Advertisement pushing method and device based on artificial intelligence and medium

Abstract

The invention discloses an advertisement pushing method and device based on artificial intelligence and a medium, and relates to the technical field of advertisement pushing, and the method comprises the steps: collecting data information of a user, carrying out the preprocessing operation, and dividing the data into a scene data set and a user interest data set; constructing a scene recognition model, and training the scene recognition model through the scene data set; constructing a user interest model, and training the user interest model through the user interest data set; analyzing the advertisement content, extracting an advertisement keyword, defining a scene keyword for the scene according to the current scene of the user, and matching and pushing the scene keyword; and carrying out effect evaluation and optimization on the pushed advertisement. According to the method, advertisements related to the current scene can be pushed to the user through scene perception advertisement pushing, so that the click rate and the conversion rate of the advertisements are improved, the satisfaction degree of the user is improved, and the loyalty of the user to the brand is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of advertisement push, and in particular to an advertisement push method, device and medium based on artificial intelligence. Background Art

[0002] Traditional advertising push methods are mainly based on users' historical behaviors and interests, but with the development of mobile devices and IoT technology, scene-aware advertising push has gradually attracted attention. This method captures users' scene information (such as location, time, device status, etc.) in real time and pushes ads related to the current scene to users, thereby improving the click-through rate and conversion rate of ads.

[0003] In the prior art, how to accurately identify the user's current scene and push relevant advertisements to the user based on the scene information has become a major problem. Summary of the invention

[0004] In view of the problems existing in the above-mentioned existing artificial intelligence-based advertising push methods, the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is how to provide an advertisement pushing method based on artificial intelligence.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: an artificial intelligence-based advertising push method, which includes collecting user data information and performing preprocessing operations to divide the data into a scene data set and a user interest data set; building a scene recognition model and training the scene recognition model through the scene data set; building a user interest model and training the user interest model through the user interest data set; analyzing the advertisement content, extracting advertisement keywords, defining scene keywords for the scene according to the user's current scene, and matching and pushing them; and evaluating and optimizing the effects of the pushed advertisements.

[0007] As a preferred solution of the artificial intelligence-based advertising push method described in the present invention, the user's data information includes location information, time information, and current application information, the preprocessing operation includes denoising, normalization, and classification processing, and the advertisement includes text advertisement, video advertisement, and voice advertisement.

[0008] As a preferred solution of the artificial intelligence-based advertising push method described in the present invention, the division of the scene data set includes the following steps: applying the DBSCAN algorithm to cluster the pre-processed location information data to form several candidate areas, each candidate area represents an area where users often appear; statistics are performed on the records in each candidate area to divide them into different time periods; the application usage data characteristics in each time period are analyzed; according to the user's behavior pattern in the time period, the records in the candidate area are marked with clear scene labels; when there is a conflict in the behavior patterns of different time periods in the same area, a time feature weight judgment mechanism is used to mark the records as "double records"; all records that have been clustered by location and subdivided by time and confirmed by behavior pattern comparison are classified into the scene data set; the division of the user interest data set includes the following steps: normalizing the application usage data in each record; establishing a pre-defined application category label library; matching each behavior data in the record with the label library to determine the user's interest intensity in each category.

[0009] As a preferred solution of the artificial intelligence-based advertising push method of the present invention, wherein: constructing a scene recognition model and training the scene recognition model through a scene data set includes the following steps:

[0010] Design the network structure, including input layer, convolution layer, pooling layer, fully connected layer and output layer;

[0011] Use the Softmax function as the activation function of the output layer to obtain the probability distribution of each scene;

[0012] The scene recognition model is trained through the scene dataset, and the cross entropy is used as the loss function. The formula is as follows:

[0013]

[0014] Where L is the cross entropy loss, N is the number of data, and y i is the true label, is the predicted value of the model;

[0015] Use the Adam optimizer to update the model parameters, input the real-time collected data into the trained scene recognition model, obtain the scene probability distribution output by the model, and select the scene with the highest probability as the current scene.

[0016] As a preferred solution of the artificial intelligence-based advertising push method of the present invention, wherein: constructing a user interest model and training the user interest model through a user interest data set includes the following steps:

[0017] Use the word segmentation tool to segment the user interest dataset;

[0018] Use the TF-IDF algorithm to calculate the weight of the user interest dataset after word segmentation. The formula is as follows:

[0019]

[0020] Where TF(t,d) is the frequency of word t in document d, and IDF(t) is the inverse document frequency of word t;

[0021] The BERT model is used as the basic model for user interest modeling and optimized. The optimized model is as follows: , where w i is the i-th word in the sequence,

[0022] Token embedding layer: ;

[0023] Position Embedding Layer: ;

[0024] BERT Input Layer: ;

[0025] Transformer layer: ;

[0026] Output layer: ;

[0027] In the formula, I (w i ) is the vocabulary w i One-hot encoding, W token is the token embedding matrix, I(i) is the one-hot encoding of position i, and W position is the position embedding matrix, b is the number of Transformer layers, is the output of layer b, W O and B O are the weights and biases of the output layer, is the output of the last Transformer layer;

[0028] Use labeled user interest datasets for model training and cross entropy as the loss function;

[0029] Use the Adam optimizer to update the model parameters, input the real-time collected data into the trained BERT model, obtain the user interest distribution output by the model, and select the interest with the highest probability as the user's current interest.

[0030] As a preferred solution of the artificial intelligence-based advertising push method of the present invention, the following steps are included: analyzing the advertising content, extracting advertising keywords, defining scene keywords for the scene according to the user's current scene, and matching and pushing.

[0031] Conduct text analysis on each advertisement content to extract keywords and themes;

[0032] Use the TF-IDF algorithm to calculate the weight of keywords in the advertising content;

[0033] Based on the user's current scenario, define a set of keywords and weights for the scenario;

[0034] Use cosine similarity to calculate the similarity between scene keywords and advertising content keywords, as well as the similarity between user interests and advertising content keywords. The formula is as follows:

[0035]

[0036]

[0037] In the formula, A is the scene keyword vector, B is the keyword vector of the advertising content, and C is the user interest keyword vector. is the similarity between the scene keywords and the advertising content keywords, The similarity between user interests and keywords in advertising content;

[0038] Combine the similarity between the scene and the advertisement and the similarity between the user's interest and the advertisement to calculate the comprehensive similarity. The formula is as follows:

[0039]

[0040] Where α is a weight parameter used to adjust the importance of scene similarity and interest similarity;

[0041] Sort ads according to comprehensive similarity, select the ads with the highest similarity for push, and record the time, content, and user feedback of pushed ads for subsequent model optimization.

[0042] As a preferred solution of the artificial intelligence-based advertising push method of the present invention, when pushing advertisements, the following steps are also included:

[0043] Collect the user's daily active time through the user's application usage records;

[0044] For each user, record their activity level every hour of the day to form a 24-dimensional vector;

[0045] Use the survival analysis model to predict the user's scenario end time. The formula is as follows:

[0046]

[0047] Where S(t) is the probability that the event has not occurred after time t, and T is the time when the event occurs;

[0048] The Cox proportional hazard model is used to consider the impact of multiple covariates on the time of event occurrence. The formula is as follows:

[0049]

[0050] Where h(t, X) is the hazard function at time t given the covariate X, is the baseline hazard function and β is the coefficient of the covariate.

[0051] As a preferred solution of the artificial intelligence-based advertising push method of the present invention, wherein: evaluating and optimizing the effect of pushed advertisements includes the following steps:

[0052] Set a unique identifier for each ad push, collect user behavior data related to the ad, and associate the collected data with the unique identifier of the ad;

[0053] Define the evaluation index of advertising effect, the formula is as follows:

[0054]

[0055]

[0056] In the formula, click count is the total number of times users click on the ad, impression count is the total number of times the ad is shown to users, and the number of users who complete the target action is the number of users who complete the specified target action;

[0057] Calculate the effect of each ad push based on the defined evaluation indicators, compare the effects of different ad content, push timing and push strategies, and find the best combination;

[0058] Based on the user's feedback on the advertisement, the user's interest model is updated.

[0059] A computer device includes a memory and a processor, wherein the memory stores a computer program, and is characterized in that when the processor executes the computer program, it implements the steps of an artificial intelligence-based advertising push method.

[0060] A computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the computer program implements the steps of an artificial intelligence-based advertising push method.

[0061] The beneficial effects of the present invention are as follows: through scene-aware advertising push, advertisements related to the current scene can be pushed to users, thereby improving the click-through rate and conversion rate of advertisements; because the advertising content and push timing are matched with the actual needs and scenes of users, user satisfaction can be improved and user loyalty to the brand can be enhanced; by collecting user feedback information on pushed advertisements, the user's interest model and scene recognition model can be updated in real time, thereby optimizing the advertising push strategy to make it more accurate and efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0063] Figure 1 This is a flow chart of the artificial intelligence-based advertising push method in Example 1. DETAILED DESCRIPTION

[0064] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0065] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0066] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0067] Example 1

[0068] Reference Figure 1 , which is the first embodiment of the present invention, and this embodiment provides an advertisement push method based on artificial intelligence, and the advertisement push method based on artificial intelligence includes the following steps:

[0069] S1. Collect user data information and perform preprocessing operations to divide the data into a scene data set and a user interest data set. The user data information includes location information, time information, and current application information. The preprocessing operations include denoising, normalization, and classification. The advertisements include text advertisements, video advertisements, and voice advertisements.

[0070] S2, build a scene recognition model and train the scene recognition model through the scene data set;

[0071] S3, construct a user interest model and train the user interest model through the user interest dataset;

[0072] S4, analyzing the advertisement content, extracting advertisement keywords, defining scene keywords for the scene according to the user's current scene, and matching and pushing;

[0073] S5. Evaluate and optimize the effectiveness of pushed advertisements.

[0074] Specifically, dividing the scene data set includes the following steps:

[0075] S11. Clustering the preprocessed location information data to form candidate regions.

[0076] (1) Using the denoised, normalized and time-synchronized location information data completed in the preprocessing stage, the DBSCAN algorithm is used to cluster the data;

[0077] (2) During the implementation process, the key parameters of the DBSCAN algorithm, namely the radius threshold ε and the minimum number of sample points MinPts, are set to ensure that the clustering results can accurately reflect the areas where users frequently appear;

[0078] (3) After clustering, each cluster constitutes a candidate area, which represents the places that users often visit in their daily lives, such as home areas, office areas, or business district areas.

[0079] S12, performing time period statistics and division on the data records in each candidate area.

[0080] (1) For each candidate area, based on the time information attached to the records, the data records in the area are divided into several time periods according to the user's activity time (for example, they can be divided into hourly periods, morning, morning, afternoon, evening and night);

[0081] (2) The division of time periods adopts a method that combines preset fixed time periods with adaptive adjustment: first, a preliminary division is made based on regular time periods, and then the boundaries are automatically fine-tuned based on the time distribution characteristics of user behavior data in the candidate area to more accurately reflect the actual activity peak;

[0082] (3) All records are retained in each time period and provide time context information for subsequent behavior pattern analysis.

[0083] S13. Analyze application usage data characteristics in each time period.

[0084] (1) For each time period divided in each candidate area, statistical analysis is performed on the application usage data (e.g., clicks, browsing, and dwell time of various applications) contained in the data records;

[0085] (2) Based on a pre-defined application category tag library (such as "office", "entertainment", "social", "shopping", etc.), calculate the frequency of occurrence and usage intensity of each category in each time period;

[0086] (3) The statistical results serve as an important basis for measuring user behavior patterns during the time period, reflecting the types of activities that users mainly engage in during different time periods.

[0087] S14. Add clear scene labels to the candidate area records according to the user's behavior pattern within the time period.

[0088] (1) According to the data obtained in step S13, the main activity type of the user in each time period is determined, and one or more scene labels that best describe the behavior in the time period are determined, such as "office place", "leisure and entertainment place" or "shopping center", etc.;

[0089] (2) Associate each data record with the scene label of its time period to form a preliminary scene description;

[0090] (3) At this time, the records in each candidate area have a scene identifier of the corresponding time period, which serves as an important basis for subsequent advertising matching.

[0091] S15. Handle conflicts in behavior patterns at different time periods in the same area.

[0092] (1) If data records in different time periods in the same candidate area show obvious behavioral pattern conflicts (for example, both office behavior and entertainment behavior appear in the same geographical location), the preset "time feature weight judgment mechanism" is used to evaluate the records in the area;

[0093] (2) Based on the weight values ​​of the application usage data in each time period, different scene labels are compared. When the weights of two or more scene labels are similar or difficult to distinguish, the record is marked as a "double record". This mechanism ensures that when the behavior pattern is unclear, the record will not be simply classified as a single scene, but will retain the possibility of further correction based on subsequent advertising feedback.

[0094] S16. All records confirmed by location clustering, time division and behavior pattern comparison are classified into the scene data set.

[0095] (1) For all data records in the candidate area, according to the scene tags and "dual record" marks obtained in step S14 and step S15, if a record has only a single scene tag, the scene attribute of the record is directly set to the tag and marked as "single scene" in the data structure. If a record contains two or more scene tags at the same time, the system saves all candidate scene tags in the data record. For "single scene" records, they are directly stored in the standard record area of ​​the scene data set. For "dual record" records, the system stores them in a dedicated record format in the scene data set. In addition to the conventional location information, time information and behavior data, the record also has a "multi-scene tag" field attached.

[0096] (2) The records included in the scene dataset contain both candidate area and time period information and scene labels determined after behavior pattern comparison, providing complete and accurate training data for the subsequent construction of the scene recognition model.

[0097] Partitioning user interest datasets includes the following steps:

[0098] S17, normalizing the application usage data contained in each preprocessed data record to eliminate differences between different data sources and scales;

[0099] S18. Establish a predefined application category tag library, which includes but is not limited to categories such as "office", "entertainment", "social", and "shopping", and each category can be further divided into subcategories;

[0100] S19. Match the normalized behavioral data in each data record with the application category label library, and determine the interest intensity of the user in each category reflected by the record by comparing the numerical values ​​of each category. Based on the matching results, classify each data record into a user interest data set to provide input data for the subsequent construction of a user interest model.

[0101] Furthermore, building a scene recognition model and training the scene recognition model using a scene dataset includes the following steps:

[0102] Design the network structure, including input layer, convolution layer, pooling layer, fully connected layer and output layer. The formula is as follows:

[0103]

[0104] In the formula, y is the output of the network, f is the activation function, x is the input data, w is the weight, and b is the bias;

[0105] Use the Softmax function as the activation function of the output layer to obtain the probability distribution of each scene. The formula is as follows:

[0106]

[0107] In the formula, P (y i ) is the probability of category i, y i is the original output value of category i, j is the length of vector y, which is also the total number of categories, and M is the total length of vector y;

[0108] The scene recognition model is trained through the scene dataset, and the cross entropy is used as the loss function. The formula is as follows:

[0109]

[0110] Where L is the cross entropy loss, N is the number of data, and y i is the true label, is the predicted value of the model;

[0111] Use the Adam optimizer to update the model parameters, input the real-time collected data into the trained scene recognition model, obtain the scene probability distribution output by the model, and select the scene with the highest probability as the current scene.

[0112] Constructing a user interest model and training the user interest model through a user interest dataset includes the following steps:

[0113] Use the word segmentation tool to segment the user interest dataset;

[0114] Use the TF-IDF algorithm to calculate the weight of the user interest dataset after word segmentation. The formula is as follows:

[0115]

[0116] Where TF(t,d) is the frequency of word t in document d, and IDF(t) is the inverse document frequency of word t;

[0117] The BERT model is used as the basic model for user interest modeling and optimized. The optimized model is as follows: , where w i is the i-th word in the sequence,

[0118] Token embedding layer: ;

[0119] Position Embedding Layer: ;

[0120] BERT Input Layer: ;

[0121] Transformer layer: ;

[0122] Output layer: ;

[0123] In the formula, I (w i ) is the vocabulary w i One-hot encoding, W token is the token embedding matrix, I(i) is the one-hot encoding of position i, and W position is the position embedding matrix, b is the number of Transformer layers, is the output of layer b, W O and B O are the weights and biases of the output layer, is the output of the last Transformer layer;

[0124] Use labeled user interest datasets for model training and cross entropy as the loss function;

[0125] Use the Adam optimizer to update the model parameters, input the real-time collected data into the trained BERT model, obtain the user interest distribution output by the model, and select the interest with the highest probability as the user's current interest.

[0126] Preferably, analyzing the advertisement content, extracting advertisement keywords, defining scene keywords for the scene according to the user's current scene, and matching and pushing include the following steps:

[0127] For each advertisement content to be pushed (including text ads, video titles and descriptions, and text transcription results of voice ads), a preprocessing algorithm is used to perform text segmentation and semantic analysis to extract keywords and subject terms that describe the theme of the advertisement. During the extraction process, natural language processing technology (such as BERT or LSTM-based models) is used to capture contextual information to ensure that keyword extraction has a high accuracy rate. For example, for an advertisement, its segmentation results may include keywords such as "discount", "new products", "electronic products", and at the same time determine that the theme category of the advertisement is "shopping promotion";

[0128] Use the TF-IDF algorithm to calculate the weight of keywords in the advertising content;

[0129] For records with a single scene label, the scene in which the user is currently located (e.g., "office", "shopping center" or "leisure and entertainment venue") is determined based on the scene data set previously obtained through candidate area clustering, time division and behavior pattern analysis, and the historical records in the scene are counted to extract the keywords that can describe the scene and their long-term statistical weights to form a scene keyword vector; for the parts belonging to scene A and scene B in the dual records, the respective scene keyword vectors are counted according to the processing steps of a single scene, and the candidate scene keyword vectors are fused using a weighted fusion method to obtain a comprehensive scene keyword vector;

[0130] Use cosine similarity to calculate the similarity between scene keywords and advertising content keywords, as well as the similarity between user interests and advertising content keywords. The formula is as follows:

[0131]

[0132]

[0133] In the formula, A is the scene keyword vector, B is the keyword vector of the advertising content, and C is the user interest keyword vector. is the similarity between the scene keywords and the advertising content keywords, The similarity between user interests and keywords in advertising content;

[0134] Combine the similarity between the scene and the advertisement and the similarity between the user's interest and the advertisement to calculate the comprehensive similarity. The formula is as follows:

[0135]

[0136] Where α is a weight parameter used to adjust the importance of scene similarity and interest similarity;

[0137] Sort ads according to comprehensive similarity, select the ads with the highest similarity for push, and record the time, content, and user feedback of pushed ads for subsequent model optimization.

[0138] Preferably, when pushing advertisements, the following steps are also included:

[0139] Collect the user's daily active time through the user's application usage records;

[0140] For each user, record their activity level every hour of the day to form a 24-dimensional vector;

[0141] Use the survival analysis model to predict the user's scenario end time. The formula is as follows:

[0142]

[0143] Where S(t) is the probability that the event has not occurred after time t, and T is the time when the event occurs;

[0144] The Cox proportional hazard model is used to consider the impact of multiple covariates on the time of event occurrence. The formula is as follows:

[0145]

[0146] Where h(t, X) is the hazard function at time t given the covariate X, is the baseline hazard function and β is the coefficient of the covariate.

[0147] The evaluation and optimization of the pushed ads include the following steps:

[0148] Set a unique identifier for each ad push, collect user behavior data related to the ad, such as clicks, browsing time, and purchases, and associate the collected data with the ad's unique identifier;

[0149] Define the evaluation index of advertising effect, the formula is as follows:

[0150]

[0151]

[0152] In the formula, click count is the total number of times users click on the ad, impression count is the total number of times the ad is displayed to users, and the number of users who complete the target action is the number of users who complete the specified target action, where the specified target action can be purchase, click, registration, etc.

[0153] Calculate the effect of each ad push based on the defined evaluation indicators, compare the effects of different ad content, push timing and push strategies, and find the best combination;

[0154] Based on the user's feedback on the advertisement, such as clicks and purchases, the user's interest model is updated.

[0155] Example 2

[0156] The second embodiment of the present invention is different from the previous embodiment in that:

[0157] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0158] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0159] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0160] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0161] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An artificial intelligence-based advertising push method, characterized in that: include, Collect user data information and perform preprocessing operations to divide the data into scenario data sets and user interest data sets; Build a scene recognition model and train it using scene datasets; Build a user interest model and train it using a user interest dataset; Analyze the advertising content, extract advertising keywords, define scenario keywords for the scenario based on the user's current scenario, and perform matching and push; Evaluate and optimize the effectiveness of pushed advertisements.

2. The artificial intelligence-based advertising push method according to claim 1, characterized in that: The user data information includes location information, time information, and current application information. The preprocessing operation includes normalization processing and classification processing for denoising. The advertisement includes text advertisement, video advertisement, and voice advertisement.

3. The artificial intelligence-based advertising push method according to claim 2, characterized in that: Dividing the scene data set comprises the following steps: The DBSCAN algorithm is applied to cluster the preprocessed location information data to form several candidate areas, each of which represents an area where users often appear; Count the records in each candidate area and divide them into different time periods; Analyze application usage data characteristics in each time period; According to the user's behavior pattern within the time period, the records in the candidate area are marked with clear scene labels; When there are conflicts in the behavior patterns in different time periods in the same area, the time feature weight judgment mechanism is adopted to mark the records as "double records"; All records that have been clustered by location and subdivided by time and confirmed by behavioral pattern comparison are classified into the scene dataset; Dividing the user interest data set includes the following steps: Normalize the application usage data in each record; Establish a library of predefined application category tags; Match the behavioral data in the record with the tag library to determine the user's interest intensity in each category.

4. The artificial intelligence-based advertising push method according to claim 3, characterized in that: Building a scene recognition model and training the scene recognition model through a scene dataset includes the following steps: Design the network structure, including input layer, convolution layer, pooling layer, fully connected layer and output layer; Use the Softmax function as the activation function of the output layer to obtain the probability distribution of each scene; The scene recognition model is trained using the scene dataset and cross entropy is used as the loss function; Use the Adam optimizer to update the model parameters, input the real-time collected data into the trained scene recognition model, obtain the scene probability distribution output by the model, and select the scene with the highest probability as the current scene.

5. The artificial intelligence-based advertising push method according to claim 4, characterized in that: Constructing a user interest model and training the user interest model through a user interest dataset includes the following steps: Use the word segmentation tool to segment the user interest dataset; Use the TF-IDF algorithm to calculate the weight of the user interest dataset after word segmentation; Use the BERT model as the basic model for user interest modeling and optimize it; Use labeled user interest datasets for model training and cross entropy as the loss function; Use the Adam optimizer to update the model parameters, input the real-time collected data into the trained BERT model, obtain the user interest distribution output by the model, and select the interest with the highest probability as the user's current interest.

6. The artificial intelligence-based advertising push method according to claim 5, characterized in that: Analyze the advertisement content, extract advertisement keywords, define scene keywords for the scene according to the user's current scene, and match and push them, including the following steps: Conduct text analysis on each advertisement content to extract keywords and themes; Use the TF-IDF algorithm to calculate the weight of keywords in the advertising content; Based on the user's current scenario, define a set of keywords and weights for the scenario; Use cosine similarity to calculate the similarity between scene keywords and advertising content keywords, as well as the similarity between user interests and advertising content keywords. The formula is as follows: In the formula, A is the scene keyword vector, B is the keyword vector of the advertising content, and C is the user interest keyword vector. is the similarity between the scene keywords and the advertising content keywords, The similarity between user interests and keywords in advertising content; Combine the similarity between the scene and the advertisement and the similarity between the user's interest and the advertisement to calculate the comprehensive similarity. The formula is as follows: Where α is a weight parameter used to adjust the importance of scene similarity and interest similarity; Sort ads according to comprehensive similarity, select the ads with the highest similarity for push, and record the time, content, and user feedback of pushed ads for subsequent model optimization.

7. The artificial intelligence-based advertising push method according to claim 6, characterized in that: When pushing advertisements, the following steps are also included: Collect the user's daily active time through the user's application usage records; For each user, record their activity level every hour of the day to form a 24-dimensional vector; Use the survival analysis model to predict the user's scenario end time. The formula is as follows: Where S(t) is the probability that the event has not occurred after time t, and T is the time when the event occurs; The Cox proportional hazard model is used to consider the impact of multiple covariates on the time of event occurrence. The formula is as follows: Where h(t, X) is the hazard function at time t given the covariate X, is the baseline hazard function and β is the coefficient of the covariate.

8. The artificial intelligence-based advertising push method according to claim 7, characterized in that: The evaluation and optimization of the pushed ads include the following steps: Set a unique identifier for each ad push, collect user behavior data related to the ad, and associate the collected data with the unique identifier of the ad; Define the evaluation index of advertising effect, the formula is as follows: In the formula, click count is the total number of times users click on the ad, impression count is the total number of times the ad is shown to users, and the number of users who complete the target action is the number of users who complete the specified target action; Calculate the effect of each ad push based on the defined evaluation indicators, compare the effects of different ad content, push timing and push strategies, and find the best combination; Based on the user's feedback on the advertisement, the user's interest model is updated.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the artificial intelligence-based advertising push method described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the artificial intelligence-based advertising push method described in any one of claims 1 to 8 are implemented.

Images