Advertisement pushing method and system based on user behavior analysis

Abstract

The invention discloses an advertisement pushing method and system based on user behavior analysis, and relates to the technical field of advertisement pushing, and the method comprises the following specific steps: data collection and preprocessing: employing a web crawler technology, setting corresponding capture rules and ranges for a social media platform, a professional interest forum and an e-commerce platform, and carrying out the data collection and preprocessing; according to the method, the whole process of user interest evolution is deeply tracked, the real-time behavior data of the user is combined, and the advertisement content is adaptively adjusted, so that the advertisement highly matched with the current interest and behavior of the user can be accurately pushed, meanwhile, a mixed feature fusion and dynamic weight distribution mechanism is introduced, and the user experience is improved. According to the method, user interest features are extracted more comprehensively, and feature weights are dynamically adjusted according to real-time feedback of user behaviors, so that feature vectors capable of reflecting real-time interests of users more accurately are constructed, and advertisement pushing accuracy is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of advertisement push, and in particular to an advertisement push method and system based on user behavior analysis. Background Art

[0002] In the current digital marketing era, advertising push technology, as a key bridge connecting consumers with goods or services, is becoming increasingly important. With the rapid development of Internet technology, user behavior data has shown explosive growth. These data contain rich information on user preferences, interests and needs, providing unprecedented precision possibilities for advertising push. Companies have used advanced technologies such as big data and artificial intelligence to conduct in-depth analysis of user behavior in order to achieve personalized customization and precise push of advertising content. This trend has not only promoted the continuous innovation of advertising push technology, but also promoted the transformation and upgrading of the entire advertising industry.

[0003] However, although traditional advertising push technology has improved the coverage of advertising to a certain extent, its inherent limitations cannot be ignored. Traditional advertising push methods are often based on static user portraits, that is, a fixed user interest model is built according to the user's past behavior data, and advertising is pushed based on this. This method ignores the diversity and dynamics of user interests, resulting in the pushed advertising content often being out of touch with the user's current actual needs. In addition, traditional technology lacks the ability to monitor and respond to user real-time behavior, and is unable to adjust advertising content in a timely manner according to changes in user current behavior, thereby reducing the accuracy and effectiveness of advertising push. This fixed push method is not only difficult to arouse user interest, but may also cause user disgust, which in turn affects user trust and loyalty to the push platform.

[0004] In view of the above problems, it is necessary to optimize the existing advertising push methods and systems, deeply analyze user behavior data, track the evolution of user interests in all directions, and realize adaptive dynamic adjustment of advertising content according to user real-time behavior, accurately grasp the changing trends of user interests, and improve the accuracy and effectiveness of advertising push. Therefore, it is of great significance to develop an advertising push method and system based on user behavior analysis that can comprehensively realize the above characteristics. Summary of the invention

[0005] The purpose of the present invention is to make up for the shortcomings of the prior art and to provide an advertising push method and system based on user behavior analysis. It can build a virtual or enhanced experimental environment through VR / AR device modules, so that the experimenter can observe and understand the structure and operation process of the experimental instrument in an immersive way. This highly realistic and immersive experience helps to enhance the experimenter's perception and understanding of the experimental process and improve the learning effect.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: On the one hand, an advertisement push method based on user behavior analysis comprises the following specific steps:

[0007] Data collection and preprocessing: Using web crawler technology, we set corresponding crawling rules and scopes for social media platforms, professional interest forums, and e-commerce platforms, collect HTML structure information of web pages, parse users' browsing history and like and comment behaviors, and obtain users' search history, purchase history, and advertising interaction-related data within the platform by calling the APIs provided by each platform. We preliminarily organize the data in a unified data format, and clean, denoise, and standardize the data.

[0008] Dynamic modeling and tracking of user interests: Use natural language processing technology to extract text features from text behavior data, use image recognition technology to obtain image features from image behavior data, and at the same time count the behavior frequency and duration quantitative features and integrate them into feature vectors. Use deep neural network architecture to achieve hybrid feature fusion and dynamic weight allocation. Dynamically adjust the weights of each feature according to user behavior changes in real time to generate a fused feature vector that accurately reflects the user's real-time interest evolution. Sampling it at fixed time intervals to construct time series data, use deep learning algorithms to build a user interest evolution model, divide the data set and set hyperparameters. After iterative training, optimization based on the validation set and evaluation of the test set, the model can predict interest evolution based on historical feature sequences and process new behavior data in real time for tracking and analysis;

[0009] Dynamic content adaptive advertising push decision-making: By capturing user operation event information in real time through key element embedding points in web pages and mobile applications, the backend integrates it with the output of the user interest evolution model to form comprehensive user status information, and searches and matches in the advertising resource library based on this information, uses natural language generation technology to adjust the key selling points of advertising copy, uses image editing technology to optimize the picture display details, and dynamically adjusts the discount information through specific algorithms based on user interest popularity and behavioral activity factors. The timing and frequency of advertising push are determined by comprehensively considering the stability of user interests, behavioral activity and business goals, forming a complete dynamic content adaptive advertising push strategy;

[0010] Advertisement push execution and feedback: According to the formulated advertisement push strategy, the corresponding interfaces of mobile application push service, web pop-up window push and email push channels are called to accurately push the advertisement information to users after packaging it in the required format of each channel, and the push status is monitored in real time. Meanwhile, the feedback information of the user's interaction with the advertisement is recorded on the advertisement display page and the back-end server, and then sorted and stored according to the specifications and sent back to the data collection link for continuous optimization of the user interest evolution model and advertisement push strategy, thereby forming a data closed loop to ensure that advertisement push closely follows the changes in user behavior.

[0011] Furthermore, in the user interest dynamic modeling and tracking step, natural language processing technology is used to extract text features from text behavior data, image recognition technology is used to obtain image features from image behavior data, and the behavior frequency and duration behavior quantitative features are statistically analyzed and integrated into feature vectors. For extracting text features from text behavior data, the formula is used. The contribution weight of each word to the user's interest is determined by comprehensively considering the topic-related importance and frequency of occurrence, and then the text feature vector is constructed, where TF weigh ted (w,D) represents the feature weight value of word w in document D after considering the comprehensive weight of topic and frequency. TF(w,D) is the frequency of word w in document D. P(topic i |w) indicates that given word n, the document belongs to the topic topic i The conditional probability of w′∈D P(topic i |w′) is the total number of words in document D about the topic i The conditional probability sum of , λ is the weight adjustment parameter, IDF(w) is the inverse document frequency of word w, and the calculation formula is Where N is the total number of documents and DF(w) is the number of documents containing word w.

[0012] Furthermore, in the user interest dynamic modeling and tracking step, the algorithm formula used for obtaining image features from image behavior data using image recognition technology is: image =∑ r∈R α r ·(∑ f∈F ,β f ·f)·Attn(r), where F image represents the final fused image feature vector, r is the image region index, R represents the set of all regions after the image is divided, α r is the regional weight parameter, F r is the set of all features extracted from the image region r, β fis the importance weight coefficient of feature f in region r, f represents a specific feature vector in region r, Attn(r) is the regional attention score, which is calculated by the following formula: Among them, T is the image pattern set related to user interests, Sim(I r ,P t ) is to calculate the visual features of the image region r and the image pattern P t The similarity function of t It is the picture mode P t Through the above calculation, the features of each area of ​​the image are integrated according to their importance and user attention, and an image feature vector that fully fits the user's interests is obtained.

[0013] Furthermore, in the step of dynamic modeling and tracking of user interests, the frequency and duration of statistical behavior are quantitatively characterized and integrated into a feature vector, using the formula Among them, Q feature (t) represents the quantitative feature value of a behavior at time t, i is the behavior event index, n is the total number of behavior events related to the topic of interest, and f i (t) is the basic behavior index corresponding to behavior event i at time t, ω i (t) is the interest intensity weight of behavioral event i at time t, and its calculation formula is: Among them, t i is the time when behavior event i occurs, Intensity(i) is the initial interest intensity evaluation value of behavior event i, σ is the time decay coefficient used to control the decay speed of the behavior interest intensity weight over time. The longer the time, the more the weight decays, indicating that the influence of behaviors in the distant past on the current interest gradually decreases. δ(tt i ) is the time decay function, which is used to further reflect the influence of behavioral events on current interests over time. The calculation formula is: Among them, τ is the time threshold, ρ is the time decay adjustment parameter. When the time of the behavior event is within the threshold τ from the current time t, it is considered to have a greater impact on the current interest, and the decay function value is 1. After exceeding the threshold, its influence is reduced according to the exponential decay law, thereby highlighting the importance of recent behavior to the current interest, so that the quantitative characteristics of behavior can better fit the real-time evolution of user interests.

[0014] Furthermore, in the user interest dynamic modeling and tracking step, hybrid feature fusion and dynamic weight allocation are realized with the help of a deep neural network architecture, and the weights of each feature are dynamically adjusted in real time according to changes in user behavior to generate a fused feature vector that accurately reflects the evolution of the user's real-time interests. The feature fusion formula is: Among them, F fusion (t) represents the fused user interest feature vector at time t, w d (t) is the weight value of the dth feature dimension at time t, and its calculation formula is: Among them, w d (t-1) is the weight value of the d-th feature dimension at the previous moment, η is the learning rate parameter, ΔB d (t) is the change in user behavior related to the dth feature dimension at time t, D is the total number of feature dimensions, and F d (t) is the feature vector corresponding to the dth feature dimension at time t.

[0015] Furthermore, in the user interest dynamic modeling and tracking step, a deep learning algorithm is selected to construct a user interest evolution model, and the data set is divided and hyperparameters are set. After iterative training, optimization based on the validation set and evaluation based on the test set, the model can predict interest evolution based on the historical feature sequence and process new behavior data in real time for tracking analysis. Specifically, the forget gate weight matrix W in the LSTM model is randomly initialized. f , input gate weight matrix W i , candidate cell state weight matrix W c , output gate weight matrix W o and the forget gate bias vector b f , input gate bias vector b i , candidate cell state bias vector b c , output gate bias vector b o , set the number of hidden layer neurons to h dimensions, determine the initial state of the model, and initialize the hidden state h0 and cell state c0 of each time step to zero vector. For each training sample, time step t = 1 to N, according to formula f t =σ(W f ·[h t-1 ,F fusion (t)]+b f ) calculates the forget gate output, which is used to determine the degree of forgetting the cell state at the previous moment, according to formula i t =σ(W i ·[h t-1 ,F fusion (t)]+b i ) Calculate the input gate output value i t , which determines the degree to which the current new information is integrated into the cell state, according to the formula Calculating candidate cell states Generate potential new information candidates using the formula The forget gate controls the cell state c at the previous momentt-1 The retained information and the new information after the input gate filtering Perform element-wise multiplication and then add to get the updated cell state c at the current time step. t , used to store the historical information of user interest evolution for a long time, through formula o t =σ(W o ·[h t-1 ,F fusion (t)]+b o ) Calculate the output value o of the output gate t , to determine the amount of information output by the current hidden state, according to the formula h t =o t ⊙tanh(c t ), use the output gate to output the value o t For the updated cell state c t After performing the hyperbolic tangent transformation, perform element-wise multiplication to obtain the updated hidden state h at the current time step. t , based on the current hidden state h t , it is converted into a prediction result of the user interest evolution in the current time step through a fully connected layer, compared with the actual label, and the predicted loss value is calculated using the loss function. According to the calculated loss value, the back propagation algorithm is used to start from the last time step, and the gradient of each weight matrix and bias vector with respect to the loss function is calculated in turn. The optimization algorithm is used to update each parameter in the direction of the gradient to reduce the loss value, so that the model gradually fits the evolution law of user interests in the training data. After multiple rounds of iterative training, the test set data is used to evaluate the generalization ability and final performance of the model according to the same calculation process. If the evaluation indicators meet the expected requirements, the trained LSTM model can be used to receive new user behavior data input in real time in the future, track and analyze the evolution of user interests, and provide a basis for making advertising push decisions.

[0016] Furthermore, in the step of making a decision to push dynamic content-adaptive advertisements, the key selling points of advertisements are adjusted using natural language generation technology, using the formula: W s (p) = α × Rel(p, I) + β × Attn(p, B) + γ × Nov(p), where Ws(p) represents the comprehensive weight of the key selling point p when generating advertising copy, α, β, and γ are weight adjustment parameters, α focuses on reflecting the weight affected by the relevance of the selling point and the user's interest topic, β focuses on reflecting the weight affected by the attention based on the user's real-time behavior, and γ considers the contribution of the novelty of the selling point itself to the weight. Rel(p, I) is the relevance score between the key selling point p and the user's current interest topic I, and the calculation formula is: Where w is the word in the key selling point p, |p| represents the number of words contained in the key selling point p, Sim(w,I) is the similarity between the word w and the user's interest topic I, and Attn(p,B) is the attention score of the key selling point p based on the user's recent behavior B. The calculation formula is: Where b represents the user's behavior, B is the recent user behavior set, Hit(p,b) is the indicator function, when behavior b involves content related to key selling point p, Hit(p,b) = 1, otherwise it is 0, Weight(b) is the weight of behavior b, Nov(p) is the novelty score of key selling point p, and the calculation formula is: Among them, Count total (p) is the total number of times the key selling point p appears in the entire advertising material library and the advertising content that the user has been exposed to in the past. P is the set of all the key selling points that can be selected. The weight W calculated based on the above key selling points s (p) Sort, filter and combine selling points, and use natural language generation templates to generate advertising copy content that highlights and best matches the user's current interests.

[0017] Furthermore, in the dynamic content adaptive advertising push decision-making step, the preferential information is dynamically adjusted through a specific algorithm in combination with the user's interest heat and behavior activity factors, and the calculation formula is: Among them, K adj is the adjustment coefficient of the preferential strength, λ, μ, θ are weight distribution parameters, which are used to adjust the influence of user interest, behavior activity and product profit space on the preferential strength adjustment, I h is the user interest heat index, and its calculation formula is: Among them, n is the number of different factors used to measure the user's interest, Weight i is the weight of the i-th factor, which is set according to its importance in reflecting the user's interest. i is the change in the ith factor, A a Is the user behavior activity indicator, and the calculation formula is: Among them, B recent is the user's recent behavior set, B total It is the set of all user behaviors within the statistical scope, Score(b) is the activity score of behavior b, and Pro(P) is the product profit margin indicator, which is determined based on the cost, pricing and market competition of the product or service.

[0018] On the other hand, an advertisement push system based on user behavior analysis includes the following components: a data collection module, a data preprocessing module, a module for building and tracking user interest evolution based on behavior data, a dynamic content adaptive advertisement push decision module, and an advertisement push execution module;

[0019] The data acquisition module is equipped with a multi-platform data interface, which is compatible with data sources of different formats. It is responsible for widely collecting user behavior data from many network platforms and transmitting the collected data to the data preprocessing module;

[0020] The data preprocessing module performs cleaning, denoising and standardization on the raw data from the data acquisition module, outputs high-quality data with a unified format that is convenient for subsequent analysis, and provides a high-quality input data foundation for the subsequent construction of the interest evolution model;

[0021] The user interest evolution module integrates the functions of extracting user interest features, building an interest evolution model, and tracking and analyzing interest evolution. It extracts interest features from preprocessed data to build vectors, generates and trains interest evolution models through time series analysis combined with deep learning algorithms, and tracks and analyzes interest evolution in real time based on new data, and outputs the analysis results to the advertising push decision module.

[0022] The dynamic content adaptive advertising push decision module formulates a detailed and dynamically changeable advertising push strategy based on the analysis results output by the interest evolution module and the current user behavior data obtained in real time, with reference to the advertising resource library, and comprehensively considers various factors, and makes real-time adaptive adjustments to various elements of the advertising content, and passes the strategy to the advertising push execution module;

[0023] The advertisement push execution module pushes advertisements to users according to the formulated strategy using multiple push channels, and transmits the user's feedback information on the advertisements back to the data collection module, thereby realizing the recycling of data and continuous optimization of the system.

[0024] Compared with the prior art, the method and system for advertising push based on user behavior analysis have the following beneficial effects:

[0025] 1. The present invention deeply tracks the entire process of user interest evolution, combines user real-time behavior data, and adaptively adjusts the advertising content, so as to accurately push advertisements that are highly matched with the user's current interests and behaviors. At the same time, it introduces a hybrid feature fusion and dynamic weight allocation mechanism to more comprehensively extract user interest features, and dynamically adjust feature weights according to real-time feedback from user behavior, thereby constructing a feature vector that can more accurately reflect the user's real-time interests, greatly improving the accuracy of advertising push, allowing advertisements to more accurately reach potential target users, and significantly increasing the probability of advertisements being noticed, clicked, and converted by users.

[0026] 2. The present invention ensures that the advertisements received by users are always closely related to their needs by dynamically adjusting the advertisement content, effectively reducing the interference and annoyance caused by irrelevant advertisements, and improving the comfort and satisfaction of users when browsing various types of information. As the interests and behaviors of users change, the advertisement content can be adaptively adjusted in real time to ensure that the advertisements are always consistent with the current needs of users, which not only enhances the user's favorability towards the push platform, but also improves the user's loyalty.

[0027] Other advantages, objectives and features of the present invention will be set forth in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be taught from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0029] Figure 1 The present invention is a flowchart of an advertisement push method based on user behavior analysis.

[0030] Figure 2 A schematic diagram of an advertising push system based on user behavior analysis. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0032] Embodiment 1

[0033] This embodiment describes in detail a specific application of an advertisement push method and system based on user behavior analysis in the field of electronic product e-commerce. Through the present invention, the accuracy and effect of advertisement push are effectively improved, providing strong support for electronic product e-commerce platforms to achieve precision marketing.

[0034] An electronic product e-commerce platform records various user behavior data through its own server, and uses web crawler technology to collect user behavior information from relevant technology information websites, electronic product evaluation forums, etc. For example, it collects records of users browsing different electronic product pages such as mobile phones, tablets, and smart wearable devices on the e-commerce platform (including the specific product details page URL visited, visit time, etc.), checks users' likes and comments on various electronic products (such as users praising a mobile phone for its good camera effect and complaining about a smart watch's insufficient battery life, etc.), obtains users' search history on the platform, as well as purchase records (purchased electronic product models, prices, transaction times, etc.) and past responses to pushed electronic products. The advertising interaction behavior data such as clicks and dwell time of product advertisements are collected using data cleaning rules to remove duplicate page browsing records, eliminate abnormal purchase records caused by network failures (such as order data with abnormally high or low amounts), and stop word filtering is performed on text data (comments, search keywords, etc.) to facilitate subsequent text analysis and extraction of key information. For numerical data such as browsing time and number of clicks, standardization methods are used to standardize them so that these data can participate in subsequent analysis at the same scale. For example, after the time a user spends browsing a mobile phone details page is standardized, it can be better compared with the time data of other users browsing similar products, reflecting their relative attention to the product.

[0035] Through natural language processing technology, we analyze the interests of users in their comments and search keywords, and use the formula Calculate feature weights. For example, for the keyword "long battery life tablet", TF ( w,D ) is the frequency of the keyword in document D (which can be regarded as a collection of documents related to the user's recent search behavior). Assuming that the keyword has been searched 3 times in the past week, the frequency is 3, and P (topic i |w) represents the conditional probability that a document belongs to the topic of "tablet computer battery life" when given the term "long battery life tablet computer". Assume that the probability obtained through topic model analysis is 0.6, ∑ w′∈D P(topici|w′) is the sum of the conditional probabilities of all words in the document about the topic “tablet battery life”, assuming it is 5, λ is set to 0.7 (focusing on topic relevance), and IDF(w) is calculated according to the formula Calculation: Assuming the total number of search documents N is 1000 and the number of documents containing “long battery life tablet” DF(w) is 50, then Substituting into the formula, we can get the feature weight of the keyword, and use it to construct a text feature vector. If many users frequently search for words containing "long battery life" and the word is closely related to the tablet computer battery life theme, then its weight in the text feature vector will be higher, indicating that the user group is more concerned about the battery life function of the tablet computer. For the electronic product pictures browsed by users, first divide the pictures into regions (such as nine-square grid division), analyze the content displayed in each region (such as product appearance details, function buttons, screen display effects, etc.), and for the area showing battery life-related icons or parameters, according to the degree of user attention to the area in similar electronic product pictures (through past user gaze points, stay time and other behavioral data statistics when browsing pictures), the degree of correlation between the features in the area and the battery life function (such as specific electronic display icon features, etc.), use the formula Calculate the regional weights and fuse the features to obtain the image feature vector. Suppose a region displays a battery power icon, the α of the region r (Region weight parameter) is 0.8 (indicating that the user often pays attention to this region) according to the analysis of user historical behavior, β f (Feature Importance Weight Coefficient in Region) For the battery icon feature, it is set to 0.9 (because it is closely related to the battery life function), and Attn(r) is calculated by the formula Calculation, assuming that there is an image mode P of "high power display" t , γ t (Image mode attention weight coefficient) is set to 0.7, Sim(I r ,P t )(region and image pattern similarity) is calculated to be 0.8, and the calculation of other regions is similar. Finally, the image feature vector is fused. If a region is often concerned by users and has a strong correlation with key functions, its weight in the image feature vector is large, reflecting the user's visual interest in the corresponding function. The frequency and duration of users browsing different electronic product pages are counted, and the quantitative behavior of the behavior is determined by combining the user's interaction with the product function at that time (such as whether to click to view the battery life parameters, take photos, and other detailed function introductions) and then according to the formula Calculate the feature value. For example, if a user browses the details page of a certain brand of mobile phone several times in the past month and each time he / she browses the introduction of the camera function in depth, let the behavior event of browsing the introduction of the camera function be i, and f i (t) (the basic behavior indicator corresponding to behavior event i at time t) is the number of views, assuming it is 5 times, ω i (t) According to the formula Calculate, assuming that the time t at which behavioral event i occurs iFor different time points in the past month, Intensity(i) (initial interest intensity evaluation value) is 0.8 based on the comprehensive evaluation of user viewing time, click depth, etc., and σ (time decay coefficient) is set to 0.5, and ω is calculated. i (t), δ(tt i ) According to the formula Calculate, assuming that τ (time threshold) is 10 days, ρ (time decay adjustment parameter) is 5, if the current time t is the same as the last time you viewed the photo function introduction t i The interval is less than 10 days, δ(tt i )=1, otherwise the attenuation value is calculated according to the formula. The quantitative characteristic value of the behavior related to the mobile phone camera function is obtained by comprehensive calculation, reflecting a strong interest. At the same time, the weight of recent behavior will be higher than that of longer-term behavior, reflecting the timeliness of interest. The user behavior changes are monitored in real time. For example, it is found that the user's recent search and browsing behaviors related to the mobile phone camera function have increased. According to the formula Improve the weights of feature dimensions such as text, image, and behavior quantification related to the camera function. Suppose a text feature dimension d related to the camera function has a weight w at the last moment t-1. d (t-1) is 0.3, η (learning rate parameter) is set to 0.1, ΔB d (t) (the change in user behavior related to this feature dimension) is the recent increase in the frequency of searches related to the camera function, assuming it is 2 times, and D (the total number of feature dimensions) is 10. The weight w of this feature dimension at the current time t is calculated. d (t), highlighting these features that are closely related to the changes in user interests in the process of fusion, through the formula The features of different dimensions are weighted and fused into a fused feature vector that can better reflect the real-time evolution of users' interests. For example, the fused feature vector can more accurately reflect the changing trend of users' interests from focusing on the battery life of mobile phones to focusing more on the camera function.

[0036] The fused feature vectors are sampled at weekly intervals to construct time series data. For example, a feature vector is generated every week to reflect the user's interest in electronic products in that week. These vectors arranged in chronological order are used as training data. The long short-term memory network (LSTM) is used to train the user interest evolution model. During the training process, the forget gate f t =σ(W f ·[h t-1 ,F fusiom (t)]+b f ), input gate i t =σ(W i ·[h t-1 ,F fusion(t)]+b i ), candidate cell states Cell status update Output gate o t =σ ( W o ·[h t-1 ,F fusiom (t)]+b o ), hidden state update h t =o t ⊙tanh(c t ) and other related formulas to continuously adjust the model parameters (weight matrices and bias vectors, etc.). Assume that at a certain time step t, the previous hidden state h t-1 is [0.2, 0.3, …, 0.1] (the dimension is set to 10), and the fused feature vector F fusiom (t) is [0.5, 0.4, …, 0.3] (the dimension is set to 10), W f (Forget gate weight matrix), W i (input gate weight matrix), W c (candidate cell state weight matrix), W o (Output gate weight matrix) is continuously adjusted according to the initialization and training process, b f (forget gate bias vector), b i (input gate bias vector), b ε (candidate cell state bias vector), b o (output gate bias vector) is also adjusted accordingly, and the forget gate output f is calculated by these formulas t , input gate output i t , candidate cell states Update cell state c t 、Yu goes out to output o t and the hidden state h t , so that the model can accurately capture the evolution of users' interest in electronic products from germination, development to transfer (such as from focusing on tablets to preferring mobile phones) according to the historical interest feature sequence, and can receive new user behavior data in real time. For example, if a user suddenly starts searching for a large amount of smart watch-related information today, the model can promptly track the new changing trend of users' interest in the field of smart watches.

[0037] On the web page of the e-commerce platform, by embedding points on key elements of the electronic product details page (such as product function introduction module, purchase button, picture display area, etc.) on mobile phones, tablet computers, etc., the user's operation behavior information (such as clicking to view photo samples, hovering the mouse over the battery life parameter introduction, and other operations and operation time) is captured in real time. In the mobile application, the native interactive monitoring interface is used to collect the user's browsing of different electronic product categories, viewing product reviews, and other operation behaviors in the application. The back-end server integrates these real-time behavior data with the output of the user interest evolution model, such as the current user's interest in the mobile phone camera function is high and the interest is rising, to form comprehensive user status information. For example, it is known that the user is not only interested in the camera function, but also is viewing the camera samples of a certain mobile phone, indicating that he has a further understanding of the camera performance of the specific product. According to the integrated comprehensive user status information, the mobile phone advertising copy materials are selected in the advertising resource library, and the formula W is used to select the mobile phone advertising copy materials. s (p) = α × Rel (p, I) + β × Attn (p, B) + γ × Nov (p) to determine the weight of each key selling point (such as high-pixel lens, multiple camera modes, night scene shooting optimization, etc.) in the copy. Assume that for the key selling point p of "night scene shooting optimization", α (focusing on the correlation between the selling point and the user's interest topic) is set to 0.5, and Rel (p, I) is calculated according to the formula Note that, assuming that the similarity between the word "night scene shooting optimization" in selling point p and the related words in the user's interest topic "mobile phone camera function" (such as "night scene shooting low noise", etc.) Sim(w,I) is calculated by the word vector model to obtain a higher value, |p| (the number of words contained in the key selling point p) is 3, and the value of Rel(p,I) is calculated, β (focusing on reflecting the influence weight of attention based on user real-time behavior) is set to 0.3, Attn(p,B) is calculated according to the formula Calculation: Assuming that the user has viewed the night scene samples taken by the mobile phone many times recently (behavior b), Hit(p,b) = 1, Weight(b) (the weight of behavior b) is set to 0.8 according to the importance of behavior, and the value of Attn(p,B) is calculated. γ (considering the contribution of the novelty of the selling point itself to the weight) is set to 0.2. Nov(p) is calculated according to the formula Calculate, assuming that the number of times "night scene shooting optimization" appears in the entire advertising material library and past advertisements is relatively small, calculate the value of Nov(p), and comprehensively calculate the weight W of this selling point s(p), if the user is currently interested in the camera function and has recently frequently viewed night scene photography-related content, the weight of the selling point of "night scene photography optimization" will increase. This selling point will be highlighted when the copy is generated, and advertising copy content that fits the user's interests will be generated, such as "This phone is equipped with an ultra-sensitive lens and has a powerful night scene photography optimization function, allowing you to easily capture the dazzling night scene." Image editing technology is used to optimize the picture display details, combined with the user's interest heat (currently the interest in the mobile phone camera function is strong, and the interest heat index value is high), behavioral activity (such as frequent browsing of different mobile phone camera function introduction pages recently, the behavioral activity index value is also high) and product profit space (mobile phone products have a certain profit margin based on cost and pricing), through the formula Calculate the adjustment coefficient of the preferential strength, assuming that λ (the weight of adjusting the influence of user interest on the preferential strength adjustment) is set to 0.4, I h According to the formula Calculation, assuming n (the number of factors that stimulate user interest) is 3, including the duration of attention to the camera function ( Set to 0.4, Interest i Quantified by the number of days of duration), the frequency of recent related behaviors ( Set to 0.3, Interest i is the actual frequency), the weight of the interest feature in the fusion feature vector (Weight, set to 0.3, is the actual proportion), and I h The value of μ (the influence weight of the activity of the adjustment behavior on the adjustment of the preferential intensity) is set to 0.3, and A a According to the formula Calculate, assuming B recemt (Recent behavior set) includes browsing the camera function page, clicking to view sample photos, etc. Score (b) (behavior activity score) is set according to the complexity and depth of the behavior, and A is calculated. a The value of θ (the weight of adjusting the product profit margin on the discount adjustment) is set to 0.3, and Pro (P) (product profit margin index) is determined according to the mobile phone cost, pricing and market competition, and is assumed to be 0.6. K is calculated. adj, assuming that the calculated coefficient value is moderate, adjust the original discount intensity (such as from 10% off to 20% off), and appropriately increase the gifts related to photography (such as giving away online courses on mobile photography tutorials), adjust the service duration, such as extending the free cloud storage. Analyze the user's historical behavior data and find that users are usually more active in browsing electronic product information between 7pm and 9pm, and the current user's interest in the mobile phone camera function is in a rising and stable period. Combined with these factors, determine to push mobile phone advertisements containing the above-mentioned adaptively adjusted advertising content (copy, pictures, discount information) at a higher frequency (such as once every two days) during this time period to increase the probability of the advertisement being noticed and clicked by users, thereby promoting product sales.

[0038] According to the formulated advertising push strategy, advertising information is pushed through multiple channels such as mobile application push (using the push service of the mobile operating system to push advertising notifications to the user's mobile phone), web pop-up push (when the user visits the e-commerce platform webpage and meets the triggering conditions, such as staying on the electronic product classification page for more than a certain period of time, a mobile phone advertising pop-up window will pop up), and email push (for registered members and users with purchasing intentions, emails containing mobile phone advertising content are sent). During the push process, the push status is monitored in real time, and relevant information such as whether the push of each channel is successful, the push time, and the delivered user are recorded. For example, the mobile push message delivery receipt shows that a user has successfully received the advertising notification, and the web pop-up push statistics the number of display times, etc., to facilitate the subsequent analysis of the push effect, and the user's The back-end server also records the transaction-related data in detail (such as the model and price of the purchased mobile phone). After collecting and collating this feedback information, it is sent back to the data collection link to continuously optimize the user interest evolution model (for example, according to the user's click feedback on different mobile phone advertisements, adjust the weight distribution of the camera-related features in the model) and the advertising push strategy (according to the user's feedback on the push channel, push time, etc., adjust the subsequent push channel selection and push time arrangement, etc.), form a data closed loop, and continuously improve the accuracy and effectiveness of advertising push, and store preferential information such as photo storage time) to make the Youjia package more attractive and motivate users to buy.

[0039] In summary, the present invention effectively improves the accuracy and effectiveness of e-commerce platform advertising push, provides strong support for electronic product e-commerce platforms to achieve precision marketing, and is expected to significantly improve user engagement, purchase conversion rate, and user satisfaction, while optimizing the configuration and utilization efficiency of advertising resources.

[0040] Embodiment 2

[0041] This embodiment describes in detail a specific application of an advertisement push method and system based on user behavior analysis in the field of travel services. Through the present invention, not only can advertisement push be more in line with user needs, user participation and booking conversion rate can be improved, but also travel companies can be helped to better understand user needs and optimize products and services.

[0042] The travel booking platform records users' behavioral data through its own system, and uses API to collect relevant information from travel guide sharing websites, social media travel topic sections, etc. The collected content includes the travel destinations searched by users on the booking platform (such as keywords such as "coastal city tourism" and "ancient town tour" and search time), the travel route details pages browsed (specific route URL visited, browsing time, etc.), the booked hotels and transportation service information (booked hotel star rating, room type, transportation travel time and method, etc.), the travel guide articles browsed on the guide website, the pictures and text content of tourist attractions liked and commented on, as well as the past clicks on pushed travel advertisements (such as travel route advertisements, hotel discount advertisements, etc.), the length of stay and other advertising interaction behavior data, remove invalid search terms caused by misoperation, delete repeated travel route browsing records (data of repeatedly viewing the same route multiple times in a short period of time is only retained once), and unify the encoding and filter stop words for text data (guide articles, comment content, etc.) to facilitate subsequent text analysis to extract key travel points of interest.

[0043] Natural language processing technology is used to analyze the travel guide articles and comments browsed by users, and text features are extracted based on the text feature representation formula of comprehensive theme and frequency weight. For example, for words such as "beach bathing beach" and "seafood cuisine" that frequently appear in articles browsed by users and are closely related to the theme of "coastal city tourism", weights are calculated based on factors such as their word frequency, conditional probability under the theme, and inverse document frequency, and a text feature vector is constructed to reflect the user's interest in specific elements of coastal city tourism. If many guides mention it and it is strongly related to the seaside theme, the weight of the relevant words in the vector will be higher, reflecting the user group. Based on the user's attention to seaside tourism-related content, the image regions are divided for the tourist attraction pictures browsed by users, and the tourist features displayed in each region are analyzed. According to the degree of attention paid to each region when users browsed pictures in the past and the degree of correlation between the features in the region and the specific tourist theme, the image feature representation formula that integrates the regional attention and feature importance is used to calculate the regional weight and fuse the features to obtain the image feature vector, highlighting the tourist feature areas that the user is visually interested in. For example, if the seaside sunrise beauty area is often paid attention to and is strongly associated with the seaside theme, its weight in the image feature vector is large, reflecting the user's interest in this type of scenery. The interest is determined by the frequency and duration of the user's search for tourist destinations, browsing tourist routes, booking tourist services, etc. The interest intensity weight of the behavior is determined in combination with the user's attention to the attractions and hotel features in the route at that time. Then, the characteristic value is calculated according to the behavior quantitative feature representation formula that takes into account the interest intensity and time decay. For example, if a user has searched for coastal city tourism many times in recent months and has focused on luxury coastal hotels with private beaches each time, then the quantitative characteristic value of the behavior related to high-end hotels in coastal city tourism will be higher, reflecting a stronger interest preference. At the same time, the weight of recent behavior is higher than that of long-term behavior, reflecting the dynamic change of interest over time. The user behavior changes are monitored in real time. For example, it is found that the user's recent search and browsing behaviors related to ancient town tourism have increased. According to the dynamic weight calculation formula based on user behavior feedback, the weights of feature dimensions such as text, image, and behavior quantification related to ancient town tourism are increased. In the feature fusion process, these features that reflect the change of user interest are highlighted. Through the multi-dimensional feature fusion formula based on dynamic weights, the features of different dimensions are weighted and fused into a fusion feature vector that can better reflect the real-time evolution of the user's interest. For example, it can accurately present the trend of users' interest transfer from paying attention to coastal city tourism to being more enthusiastic about ancient town tourism.

[0044] The fused feature vectors are sampled at monthly time intervals to construct time series data, and a feature vector that reflects the user's travel interest status for that month is formed every month. These vectors are used as training data, and the long short-term memory network (LSTM) is used to train the user interest evolution model. During the training process, the model parameters are continuously adjusted according to relevant formulas such as the forgetting gate, input gate, candidate cell state, cell state update, and output gate, so that the model can accurately capture the evolution of users' travel interests from exploring different destinations, determining preferences to possible interest expansion, and can receive new user behavior data in real time. For example, if a user suddenly starts to search for a large amount of grassland tourism-related information, the model can promptly track the new changing trend of users' interests in the field of grassland tourism.

[0045] On the web page of the travel booking platform, JavaScript embedding is performed on key elements such as the travel destination recommendation page, travel route display area, and hotel booking section to capture user operation behavior information in real time. The native interactive monitoring interface is used in the mobile application to collect user operation behaviors such as browsing different travel categories and viewing user reviews in the application. The back-end server integrates these real-time behavior data with the results output by the user interest evolution model, such as the current user interest in ancient town tours is high and in a stable stage, to form comprehensive user status information. For example, it knows that the user is not only interested in ancient town tours, but is also checking out the special folk activity routes of a certain ancient town, indicating that he has a deep understanding of the specific content of the ancient town tour.

[0046] According to the integrated user status information, the advertising copy materials for the ancient town tour are screened in the advertising resource library, and the weights of the key selling points (such as the historical charm of the thousand-year-old ancient town, the unique folk performances, the special food experience, etc.) in the copy are determined by using the copy generation weight calculation formula based on the theme relevance and user attention. If the user is currently very interested in the folk performances in the ancient town and has frequently viewed the relevant introduction content of the folk performances recently, the weight of the selling point of "unique folk performances" will increase. When the copy is generated, this selling point will be highlighted to generate advertising copy content that fits the user's interests, such as "Come to this thousand-year-old ancient town, immerse yourself in the unique folk performances, and feel the unique charm of traditional culture." For advertising pictures, the pictures to be displayed are determined according to the user's interests (such as choosing to display pictures of the ancient town folk performance scene, the ancient street night view, etc.), and the image cropping and scaling coefficient calculation formula based on the focus on the user's interest area is used to analyze the importance of areas such as folk performances and ancient street features in the picture, changes in user attention to the area, and other factors, and calculate the cropping coefficient and scaling coefficient. Coefficient, crop the picture (focus on the wonderful moments of folk performances and the characteristic buildings of the ancient street) and zoom (appropriately enlarge the details), add annotations to explain the performance time, architectural history and other information, optimize the picture display effect, attract users to pay attention to the characteristic tourism elements of the ancient town, combine the user's interest heat (currently there is a strong interest in ancient town tours, and the interest heat index value is relatively high), behavioral activity (such as frequent browsing of different ancient town tourism route introduction pages recently, and the behavioral activity index value is also relatively high) and product profit space (tourism route products have a certain profit space based on cost and pricing), calculate the discount intensity adjustment coefficient through the calculation formula of the discount intensity adjustment coefficient based on comprehensive factors of interest and behavior, assuming that the calculated coefficient value is moderate, adjust the original discount intensity (such as from 10% off to 20% off), and appropriately increase the gifts related to the ancient town tour (such as giving away characteristic handicrafts of the ancient town), adjust the service time (such as extending the service time of the ancient town tour guide), and other preferential information to make the discount package more attractive and encourage users to book travel services.

[0047] According to the formulated advertising push strategy, advertising information is pushed through multiple channels such as mobile application push, web pop-up push, and email push. During the push process, the push status is monitored in real time, and relevant information such as whether the push of each channel is successful, the push time, and the delivered users are recorded. For example, the delivery report returned by the mobile push service shows that a user has successfully received the ancient town tour advertising push notification, and the web pop-up push statistics module records the number of displays and clicks on each pop-up window. The email push platform feedbacks the sending status of the email (whether it is successfully sent, whether it is opened by the user, etc.), which is convenient for subsequent in-depth analysis of the push effect and evaluation of the effectiveness of different channels for different user groups. The user's interactive behavior with the advertisement is recorded through JavaScript code on the advertisement display page, such as whether to click on the advertisement to enter the ancient town tour route details page, the length of time to stay on the details page (if the user stays for a long time, it may indicate a high interest in the route), whether to make inquiries (such as inquiring about the arrangements of folk activities in the ancient town, accommodation conditions, etc.), and whether the reservation is completed The back-end server records transaction-related data in detail, including the booked ancient town tour route, hotel room type, travel date, and payment amount. At the same time, it collects user evaluation feedback after booking (such as satisfaction with the ancient town tour experience, suggestions for route arrangements, etc.), collects and organizes these feedback information and sends them back to the data collection link for continuous optimization of the user interest evolution model (for example, if it is found that users have more inquiries about a certain type of ancient town folk activities, the weight of the features related to the folk activities will be increased accordingly in the model) and advertising push strategies (if it is found that users in a certain area have good feedback on the ancient town tour advertisements pushed by email, increase the frequency of email push to users in that area, and adjust the route recommendations and discount combinations pushed according to the user's booking status and feedback on different routes), form a data closed loop, and continuously improve the accuracy and effectiveness of advertising push, provide users with travel advertising recommendation services that are more in line with their needs, and help tourism companies improve marketing effectiveness and customer satisfaction.

[0048] To sum up, through the present invention, not only can advertising push be more in line with user needs, user engagement and booking conversion rate be improved, but it can also help tourism companies better understand user needs, optimize products and services, enhance brand image and market competitiveness, and show strong application value and broad development prospects in the field of tourism travel services.

[0049] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. An advertisement push method based on user behavior analysis, characterized in that: The method comprises the following specific steps: Data collection and preprocessing: Using web crawler technology, we set corresponding crawling rules and scopes for social media platforms, professional interest forums, and e-commerce platforms, collect HTML structure information of web pages, parse users' browsing history and like and comment behaviors, and obtain users' search history, purchase history, and advertising interaction-related data within the platform by calling the APIs provided by each platform. We then clean, denoise, and standardize the data. Dynamic modeling and tracking of user interests: Use natural language processing technology to extract text features from text behavior data, use image recognition technology to obtain image features from image behavior data, and at the same time count the behavior frequency and duration behavior quantitative features and integrate them into feature vectors. Use deep neural network architecture to achieve hybrid feature fusion and dynamic weight allocation. Dynamically adjust the weights of each feature according to user behavior changes in real time to generate a fused feature vector that reflects the user's real-time interest evolution, and sample it at fixed time intervals to construct time series data. Use deep learning algorithms to build a user interest evolution model, divide the data set and set hyperparameters. After iterative training, optimization based on the validation set and evaluation of the test set, the model can predict interest evolution based on historical feature sequences and process new behavior data in real time for tracking and analysis; Dynamic content adaptive advertising push decision-making: By capturing user operation event information in real time through key element embedding points in web pages and mobile applications, the backend integrates it with the output of the user interest evolution model to form comprehensive user status information, and searches and matches in the advertising resource library based on this information, uses natural language generation technology to adjust the key selling points of advertising copy, uses image editing technology to optimize the picture display details, and dynamically adjusts the discount information through specific algorithms based on user interest popularity and behavioral activity factors. The timing and frequency of advertising push are determined by comprehensively considering the stability of user interests, behavioral activity and business goals, forming a complete dynamic content adaptive advertising push strategy; Advertisement push execution and feedback: According to the formulated advertisement push strategy, the corresponding interfaces of mobile application push service, web pop-up window push and email push channels are called to accurately push the advertisement information to users after packaging it in the required format of each channel, and the push status is monitored in real time. Meanwhile, the feedback information of the user's interaction with the advertisement is recorded on the advertisement display page and the back-end server, and then sorted and stored according to the specifications and sent back to the data collection link for continuous optimization of the user interest evolution model and advertisement push strategy, thereby forming a data closed loop to ensure that advertisement push closely follows the changes in user behavior.

2. The method for pushing advertisements based on user behavior analysis according to claim 1, characterized in that: In the user interest dynamic modeling and tracking step, natural language processing technology is used to extract text features from text behavior data, and image recognition technology is used to obtain image features from image behavior data. At the same time, the behavior frequency and duration behavior quantitative features are counted and integrated into feature vectors. For extracting text features from text behavior data, the formula is used. The contribution weight of each word to the user's interest is determined by comprehensively considering the topic-related importance and frequency of occurrence, and then the text feature vector is constructed, where TF weighted ( w,D ) It represents the feature weight value of word w in document D after considering the comprehensive weight of topic and frequency, TF ( w,D ) is the frequency of word w in document D, P ( topic i ∣w ) Indicates that given word n, the document belongs to the topic topic i The conditional probability of w′∈D P ( topic i ∣w ′) is a list of all words in document D about the topic i The conditional probability summation, λ is the weight adjustment parameter, IDF ( w is the inverse document frequency of word w, calculated as Where N is the total number of documents and DF(w) is the number of documents containing word w.

3. The method for pushing advertisements based on user behavior analysis according to claim 1, characterized in that: In the user interest dynamic modeling and tracking step, the algorithm formula used to obtain image features from image behavior data using image recognition technology is: image =∑ r∈R α r ·(∑ f∈F, β f ·f)·Attn(r), where F image represents the final fused image feature vector, r is the image region index, R represents the set of all regions after the image is divided, α r is the regional weight parameter, F r is the set of all features extracted from the image region r, β f is the importance weight coefficient of feature f in region r, f represents a specific feature vector in region r, Attn(r) is the regional attention score, which is calculated by the following formula: Among them, T is the image pattern set related to user interests, Sim(I r ,P t ) is to calculate the visual features of the image region r and the image pattern P t The similarity function of t It is the picture mode P t Through the above calculation, the features of each area of ​​the image are integrated according to their importance and user attention, and an image feature vector that fully fits the user's interests is obtained.

4. The method for pushing advertisements based on user behavior analysis according to claim 1, characterized in that: In the user interest dynamic modeling and tracking step, the frequency and duration of statistical behavior are quantified and integrated into feature vectors, using the formula Among them, Q feature (t) represents the quantitative feature value of a behavior at time t, i is the behavior event index, n is the total number of behavior events related to the topic of interest, and f i (t) is the basic behavior index corresponding to behavior event i at time t, ω i( t ) is the interest intensity weight of behavioral event i at time t, and its calculation formula is: Among them, t i is the time when behavioral event i occurs, Intensity ( i ) is the initial interest intensity evaluation value of behavior event i, σ is the time decay coefficient used to control the decay speed of the behavior interest intensity weight over time. The longer the time, the more the weight decays, indicating that the influence of the behavior in the distant past on the current interest gradually decreases, δ ( tt i) is a time decay function, which is used to further reflect the influence of behavioral events on current interests over time. The calculation formula is: Among them, τ is the time threshold, ρ is the time decay adjustment parameter. When the time of the behavior event is within the threshold τ from the current time t, it is considered to have a greater impact on the current interest, and the decay function value is 1. After exceeding the threshold, its influence is reduced according to the exponential decay law, thereby highlighting the importance of recent behavior to the current interest, so that the quantitative characteristics of behavior can better fit the real-time evolution of user interests.

5. The method for pushing advertisements based on user behavior analysis according to claim 1, characterized in that: In the user interest dynamic modeling and tracking step, hybrid feature fusion and dynamic weight allocation are realized with the help of deep neural network architecture, and the weights of each feature are dynamically adjusted in real time according to the changes in user behavior to generate a fused feature vector that accurately reflects the real-time evolution of the user's interests. The feature fusion formula is: Among them, F fusion( t ) represents the fused user interest feature vector at time t, w d( t ) is the weight value of the dth feature dimension at time t, and its calculation formula is: Among them, w d( t-1 ) is the weight value of the dth feature dimension at the previous moment, η is the learning rate parameter, ΔB d( t ) is the change in user behavior related to the dth feature dimension at time t, D is the total number of feature dimensions, and F d( t is the feature vector corresponding to the dth feature dimension at time t.

6. The method for pushing advertisements based on user behavior analysis according to claim 1, characterized in that: In the user interest dynamic modeling and tracking step, a deep learning algorithm is selected to build a user interest evolution model, and the data set is divided and hyperparameters are set. After iterative training, optimization based on the validation set and evaluation based on the test set, the model can predict interest evolution based on the historical feature sequence and process new behavior data in real time for tracking analysis. Specifically, the forget gate weight matrix W in the LSTM model is randomly initialized. f , input gate weight matrix W i , candidate cell state weight matrix W c , output gate weight matrix W o and the forget gate bias vector b f , input gate bias vector b i , candidate cell state bias vector b c , output gate bias vector b o , set the number of hidden layer neurons to h dimensions, determine the initial state of the model, and initialize the hidden state h0 and cell state c0 of each time step to zero vector. For each training sample, time step t = 1 to N, according to formula f t =σ(W f ·[h t-1 ,F fusion (t)]+b f ) calculates the forget gate output, which is used to determine the degree of forgetting the cell state at the previous moment, according to formula i t =σ(W i ·[h t-1 ,F fusion (t)]+b i ) Calculate the input gate output value i t , which determines the degree to which the current new information is integrated into the cell state, according to the formula Calculating candidate cell states Generate potential new information candidates using the formula The forget gate controls the cell state c at the previous moment t-1 The retained information and the new information after the input gate filtering Perform element-wise multiplication and then add to get the updated cell state c at the current time step. t , used to store the historical information of user interest evolution for a long time, through formula o t =σ(W o ·[h t-1 ,F fusion (t)]+b o ) Calculate the output value o of the output gate t , to determine the amount of information output by the current hidden state, according to the formula h t =o t ⊙tanh(c t ), use the output gate to output the value o t For the updated cell state c t After performing the hyperbolic tangent transformation, perform element-wise multiplication to obtain the updated hidden state h at the current time step. t , based on the current hidden state h t , it is converted into a prediction result of the user interest evolution in the current time step through a fully connected layer, compared with the actual label, and the predicted loss value is calculated using the loss function. According to the calculated loss value, the back propagation algorithm is used to start from the last time step, and the gradient of each weight matrix and bias vector with respect to the loss function is calculated in turn. The optimization algorithm is used to update each parameter in the direction of the gradient to reduce the loss value, so that the model gradually fits the evolution law of user interests in the training data. After multiple rounds of iterative training, the test set data is used to evaluate the generalization ability and final performance of the model according to the same calculation process. If the evaluation indicators meet the expected requirements, the trained LSTM model can be used to receive new user behavior data input in real time in the future, track and analyze the evolution of user interests, and provide a basis for making advertising push decisions.

7. The method for pushing advertisements based on user behavior analysis according to claim 1, characterized in that: In the dynamic content adaptive advertising push decision-making step, natural language generation technology is used to adjust the key selling points of the advertising copy, using the formula: W s( p ) =α×Rel ( p,I ) +β×Attn ( p,B ) +γ×Nov ( p ) , where W s( p ) Represents the comprehensive weight of the key selling point p when generating advertising copy. α, β, and γ are weight adjustment parameters. α focuses on reflecting the weight affected by the relevance of the selling point to the user's interest topic, β focuses on reflecting the weight affected by the attention based on the user's real-time behavior, and γ considers the contribution of the novelty of the selling point itself to the weight. ( p,I ) is the relevance score between the key selling point p and the user's current interest topic I, and the calculation formula is: Among them, w is the word in the key selling point p, | p | Indicates the number of words contained in the key selling point p, Sim ( w,I ) is the similarity between word w and user interest topic I, Attn ( p,B ) The key selling point p is the attention score based on the user's recent behavior B, and the calculation formula is: Among them, b represents the user's behavior, B is the recent user behavior set, Hit ( p,b ) Is an indicator function. When behavior b involves content related to key selling point p, Hit ( p,b ) =1, otherwise 0, Weight ( b ) is the weight of action b, Nov ( p ) is the novelty score of the key selling point p, calculated as: Among them, Count total( p ) is the total number of times the key selling point p appears in the entire advertising material library and the advertising content that the user has been exposed to in the past. P is the set of all the optional key selling points. The weight W is calculated based on the above key selling points. s( p ) , sort, filter and combine selling points, and use natural language generation templates to generate advertising copy content that highlights and best matches the user's current interests.

8. The method for pushing advertisements based on user behavior analysis according to claim 1, characterized in that: In the dynamic content adaptive advertising push decision-making step, the preferential information is dynamically adjusted through a specific algorithm in combination with the user's interest heat and behavior activity factors, and the calculation formula is: Among them, K adj is the adjustment coefficient of the preferential strength, λ, μ, θ are weight distribution parameters, which are used to adjust the influence of user interest, behavior activity and product profit space on the preferential strength adjustment, I h is the user interest heat index, and its calculation formula is: Among them, n is the number of different factors used to measure the user's interest, Weight i is the weight of the i-th factor, which is set according to its importance in reflecting the user's interest. i is the change in the ith factor, A a Is the user behavior activity indicator, and the calculation formula is: Among them, B recent is the user's recent behavior set, B total It is the set of all user behaviors within the statistical range. Score ( b ) is the activity score of behavior b, Pro ( P ) It is an indicator of product profit margin, which is determined based on factors such as product or service cost, pricing, and market competition.

9. An advertisement push system based on user behavior analysis, applicable to an advertisement push method based on user behavior analysis according to any one of claims 1 to 8, characterized in that: The system includes the following components: a data collection module, a data preprocessing module, a module for building and tracking user interest evolution based on behavior data, a dynamic content adaptive advertising push decision module, and an advertising push execution module; The data acquisition module is equipped with a multi-platform data interface, which is compatible with data sources of different formats. It is responsible for widely collecting user behavior data from many network platforms and transmitting the collected data to the data preprocessing module; The data preprocessing module performs cleaning, denoising and standardization on the raw data from the data acquisition module, outputs high-quality data with a unified format that is convenient for subsequent analysis, and provides a high-quality input data foundation for the subsequent construction of the interest evolution model; The user interest evolution module integrates the functions of extracting user interest features, building an interest evolution model, and tracking and analyzing interest evolution. It extracts interest features from preprocessed data to build vectors, generates and trains interest evolution models through time series analysis combined with deep learning algorithms, and tracks and analyzes interest evolution in real time based on new data, and outputs the analysis results to the advertising push decision module. The dynamic content adaptive advertising push decision module formulates a detailed and dynamically changeable advertising push strategy based on the analysis results output by the interest evolution module and the current user behavior data obtained in real time, with reference to the advertising resource library, and comprehensively considers various factors, and makes real-time adaptive adjustments to various elements of the advertising content, and passes the strategy to the advertising push execution module; The advertisement push execution module pushes advertisements to users according to the formulated strategy using multiple push channels, and transmits the user's feedback information on the advertisements back to the data collection module, thereby realizing the recycling of data and continuous optimization of the system.

Images