), and the attention weight feature is (the camera element attention weight is 0.7, and the other elements attention weight is 0.3). The above features are spliced ​​in a certain order, for example, the behavior density features and the interaction intensity features are spliced ​​first, and then the semantic topic distribution features and the attention weight features are spliced ​​to obtain a multi-dimensional combined behavior feature set. This combined behavior feature set contains information on multiple aspects of the user's interaction with the advertisement, and can more comprehensively reflect the user's interactive behavior characteristics for the advertisement.

[0023] Step S125: calling the advertisement knowledge base to match the product category tree corresponding to the advertisement content identifier, extracting the hierarchical association features of the category path and the user's historical preference matching degree, and generating the advertisement content association feature set.

[0024] In this embodiment, taking the advertisement "AD001" as an example, the advertisement knowledge base is called to match the product category tree corresponding to the advertisement content identifier. The advertisement knowledge base stores product category information corresponding to various advertisements, and the product category tree corresponding to the "AD001" mobile phone advertisement can be found through matching. The product category tree may contain categories of different levels such as "electronic products", "mobile phones", and "smart phones".

[0025] Step S1251: parse the parent-child node relationship of the product category tree to generate category graph structure data including a hierarchical connection relationship.

[0026] In this embodiment, the parent-child node relationship of the product category tree is parsed. For example, "electronic products" is the parent node of "mobile phone", and "mobile phone" is the parent node of "smart phone". The above node relationship is represented by a graph to form category graph structure data. In the category graph structure data, nodes represent categories and edges represent parent-child relationships. In this way, the hierarchical connection relationship between categories can be clearly seen, providing a basis for the subsequent extraction of hierarchical association features.

[0027] Step S1252: Input the category graph structure data into the graph neural network, and generate a hierarchical-aware embedding vector for each category node through feature propagation of adjacent nodes.

[0028] In this embodiment, the generated category graph structure data is input into the graph neural network. The graph neural network can learn the characteristics of each category node by propagating the features of adjacent nodes. For example, the "smartphone" node will obtain feature information from its parent node "mobile phone" and its other adjacent "smartphone" subcategory nodes. After calculation by the graph neural network, a hierarchical-aware embedding vector is generated for each category node. The above hierarchical-aware embedding vector contains the hierarchical information of the category node in the category tree and the association information with other nodes, which can better represent the characteristics of the category node.

[0029] Step S1253: Count the historical exposure times and click-through conversion rate of each category node in the target user's historical interaction data set to generate a user category preference distribution vector.

[0030] In this embodiment, the historical exposure times and click-through conversion rate of each category node in the historical interaction data set of the target user (such as user A) are counted. For example, in the historical interaction data of user A, the "electronic products" category was exposed 10 times, of which there were 3 clicks, and the click-through conversion rate was 30%; the "mobile phone" category was exposed 8 times, with 2 clicks, and the click-through conversion rate was 25%; the "smart phone" category was exposed 6 times, with 2 clicks, and the click-through conversion rate was 33.3%. The above historical exposure times and click-through conversion rates are sorted according to the order of category nodes to generate a user category preference distribution vector, which reflects the user's preference for different category nodes.

[0031] Step S1254: Perform a matrix product operation on the hierarchical-aware embedding vector and the user category preference distribution vector to generate a category node matching weight.

[0032] In this embodiment, the previously generated hierarchical-aware embedding vector and the user category preference distribution vector are subjected to a matrix product operation. Assume that the hierarchical-aware embedding vector is a 3-dimensional vector (corresponding to the three category nodes of "electronic products", "mobile phones" and "smart phones"), and the user category preference distribution vector is also a 3-dimensional vector. Multiply the corresponding elements of these two vectors to obtain the matching weight of each category node. For example, the hierarchical-aware embedding vector element of the "electronic products" category is 0.2, and the user category preference distribution vector element is 0.3, then the matching weight of the category is 0.2×0.3=0.06. In this way, the hierarchical information of the category nodes and the user's preference information are comprehensively considered to generate a more accurate category node matching weight.

[0033] Step S1255: concatenating the hierarchical-aware embedding vector and the category node matching weight to generate a dimension-aligned feature matrix of the advertising content-related feature set.

[0034] In this embodiment, the hierarchical-aware embedding vector and the category node matching weight are spliced. For example, the hierarchical-aware embedding vector is (0.2, 0.3, 0.4), and the category node matching weight is (0.06, 0.08, 0.1). They are spliced ​​together in order to obtain a new vector (0.2, 0.3, 0.4, 0.06, 0.08, 0.1). This vector is part of the dimension-aligned feature matrix of the advertising content-related feature set. All relevant hierarchical-aware embedding vectors and category node matching weights are spliced ​​in this way, and finally a dimension-aligned feature matrix of the advertising content-related feature set is generated. The dimension-aligned feature matrix contains the association information between the advertising content and the product category tree and the user's preference information for the above categories.

[0035] Step S130: calling a preset deep prediction model to perform temporal dependency modeling on the combined behavior feature set to generate an advertising content preference prediction result, wherein the advertising content preference prediction result includes a user's click probability distribution for the candidate advertising content and an interaction intention classification label.

[0036] In this embodiment, after the combined behavior feature set is generated, a preset deep prediction model is called to perform temporal dependency modeling on it to predict the user's advertising content preference. Continuing with the online shopping platform as an example, the combined behavior feature set of user A is input into the deep prediction model.

[0037] Step S131: input the combined behavior feature set into the temporal convolutional network of the deep prediction model to extract local patterns, and generate primary temporal features of each behavior segment unit.

[0038] In this embodiment, the temporal convolutional network of the deep prediction model is used to extract local patterns in the combined behavior feature set. Taking the combined behavior feature set of user A as an example, the combined behavior feature set contains the features of multiple behavior fragment units. The temporal convolutional network processes the above features through convolution operations. For example, the convolution kernel slides on the combined behavior feature set, and convolution calculations are performed on the features of each behavior fragment unit. Assuming that the combined behavior feature set is a multidimensional vector, the convolution kernel size of the temporal convolutional network is 3×3, and the convolution kernel slides on the vector, selecting 3 consecutive feature elements each time for convolution operations. After the convolution operation, the primary temporal features of each behavior fragment unit are generated, and the primary temporal features reflect the local temporal patterns within each behavior fragment unit.

[0039] Step S132: Perform multi-head self-attention mechanism processing on the primary temporal features to capture the long-range dependencies between different behavior fragment units and generate enhanced temporal features.

[0040] In this embodiment, a multi-head self-attention mechanism is performed on the generated primary time series features. The multi-head self-attention mechanism allows the model to pay attention to the long-range dependencies between different behavior fragment units. Taking the primary time series features of user A as an example, the feature contains feature information of multiple behavior fragment units. The multi-head self-attention mechanism divides the primary time series features into multiple heads, and each head independently calculates the attention score. For example, suppose there are 3 heads, each head performs a different linear transformation on the primary time series features, and then calculates the attention score between each behavior fragment unit and other behavior fragment units. The attention score represents the degree of association between different behavior fragment units. The attention score is combined with the primary time series features by weighted summation to obtain enhanced time series features. The enhanced time series features can better reflect the long-range dependencies between different behavior fragment units and improve the model's ability to understand user behavior patterns.

[0041] Step S133: input the enhanced temporal features into the gated recurrent unit network of the deep prediction model for state update to generate a dynamic hidden state sequence of user behavior evolution.

[0042] In this embodiment, the enhanced timing features are input into the gated recurrent unit network of the deep prediction model. The gated recurrent unit network can perform state updates based on the input enhanced timing features to generate a dynamic hidden state sequence of the user behavior evolution. Taking the enhanced timing features of user A as an example, the gated recurrent unit network includes an input gate, a forget gate, and an output gate. The input gate controls the degree to which new input information enters the unit, the forget gate controls the degree to which old information is retained in the unit, and the output gate controls the degree to which the unit outputs information. At each time step, the gated recurrent unit network updates the current hidden state based on the current enhanced timing features and the hidden state of the previous time step through the calculation of the input gate, the forget gate, and the output gate. After calculations for multiple time steps, a dynamic hidden state sequence of the user behavior evolution is generated, which records the changes in the user's behavior state at different time points.

[0043] Step S134: performing maximum pooling processing on the dynamic hidden state sequence, extracting a global behavior representation vector, and performing feature cross calculation on the global behavior representation vector and a feature set associated with the advertisement content to obtain a feature cross result.

[0044] In this embodiment, maximum pooling is performed on the dynamic hidden state sequence to extract the global behavior representation vector.

[0045] Step S1341: Divide the dynamic hidden state sequence into sliding windows along the time dimension to generate multiple local state subsequences.

[0046] In this embodiment, taking the dynamic hidden state sequence of user A as an example, sliding window division is performed along the time dimension. Assume that the size of the sliding window is 3 time steps, and the window slides 1 time step each time. Then the dynamic hidden state sequence will be divided into multiple local state subsequences. For example, the dynamic hidden state sequence contains hidden states of 10 time steps, and 8 local state subsequences will be obtained through sliding window division. Each local state subsequence contains hidden states of 3 consecutive time steps, and the above local state subsequences provide the basis for the subsequent maximum pooling operation.

[0047] Step S1342: Perform element-by-element maximum selection on the hidden state vector in each local state subsequence to generate a local pooling feature.

[0048] In this embodiment, for each local state subsequence, the hidden state vector therein is subjected to element-by-element maximum selection. For example, a local state subsequence contains hidden state vectors of 3 time steps, and each vector is a multidimensional vector. The corresponding elements of these 3 vectors are compared, and the maximum value is selected as the corresponding element of the local pooling feature. Assuming that the dimension of the hidden state vector is 5-dimensional, a 5-dimensional local pooling feature vector is obtained after element-by-element maximum selection. Each local state subsequence performs such an operation to generate multiple local pooling features.

[0049] Step S1343: All local pooling features are connected in series and compressed to a preset dimension through a fully connected layer to generate an intermediate aggregation vector.

[0050] In this embodiment, all the generated local pooling features are connected in series. Assuming that 8 local pooling feature vectors are generated, each with a dimension of 5 dimensions, they are connected in sequence to obtain a 40-dimensional vector, and then the vector is input into the fully connected layer. The function of the fully connected layer is to compress the high-dimensional vector to a preset dimension. For example, the preset dimension is 10 dimensions, and the fully connected layer compresses the 40-dimensional vector into a 10-dimensional intermediate aggregation vector through linear transformation and activation function calculation. The intermediate aggregation vector contains the global information of the dynamic hidden state sequence.

[0051] Step S1344: perform residual connection on the intermediate aggregation vector and the terminal temporal features output by the temporal convolutional network, and perform layer normalization on the residual connection result to generate a final global behavior representation vector for downstream prediction tasks.

[0052] In this embodiment, the intermediate aggregation vector is residually connected to the terminal time series features output by the temporal convolutional network. Residual connection can help the model better learn feature information at different levels. Assuming that the dimension of the intermediate aggregation vector is 10 dimensions, and the dimension of the terminal time series features output by the temporal convolutional network is also 10 dimensions, their corresponding elements are added to obtain a new vector. The new vector is then layer-normalized, and layer normalization can make the element distribution of the vector more stable. After layer normalization, the final global behavior representation vector is generated. The final global behavior representation vector contains the global feature information of the user behavior, which is used for subsequent downstream prediction tasks.

[0053] Next, perform feature cross calculation on the global behavior characterization vector and the feature set associated with the advertising content. Taking the global behavior characterization vector of user A and the feature set associated with the advertising content of the advertisement "AD001" as an example, feature cross calculation can be achieved by multiplying the corresponding elements of the two vectors, or performing other nonlinear transformations. Assuming that the global behavior characterization vector is (0.2, 0.3, 0.4) and the feature set associated with the advertising content is (0.1, 0.2, 0.3), multiply their corresponding elements and obtain the feature cross result of (0.02, 0.06, 0.12). This feature cross result combines the global features of user behavior and the associated features of advertising content, providing richer information for subsequent predictions.

[0054] Step S135: Perform a nonlinear transformation on the feature cross-reflection result through the multi-layer perceptron network of the deep prediction model, and output the advertising content preference prediction result, wherein the click probability distribution is normalized by the Sigmoid function, and the interaction intention classification label is generated by the Softmax function.

[0055] In this embodiment, after the feature crossover result is obtained, it is input into the multilayer perceptron network of the deep prediction model for nonlinear transformation. The multilayer perceptron network consists of multiple fully connected layers, each of which contains multiple neurons, and the neurons are connected by weights. Taking the feature crossover result (0.02, 0.06, 0.12) of user A as an example, it is input into the input layer of the multilayer perceptron network. The neurons in the input layer receive the above-mentioned feature values ​​and pass the results to the next layer by weighted summation and activation function calculation. Assume that there are 3 neurons in the input layer, corresponding to the 3 elements of the feature crossover result, 5 neurons in the middle layer, and 2 neurons in the output layer, corresponding to the click probability and interaction intention classification, respectively. The weighted summation calculation of each neuron can be expressed as multiplying the input value by the corresponding weight, then adding all the products, and adding a bias term. For example, the first neuron in the middle layer receives the outputs of the three neurons in the input layer. Assuming the corresponding weights are 0.1, 0.2, and 0.3, and the bias term is 0.05, the input value of the neuron is 0.02×0.1+0.06×0.2+0.12×0.3+0.05=0.05+0.002+0.012+0.036=0.099. After being processed by the activation function (such as the ReLU function), the neuron outputs a new value. Through the layer-by-layer calculation of the multilayer perceptron network, the original prediction result is finally obtained at the output layer.

[0056] For the click probability distribution, the Sigmoid function is used for normalization. The Sigmoid function can map the output value to the interval [0,1], indicating the probability of click. Assuming that the output value of the neuron corresponding to the click probability in the output layer is 0.8, after the Sigmoid function calculation, the click probability = 1 / (1+e^(-0.8))≈0.69. For the interaction intent classification label, the Softmax function is used to generate it. Assuming that the output values ​​of the neurons corresponding to the interaction intent classification in the output layer are 0.2 and 0.3 respectively, the Softmax function will convert the above values ​​into probability distribution. First, the exponent value is calculated, e^0.2≈1.22, e^0.3≈1.35, and then the probability is calculated. The probability of the first classification is 1.22 / (1.22+1.35)≈0.47, and the probability of the second classification is 1.35 / (1.22+1.35)≈0.53. The classification with the highest probability is the predicted interaction intent classification label. In this way, we obtain the advertising content preference prediction results including the user's click probability distribution for the candidate advertising content and the interaction intention classification labels.

[0057] Step S140: performing dynamic weight matching based on the advertisement content preference prediction result and the advertisement content associated feature set to determine a priority ranking list of candidate advertisement contents.

[0058] In this embodiment, after obtaining the advertising content preference prediction result and the advertising content associated feature set, it is necessary to perform dynamic weight matching to determine the priority ranking list of candidate advertising content. Continuing to take the online shopping platform as an example, the advertising content library of the platform contains a large number of advertisements, from which it is necessary to screen out and sort the advertisements suitable for user A.

[0059] Step S141: Filtering a candidate advertisement content subset that matches the region and terminal type of the target user from the advertisement content library.

[0060] In this embodiment, it is assumed that the region to which user A belongs is a certain city area, and the terminal type used is a mobile phone. The advertising content library stores the geographical scope and terminal type information applicable to each advertisement. Therefore, this embodiment can filter out advertisements that are applicable to the region where user A is located and support mobile terminals from the advertising content library based on the above information. For example, there are 100 advertisements in the advertising content library, of which only 30 are applicable to the region where user A is located, and 20 of these 30 advertisements support mobile terminals, then these 20 advertisements constitute the candidate advertising content subset. The purpose of such screening is to ensure that the recommended advertisements match the actual situation of the user and to improve the relevance and effectiveness of the advertisements.

[0061] Step S142: Calculate the cosine similarity between each candidate advertisement content and the global behavior representation vector to generate a content relevance score.

[0062] In this embodiment, in order to measure the relevance between each candidate advertisement content and the user's global behavior representation vector, the cosine similarity between them is calculated.

[0063] Step S1421: extracting an embedding vector of the candidate advertising content from the advertising content library, wherein the embedding vector is generated by jointly encoding the advertising text and visual features using a pre-trained dual-tower model.

[0064] In this embodiment, for each candidate advertising content, its embedding vector is extracted from the advertising content library. The pre-trained dual-tower model encodes the advertising text and visual features respectively. Taking the candidate advertisement "AD002" as an example, its advertising text is "Fashionable sports shoes, comfortable and breathable", and the visual feature is a picture of sports shoes. One tower of the dual-tower model processes the advertising text and converts the text into a vector representation through operations such as word embedding and convolutional neural networks. The other tower processes the visual features, extracts the features of the picture through a convolutional neural network, and also converts it into a vector representation. The two vectors are then spliced ​​or otherwise fused to generate the embedding vector of the advertisement "AD002". Assume that the embedding vector is (0.1, 0.2, 0.3).

[0065] Step S1422: perform L2 normalization processing on the global behavior representation vector, and calculate the dot product between the normalized global behavior representation vector and each candidate advertisement content embedding vector to generate an original similarity score.

[0066] In this embodiment, the global behavior characterization vector of user A is L2 normalized. The purpose of L2 normalization is to normalize the length of the vector to 1 so that the vector only represents the direction. Assuming that the global behavior characterization vector is (0.2, 0.3, 0.4), its L2 norm = √(0.2^2+0.3^2+0.4^2) = √(0.04+0.09+0.16) = √0.29≈0.54. The normalized global behavior characterization vector is (0.2 / 0.54, 0.3 / 0.54, 0.4 / 0.54)≈(0.37, 0.56, 0.74). Then the dot product between the normalized global behavior representation vector and the embedding vector (0.1, 0.2, 0.3) of the candidate advertisement “AD002” is calculated, and the original similarity score = 0.37×0.1+0.56×0.2+0.74×0.3=0.037+0.112+0.222=0.371.

[0067] Step S1423: Perform piecewise linear transformation on the original similarity score, and map the score interval to the range of [0, 1] to generate a content relevance score.

[0068] In this embodiment, in order to convert the original similarity score into a more intuitive score, a piecewise linear transformation is performed on it. Assuming that the range of the original similarity score is between [-1, 1], this embodiment hopes to map it to the range of [0, 1]. The conversion can be performed by a linear function, for example, content relevance score = (original similarity score + 1) / 2. For the original similarity score of 0.371, the content relevance score = (0.371 + 1) / 2 = 0.6855.

[0069] Step S1424: and downgrading the content relevance score according to the historical exposure frequency of the candidate advertising content.

[0070] In this embodiment, considering that the historical exposure frequency of the candidate advertising content may reduce its attractiveness to users, the content relevance score is downgraded. Assuming that the historical exposure frequency of the candidate advertisement "AD002" is 5 times, according to the preset downgrading rules, the higher the exposure frequency, the greater the downgrading coefficient. Assuming that the downgrading coefficient is 0.8, the downgraded content relevance score = 0.6855 × 0.8 = 0.5484. Through such downgrading, it is possible to avoid over-recommendation of advertisements that have been exposed multiple times and improve the diversity of advertisement recommendations.

[0071] Step S143: Matching the category path hierarchical association features in the advertisement content association feature set according to the interaction intention classification label to generate a category fitness score.

[0072] In this embodiment, the category path hierarchical association features in the advertising content association feature set are matched according to the interaction intention classification label in the advertising content preference prediction result. Assuming that the interaction intention classification label is "buy a mobile phone", for the candidate advertisement "AD001" (mobile phone advertisement), its advertising content association feature set contains category path hierarchical association features, such as "electronic products-mobile phones-smartphones". Then, a category suitability score can be generated based on the degree of match between the interaction intention classification label and the category path hierarchical association features. If the match is high, for example, the interaction intention classification label completely matches a node in the category path, then the category suitability score is high; if the match is low, the score is low. Assuming the match is 80%, the category suitability score can be set to 0.8.

[0073] Step S144: Dynamically weight the click probability distribution, content relevance score and category suitability score to generate a comprehensive recommendation score, wherein the weight coefficient of the dynamic weighted sum is dynamically adjusted according to the click rate in the real-time feedback data.

[0074] In this embodiment, the click probability distribution, content relevance score and category fitness score are dynamically weighted and summed. Assume that the click probability distribution is 0.69, the content relevance score (after de-weighting) is 0.5484, and the category fitness score is 0.8. The initial weight coefficient can be set based on experience. Assume that the weight of the click probability distribution is 0.4, the weight of the content relevance score is 0.3, and the weight of the category fitness score is 0.3. Then the comprehensive recommendation score = 0.69 × 0.4 + 0.5484 × 0.3 + 0.8 × 0.3 = 0.276 + 0.16452 + 0.24 = 0.68052. As the real-time feedback data is updated, the system will dynamically adjust the weight coefficient according to the click rate. For example, if it is found that the click rate is more correlated with the content relevance score, the weight of the content relevance score will be appropriately increased. Assuming that after adjustment, the weight of the click probability distribution becomes 0.3, the weight of the content relevance score becomes 0.4, and the weight of the category fit score becomes 0.3, then the new comprehensive recommendation score = 0.69×0.3+0.5484×0.4+0.8×0.3=0.207+0.21936+0.24=0.66636.

[0075] Step S145: Arrange the candidate advertisement content subsets in descending order according to the comprehensive recommendation scores, generate the priority ranking list and filter the advertisement contents with scores lower than a preset threshold.

[0076] In this embodiment, the comprehensive recommendation score of each advertisement in the candidate advertisement content subset is arranged in descending order. Assume that there are 5 advertisements in the candidate advertisement content subset, and their comprehensive recommendation scores are 0.68052, 0.66636, 0.5, 0.3, and 0.2 respectively. After being arranged in descending order, they are 0.68052, 0.66636, 0.5, 0.3, and 0.2. The preset threshold can be set according to the actual situation. Assuming that the preset threshold is 0.4, then the advertisements with a score lower than 0.4 are filtered out, and the remaining advertisements "AD001" (score 0.68052), "AD002" (score 0.66636), and "AD003" (score 0.5) form a priority sorting list. The advertisements in the priority sorting list are arranged from high to low according to the comprehensive recommendation score, and the higher the score, the more preferentially the advertisement is recommended to the user.

[0077] Step S150: Pushing an adapted advertising content set to target users based on the priority sorting list, and updating parameter configuration of the depth prediction model according to real-time feedback data.

[0078] In this embodiment, after obtaining the priority sorting list, an adapted advertising content set is pushed to the target user, and the parameter configuration of the deep prediction model is updated according to the real-time feedback data to improve the accuracy and effect of the advertising recommendation.

[0079] Step S151: Select the top N candidate advertisement contents from the priority ranking list to generate an initial push set, wherein N is dynamically set according to the screen size of the user terminal.

[0080] In this embodiment, the number of selected advertisements N is dynamically set according to the screen size of the user terminal. Assuming that the screen size of the mobile phone used by user A is small, the system determines N to be 3 according to the preset rules. The top three candidate advertisement contents, namely "AD001" (score 0.68052), "AD002" (score 0.66636), and "AD003" (score 0.5), are selected from the priority sorting list to form the initial push set. The above advertisements are selected based on the comprehensive recommendation score and are most likely to meet the user's preferences and needs.

[0081] Step S152: monitor the real-time interactive behavior of users on the advertising content in the initial push set, record click events, exposure duration and secondary forwarding operations to generate a real-time feedback data set.

[0082] In this embodiment, after pushing the advertisements in the initial push set to user A, the system will monitor the user's real-time interactive behavior. For example, after seeing the advertisement "AD001", user A clicks on the advertisement, stays on the advertisement page for 2 minutes, and shares the advertisement with friends, thereby recording the above interactive behavior, the click event is recorded as "clicked AD001", the exposure time is recorded as 2 minutes, and the secondary forwarding operation is recorded as "shared AD001 with friends". For advertisements "AD002" and "AD003", the user's interactive behavior is also recorded, and the above records are integrated to generate a real-time feedback data set, which contains the user's actual response information to the pushed advertisements.

[0083] Step S153: Incrementally merge the real-time feedback data set and the historical interaction data set, reconstruct the training sample set and re-divide the validation set and the test set.

[0084] In this embodiment, the generated real-time feedback data set is incrementally merged with the previously acquired historical interaction data set. Assuming that the historical interaction data set contains 1000 records and the real-time feedback data set contains 10 records, these 10 records are added to the historical interaction data set to obtain a new data set containing 1010 records, and then the new data set is reconstructed into a training sample set, and the verification set and the test set are re-divided. For example, it can be divided in a ratio of 80%, 10%, and 10%, that is, 808 records are used as the training sample set, 101 records are used as the verification set, and 101 records are used as the test set. In this way, the training sample set is updated with the latest user interaction data, so that the model can learn the user's latest behavior patterns.

[0085] Step S154: Use a sliding window mechanism to perform online parameter fine-tuning on the deep prediction model, wherein the model loss function of the deep prediction model integrates the click-through rate prediction error and the intent classification cross entropy loss.

[0086] In this embodiment, a sliding window mechanism is used to perform online parameter fine-tuning on the depth prediction model.

[0087] Step S1541: setting a sliding window interval of a fixed time window length, and extracting the most recently generated interaction record within the sliding window interval from the real-time feedback data set as the current fine-tuning sample.

[0088] In this embodiment, the length of the sliding window is set to 1 day. The interaction records generated in the last day are extracted from the real-time feedback dataset as the current fine-tuning sample. Assuming that the real-time feedback dataset records the user's interaction behavior within a week, only the interaction records of the last day are extracted as the current fine-tuning sample. If there are 5 interaction records on the last day, then these 5 records constitute the current fine-tuning sample.

[0089] Step S1542: input the current fine-tuning sample into the last two fully connected layers of the depth prediction model, unfreeze the weight parameters of the last two fully connected layers and freeze the parameters of the remaining layers.

[0090] In this embodiment, the current fine-tuning sample is input into the last two fully connected layers of the depth prediction model. When fine-tuning, the weight parameters of the last two fully connected layers are unfrozen so that they can be updated according to the current fine-tuning sample, while the parameters of the remaining layers are frozen to avoid excessive impact on the overall structure and parameters of the model. For example, the depth prediction model includes an input layer, an intermediate layer, and an output layer. The intermediate layer has multiple fully connected layers. Only the weight parameters of the last two fully connected layers are unfrozen, and the parameters of the remaining layers remain unchanged.

[0091] Step S1543: Calculate the gradient direction according to the click rate prediction error and the intent classification cross entropy loss of the current fine-tuning sample, and use an adaptive momentum optimization algorithm to update the weight parameters of the last two fully connected layers along the gradient direction.

[0092] In this embodiment, the gradient direction is calculated based on the click-through rate prediction error of the current fine-tuning sample and the intent classification cross entropy loss. The click-through rate prediction error can be obtained by calculating the difference between the predicted click probability and the actual click situation. For example, if the predicted click probability is 0.6, but the actual user does not click, then the click-through rate prediction error is large. The intent classification cross entropy loss can measure the difference between the predicted interaction intent classification label and the actual interaction intent. The gradient direction is calculated by the above loss function, and the gradient direction indicates the direction in which the weight parameters need to be updated. An adaptive momentum optimization algorithm (such as the Adam algorithm) is used to update the weight parameters of the last two fully connected layers along the gradient direction. The Adam algorithm can dynamically adjust the learning rate according to the first-order moment estimation and second-order moment estimation of the gradient, so that the update of the weight parameters is more stable and efficient.

[0093] Step S1544: performing exponential moving average processing on the updated weight parameters of the last two fully connected layers and the corresponding historical weight parameters before updating to generate smoothed new weight parameters.

[0094] In this embodiment, in order to make the update of weight parameters smoother, the weight parameters of the last two fully connected layers after update are subjected to exponential moving average processing with the corresponding historical weight parameters before update. Assume that the historical weight parameter before update is W_old, the weight parameter after update is W_new, and the attenuation coefficient of exponential moving average is 0.9. Then the smoothed new weight parameter W_smooth=0.9×W_old+0.1×W_new. In this way, the stability of historical weight parameters is taken into account, and new update information is introduced, making the update of weight parameters more reasonable.

[0095] Step S1545: Load the smoothed new weight parameters into the online instance of the deep prediction model, and add the current fine-tuning sample to the validation set to calculate the click-through rate prediction accuracy.

[0096] In this embodiment, the smoothed new weight parameters are loaded into the online instance of the deep prediction model. Then the current fine-tuning sample is added to the validation set, and the click-through rate prediction accuracy is calculated using the validation set. For example, there are 100 records in the validation set, and the model predicts that users will click on 60 of them, but users actually click on 55 records, then the click-through rate prediction accuracy = 55 / 60≈0.917 (keep three decimal places). By calculating the click-through rate prediction accuracy, the performance of the model on new data can be evaluated to understand whether the prediction effect of the model has been improved.

[0097] Step S1546: When the click rate prediction accuracy is higher than the preset threshold of the current online version, a hot update operation is triggered to synchronize the smoothed new weight parameters to the online advertising push service.

[0098] In this embodiment, a threshold value of the click-through rate prediction accuracy is preset, for example, set to 0.9. The click-through rate prediction accuracy currently calculated is compared with the threshold value and the accuracy of the current online version. If the current click-through rate prediction accuracy is higher than the preset threshold value and better than the accuracy of the current online version, it means that the online parameter fine-tuning through the sliding window mechanism has achieved good results and the performance of the model has been improved. At this time, the hot update operation is triggered to synchronize the smoothed new weight parameters to the online advertising push service. In this way, the online advertising recommendation system can use the updated model parameters to recommend advertisements, thereby providing users with advertising content that is more in line with their preferences, improving the click-through rate of advertisements and the user's interactive experience. For example, after updating the weight parameters, the online advertising push service may make the originally less accurate advertising recommendations more in line with user needs, and users are more likely to click on the recommended ads, thereby improving the conversion rate of advertisements and the revenue of the platform.

[0099] Step S155: When the average click rate fluctuation in the real-time feedback data set exceeds a preset tolerance threshold, the model structure optimization operation is triggered to adjust the convolution kernel size of the temporal convolutional network of the deep prediction model and the hidden layer dimension of the gated recurrent unit network.

[0100] In this embodiment, the fluctuation of the average click rate in the real-time feedback data set is continuously monitored. A tolerance threshold is preset, for example, set to 10%. By calculating the average click rate of different time periods in the real-time feedback data set, and comparing the change in the average click rate of adjacent time periods. Assuming that the average click rate of the previous time period is 0.15, and the average click rate of the current time period is 0.12, the click rate has dropped by (0.15-0.12) / 0.15=0.2, or 20%, which exceeds the preset tolerance threshold of 10%. This indicates that the performance of the model may have fluctuated greatly, and the original model structure may not be able to adapt well to the new user behavior pattern.

[0101] At this time, the model structure optimization operation is triggered. First, adjust the convolution kernel size of the temporal convolutional network of the deep prediction model. The convolution kernel size affects the model's ability to extract local features. If the current convolution kernel size is small, it may not be able to capture time series features that are long enough; if the convolution kernel size is too large, it may result in excessive computation and easy overfitting. For example, the original convolution kernel size of the temporal convolutional network is 3×3. After analysis and experiments, you can try to adjust it to 5×5 to enhance the model's ability to capture longer time series features.

[0102] At the same time, the hidden layer dimension of the gated recurrent unit network is adjusted. The hidden layer dimension determines the complexity of information that the gated recurrent unit network can learn and represent. If the hidden layer dimension is too low, the model may not be able to learn sufficiently complex user behavior evolution patterns; if the hidden layer dimension is too high, it may lead to overfitting and waste of computing resources. Assuming that the hidden layer dimension of the original gated recurrent unit network is 64, after evaluation and experiments, it can be adjusted to 128, so that the deep prediction model can learn more complex changes in user behavior states. Through such model structure optimization operations, the deep prediction model can better adapt to new user data and behavior patterns, and improve the accuracy and stability of advertising content preference prediction.

[0103] Based on the above steps, by obtaining the historical interaction data set of the target user, cross-dimensional feature fusion processing is performed, and the deep prediction model is used to perform temporal dependency modeling to generate the advertising content preference prediction results, and then the priority ranking list of candidate advertising content is determined through dynamic weight matching. Finally, the adapted advertising content set is pushed to the target user, and the parameters and structure of the deep prediction model are updated based on real-time feedback data. The whole process forms a closed-loop advertising recommendation system that can continuously adjust the recommendation strategy based on the user's latest behavior data, improve the accuracy and effectiveness of advertising recommendations, provide users with more personalized advertising content that better meets their needs, and also bring higher conversion rates and revenue to advertisers.

[0104] Figure 2 The hardware structure of the advertising service system 100 provided in the embodiment of the present application for implementing the above-mentioned intelligent advertising content recommendation method using the application prediction model is shown. Figure 2 As shown, the advertisement serving system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0105] In one possible design, the advertising service system 100 may be a single server or a server group. The server group may be centralized or distributed (for example, the advertising service system 100 may be a distributed system). In some embodiments, the advertising service system 100 may be local or remote. For example, the advertising service system 100 may access information and / or data stored in a machine-readable storage medium 120 via a network. For another example, the advertising service system 100 may be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the advertising service system 100 may be implemented on an advertising service system. By way of example only, the advertising service system may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, or the like, or any aggregation thereof.

[0106] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions used by the advertising service system 100 to execute or use to complete the exemplary methods described in this application.

[0107] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the intelligent advertising content recommendation method based on the application prediction model of the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected via the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0108] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned advertising service system 100. The implementation principles and technical effects are similar, and will not be repeated in this embodiment.

[0109] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are set. When a processor runs the computer executable instructions, the intelligent advertising content recommendation method using the above application prediction model is implemented.

[0110] It should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof. Similarly, it should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof.

Claims

1. An intelligent advertising content recommendation method using a prediction model, characterized in that: The method comprises: Acquire a historical interaction data set of a target user in an advertisement interaction scenario, wherein the historical interaction data set includes a plurality of advertisement interaction sequences, each advertisement interaction sequence consisting of a user identifier, an advertisement content identifier, an interaction behavior type, and an interaction timestamp; Performing cross-dimensional feature fusion processing on the historical interaction data set to generate a combined behavior feature set and an advertisement content-related feature set for each advertisement interaction sequence; Calling a preset deep prediction model to perform time-series dependency modeling on the combined behavior feature set to generate an advertising content preference prediction result, wherein the advertising content preference prediction result includes a user's click probability distribution for the candidate advertising content and an interaction intention classification label; Performing dynamic weight matching based on the advertising content preference prediction result and the advertising content associated feature set to determine a priority ranking list of candidate advertising contents; Based on the priority sorting list, a set of adapted advertising content is pushed to target users, and parameter configuration of the deep prediction model is updated according to real-time feedback data.

2. The intelligent advertising content recommendation method using a prediction model according to claim 1, characterized in that: The performing cross-dimensional feature fusion processing on the historical interaction data set to generate a combined behavior feature set and an advertisement content-related feature set for each advertisement interaction sequence includes: Performing time window segmentation processing on the interactive behavior types in the advertisement interactive sequence to generate a plurality of behavior segment units, each behavior segment unit including continuous interactive behaviors within a preset time length; Extracting the behavior density feature and interaction strength feature of each behavior segment unit, wherein the behavior density feature is determined by the interaction frequency per unit time, and the interaction strength feature is determined by the weighted value of the interaction duration and the operation depth; Performing semantic analysis on the advertisement material corresponding to the advertisement content identifier to extract semantic topic distribution features of the advertisement text and attention weight features of the visual elements; The behavior density feature, interaction intensity feature, semantic topic distribution feature and attention weight feature corresponding to the same advertisement interaction sequence are concatenated to generate the combined behavior feature set; The advertisement knowledge base is called to match the product category tree corresponding to the advertisement content identifier, the hierarchical association features of the category path and the user history preference matching degree are extracted, and the advertisement content association feature set is generated.

3. The intelligent advertising content recommendation method using a prediction model according to claim 1, characterized in that: The calling of a preset deep prediction model to perform temporal dependency modeling on the combined behavior feature set to generate an advertisement content preference prediction result includes: Inputting the combined behavior feature set into the temporal convolutional network of the deep prediction model to extract local patterns and generate primary temporal features of each behavior segment unit; Performing a multi-head self-attention mechanism on the primary temporal features to capture the long-range dependencies between different behavior fragment units and generate enhanced temporal features; Inputting the enhanced time series features into the gated recurrent unit network of the deep prediction model for state update to generate a dynamic hidden state sequence of user behavior evolution; Performing maximum pooling processing on the dynamic hidden state sequence to extract a global behavior representation vector, and performing feature cross calculation on the global behavior representation vector and a feature set associated with the advertisement content to obtain a feature cross result; The feature cross-reaction result is nonlinearly transformed through the multi-layer perceptron network of the deep prediction model to output the advertising content preference prediction result, wherein the click probability distribution is normalized by the Sigmoid function and the interaction intention classification label is generated by the Softmax function.

4. The intelligent advertising content recommendation method using a prediction model according to claim 3, characterized in that: The step of performing dynamic weight matching based on the advertisement content preference prediction result and the advertisement content associated feature set to determine a priority ranking list of candidate advertisement contents includes: Filtering a subset of candidate advertising content that matches the region and terminal type of the target user from the advertising content library; Calculating the cosine similarity between each candidate advertisement content and the global behavior representation vector to generate a content relevance score; Matching the category path hierarchical association features in the advertisement content association feature set according to the interaction intention classification label to generate a category fitness score; Dynamically weighting and summing the click probability distribution, content relevance score and category suitability score to generate a comprehensive recommendation score, wherein the weight coefficient of the dynamic weighted summation is dynamically adjusted according to the click rate in the real-time feedback data; The candidate advertisement content subsets are arranged in descending order according to the comprehensive recommendation scores, the priority ranking list is generated, and advertisement contents with scores lower than a preset threshold are filtered out.

5. The intelligent advertising content recommendation method using a prediction model according to claim 1, characterized in that: The step of pushing an adapted advertising content set to a target user based on the priority sorting list and updating parameter configuration of the deep prediction model according to real-time feedback data includes: Selecting the top N candidate advertisement contents from the priority ranking list to generate an initial push set, wherein N is dynamically set according to the screen size of the user terminal; Monitor users' real-time interactive behaviors on the advertising content in the initial push set, record click events, exposure duration, and secondary forwarding operations to generate a real-time feedback data set; Incrementally merge the real-time feedback data set with the historical interaction data set, reconstruct the training sample set, and re-divide the validation set and test set; A sliding window mechanism is used to fine-tune the parameters of the deep prediction model online, wherein the model loss function of the deep prediction model integrates the click-through rate prediction error and the intent classification cross entropy loss; When the average click rate fluctuation in the real-time feedback data set exceeds a preset tolerance threshold, a model structure optimization operation is triggered to adjust the convolution kernel size of the temporal convolutional network and the hidden layer dimension of the gated recurrent unit network of the deep prediction model.

6. The intelligent advertising content recommendation method using a prediction model according to claim 5, characterized in that: The online parameter fine-tuning of the depth prediction model using a sliding window mechanism includes: Setting a sliding window interval with a fixed time window length, and extracting the latest interaction record generated within the sliding window interval from the real-time feedback data set as a current fine-tuning sample; Input the current fine-tuning sample into the last two fully connected layers of the depth prediction model, unfreeze the weight parameters of the last two fully connected layers and freeze the parameters of the remaining layers; Calculating a gradient direction according to the click rate prediction error and the intent classification cross entropy loss of the current fine-tuning sample, and updating the weight parameters of the last two fully connected layers along the gradient direction using an adaptive momentum optimization algorithm; Performing exponential moving average processing on the updated weight parameters of the last two fully connected layers and the corresponding historical weight parameters before updating to generate smoothed new weight parameters; Loading the smoothed new weight parameters into the online instance of the deep prediction model, and adding the current fine-tuning sample into the validation set to calculate the click-through rate prediction accuracy; When the click rate prediction accuracy is higher than a preset threshold of the current online version, a hot update operation is triggered to synchronize the smoothed new weight parameters to the online advertising push service.

7. The intelligent advertising content recommendation method using a prediction model according to claim 4, characterized in that: The step of calculating the cosine similarity between each candidate advertisement content and the global behavior representation vector to generate a content relevance score includes: Extracting an embedding vector of a candidate advertisement content from an advertisement content library, wherein the embedding vector is generated by jointly encoding advertisement text and visual features through a pre-trained dual-tower model; Perform L2 normalization on the global behavior representation vector, and calculate the dot product between the normalized global behavior representation vector and each candidate ad content embedding vector to generate the original similarity score; Perform piecewise linear transformation on the original similarity scores and map the score interval to the range of [0, 1] to generate content relevance scores; In addition, the content relevance score is downgraded according to the historical exposure frequency of the candidate advertising content.

8. The intelligent advertising content recommendation method using a prediction model according to claim 3, characterized in that: The performing maximum pooling processing on the dynamic hidden state sequence to extract a global behavior representation vector includes: Divide the dynamic hidden state sequence into sliding windows along the time dimension to generate multiple local state subsequences; Perform element-by-element maximum selection on the hidden state vector in each local state subsequence to generate local pooling features; All local pooling features are concatenated and compressed to the preset dimension through a fully connected layer to generate an intermediate aggregation vector; The intermediate aggregation vector is residually connected with the terminal temporal features output by the temporal convolutional network, and the residual connection result is layer-normalized to generate the final global behavior representation vector for downstream prediction tasks.

9. The intelligent advertising content recommendation method using a prediction model according to claim 2, characterized in that: The step of extracting the hierarchical association features of the category path and the matching degree of the user's historical preferences to generate the advertisement content association feature set includes: Parsing the parent-child node relationship of the product category tree to generate category graph structure data including a hierarchical connection relationship; Inputting the category graph structure data into a graph neural network, and generating a hierarchical perception embedding vector for each category node through feature propagation of adjacent nodes; Counting the historical exposure times and click-through conversion rate of each category node in the target user's historical interaction data set to generate a user category preference distribution vector; Performing a matrix product operation on the hierarchical perception embedding vector and the user category preference distribution vector to generate a category node matching weight; The hierarchical-aware embedding vector and the category node matching weight are concatenated to generate a dimension-aligned feature matrix of the advertising content-related feature set.

10. An advertising service system, characterized in that: The advertising service system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to run the programs, instructions or codes in the memory to implement the intelligent advertising content recommendation method using the application prediction model described in any one of claims 1 to 9.

Images