An accurate placement method and system for Internet advertising data based on artificial intelligence

Abstract

The present invention relates to the technical field of advertising delivery, and specifically provides an accurate delivery method and system for Internet advertising data based on artificial intelligence, including: obtaining the behavior sequence of a target user and the advertising data sequence watched; dividing the behavior sequence based on the viewing duration of the target user for different advertising data to obtain three sub-behavior sequences corresponding to the behavior sequence. By dividing the behavior sequence based on the viewing duration of the user for different advertisements, the present invention can more accurately identify the interest changes and behavior patterns of the user, so as to provide personalized advertising delivery for each user, making the advertising delivery more targeted rather than simply unified delivery. Moreover, by predicting the behavior sequence at the next moment based on the historical behavior data of the user, accurate advertising delivery preparation can be made before the user's behavior occurs, avoiding the display of outdated or irrelevant advertisements.

Description

Technical Field

[0001] The present invention relates to the technical field of advertising delivery, and particularly relates to an accurate delivery method and system for Internet advertising data based on artificial intelligence. Background Art

[0002] Artificial intelligence (AI) refers to technologies and methods that simulate and expand human intelligence, enabling machines to perform intelligent tasks such as perception, reasoning, learning, and decision-making through algorithms, computer models, machine learning, etc. Internet advertising data refers to data collected regarding aspects such as ad display, user interaction, and advertising effectiveness when advertising is delivered through Internet channels.

[0003] Traditional methods usually adopt a unified advertising recommendation strategy, lacking personalization and being unable to accurately identify changes in user interests, resulting in poor advertising delivery effects. Secondly, traditional methods are difficult to predict and adjust user behavior in real time, often showing ads that do not match the user's current interests, increasing the user's advertising fatigue. Moreover, traditional methods often rely on manual intervention to optimize advertising delivery, with low efficiency and a lack of automation, being unable to quickly adapt to changes in user behavior, resulting in poor timeliness and accuracy of advertising delivery effects, thereby affecting the click-through rate and conversion rate of ads. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the above-mentioned drawbacks of the prior art and provide an accurate delivery method and system for Internet advertising data based on artificial intelligence.

[0005] The technical solution adopted to solve the above technical problem is: An accurate delivery method for Internet advertising data based on artificial intelligence, comprising:

[0006] Obtain the behavior sequence of the target user and the sequence of viewed advertising data;

[0007] Based on the viewing duration of the target user for different advertising data, divide the behavior sequence to obtain three sub-behavior sequences corresponding to the behavior sequence;

[0008] Based on a pre-trained behavior prediction model, perform behavior prediction through the three sub-behavior sequences to obtain the behavior sequence at the next preset moment;

[0009] Based on the behavior sequence at the next preset moment, construct an interactive bipartite weighted graph, and based on a pre-trained interactive score prediction model, perform interactive score prediction through the interactive bipartite weighted graph to obtain a list of predicted interactive scores corresponding to the behavior sequence at the next preset moment;

[0010] Generate an advertising data delivery list based on the list of predicted interactive scores corresponding to the behavior sequence at the next preset moment.

[0011] Preferably, the behavior sequence includes the interaction information of the target user with different advertisement data, where the expression of the behavior sequence is as follows:

[0012] ;

[0013] Wherein, represents the interaction information of the target user with different advertisement data, represents the th advertisement data ID, represents the advertisement feature of the th advertisement data, represents the number of advertisement data viewed by the target user.

[0014] Preferably, the behavior sequence is divided based on the viewing duration of the target user for different advertisement data to obtain three sub-behavior sequences corresponding to the behavior sequence, including:

[0015] Dividing different advertisement data into multiple advertisement groups based on the advertisement duration;

[0016] Comparing the number of samples in the multiple advertisement groups with a preset sample number threshold. If the number of samples in the multiple advertisement groups is higher than the preset sample number threshold, the advertisement group is determined as a label with sufficient sample size, otherwise, the advertisement group is determined as a label with insufficient sample size;

[0017] Obtaining the viewing duration of multiple advertisement data in the advertisement group with the label of sufficient sample size to obtain the viewing duration sequence corresponding to the advertisement group with the label of sufficient sample size;

[0018] Sorting the viewing duration sequence corresponding to the advertisement group with the label of sufficient sample size to obtain the viewing duration order sequence corresponding to the advertisement group with the label of sufficient sample size;

[0019] Dividing the advertisement group with the label of sufficient sample size into three sub-advertisement groups based on the viewing duration order sequence corresponding to the advertisement group with the label of sufficient sample size.

[0020] Preferably, dividing the behavior sequence based on the viewing duration of the target user for different advertisement data to obtain three sub-behavior sequences corresponding to the behavior sequence further includes:

[0021] Obtaining the viewing duration of multiple advertisement data in the advertisement group with the label of insufficient sample size to obtain the viewing duration sequence corresponding to the advertisement group with the label of insufficient sample size;

[0022] Match the viewing duration sequence corresponding to the advertisement group with insufficient sample size tags and the viewing duration sequence corresponding to the advertisement group with sufficient sample size tags based on the K-nearest neighbor algorithm to find the closest neighbor group, where the neighbor group belongs to the advertisement group with sufficient sample size tags;

[0023] Divide the advertisement group with insufficient sample size tags into three sub-advertisement groups based on the neighbor group;

[0024] Divide the behavior sequence based on the three sub-advertisement groups corresponding to the advertisement groups with insufficient sample size tags and sufficient sample size tags to obtain three sub-behavior sequences corresponding to the behavior sequence, where the three sub-behavior sequences include a non-viewing sub-behavior sequence, an interested sub-behavior sequence, and a non-interested sub-behavior sequence.

[0025] Preferably, the behavior prediction model includes an advertisement embedding module, a feature extraction module, and a behavior prediction module. The advertisement embedding module is used to convert the behavior sequence of the target user and the viewed advertisement data sequence into embedding vectors. The feature extraction module is used to extract the fine-grained interest expression of the target user from the three sub-behavior sequences based on the embedding vectors. The behavior prediction module is used to predict the behavior sequence at the next preset moment based on the fine-grained interest expression of the target user.

[0026] Preferably, the advertisement embedding module is used to embed the advertisement data ID, advertisement data author ID, advertisement data feedback tag, and location information to obtain an advertisement data ID embedding vector, an advertisement data author ID embedding vector, an advertisement data feedback tag embedding vector, and a location information embedding vector. The advertisement data ID embedding vector, advertisement data author ID embedding vector, advertisement data feedback tag embedding vector, and location information embedding vector are fused to obtain the embedding vector of the advertisement data. Based on the embedding vector of the advertisement data, the embedding vectors corresponding to the three sub-behavior sequences are obtained. The feature extraction module uses a Transformer network to input the embedding vectors corresponding to the three sub-behavior sequences into different Transformer networks to obtain the feature representations corresponding to the three sub-behavior sequences. The behavior prediction module uses a long short-term memory network to input the feature representations corresponding to the three sub-behavior sequences into different long short-term memory networks to obtain the behavior sequence at the next preset moment.

[0027] Preferably, the interactive bipartite weighted graph includes a user node set, an advertising data node set, a connection edge set, and an edge weight set. The user nodes correspond to target users, the advertising data nodes correspond to advertising data, the connection edges are used to connect the user nodes and the advertising data nodes. When a target user views advertising data, there is a connection edge between the corresponding target user node and the advertising data node. The edge weight of the connection edge is defined based on the viewing duration of the target user viewing the advertising data. When the viewing duration is higher than a preset viewing time threshold, the edge weight of the connection edge is defined as 1; otherwise, the edge weight of the connection edge is defined as 0.

[0028] Preferably, the interactive score prediction model includes an advertising feature extraction module, a feature combination module, and a score prediction module. The advertising feature extraction module is used to extract the text feature vector, visual feature vector, and audio feature vector of the advertising data. The feature combination module is used to splice the text feature vector, visual feature vector, and audio feature vector of the advertising data pairwise to obtain three cross features, and perform pairwise Hadamard product operations on the three cross features respectively to obtain three combined features of the advertising data. The three combined features of the advertising data are used as the row vectors of the advertising data nodes in the interactive bipartite weighted graph. The score prediction module uses a multi-layer perceptron and calculates the sum of the edge weights of the interactive bipartite weighted graph based on the interactive bipartite weighted graph and the row vectors of the advertising data nodes in the interactive bipartite weighted graph. The sum of the edge weights of the interactive bipartite weighted graph is the interactive score of the interactive bipartite weighted graph.

[0029] Preferably, the multi-layer perceptron is composed of four fully connected layers.

[0030] The technical solution adopted to solve the above technical problems is: An Internet advertising data precise placement system based on artificial intelligence, which is applicable to the above-mentioned Internet advertising data precise placement method based on artificial intelligence, includes:

[0031] A data collection unit, which is used to obtain the behavior sequence of the target user and the viewed advertising data sequence;

[0032] A behavior division unit, which is used to divide the behavior sequence based on the viewing duration of the target user for different advertising data to obtain three sub-behavior sequences corresponding to the behavior sequence;

[0033] A behavior prediction unit, which is used to perform behavior prediction through the three sub-behavior sequences based on a pre-trained behavior prediction model to obtain the behavior sequence at the next preset moment;

[0034] A score prediction unit, which is used to construct an interactive bipartite weighted graph based on the behavior sequence at the next preset moment, and predict the interactive score through the interactive bipartite weighted graph based on a pre-trained interactive score prediction model, so as to obtain a list of predicted interactive scores corresponding to the behavior sequence at the next preset moment;

[0035] An advertisement placement unit, which is used to generate an advertisement data placement list based on the list of predicted interactive scores corresponding to the behavior sequence at the next preset moment.

[0036] The beneficial effects of the present invention are as follows: (1) By dividing the behavior sequence based on the viewing duration of different advertisements by the user, the present invention can more accurately identify the interest changes and behavior patterns of the user, so as to provide personalized advertisement placement for each user. This makes the advertisement placement more targeted instead of simple unified placement, and by predicting the behavior sequence at the next moment based on the user's behavior history data, it is possible to prepare for the precise placement of advertisements before the user's behavior occurs, avoiding the display of outdated or irrelevant advertisements; (2) By constructing an interactive bipartite weighted graph and based on the interaction between the user and the advertisement, the present invention can better capture the interaction relationship between the user and the advertisement. This method can effectively judge which advertisements are more attractive to a specific user, thus avoiding the display of ineffective advertisements and improving the relevance of the advertisements. And by predicting the interactive score of the advertisement through the interactive score prediction model, the placement strategy of the advertisement can be dynamically adjusted to ensure that the advertisement seen by the user best matches their current needs and interests, enhancing the acceptance and experience of the user; (3) By predicting based on the behavior sequence and calculating the interactive score, the present invention automatically generates an advertisement data placement list, which can greatly improve the efficiency of advertisement placement, reduce manual intervention, improve the timeliness and accuracy of placement, and based on behavior prediction and interaction prediction, it can adapt to the changes in user behavior in real time. When the user's interest or behavior pattern changes, the advertisement system can automatically adjust to avoid the display of outdated advertisements and improve the advertisement conversion rate. Description of the Drawings

[0037] Figure 1 It is a schematic flow chart of the steps of the overall method in an embodiment proposed by the present invention;

[0038] Figure 2 It is a schematic system architecture diagram of the overall system in an embodiment proposed by the present invention.

[0039] Reference Signs: 1, data acquisition unit; 2, behavior division unit; 3, behavior prediction unit; 4, score prediction unit; 5, advertisement placement unit. Detailed Embodiments

[0040] Embodiment 1, as Figure 1As shown in the figure, a method for precise placement of Internet advertising data based on artificial intelligence proposed by the present invention includes:

[0041] S1. Obtain the behavior sequence of the target user and the advertising data sequence viewed;

[0042] S2. Divide the behavior sequence based on the viewing duration of the target user for different advertising data to obtain three sub-behavior sequences corresponding to the behavior sequence;

[0043] S3. Perform behavior prediction through the three sub-behavior sequences based on a pre-trained behavior prediction model to obtain the behavior sequence at the next preset moment;

[0044] S4. Construct an interactive bipartite weighted graph based on the behavior sequence at the next preset moment, and perform interactive score prediction through the interactive bipartite weighted graph based on a pre-trained interactive score prediction model to obtain a list of predicted interactive scores corresponding to the behavior sequence at the next preset moment;

[0045] S5. Generate an advertising data placement list based on the list of predicted interactive scores corresponding to the behavior sequence at the next preset moment.

[0046] In the present invention, the behavior sequence division analyzes the viewing duration and interaction of the user for different advertisements, and splits the behavior sequence of the user into multiple sub-sequences, representing the user's behavioral responses to different advertisements; the generation of the advertising placement list sorts the advertisements according to the prediction results of the interactive scores to generate an advertising list suitable for placement. Usually, the advertisements are sorted by priority according to the predicted scores. The advertisement features include but are not limited to advertisement type, advertisement content category, advertisement duration, advertisement display position, the viewing duration of the advertisement by the user, advertisement data ID, advertisement data author ID, advertisement data feedback label, and location information, etc.

[0047] Embodiment 2, a method for precise placement of Internet advertising data based on artificial intelligence proposed by the present invention. Compared with Embodiment 1, this embodiment further includes: The behavior sequence includes the interaction information of the target user for different advertising data, where the expression of the behavior sequence is as follows:

[0048] ;

[0049] Among them, represents the interaction information of the target user for different advertising data, represents the th advertisement data ID, represents the advertisement feature of the th advertisement data, represents the number of advertisement data viewed by the target user.

[0050] In this embodiment, the advertisement features of the advertisement data include text features, visual features, and sound features.

[0051] In an alternative embodiment, the behavior sequence is divided based on the viewing duration of the target user for different advertisement data to obtain three sub-behavior sequences corresponding to the behavior sequence, including:

[0052] A1. Divide different advertisement data into multiple advertisement groups based on the advertisement duration;

[0053] A2. Compare the number of samples in multiple advertisement groups with a preset sample number threshold. If the number of samples in multiple advertisement groups is higher than the preset sample number threshold, then determine the advertisement group as a label with sufficient sample size; otherwise, determine the advertisement group as a label with insufficient sample size;

[0054] A3. Obtain the viewing durations of multiple advertisement data in the advertisement groups with the label of sufficient sample size to obtain a viewing duration sequence corresponding to the advertisement groups with the label of sufficient sample size;

[0055] A4. Sort the viewing duration sequence corresponding to the advertisement groups with the label of sufficient sample size to obtain a viewing duration order sequence corresponding to the advertisement groups with the label of sufficient sample size;

[0056] A5. Divide the advertisement groups with the label of sufficient sample size into three sub-advertisement groups based on the viewing duration order sequence corresponding to the advertisement groups with the label of sufficient sample size.

[0057] It should be noted that the advertisement duration refers to the time length of the advertisement display, usually the time span from the start to the end of the advertisement, and this time information is calculated based on the advertisement exposure time; the advertisement group refers to classifying advertisement data into different groups according to the advertisement duration; the viewing duration sequence represents the sequence of the viewing durations of all users for the advertisement within a certain advertisement group.

[0058] In an alternative embodiment, the behavior sequence is divided based on the viewing duration of the target user for different advertisement data to obtain three sub-behavior sequences, and it further includes:

[0059] A6. Obtain the viewing durations of multiple advertisement data in the advertisement groups with the label of insufficient sample size to obtain a viewing duration sequence corresponding to the advertisement groups with the label of insufficient sample size;

[0060] A7. Match the viewing duration sequence corresponding to the advertisement groups with the label of insufficient sample size with the viewing duration sequence corresponding to the advertisement groups with the label of sufficient sample size based on the K-nearest neighbor algorithm to find the closest neighbor group, where the neighbor group belongs to the advertisement groups with the label of sufficient sample size;

[0061] A8. Group the advertisements with insufficient sample size labels into three sub - advertisement groups based on neighbor groups;

[0062] A9. Divide the behavior sequence based on the three sub - advertisement groups corresponding to the advertisement groups with insufficient sample size labels and sufficient sample size labels to obtain three sub - behavior sequences corresponding to the behavior sequence, where the three sub - behavior sequences include a non - viewing sub - behavior sequence, an interested sub - behavior sequence, and an uninterested sub - behavior sequence.

[0063] It should be noted that the K - Nearest Neighbor (K - NN) algorithm is a commonly used supervised learning algorithm mainly for classification and regression problems. K - NN is used to find the advertisement group most similar to the advertisement group with insufficient sample size labels by comparing the viewing duration sequences of different advertisement groups. Specifically, K - NN will judge which advertisement groups have similar behavior patterns according to the similarity of the viewing duration sequences (such as Euclidean distance or cosine similarity, etc.) and regard these groups as "neighbors"; Neighbor Groups refer to the advertisement groups found by the K - Nearest Neighbor algorithm that are most similar to the advertisement group with insufficient sample size labels. Neighbor groups usually come from the advertisement groups with sufficient sample size labels because the data of these groups is more sufficient and the analysis results are more reliable.

[0064] In an optional embodiment, the behavior prediction model includes an advertisement embedding module, a feature extraction module, and a behavior prediction module. The advertisement embedding module is used to convert the behavior sequence of the target user and the viewed advertisement data sequence into embedding vectors. The feature extraction module is used to extract the fine - grained interest expression of the target user from the three sub - behavior sequences based on the embedding vectors. The behavior prediction module is used to predict the behavior sequence at the next preset moment based on the fine - grained interest expression of the target user.

[0065] In an optional embodiment, the advertisement embedding module is used to perform embedding on the advertisement data ID, advertisement data author ID, advertisement data feedback label, and location information to obtain an advertisement data ID embedding vector, an advertisement data author ID embedding vector, an advertisement data feedback label embedding vector, and a location information embedding vector. The advertisement data ID embedding vector, the advertisement data author ID embedding vector, the advertisement data feedback label embedding vector, and the location information embedding vector are fused to obtain the embedding vector of the advertisement data. Based on the embedding vector of the advertisement data, the embedding vectors corresponding to the three sub - behavior sequences are obtained. The feature extraction module uses a Transformer network and inputs the embedding vectors corresponding to the three sub - behavior sequences into different Transformer networks to obtain the feature representations corresponding to the three sub - behavior sequences. The behavior prediction module uses a long short - term memory network and inputs the feature representations corresponding to the three sub - behavior sequences into different long short - term memory networks to obtain the behavior sequence at the next preset moment.

[0066] It should be noted that position information embedding refers to converting the relevant information of the advertisement display position (such as the display position of the advertisement on the page, the geographical location targeted by the advertisement, etc.) into an embedded vector, and these positions may affect the click-through rate of the advertisement and the way users interact. By mapping the position information to a low-dimensional vector space, the model can learn the impact of different positions on the advertisement effect; The Transformer network is a deep learning architecture based on the self-attention mechanism. Compared with the traditional recurrent neural network (RNN), the Transformer can process input data in parallel and more effectively capture long-distance dependencies. It learns the correlations in the input sequence through multiple layers of self-attention mechanisms and generates a weighted representation of each input element. In the feature extraction module, the Transformer network is used to extract feature representations from three sub-behavior sequences; The long short-term memory network (LSTM) is a special type of recurrent neural network (RNN) specifically designed to address the problems of vanishing and exploding gradients that traditional RNNs may encounter when processing long sequences. The LSTM retains long-term dependency information by introducing memory cells, making it perform excellently in sequence prediction tasks; There are various ways to fuse the LSTM outputs, specifically including the following: Assign a weight to the output of each LSTM and perform weighting according to the importance of different sub-behavior sequences; The outputs of the three LSTMs are vectors, and they can be concatenated. After concatenation, subsequent processing can be performed through a fully connected layer (such as a neural network layer) to finally obtain the behavior sequence at the next moment; The outputs of the three LSTMs can be fused through a multi-layer perceptron (MLP). The output vectors of the three LSTMs can be used as inputs and fed into a deep network (such as an MLP). Through the network, the relationships between the various sub-behavior sequences are learned, and finally the behavior sequence at the next moment is output.

[0067] In an optional embodiment, the interactive bipartite weighted graph includes a user node set, an advertisement data node set, a connection edge set, and an edge weight set. The user nodes correspond to the target users, the advertisement data nodes correspond to the advertisement data, the connection edges are used to connect the user nodes and the advertisement data nodes. When the target user views the advertisement data, there is a connection edge between the corresponding target user node and the advertisement data node. The edge weight of the connection edge is defined based on the viewing duration of the target user viewing the advertisement data. When the viewing duration is higher than the preset viewing time threshold, the edge weight of the connection edge is defined as 1; otherwise, the edge weight of the connection edge is defined as 0.

[0068] It should be noted that the viewing duration of the target user's ad viewing data belongs to the ad features. The specific acquisition methods include: Front-end code monitoring: When the ad is played, the front-end code of the application or web page can be monitored through JavaScript, SDK, etc. to record the start time and end time of the ad playback; Ad player events: If the ad is played in a player, the events of the player can be listened to, such as the ad start playback (onAdStart) and the ad end playback (onAdEnd), and the duration can be calculated.

[0069] In an optional embodiment, the interaction score prediction model includes an ad feature extraction module, a feature combination module, and a score prediction module. The ad feature extraction module is used to extract the text feature vector, visual feature vector, and sound feature vector of the ad data. The feature combination module is used to splice the text feature vector, visual feature vector, and sound feature vector of the ad data pairwise to obtain three cross features, and perform pairwise Hadamard product operations on the three cross features respectively to obtain three combined features of the ad data. The three combined features of the ad data are used as the row vectors of the ad data nodes in the interaction bipartite weighted graph. The score prediction module uses a multi-layer perceptron and calculates the sum of the edge weights of the interaction bipartite weighted graph based on the interaction bipartite weighted graph and the row vectors of the ad data nodes in the interaction bipartite weighted graph. The sum of the edge weights of the interaction bipartite weighted graph is the interaction score of the interaction bipartite weighted graph.

[0070] It should be noted that the pairwise Hadamard product operation must be for matrices of the same order. The following are the methods to make the matrices of the same order: (1) Usually, when dealing with features of different modalities, the embedding layer or fully connected layer in a deep learning model (such as a neural network) is used to map the text, visual, and sound features to the same dimension. For example, the text features may be converted into vectors of a fixed dimension through a text embedding model (such as BERT, Word2Vec, etc.), the visual features are extracted through a convolutional neural network (CNN), and the sound features are extracted through an acoustic model. Through appropriate layers, they are converted into the same dimension (n = m = p); (2) Use dimensionality reduction techniques (such as PCA, t-SNE, Autoencoder, etc.) to project them into the same low-dimensional space to ensure that they have the same dimension; (3) If the dimensions of the features are different, the feature vectors with different dimensions can be converted into the same dimension through methods such as padding or interpolation.

[0071] It should be noted that the main objective of the advertisement feature extraction module is to extract features from various data types of advertisements (such as text, vision, sound, etc.). These features are processed through methods such as embedding, convolutional neural network (CNN), recurrent neural network (RNN), or pre-trained models (such as BERT) to obtain feature vectors of each type; the interactive bipartite weighted graph refers to modeling advertisement data and user behavior data as a bipartite graph. In this graph, one part of the nodes represents advertisements, and the other part represents users or other behavioral entities. The weight of the edge represents the interaction intensity between the advertisement and the user; the row vector of the advertisement data node represents the feature information of the advertisement, including the text, vision, and sound features of the advertisement, and the combined features formed through concatenation and Hadamard product. These features provide the specific content information of the advertisement; the relationship between the edge weight and the row vector of the advertisement data node is mainly reflected in the score prediction module. The score prediction module uses a multi-layer perceptron (MLP) to process the node information of the interactive bipartite weighted graph. By interacting the row vector of the advertisement data node with the features of the target user, the interaction intensity between the advertisement and the user, that is, the edge weight, is calculated; the predicted interaction score list represents the potential interaction intensity between the advertisement and the user, indicating the degree of interest of the user in the advertisement sequence. However, a user views multiple advertisement data, and an advertisement data is also viewed by multiple users. Then, the interactive bipartite weighted graph of the target user includes the remaining user nodes and an advertisement data node viewed by the target user. Based on the advertisement data sequence viewed by the target user, the predicted interaction score list can be obtained.

[0072] In an optional embodiment, the multi-layer perceptron is composed of four fully connected layers.

[0073] It should be noted that the multi-layer perceptron is a most basic feedforward neural network. The fully connected layer is a basic layer in the neural network, also called the "Dense Layer". In the fully connected layer, each input node is connected to each output node, and all inputs affect the output.

[0074] Embodiment 3, as Figure 2 shown, an Internet advertisement data precise placement system based on artificial intelligence proposed by the present invention is applicable to the described method for precise placement of Internet advertisement data based on artificial intelligence, and includes:

[0075] The data acquisition unit 1 is used to obtain the behavior sequence of the target user and the viewed advertisement data sequence;

[0076] The behavior division unit 2 is used to divide the behavior sequence based on the viewing duration of the target user for different advertisement data to obtain three sub-behavior sequences corresponding to the behavior sequence;

[0077] A behavior prediction unit 3, which is used to predict behaviors through three sub-behavior sequences based on a pre-trained behavior prediction model to obtain a behavior sequence at the next preset moment;

[0078] A score prediction unit 4, which is used to construct an interactive bipartite weighted graph based on the behavior sequence at the next preset moment, and predict the interactive score through the interactive bipartite weighted graph based on a pre-trained interactive score prediction model to obtain a list of predicted interactive scores corresponding to the behavior sequence at the next preset moment;

[0079] An advertisement placement unit 5, which is used to generate an advertisement data placement list based on the list of predicted interactive scores corresponding to the behavior sequence at the next preset moment.

[0080] The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those skilled in the art.

Claims

1. A method for accurately delivering Internet advertising data based on artificial intelligence, characterized in that: include: Obtain the target user's behavior sequence and the viewed advertisement data sequence; Dividing the behavior sequence based on the viewing time of different advertisement data by the target user to obtain three sub-behavior sequences corresponding to the behavior sequence; Based on the pre-trained behavior prediction model, behavior prediction is performed through the three sub-behavior sequences to obtain the behavior sequence at the next preset moment; Constructing an interaction bipartite weighted graph based on the behavior sequence at the next preset moment, and performing interaction score prediction through the interaction bipartite weighted graph based on a pre-trained interaction score prediction model to obtain a predicted interaction score list corresponding to the behavior sequence at the next preset moment; Generate an advertisement data delivery list based on the predicted interaction score list corresponding to the behavior sequence at the next preset moment; The interactive bipartite weighted graph includes a user node set, an advertisement data node set, a connection edge set and an edge weight set. The user node corresponds to a target user, the advertisement data node corresponds to advertisement data, and the connection edge is used to connect the user node and the advertisement data node. When the target user watches the advertisement data, a connection edge exists between the corresponding target user node and the advertisement data node. The edge weight of the connection edge is defined based on the viewing time of the target user watching the advertisement data. When the viewing time is higher than a preset viewing time threshold, the edge weight of the connection edge is defined as 1, otherwise, the edge weight of the connection edge is defined as 0. The interactive score prediction model includes an advertising feature extraction module, a feature combination module and a score prediction module. The advertising feature extraction module is used to extract the text feature vector, visual feature vector and sound feature vector of the advertising data. The feature combination module is used to respectively concatenate the text feature vector, visual feature vector and sound feature vector of the advertising data in pairs to obtain three cross features, perform pairwise Hadamard product operations on the three cross features to obtain three combined features of the advertising data, and use the three combined features of the advertising data as the row vectors of the advertising data nodes in the interactive bipartite weighted graph. The score prediction module adopts a multilayer perceptron and calculates the sum of the edge weights of the interactive bipartite weighted graph based on the interactive bipartite weighted graph and the row vectors of the advertising data nodes in the interactive bipartite weighted graph. The sum of the edge weights of the interactive bipartite weighted graph is the interactive score of the interactive bipartite weighted graph.

2. According to claim 1, a method for accurate delivery of Internet advertising data based on artificial intelligence is characterized in that: The behavior sequence includes the target user's interactive information on different advertisement data, wherein the expression of the behavior sequence is as follows: ； in, Indicates the target user’s interactive information on different advertising data. Indicates Advertising data ID, Indicates advertising features of advertising data, Indicates the amount of advertising data viewed by the target user.

3. The method for accurately delivering Internet advertising data based on artificial intelligence according to claim 2, characterized in that: The behavior sequence is divided based on the viewing time of different advertisement data by the target user to obtain three sub-behavior sequences corresponding to the behavior sequence, including: Divide different advertisement data into multiple advertisement groups based on advertisement duration; Comparing the number of samples in the multiple advertisement groups with a preset sample number threshold, if the number of samples in the multiple advertisement groups is higher than the preset sample number threshold, determining the advertisement group as a label with sufficient sample size, otherwise, determining the advertisement group as a label with insufficient sample size; Obtaining the viewing duration of multiple advertisement data in the advertisement group with the sufficient sample size label to obtain a viewing duration sequence corresponding to the advertisement group with the sufficient sample size label; Sorting the viewing time sequences corresponding to the advertisement groups with the sufficient sample size label to obtain the viewing time order sequences corresponding to the advertisement groups with the sufficient sample size label; The advertisement group with the sufficient sample size label is divided into three sub-advertisement groups based on the viewing time sequence corresponding to the advertisement group with the sufficient sample size label.

4. The method for accurately delivering Internet advertising data based on artificial intelligence according to claim 3 is characterized in that: The behavior sequence is divided based on the viewing time of different advertisement data by the target user to obtain three sub-behavior sequences corresponding to the behavior sequence, and further includes: Obtaining the viewing duration of multiple advertisement data in the advertisement group with the insufficient sample size label, so as to obtain a viewing duration sequence corresponding to the advertisement group with the insufficient sample size label; Matching the viewing time sequence corresponding to the advertisement group with the insufficient sample size label with the viewing time sequence corresponding to the advertisement group with the sufficient sample size label based on the K-nearest neighbor algorithm to find the closest neighbor group, wherein the neighbor group belongs to the advertisement group with the sufficient sample size label; Dividing the advertisement group with insufficient sample size labels into three sub-advertisement groups based on the neighbor grouping; The behavior sequence is divided into three sub-advertisement groups corresponding to the advertisement group based on the insufficient sample size label and the sufficient sample size label to obtain three sub-behavior sequences corresponding to the behavior sequence, wherein the three sub-behavior sequences include a not viewing sub-behavior sequence, an interested sub-behavior sequence and a not interested sub-behavior sequence.

5. According to claim 4, a method for accurate delivery of Internet advertising data based on artificial intelligence is characterized in that: The behavior prediction model includes an advertising embedding module, a feature extraction module and a behavior prediction module. The advertising embedding module is used to convert the behavior sequence of the target user and the viewed advertisement data sequence into an embedding vector. The feature extraction module is used to extract the fine-grained interest expression of the target user from the three sub-behavior sequences based on the embedding vector. The behavior prediction module is used to predict the behavior sequence at the next preset moment based on the fine-grained interest expression of the target user.

6. The method for accurate delivery of Internet advertising data based on artificial intelligence according to claim 5, characterized in that: The advertisement embedding module is used to embed the advertisement data ID, the advertisement data author ID, the advertisement data feedback label and the location information to obtain the advertisement data ID embedding vector, the advertisement data author ID embedding vector, the advertisement data feedback label embedding vector and the location information embedding vector, fuse the advertisement data ID embedding vector, the advertisement data author ID embedding vector, the advertisement data feedback label embedding vector and the location information embedding vector to obtain the embedding vector of the advertisement data, and obtain the embedding vector corresponding to the three sub-behavior sequences based on the embedding vector of the advertisement data. The feature extraction module adopts the Transformer network to input the embedding vectors corresponding to the three sub-behavior sequences into different Transformer networks to obtain the feature representations corresponding to the three sub-behavior sequences. The behavior prediction module adopts the long short-term memory network to input the feature representations corresponding to the three sub-behavior sequences into different long short-term memory networks to obtain the behavior sequence at the next preset moment.

7. The method for accurate delivery of Internet advertising data based on artificial intelligence according to claim 6, characterized in that: The multilayer perceptron consists of four fully connected layers.

8. An artificial intelligence-based Internet advertising data precision delivery system, which is applicable to an artificial intelligence-based Internet advertising data precision delivery method according to any one of claims 1 to 7, characterized in that: include: A data collection unit (1), the data collection unit (1) is used to obtain a behavior sequence of a target user and a viewed advertisement data sequence; A behavior division unit (2), the behavior division unit (2) being used to divide the behavior sequence based on the viewing time of different advertisement data by the target user, so as to obtain three sub-behavior sequences corresponding to the behavior sequence; A behavior prediction unit (3), the behavior prediction unit (3) being used to perform behavior prediction based on the pre-trained behavior prediction model through the three sub-behavior sequences to obtain a behavior sequence at the next preset moment; A score prediction unit (4), the score prediction unit (4) being used to construct an interaction bipartite weighted graph based on the behavior sequence at the next preset moment, and to perform interaction score prediction through the interaction bipartite weighted graph based on a pre-trained interaction score prediction model, so as to obtain a predicted interaction score list corresponding to the behavior sequence at the next preset moment; An advertisement delivery unit (5), the advertisement delivery unit (5) being used to generate an advertisement data delivery list based on a predicted interaction score list corresponding to the behavior sequence at the next preset moment.

Images