An advertisement pushing method and system based on user browsing behavior

Abstract

This application provides a method and system for ad push based on user browsing behavior, belonging to the field of data processing technology. First, this application simultaneously acquires the user's historical search terms and current page interaction behavior, including click weights and dwell time, and constructs an intent vector representing long-term preferences and a behavior vector reflecting current interests accordingly. Next, the two vectors are combined into an observation matrix, and a QR decomposition technique is used to extract an orthogonal basis vector representing the user's independent, core intent as a reference standard. Subsequently, the grey relational coefficient and correlation degree between this reference vector and the attribute vectors of each candidate ad are calculated to quantify the degree of matching between the two. Finally, the ads with the highest correlation are selected and accurately pushed to the user. This application achieves accurate mining of the user's core intent and matching and pushing highly relevant ads in complex page environments.

Description

Technical Field

[0001] This application belongs to the field of data processing technology, and in particular relates to an advertising push method and system based on user browsing behavior. Background Technology

[0002] User browsing behavior-based ad delivery is a crucial technology in digital marketing. It recommends relevant ads by analyzing user actions on web pages. This method can improve ad click-through rates and conversion rates, bringing significant commercial value to e-commerce platforms and demonstrating broad application prospects.

[0003] Existing ad delivery methods typically collect user browsing data, such as page clicks, dwell time, and historical search records. These methods often employ weighted summation or user profiling to linearly overlay users' real-time behavior with historical preferences, forming a comprehensive interest profile, and then matching ads accordingly.

[0004] However, in complex web page environments, there may be interference or even conflict between users' immediate interactive behaviors and long-term historical preferences. Simple linear superposition is insufficient to effectively distinguish between users' core intentions and secondary or incidental browsing behaviors. Therefore, existing technologies suffer from the technical problem of low relevance in ad delivery due to the inability to accurately separate and extract users' independent core intentions in complex browsing scenarios. Summary of the Invention

[0005] The purpose of this application is to provide an advertising push method and system based on user browsing behavior, so as to solve the problem of low relevance of advertising push in the prior art.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides an advertising push method based on user browsing behavior, comprising:

[0007] Simultaneously acquire the user's historical search term sequence and the click weight and dwell time percentage of each interactive element in the target page. The dwell time percentage is determined based on the ratio of the dwell time of the interactive element in the visible area to the total page browsing time.

[0008] Based on click weight and dwell time ratio, construct the behavior vector of each interactive element in the preset semantic space, and obtain the intent vector in the preset semantic space by discretizing the occurrence frequency of each search term in the historical search term sequence.

[0009] An observation matrix is ​​constructed based on behavior vectors and intention vectors. An orthogonal matrix is ​​obtained by performing QR decomposition on the observation matrix, and an orthogonal basis vector is extracted from the orthogonal matrix as a reference vector.

[0010] Extract the attribute vector of each advertisement to be pushed in the preset semantic space from the preset candidate advertisement library, calculate the absolute difference between the reference vector and each attribute vector in each dimension to obtain the difference sequence, and use the preset resolution coefficient to normalize the difference sequence to obtain the gray relation coefficient sequence of each advertisement to be pushed.

[0011] Based on the grey relation coefficient sequence, the grey relation degree is obtained through weighted mapping. The advertisement with the highest grey relation degree is selected from the candidate advertisement library as the target advertisement for push to the target page.

[0012] Optionally, based on click weight and dwell time ratio, a behavior vector for each interactive element in a preset semantic space is constructed, including:

[0013] Extract the business attributes of each interactive element and determine the coordinate positions corresponding to the business attributes in a preset semantic space;

[0014] The attention intensity value of each interactive element is obtained by calculating the product of click weight and dwell time ratio.

[0015] The attention intensity value is filled into the feature bits corresponding to the coordinate positions in the preset baseline behavior sequence, and the other positions in the preset baseline behavior sequence except for the feature bits are set to zero to obtain the behavior vector.

[0016] Optionally, by discretizing and mapping the frequency of occurrence of each search term in the historical search term sequence, an intent vector in a preset semantic space is obtained, including:

[0017] Calculate the cumulative frequency of different search terms in the historical search term sequence to obtain the term frequency score for each search term;

[0018] Extract the business category to which each search term belongs, and determine the coordinate position corresponding to the business category in the preset semantic space;

[0019] The word frequency score is filled into the feature bits corresponding to the coordinate positions in the preset baseline intent sequence, and zeroing is performed on the other positions in the preset baseline intent sequence except for the feature bits to obtain the intent vector.

[0020] Optionally, an observation matrix is ​​constructed based on the behavior vector and the intention vector. An orthogonal matrix is ​​obtained by performing QR decomposition on the observation matrix, and an orthogonal basis vector is extracted from the orthogonal matrix as a reference vector, including:

[0021] The behavior vector and intent vector of each interactive element are combined in a pre-defined order to obtain the observation matrix;

[0022] By performing an orthogonal projection transformation on the observation matrix, an orthogonal matrix is ​​obtained, consisting of multiple mutually perpendicular unit vectors with a magnitude of one.

[0023] Extract the unit vector located in the first column of the orthogonal matrix as the orthogonal basis vector, and determine the orthogonal basis vector as the reference vector.

[0024] Optionally, the attribute vector of each advertisement to be pushed in a preset semantic space is extracted from a preset candidate advertisement library, and the absolute difference between the reference vector and each attribute vector in each dimension is calculated to obtain a difference sequence, including:

[0025] Extract the attribute tags of each advertisement to be pushed from the candidate advertisement library. Based on the attribute tags, according to the preset mapping relationship between attribute tags and component values, obtain the component values ​​corresponding to each attribute tag in the preset semantic space. Combine all component values ​​according to the dimensional arrangement order of the preset semantic space to obtain the attribute vector of each advertisement to be pushed.

[0026] Calculate the absolute difference between the reference vector and the attribute vector in the corresponding dimension to obtain the deviation value in the corresponding dimension;

[0027] All deviation values ​​are combined according to the dimensional order to obtain the difference sequence for each advertisement to be pushed.

[0028] Optionally, the difference sequence is normalized using a preset resolution coefficient to obtain a gray relational coefficient sequence for each advertisement to be pushed, including:

[0029] Extract the minimum and maximum deviations from the difference sequence of each ad to be pushed;

[0030] The adjustment value is obtained by multiplying the resolution coefficient and the maximum deviation, and the mapping numerator is obtained by summing the minimum deviation and the adjustment value.

[0031] By summing the deviation and adjustment values ​​of each dimension in the difference sequence, the mapping denominator of each dimension is obtained. Based on the ratio of the mapping numerator to each mapping denominator, the correlation coefficient of each dimension is obtained.

[0032] By combining all correlation coefficients according to the dimensional order of the difference sequence, a gray correlation coefficient sequence for each advertisement to be pushed is obtained.

[0033] Optionally, based on the grey relational coefficient sequence, a grey relational degree is obtained through weighted mapping. The advertisement with the highest grey relational degree is determined from the candidate advertisement library as the target advertisement for push to the target page, including:

[0034] Based on the pre-defined correspondence between semantic space dimensions and weights, determine the weight value for each dimension;

[0035] The gray correlation degree of each advertisement to be pushed is obtained by summing the correlation coefficient of each dimension in the gray correlation coefficient sequence and the product of the corresponding weight value.

[0036] The ad with the highest gray correlation from the candidate ad library is selected as the target ad for push to the target page.

[0037] Secondly, this application provides an advertising push system based on user browsing behavior, including:

[0038] The acquisition module is used to synchronously acquire the user's historical search term sequence and the click weight and dwell time percentage of each interactive element in the target page. The dwell time percentage is determined based on the ratio of the dwell time of the interactive element in the visible area to the total page browsing time.

[0039] The generation module is used to construct the behavior vector of each interactive element in the preset semantic space based on the click weight and dwell time ratio, and to obtain the intent vector in the preset semantic space by discretizing the occurrence frequency of each search term in the historical search term sequence.

[0040] The generation module is also used to construct an observation matrix based on behavior vectors and intent vectors. By performing QR decomposition on the observation matrix, an orthogonal matrix is ​​obtained, and an orthogonal basis vector is extracted from the orthogonal matrix as a reference vector.

[0041] The calculation module is used to extract the attribute vector of each advertisement to be pushed in the preset semantic space from the preset candidate advertisement library, calculate the absolute difference between the reference vector and each attribute vector in each dimension to obtain the difference sequence, and use the preset resolution coefficient to normalize the difference sequence to obtain the gray relation coefficient sequence of each advertisement to be pushed.

[0042] The push module is used to obtain the gray correlation degree based on the gray correlation coefficient sequence through weighted mapping, and to determine the target push ad with the highest gray correlation degree from the candidate ad library, so as to push it to the target page.

[0043] Thirdly, this application provides an electronic device, comprising:

[0044] Memory, used to store computer programs;

[0045] A processor is configured to execute the computer program to implement the steps of the advertising push method based on user browsing behavior as described in the first aspect above.

[0046] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the advertising push method based on user browsing behavior as described in the first aspect above.

[0047] The advertising push method based on user browsing behavior provided in this application first constructs an observation matrix by combining the user's real-time interaction behavior and historical search preferences. In response to the problem of intent confusion caused by simply superimposing data in existing technologies, this application performs QR decomposition on the matrix and can effectively separate the intertwined interference information between real-time interests and long-term preferences by using orthogonal transformation. Thus, it can accurately extract the core and independent intent features of users in complex page environments.

[0048] Subsequently, using this independent intent feature as a benchmark, the degree of matching between it and candidate ads is quantified through grey relational analysis, ensuring that the recommendation is based on the user's most genuine intrinsic needs, rather than superficial behavior influenced by redundant information. Therefore, this application solves the technical problem of low ad relevance caused by the inability of existing technologies to accurately extract the user's independent core intent in complex browsing scenarios, significantly improving the accuracy of ad delivery.

[0049] Furthermore, this application first maps each interactive element on the page to a specific coordinate position in a preset semantic space, giving the user's specific interactive behavior a clear semantic dimension. Then, by combining click weight and dwell time percentage to calculate the attention intensity value, it can more reliably measure the user's true level of attention to a specific element, effectively avoiding misjudgments that may arise from a single indicator. This method of transforming the user's vague browsing behavior into a precise, structured behavioral vector ensures that the data used for analysis accurately reflects the user's current interests, improving the accuracy of ad delivery. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A flowchart illustrating an advertising push method based on user browsing behavior provided in an embodiment of this application;

[0052] Figure 2 A flowchart illustrating a method for generating reference vectors provided in an embodiment of this application;

[0053] Figure 3 A flowchart illustrating a method for generating targeted push advertisements provided in an embodiment of this application;

[0054] Figure 4 This is a schematic diagram of the structure of an advertising push system based on user browsing behavior, provided as an embodiment of this application. Detailed Implementation

[0055] Current ad-push technologies commonly construct user interest models by simply fusing or linearly superimposing users' immediate behaviors and historical preferences. However, this approach inherently flaws because it assumes that users' various behavioral intentions are linearly additive, neglecting the significant conflicts and interferences that may exist between users' accidental, immediate interactions (such as accidental clicks or brief browsing of irrelevant content) and their long-term, stable historical preferences on complex web pages. This data-level obfuscation results in interest characteristics that resemble a vague average rather than the user's most genuine and core independent intent at that moment, thus limiting the accuracy of ad matching.

[0056] To address the aforementioned issues, this application proposes an ad push method based on user browsing behavior. Its core lies in moving away from simply linearly superimposing users' real-time behavioral data and historical preference data. Instead, it constructs a comprehensive observation matrix from both, and innovatively employs a matrix QR orthogonal decomposition algorithm for processing. Through orthogonalization transformation, this method can effectively decouple and extract an orthogonal basis vector representing the user's independent core intent from intertwined and correlated data.

[0057] This method abandons the traditional approach of fusion and overlay. It separates the user's core needs from secondary behavioral interference through mathematical transformation, and uses the purified independent intent as the sole reference standard. Then, it uses gray relational analysis to measure and filter candidate ads. This solves the problem that existing technologies cannot effectively separate the interference between the user's real-time behavior and historical preferences, resulting in inaccurate identification of core intent and low relevance of ad push. It significantly improves the accuracy of ad push and user experience.

[0058] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0059] To address the problems of the prior art, embodiments of this application provide a method, apparatus, device, computer storage medium, and computer program product for pushing advertisements based on user browsing behavior. The method for pushing advertisements based on user browsing behavior provided in this application embodiment will be described first below.

[0060] Figure 1 This illustration shows a flowchart of an advertising push method based on user browsing behavior according to an embodiment of this application. Figure 1 As shown, the method includes:

[0061] S101. Simultaneously acquire the user's historical search term sequence and the click weight and dwell time percentage of each interactive element on the target page. The dwell time percentage is determined based on the ratio of the dwell time of the interactive element in the visible area to the total page browsing time.

[0062] Historical search term sequences refer to a collection of search queries recorded chronologically within a preset time period, reflecting a user's long-term stable interests and preferences. Interactive elements refer to visual objects on the target page that users can interact with, such as product images, function buttons, or text links.

[0063] Click weight refers to a pre-defined quantitative value for different types of interactive elements, used to represent the importance of a single click in determining user intent. Dwell time ratio refers to the ratio of the effective display time of a specific interactive element within the user's visible area to the user's total browsing time on the page, used to measure the intensity of the user's visual attention to that element.

[0064] In practice, the historical search records of a specified user are first retrieved from the backend database to obtain a sequence of historical search terms. ,in Indicates the first 1 or less historical search terms ≤n. ​​Meanwhile, user interaction behavior on the current target page, which includes multiple interactive elements, is monitored in real time using client-side event tracking technology. .

[0065] When a click event is detected, retrieve the interactive element that was clicked. Corresponding preset click weight Click weights are pre-set based on the type of interactive element, with specific values ​​as follows: the click weight for the main product image element is 1.0, the click weight for the product details link is 0.8, the click weight for the function button is 0.6, and the click weight for the plain text link is 0.4.

[0066] For interactive elements that are not clicked but have a certain duration of user interaction, a basic weight is assigned to them to avoid completely ignoring the user's visual attention. For elements that were neither clicked nor recorded in terms of dwell time, their weight was set to 0 and they were not included in the construction of the subsequent observation matrix. The final result is the click weight set. ,in , This indicates the number of all interactive elements that were clicked or left in the user's field of view.

[0067] At the same time, the effective dwell time of each interactive element in the visible area is recorded. Total page view time Through calculation The set of stay duration percentages is obtained. , where 1≤ ≤k.

[0068] S102. Based on the click weight and dwell time ratio, construct the behavior vector of each interactive element in the preset semantic space, and obtain the intent vector in the preset semantic space by discretizing the occurrence frequency of each search term in the historical search term sequence.

[0069] Optionally, step S102, which involves constructing the behavior vector of each interactive element in a preset semantic space based on click weight and dwell time ratio, may specifically include:

[0070] S1021. Extract the business attributes of each interactive element and determine the coordinate positions corresponding to the business attributes in the preset semantic space.

[0071] Business attributes refer to the specific category tags to which interactive elements belong within the platform's business logic, and can be category identifiers such as product categories or functional areas. The predefined semantic space is a multi-dimensional vector space predefined by the platform, with each major business category as its dimension. Each dimension corresponds to an independent business category, and the number of dimensions is determined by the total number of actual business categories on the platform. Coordinate position refers to the dimension index corresponding to a specific business attribute within the predefined semantic space, and is used to convert the business attribute into a positional index that can participate in numerical calculations.

[0072] In practice, the first step is to load the pre-maintained business category list on the platform and construct a pre-defined semantic space. .

[0073] The process of constructing the pre-defined semantic space is as follows: First, the platform extracts primary business categories as dimensions of the semantic space based on the category system in the product management system. Taking a typical e-commerce platform as an example, primary categories typically include electronics, apparel and footwear, home furnishings, books and audio-visual products, food and beverages, beauty and personal care, sports and outdoor products, and baby and toys, etc., with a number of dimensions... The value typically ranges from 10 to 50. Small vertical e-commerce platforms The value may be between 10 and 20, depending on the comprehensive e-commerce platform. The value is usually between 30 and 50, and for very large platforms that include more detailed category dimensions, it can reach 50 to 100.

[0074] The second step is to assign a unique dimension number to each first-level category, forming a mapping table from category ID to dimension index. For example, electronic products correspond to the first dimension, clothing, shoes, and bags correspond to the second dimension, and so on. The third step is to define... 3D vector space Any vector in space is represented as ,in Indicates the first The values ​​are based on the business category dimension.

[0075] The fourth step is to establish supporting data structures, including a mapping dictionary from category IDs to dimension indices, a mapping dictionary from category names to dimension indices, and a mapping dictionary from search terms to categories. In the embodiments of this application, for the sake of simplicity, it is assumed that... to A specific value between these ranges, the exact value of which is determined by the number of business categories during actual platform deployment. The example above uses... The small semantic space facilitates numerical demonstrations, but in practical applications... Typically, 20 to 40 categories are selected to cover the main category systems of mainstream e-commerce platforms.

[0076] Assume the space is Dimension, of which the first Dimension corresponding to the first Each business category Next, regarding the interactive elements obtained in S101... By parsing its metadata, the business attributes to which each interactive element belongs are extracted, and then placed in space. The coordinates corresponding to the attribute are obtained by matching the values ​​in the text. For example, interactive elements. The extracted business attributes are matched to the corresponding space. coordinate position .

[0077] S1022. The attention intensity value of each interactive element is obtained by calculating the product of the click weight and the dwell time ratio.

[0078] Attention intensity value is a quantitative value calculated by combining the user's active click intention and passive visual attention, used to represent the depth of the user's interest in a single interactive element on the current page.

[0079] In specific implementation, it is based on the click weight set already obtained in S101. and the percentage of stay For each interactive element Calculate its click weight Percentage of stay The product of these values ​​yields the intensity of interest. ,Right now Ultimately, we obtain the set of attention intensity values ​​for all interactive elements. .

[0080] S1023. Fill the feature bits corresponding to the coordinate positions in the preset baseline behavior sequence with the attention intensity value, and set the other positions in the preset baseline behavior sequence to zero to obtain the behavior vector.

[0081] Preset baseline behavior sequence refers to the dimension and preset semantic space A consistent vector template, with each component initialized to zero, is used to carry the attention intensity information of a single interactive element. A behavior vector is a structured feature vector that expresses the user's real-time attention intensity to an element in a unified semantic space, formed by filling the attention intensity value of an interactive element into the corresponding coordinate position of the baseline sequence.

[0082] In practice, firstly, for each interactive element Initialize a The zero-dimensional vector is used as a preset baseline behavior sequence. Next, the attention intensity value calculated in S1022 is... Fill to the coordinate positions determined in S1021 The corresponding feature bits are set, and the remaining bits are kept at zero.

[0083] For example, for interactive elements coordinates Attention intensity value Then fill it to The second position of the zero-dimensional vector yields the behavior vector. Its second component is The remaining bits are 0, that is Therefore, a corresponding behavior vector is generated for each interactive element on the target page.

[0084] This embodiment transforms the originally discrete and heterogeneous real-time user interaction behaviors into structured behavior vectors that are precisely located in a unified semantic space. This not only ensures the comparability of different types of interactive elements within the same dimensional system but also avoids the bias caused by a single indicator, accurately reflecting the user's true real-time interest intensity in each element on the page.

[0085] Optionally, step S102, which involves discretizing and mapping the frequency of occurrence of each search term in the historical search term sequence to obtain the intent vector in the preset semantic space, may specifically include:

[0086] S1024. Calculate the cumulative frequency of different search terms in the historical search term sequence to obtain the term frequency score for each search term.

[0087] The term frequency score refers to the total number of times a specific search term appears in the sequence of historical search terms, and is used to quantify the long-term attention frequency of users to the semantic content represented by that term.

[0088] In practice, the historical search term sequence obtained in S101 is first processed. Perform deduplication and counting to calculate the cumulative occurrences of each individual term. For example, if the historical search term sequence is... By iterating and counting, we can obtain the entries. The word frequency score is 2, and the entry The word frequency score is 1, and the entry The word frequency score is 1, thus obtaining the word frequency score mapping relationship. .

[0089] S1025. Extract the business category to which each search term belongs, and determine the coordinate position corresponding to the business category in the preset semantic space.

[0090] Business categories refer to the standardized classifications to which search terms belong in the platform's business logic, and whose definitions share the same set of preset semantic spaces with the business attributes of interactive elements. The dimensional system.

[0091] In practice, the business category to which each independent search term belongs in S1024 is extracted by searching the pre-maintained dictionary of terms and business categories on the platform, and then placed in the same preset semantic space as S1021. Determine its coordinates. For example, search for terms. After retrieval and matching, the corresponding space for its business category coordinate position Search terms Business category corresponding coordinate position Search terms Business category corresponding coordinate position .

[0092] S1026. Fill the word frequency score into the feature bits corresponding to the coordinate positions in the preset baseline intent sequence, and perform zeroing on the other positions in the preset baseline intent sequence except for the feature bits to obtain the intent vector.

[0093] Preset baseline intent sequence refers to the dimension and preset semantic space A consistent vector template, with each component initially set to zero, is used to carry the term frequency preference information of all the user's historical search terms. The intent vector refers to a comprehensive feature vector that expresses the user's long-term interest distribution in a unified semantic space, formed by filling in the term frequency scores of each historical search term according to its business category coordinate position.

[0094] In practice, first initialize a The zero-dimensional vector is used as a preset baseline intention sequence. Next, the word frequency score mapping obtained in S1024 is traversed. The frequency score of each search term is filled into the coordinate position determined by S1025.

[0095] For example, the entry The word frequency score of 2 is filled into coordinate position 1, and the word is... Fill the score 1 at coordinate position 2 to fill the entry. The score of 1 is filled to coordinate position 4, while the remaining dimensions without historical records are kept at position 0. Finally, the intent vector is obtained. And intention vector With behavior vector All in the same Dimensional Preset Semantic Space Defined in the Chinese.

[0096] This embodiment transforms users' unstructured historical search records into quantifiable and comparable intent vectors within a unified semantic space. This not only allows for the precise characterization of the strength of users' long-term preferences and ensures the alignment and comparability of historical search data and real-time interaction data within the same dimensional system, but also stably represents the distribution of users' long-term interests.

[0097] S103. Construct an observation matrix based on behavior vectors and intention vectors. Obtain an orthogonal matrix by performing QR decomposition on the observation matrix, and extract an orthogonal basis vector from the orthogonal matrix as a reference vector.

[0098] Optionally, step S103, which involves constructing an observation matrix based on the behavior vector and intention vector, obtaining an orthogonal matrix by performing QR decomposition on the observation matrix, and extracting an orthogonal basis vector from the orthogonal matrix as a reference vector, may specifically include:

[0099] Figure 2 A flowchart illustrating a method for generating reference vectors according to an embodiment of this application is shown. Figure 2 As shown, the method includes:

[0100] S1031. Combine the behavior vector and intent vector of each interactive element in a pre-defined order to obtain the observation matrix.

[0101] The observation matrix refers to the data carrier formed by integrating users' real-time interactive behavior on the current page with long-term historical preference data into a unified matrix structure. It carries all user interest features for subsequent orthogonal transformation processing. The preset arrangement order is shown in Table 1 below:

[0102] Table 1: Preset Arrangement Order Comparison Table

[0103]

[0104] As shown in Table 1, Table 1 illustrates the preset arrangement order followed during the row-direction combination process, where the first row... The rows correspond to the target page in sequence. The behavior vectors of each interactive element, the th The row corresponds to the intent vector generated from the sequence of historical search terms, and all vectors are in the same... Dimensional Preset Semantic Space Define.

[0105] In specific implementation, firstly, according to the preset arrangement order defined in Table 1, the behavior vectors generated in S1023 are... Compared with the intent vector generated in S1026 Stacked row by row. Each vector has a dimension of 1. Therefore, the final observation matrix is ​​obtained. for OK The matrix of columns has the following structure:

[0106] ;

[0107] in Representation matrix No. Line 1 The elements of the column, the first The rows correspond to the behavior vectors of each interactive element, the first row being the first row. The row corresponds to the intent vector.

[0108] S1032. By performing an orthogonal projection transformation on the observation matrix, an orthogonal matrix is ​​obtained, consisting of multiple mutually perpendicular unit vectors with a magnitude of one.

[0109] An orthogonal matrix refers to a matrix obtained by analyzing the observation matrix. The orthogonal factors obtained after QR decomposition are each a unit vector with a magnitude of 1 in each column and any two columns are perpendicular to each other. They are used to extract the independent orthogonal basis directions in the observation matrix.

[0110] In practice, the observation matrix constructed in S1031 is used. As input, the QR decomposition algorithm is applied to perform an orthogonal projection transformation, which will... Decompose into orthogonal matrices With upper triangular matrix The product of , that is, satisfying Among them, orthogonal matrix for A square matrix of order 1, the structure of which is shown below:

[0111] ;

[0112] in Represents an orthogonal matrix No. Line 1 Column elements, Each column satisfies the orthogonal constraint condition that the modulus is 1 and that it is perpendicular to other columns.

[0113] S1033. Extract the unit vector located in the first column from the orthogonal matrix as the orthogonal basis vector, and determine the orthogonal basis vector as the reference vector.

[0114] Orthogonal basis vectors refer to vectors derived from orthogonal matrices. The first unit vector extracted corresponds to the orthogonal direction with the largest variance contribution in the observation matrix and represents the user's most significant and independent core interest direction after eliminating redundant correlations. The reference vector is the baseline feature vector determined for calculating the correlation with candidate ad attribute vectors, and its physical meaning is the directional expression of the core intent reflected by the user's current comprehensive browsing behavior in the preset semantic space.

[0115] In specific implementation, the orthogonal matrix obtained first from S1032... Extract its first column vector. The vector is 1-dimensional unit vector and satisfying .because Corresponding to the principal direction that is first orthogonalized in the QR decomposition process, it integrates the interest components that contribute the most from the behavior vector and intent vector, and reflects the user's most core independent intent at present. Next, Directly determined as the reference vector, each of its components Indicates the first The projection weight of the original vector corresponding to the row in the principal orthogonal direction.

[0116] This embodiment effectively eliminates the potential linear correlation interference between behavior vectors and intent vectors, and can accurately separate the most contributing independent interest directions from mixed user behavior data to obtain a pure benchmark that can truly represent the user's core intent.

[0117] S104. Extract the attribute vector of each advertisement to be pushed in the preset semantic space from the preset candidate advertisement library, calculate the absolute difference between the reference vector and each attribute vector in each dimension to obtain the difference sequence, and use the preset resolution coefficient to normalize the difference sequence to obtain the gray relational coefficient sequence of each advertisement to be pushed.

[0118] Optionally, step S104, which involves extracting the attribute vector of each advertisement to be pushed from the preset candidate advertisement library in the preset semantic space, calculating the absolute difference between the reference vector and each attribute vector in each dimension, and obtaining the difference sequence, may specifically include:

[0119] S1041. Extract the attribute tags of each advertisement to be pushed from the candidate advertisement library. Based on the attribute tags, according to the preset mapping relationship between attribute tags and component values, obtain the component values ​​corresponding to each attribute tag in the preset semantic space. Combine all component values ​​according to the dimensional arrangement order of the preset semantic space to obtain the attribute vector of each advertisement to be pushed.

[0120] The candidate ad library refers to a data warehouse pre-stored by the platform that includes all available ad information, with each ad to be pushed accompanied by attribute tags describing the characteristics of its product or service. Attribute tags are text marks used to identify the business category or characteristics to which an ad belongs, such as category identification and product category tags.

[0121] Component value refers to the value of a certain attribute tag within a predefined semantic space. A quantitative value corresponding to a specific dimension, used to represent the strength or degree of belonging of the attribute in that dimension. An attribute vector is formed by arranging all component values ​​of the advertisement to be pushed according to a preset semantic space dimension order. A feature vector is used to represent the content features of the advertisement in the same semantic space as the reference vector.

[0122] The mapping relationship between attribute labels and component values ​​refers to the pre-established rule table on the platform that converts different attribute labels into uniform quantified values, and is used to transform unstructured label information into computable numerical components. The pre-defined mapping relationship between attribute labels and component values ​​is shown in Table 2 below:

[0123] Table 2: Mapping Relationship between Attribute Labels and Component Values

[0124]

[0125] As shown in Table 2, Table 2 illustrates the mapping relationship between attribute labels and component values ​​in each dimension of the preset semantic space, where each row corresponds to a space. One dimension identifies the business category, possible attribute tags, and mapped component values ​​corresponding to that dimension.

[0126] In practice, the first step is to extract advertisements from the candidate ad library, including... A collection of ads to be pushed. Next, for each advertisement... Extract its attached attribute tags, where Then, based on the mapping relationships defined in Table 2, in the preset semantic space... of Search for the component value corresponding to each attribute label one by one along each dimension. For example, suppose an advertisement... The attribute label is of type B. Looking up Table 2, its corresponding component value in dimension 2 is... If there is no matching label in other dimensions, the component value is set to 0.

[0127] Finally, according to space The dimensional arrangement order combines all component values ​​into Dimensional vector, to obtain the advertisement attribute vector Similarly, for each ad to be pushed... Each generates its corresponding attribute vector. This forms a set of attribute vectors. .

[0128] S1042. Calculate the absolute difference between the reference vector and the attribute vector in the corresponding dimension to obtain the deviation value in the corresponding dimension.

[0129] The deviation value refers to the absolute value of the difference between the reference vector and the attribute vector at the same position in the same dimension, and is used to quantify the degree of difference between the user's core intent and the characteristics of the advertising content in that dimension.

[0130] In practice, the reference vector obtained from S1033 is first extracted. Due to the reference vector The dimension is , while attribute vector The dimension is The reference vector needs to be projected to map it to the same plane as the attribute vector. 3D semantic space middle.

[0131] Specifically, using the observation matrix With the pre-defined semantic space The correspondence is transformed in dimension. Due to the observation matrix... of The rows correspond to the previous Each interaction element has a behavior vector and an intent vector, and each behavior vector and intent vector are themselves... dimensional vector, therefore reference vector Each component ( In essence, it represents the first Original The projection weights of a dimensional vector along the principal orthogonal directions.

[0132] Constructed by weighted aggregation dimensional reference vector , its first Each component ,in Observation matrix No. Line 1 The element of the column is the first The original vector at the th... Dimensional component values, Reference vector The Each component.

[0133] Next, for each ad to be pushed... attribute vector Calculate its relationship with the reference vector In each dimension The absolute difference on, where This yields the deviation value for the corresponding dimension. For example, regarding advertising. The deviation value in dimension 1 is The deviation value in dimension 2 is .

[0134] S1043. Combine all deviation values ​​according to the dimensional order to obtain the difference sequence for each advertisement to be pushed.

[0135] The difference sequence refers to a one-dimensional sequence formed by arranging the deviation values ​​of the advertisement to be pushed and the reference vector in all dimensions according to the preset semantic space dimension order.

[0136] In practice, the first step is to follow the preset semantic space. From the first dimension to the second The order of dimensions is used to combine the deviation values ​​of each dimension calculated in S1042. For advertising... , and put it in Deviation values ​​in each dimension Arrange them sequentially to obtain the difference sequence of the advertisement. Finally, the collection of ads to be pushed. middle Each advertisement generates its own difference sequence, resulting in a set of difference sequences. .

[0137] This embodiment ensures the comparability of user intent and advertising content within the same dimensional system, and can precisely characterize the degree of deviation between the user's core intent and the matching of each candidate advertisement in each business category, so that the advertising recommendation process is based on accurate quantitative difference measurement.

[0138] Optionally, the process of normalizing the difference sequence using preset resolution coefficients in step S104 to obtain the gray relational coefficient sequence of each advertisement to be pushed may specifically include:

[0139] S1044. Extract the minimum and maximum deviations from the difference sequence of each advertisement to be pushed.

[0140] Minimum deviation refers to the minimum value of all dimensional deviations in the difference sequence of a single ad to be pushed, and is used to represent the degree of difference between the ad and the user reference vector in the closest dimension. Maximum deviation refers to the maximum value of all dimensional deviations in the difference sequence, and is used to represent the degree of difference between the ad and the user reference vector in the most distant dimension.

[0141] In practice, each advertisement to be pushed is first extracted from S1043. difference sequence Next, iterate through all... The deviation values ​​in each dimension are compared and the minimum value is extracted and recorded as the minimum deviation. The maximum value is denoted as the maximum deviation. .

[0142] S1045. The adjustment value is obtained by multiplying the resolution coefficient and the maximum deviation, and the mapping numerator is obtained by summing the minimum deviation and the adjustment value.

[0143] The resolution coefficient is a preset parameter used to adjust the sensitivity of grey relational coefficient calculation. The adjustment value is an intermediate calculated value obtained by multiplying the resolution coefficient by the maximum deviation. The mapping numerator is the sum of the minimum deviation and the adjustment value; it is the numerator part in the grey relational coefficient calculation formula and remains constant across all dimensions. The preset resolution coefficients are shown in Table 3 below:

[0144] Table 3: Preset Resolution Coefficient Comparison Table

[0145]

[0146] As shown in Table 3, Table 3 presents the resolution coefficients. Value recommendations for different application scenarios: in high-precision matching scenarios, choose smaller values ​​to improve discrimination; in general recommendation scenarios, choose medium values ​​to balance performance; and in broad coverage scenarios, choose larger values ​​to expand the matching range.

[0147] The value ranges and recommended values ​​for each scenario in Table 3 were determined based on the following principles: resolution coefficient The theoretical function of this is to modulate the sensitivity of the grey relational coefficient to differences. The smaller the value, the stronger the ability of the correlation coefficient to distinguish deviation values. Through grid search experiments conducted on a standard test set including 10,000 user samples and 5,000 candidate ads, using ad click-through rate (CTR) as the evaluation metric, the following results were obtained:

[0148] When the platform requires a CTR improvement rate of ≥15%, corresponding to a high-precision matching scenario, it is necessary to strictly distinguish between subtle differences between user intent and advertising. Experiments show that... The CTR improvement is significant, and the use of boundary values ​​is recommended. To achieve maximum differentiation. When the platform requires a CTR improvement rate between 8% and 15%, corresponding to a typical recommendation scenario, a balance needs to be struck between accuracy and coverage. Experiments show... For stable performance, the intermediate value is recommended. .

[0149] When the platform requires a CTR improvement rate of ≥5% and places greater emphasis on ad impressions, corresponding to a broad coverage scenario, a larger matching deviation can be tolerated. Experiments show that... Optimal coverage at that time, recommended for use In practical applications, the platform can select the corresponding scenario based on the core KPI of the current business stage, namely precise conversion or traffic expansion, and further fine-tune it through A / B testing within the value range of that scenario to adapt to the characteristics of specific user groups.

[0150] In practice, first select the corresponding application scenario from Table 3 based on actual business needs, and obtain the corresponding resolution coefficient. Assuming a typical recommendation scenario is used, the recommended value is taken. Next, the resolution coefficient is compared with the maximum deviation extracted from S1044. Multiply to obtain the adjustment value. Then, the minimum deviation extracted from S1044 is... With adjustment value Adding them together yields the mapping molecule. .

[0151] S1046. By summing the deviation value and adjustment value of each dimension in the difference sequence, the mapping denominator of each dimension is obtained. Based on the ratio of the mapping numerator to each mapping denominator, the correlation coefficient of each dimension is obtained.

[0152] The mapping denominator refers to the sum of the deviation value and the adjustment value for a specific dimension in the difference sequence, and it serves as the denominator for that dimension in the grey relational coefficient calculation formula. The correlation coefficient is the ratio of the mapping numerator to the mapping denominator, used to quantify the degree of correlation between the user reference vector and the advertisement to be pushed in a specific dimension; the larger the value, the higher the matching degree in that dimension.

[0153] In practice, the difference sequence obtained from S1043 is first extracted. Each dimension deviation value ,in Next, this deviation value is compared with the adjustment value calculated in S1045. Add them together to get the mapping denominator for that dimension. .

[0154] Then, using the mapping molecule obtained in S1045 Calculate the correlation coefficient for this dimension. ,in Indicates the first The number of ads to be pushed is in the first... Correlation coefficients across dimensions This represents the minimum deviation of the advertising difference sequence. For adjustment value, For the advertisement in the The deviation value of the dimension.

[0155] S1047. Combine all correlation coefficients according to the dimensional order of the difference sequence to obtain the gray correlation coefficient sequence of each advertisement to be pushed.

[0156] The grey relational coefficient sequence refers to a one-dimensional sequence formed by arranging the relational coefficients of the advertisement to be pushed in all dimensions according to the preset semantic space dimension order.

[0157] In practice, the first step is to follow the preset semantic space. From the first dimension to the second The order of dimensions is used to combine the correlation coefficients of each dimension calculated in S1046. For advertising... , and put it in Correlation coefficients across dimensions Arranged sequentially, the grey relational coefficient sequence of the advertisement is obtained. Finally, the collection of ads to be pushed. middle Each ad to be pushed generates its own grey relational coefficient sequence, resulting in a set of grey relational coefficient sequences. .

[0158] This embodiment transforms the original difference sequence into a grey relational coefficient sequence that can accurately quantify the degree of matching between user intent and advertisement, avoiding the lack of comparability caused by differences in the difference scale of different advertisements. It can not only flexibly adjust the calculation sensitivity according to different application scenarios, but also accurately reflect the degree of matching tightness of each dimension under a unified standard.

[0159] S105. Based on the grey relational coefficient sequence, the grey relational degree is obtained through weighted mapping. The advertisement with the highest grey relational degree is determined from the candidate advertisement library as the target advertisement for push to the target page.

[0160] Optionally, step S105, based on the grey relation coefficient sequence, obtains the grey relation degree through weighted mapping, and determines the advertisement with the highest grey relation degree from the candidate advertisement library as the target advertisement for push to the target page. This process may specifically include:

[0161] Figure 3 A flowchart illustrating a method for generating targeted push advertisements according to an embodiment of this application is shown. Figure 3 As shown, the method includes:

[0162] S1051. Determine the weight value of each dimension based on the preset correspondence between semantic space dimensions and weights.

[0163] The weight value refers to the relative importance of a specific dimension in the comprehensive evaluation of the user's intent and the ad's match. The larger the weight value, the more significant the impact of that dimension on the final recommendation result.

[0164] The correspondence between semantic space dimensions and weights refers to the pre-configured semantic space by the platform. A rule table mapping each dimension to its importance weight coefficient is provided. The predefined correspondence between semantic space dimensions and weights is shown in Table 4 below:

[0165] Table 4: Correspondence between Semantic Space Dimensions and Weights

[0166]

[0167] As shown in Table 4, Table 4 presents the mapping relationship between each dimension of the preset semantic space and its corresponding weight values, where each row corresponds to a space. One dimension identifies the business category corresponding to that dimension and its weight coefficient in the weighted calculation, and the sum of all weight values ​​equals 1.

[0168] The weights of each dimension in Table 4 were determined using the following method: First, the business categories involved in all successfully converted orders (i.e., orders where users clicked on ads and completed purchases) within the past 30 days were statistically analyzed, and the order volume percentage of each category was calculated. ,in This refers to the semantic space dimension number corresponding to the category. , Category The number of orders.

[0169] Subsequently, the order volume percentages were smoothed to avoid extreme weightings using a smoothing coefficient. Calculate the initial weights Based on the average order value for each category After weighting adjustments, the final weights are obtained. .

[0170] In practice, the predefined correspondence between semantic space dimensions and weights in Table 4 is first loaded. Then, according to the preset semantic space... From the first dimension to the second The weight values ​​for each dimension are extracted sequentially according to their order, forming a set of weight values. ,in Indicates the first The weights of each dimension satisfy the constraints. .

[0171] S1052. Accumulate the product of the correlation coefficient of each dimension and the corresponding weight value in the gray correlation coefficient sequence to obtain the gray correlation degree of each advertisement to be pushed.

[0172] Grey relational degree is a comprehensive evaluation index obtained by weighted aggregation calculation of grey relational coefficient sequence. It is used to represent the overall matching degree between the advertisement to be pushed and the user's core intent in all dimensions, and the larger the value, the higher the recommendation priority.

[0173] In practice, each advertisement to be pushed is first extracted from S1047. Grey relational coefficient sequence Next, the set of weight values ​​obtained in S1051 is combined. For each dimension in the sequence correlation coefficient With corresponding weight values Perform a product operation, then sum the product results of all dimensions to obtain the gray relational degree of the advertisement. .

[0174] in Indicates the first The gray correlation of an ad to be pushed. For the first Weight values ​​for each dimension, For the advertisement in the Correlation coefficients across each dimension. And so on, the set of ads to be pushed. middle Each ad to be pushed has its own grey relational degree calculated, resulting in a grey relational degree set. .

[0175] S1053. Identify the advertisement with the highest gray correlation from the candidate advertisement library as the target advertisement for push to the target page.

[0176] Targeted push advertising refers to the advertisements that are selected from the candidate ad library after being sorted by gray relational analysis and have the highest degree of matching with the user's core intent. These ads will then be sent to the target page that the user is currently browsing for display.

[0177] In practice, the gray relational set of all ads to be pushed obtained in S1052 is first processed. The ads are compared and sorted. Next, the ads with the highest gray relational scores are extracted and identified as the target ads. That is, the corresponding grey relational degree Finally, target ads to the desired audience. The content information is packaged and sent to the designated ad slot on the target page currently being viewed by the user through the push interface, thus completing the process of precise ad push based on the user's browsing behavior.

[0178] This embodiment differentiates the importance of different business categories in recommendation decisions, ensuring that the recommendation results align with the platform's business orientation. It completes a closed loop from user behavior data collection, intent extraction, orthogonal decomposition, correlation calculation to precise push notifications, achieving accurate response to core user needs and highly relevant ad matching in complex page environments.

[0179] Example 2

[0180] Figure 4 This application provides a schematic diagram of a specific implementation of an advertising push system based on user browsing behavior, referring to... Figure 4 The system may include:

[0181] The acquisition module 410 is used to synchronously acquire the user's historical search term sequence and the click weight and dwell time ratio of each interactive element in the target page. The dwell time ratio is determined based on the ratio of the dwell time of the interactive element in the visible area to the total page browsing time.

[0182] The generation module 420 is used to construct the behavior vector of each interactive element in the preset semantic space based on the click weight and dwell time ratio, and to obtain the intent vector in the preset semantic space by discretizing the occurrence frequency of each search term in the historical search term sequence.

[0183] The generation module 420 is also used to construct an observation matrix based on the behavior vector and the intention vector, obtain an orthogonal matrix by performing QR decomposition on the observation matrix, and extract an orthogonal basis vector from the orthogonal matrix as a reference vector;

[0184] The calculation module 430 is used to extract the attribute vector of each advertisement to be pushed in the preset semantic space from the preset candidate advertisement library, calculate the absolute difference between the reference vector and each attribute vector in each dimension to obtain the difference sequence, and use the preset resolution coefficient to normalize the difference sequence to obtain the gray relation coefficient sequence of each advertisement to be pushed.

[0185] The push module 440 is used to obtain the gray correlation degree based on the gray correlation coefficient sequence through weighted mapping, and to determine the advertisement with the highest gray correlation degree from the candidate advertisement library as the target push advertisement for push to the target page.

[0186] The advertising push system based on user browsing behavior in this application is used to implement the aforementioned advertising push method based on user browsing behavior. Therefore, the specific implementation of the advertising push system based on user browsing behavior can be found in the embodiment section of the advertising push method based on user browsing behavior above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0187] One embodiment of this application provides an electronic device that may include a processor and a memory storing computer program instructions. Specifically, the processor may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement embodiments of this application.

[0188] Memory may include mass storage for data or instructions. Memory may include read-only memory (ROM), random access memory (RAM), disk storage media devices, optical storage media devices, flash memory devices, electrical, optical, or other physical / tangible memory storage devices. Thus, generally, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method according to the first aspect of this disclosure.

[0189] The processor reads and executes computer program instructions stored in memory to implement any of the above-described methods for pushing advertisements based on user browsing behavior.

[0190] The electronic device can execute the advertising push method based on user browsing behavior in the embodiments of this application, thereby realizing the advertising push method based on user browsing behavior described in conjunction with the accompanying drawings.

[0191] Furthermore, in conjunction with the advertising push method based on user browsing behavior in the above embodiments, this application embodiment can provide a computer-readable storage medium for implementation. This computer-readable storage medium stores computer program instructions; when executed by a processor, these computer program instructions implement any of the advertising push methods based on user browsing behavior in the above embodiments.

[0192] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0193] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0194] The aspects of this application have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by dedicated hardware performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0195] The above provides a detailed description of the advertising push method and system based on user browsing behavior provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for pushing advertisements based on user browsing behavior, characterized in that, include: The system synchronously acquires the user's historical search term sequence and the click weight and dwell time percentage of each interactive element on the target page. The dwell time percentage is determined based on the ratio of the dwell time of the interactive element in the visible area to the total page browsing time. Based on the click weight and the dwell time ratio, a behavior vector of each interactive element in a preset semantic space is constructed, and an intent vector in the preset semantic space is obtained by discretizing the occurrence frequency of each search term in the historical search term sequence. An observation matrix is ​​constructed based on the behavior vector and the intention vector. An orthogonal matrix is ​​obtained by performing QR decomposition on the observation matrix, and an orthogonal basis vector is extracted from the orthogonal matrix as a reference vector. Extract the attribute vector of each advertisement to be pushed in the preset semantic space from the preset candidate advertisement library, calculate the absolute difference between the reference vector and each attribute vector in each dimension to obtain the difference sequence, and use the preset resolution coefficient to normalize the difference sequence to obtain the gray relation coefficient sequence of each advertisement to be pushed. Based on the gray relation coefficient sequence, the gray relation degree is obtained through weighted mapping. The advertisement with the highest gray relation degree is determined from the candidate advertisement library as the target advertisement for push to the target page.

2. The method according to claim 1, characterized in that, The step of constructing a behavior vector for each interactive element in a preset semantic space based on the click weight and the dwell time ratio includes: Extract the business attribute of each interactive element and determine the coordinate position corresponding to the business attribute in the preset semantic space; The attention intensity value of each interactive element is obtained by calculating the product of the click weight and the dwell time ratio. The attention intensity value is filled into the feature bit corresponding to the coordinate position in the preset benchmark behavior sequence, and the other positions in the preset benchmark behavior sequence other than the feature bit are set to zero to obtain the behavior vector.

3. The method according to claim 1, characterized in that, The step of discretizing and mapping the frequency of occurrence of each search term in the historical search term sequence to obtain the intent vector in the preset semantic space includes: Calculate the cumulative frequency of different search terms in the historical search term sequence to obtain the term frequency score for each search term; Extract the business category to which each search term belongs, and determine the coordinate position corresponding to the business category in the preset semantic space; The word frequency score is filled into the feature bit corresponding to the coordinate position in the preset baseline intent sequence, and zeroing is performed on the other positions in the preset baseline intent sequence except for the feature bit to obtain the intent vector.

4. The method according to claim 1, characterized in that, The process of constructing an observation matrix based on the behavior vector and the intention vector, obtaining an orthogonal matrix by performing QR decomposition on the observation matrix, and extracting an orthogonal basis vector from the orthogonal matrix as a reference vector includes: The behavior vector and intent vector of each interactive element are combined in a preset order to obtain the observation matrix. By performing an orthogonal projection transformation on the observation matrix, an orthogonal matrix is ​​obtained, which consists of multiple mutually perpendicular unit vectors with a magnitude of one. The unit vector located in the first column of the orthogonal matrix is ​​extracted as the orthogonal basis vector, and the orthogonal basis vector is determined as the reference vector.

5. The method according to claim 1, characterized in that, The step involves extracting the attribute vector of each advertisement to be pushed from a preset candidate ad library under the preset semantic space, calculating the absolute difference between the reference vector and each attribute vector in each dimension, and obtaining a difference sequence, including: The attribute tags of each advertisement to be pushed are extracted from the candidate advertisement library. Based on the attribute tags, according to the preset mapping relationship between attribute tags and component values, the component values ​​corresponding to each attribute tag in the preset semantic space are obtained. All component values ​​are combined according to the dimensional arrangement order of the preset semantic space to obtain the attribute vector of each advertisement to be pushed. Calculate the absolute difference between the reference vector and the attribute vector in the corresponding dimension to obtain the deviation value in the corresponding dimension; All the deviation values ​​are combined according to the dimensional order to obtain the difference sequence for each advertisement to be pushed.

6. The method according to claim 1, characterized in that, The step of normalizing the difference sequence using a preset resolution coefficient to obtain a gray relational coefficient sequence for each advertisement to be pushed includes: Extract the minimum and maximum deviations from the difference sequence of each of the ads to be pushed; The adjustment value is obtained by multiplying the resolution coefficient and the maximum deviation, and the mapping numerator is obtained by summing the minimum deviation and the adjustment value. By summing the deviation value and the adjustment value of each dimension in the difference sequence, the mapping denominator of each dimension is obtained. Based on the ratio of the mapping numerator to each mapping denominator, the correlation coefficient of each dimension is obtained. All the correlation coefficients are combined according to the dimensional order of the difference sequence to obtain the gray correlation coefficient sequence for each advertisement to be pushed.

7. The method according to claim 6, characterized in that, The step of obtaining gray correlation degree through weighted mapping based on the gray correlation coefficient sequence, and determining the advertisement with the highest gray correlation degree from the candidate advertisement library as the target advertisement for push to the target page, includes: Based on the pre-defined correspondence between semantic space dimensions and weights, determine the weight value for each dimension; The gray correlation degree of each advertisement to be pushed is obtained by summing the correlation coefficient of each dimension in the gray correlation coefficient sequence and the corresponding weight value. The advertisement with the highest gray correlation from the candidate advertisement library is determined as the target advertisement for push to the target page.

8. An advertising push system based on user browsing behavior, characterized in that, include: The acquisition module is used to synchronously acquire the user's historical search term sequence and the click weight and dwell time ratio of each interactive element in the target page. The dwell time ratio is determined based on the ratio of the dwell time of the interactive element in the visible area to the total page browsing time. The generation module is used to construct a behavior vector of each interactive element in a preset semantic space based on the click weight and the dwell time ratio, and to obtain an intent vector in the preset semantic space by discretizing the occurrence frequency of each search term in the historical search term sequence. The generation module is also used to construct an observation matrix based on the behavior vector and the intention vector, obtain an orthogonal matrix by performing QR decomposition on the observation matrix, and extract an orthogonal basis vector from the orthogonal matrix as a reference vector; The calculation module is used to extract the attribute vector of each advertisement to be pushed in the preset semantic space from the preset candidate advertisement library, calculate the absolute difference between the reference vector and each attribute vector in each dimension to obtain the difference sequence, and use the preset resolution coefficient to normalize the difference sequence to obtain the gray relation coefficient sequence of each advertisement to be pushed. The push module is used to obtain the gray correlation degree through weighted mapping based on the gray correlation coefficient sequence, and to determine the advertisement with the largest gray correlation degree from the candidate advertisement library as the target push advertisement for push to the target page.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the advertising push method based on user browsing behavior as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the advertising push method based on user browsing behavior as described in any one of claims 1 to 7.

Images