User behavior prediction system and method for Internet advertisement pushing

Abstract

The invention discloses a user behavior prediction system and method for Internet advertisement pushing, and relates to the technical field of Internet information. Historical data of interaction between a user and an advertisement in a unit period is collected, and a historical record database composed of historical records of interaction between the user and the advertisement is established; time characteristics of interaction between each user and the advertisement are collected from a historical record database, a historical interaction weight matrix is established, time intervals of interaction between the users and the advertisement in a unit period are collected, and operation behaviors of the users and the advertisement each time in the unit period are collected; based on historical interaction records of the users and the advertisements and behavior characteristics of interaction between the users and the advertisements, an interest map of the users is established, trend vectors of the users are generated through a map attention network, and recommendation weights of the users for the advertisements are calculated in combination with the historical interaction records of the users and the advertisements; and recommending the advertisement with the maximum recommendation weight to another user with the same user behavior characteristics.

Description

Technical Field

[0001] The present invention relates to the field of Internet information technology, and specifically to a user behavior prediction system and method for Internet advertising push. Background Art

[0002] With the continuous development of Internet advertising push technology, precision and personalization have become key factors in enhancing user experience and advertising effects. In the current Internet advertising push scenario, user behavior prediction is the core link for achieving precise advertising placement. However, existing advertising push methods still have significant deficiencies in user behavior prediction, making it difficult to meet the actual needs in terms of the precision and real-time nature of advertising push. Specifically, existing technologies mainly rely on static user portraits or historical behavior data for prediction, lacking the ability to capture the dynamic behavior changes of users, resulting in prediction results often lagging behind the actual behavior trends of users. In addition, existing methods are insufficient in real-time data analysis and processing, unable to quickly respond to changes in user behavior, thereby reducing the timeliness and relevance of advertising push. Summary of the Invention

[0003] The purpose of the present invention is to provide a user behavior prediction system and method for Internet advertising push to solve the problems raised in the prior art.

[0004] To achieve the above purpose, the present invention provides the following technical solution: A user behavior prediction method for Internet advertising push, the method comprising: Step 100: Collect historical data of user interactions with advertisements in a unit cycle, and establish a historical record database composed of historical records of user interactions with advertisements; Step S200: Collect time features of each user's interaction with advertisements from the historical record database, and establish a historical interaction weight matrix; Step S300: Collect the time intervals of user interactions with advertisements in a unit cycle, and calculate the long-term interest evaluation value of the user; Step S400: Collect the operation behaviors of the user with each advertisement in a unit cycle, and calculate the short-term interest evaluation value of the user; Step S500: Based on the historical interaction records of the user with advertisements and the behavioral characteristics of the user's interaction with advertisements, establish an interest map of the user, and generate a trend vector of the user through a graph attention network; Step S600: Collect the current interaction records of the user with advertisements, calculate the recommendation weight of the user for advertisements in combination with the historical interaction records of the user with advertisements, and recommend the advertisement with the largest recommendation weight to another user with the same behavioral characteristics as the user.

[0005] Further, step S100 includes: Step 101: Set a unit period with a time length of T0, and obtain the historical records of the interactions between several users and advertisements within the unit time; Step 102: Collect all the advertisement contents in the historical records to form an advertisement candidate set.

[0006] Further, step S200 includes: Step S201: Collect the historical click count c of the i-th user on the j-th advertisement i,j , and calculate the basic weight of the i-th user and the j-th advertisement ; Step S202: Obtain the average time interval ∆t between the clicks of the i-th user on the j-th advertisement ij , the advertisement conversion rate d of the j-th advertisement for the i-th user ij , and calculate the interaction weight of the i-th user on the j-th advertisement , where γ represents the interaction decay coefficient, satisfying the condition 0 < γ < 1 Step S203: Normalize the basic weight and the interaction weight to obtain the historical interaction weight matrix H of the i-th user clicking on the j-th advertisement i,j , , where n is the total number of advertisements in the advertisement candidate set.

[0007] Quantify the historical interest of users in advertisements. By collecting data such as the number of times users click on advertisements, the time interval, and the advertisement conversion rate, calculate the basic weight and the interaction weight, and perform normalization processing to finally obtain the historical interaction weight matrix; The historical interaction weight matrix can accurately reflect the historical interest of users in advertisements and provide reliable basic data for subsequent recommendation algorithms. Normalization processing ensures the comparability of weights between different users and advertisements, making the recommendation system more fair and accurate The historical interaction weight matrix reflects the importance of users' historical behaviors and takes into account the time decay factor, that is, recent behaviors can better reflect users' current interests.

[0008] Further, step S300 includes: Step S301: Obtain the maximum time interval t and the minimum time interval t of the i-th user clicking on the j-th advertisement within the unit period max and min ; Step S302: Calculate the long-term interest evaluation value , where α is a balance factor, 0 < α < 1, λ1 represents the first rate control parameter, λ2 represents the second rate control parameter, and 0 < λ2 < λ1 < 1.

[0009] The long-term interest evaluation value can provide a stable interest reference for the recommendation system, helping the system understand the user's interest preferences over a relatively long period. The use of the rate control parameter and the balance factor makes the evaluation result more reasonable and reliable, avoiding the excessive influence of a single factor on the evaluation result.

[0010] Furthermore, step S400 includes: Step S401: Set the threshold score for the operation behavior. Collect the operation behavior of the i-th user for the j-th advertisement in a unit period, and sum the operation scores corresponding to the operation behavior in the unit period to obtain the original interaction score S of the i-th user for the j-th advertisement. i,j raw ; The operation behaviors include: click, browse, and slide. Among them, set the click operation score b1, the browse operation score b2, and the slide operation score b3, satisfying the condition b1 > b2 > 0 > b3; Step S403: Obtain the maximum value of the original interaction scores from all users' original interaction scores for each advertisement, denoted as max(S raw ), and the minimum value of the original interaction scores, denoted as min(S raw ). Perform Min-Max normalization on the original interaction score of the i-th user for the j-th advertisement to obtain the short-term interest evaluation value S of the i-th user for the j-th advertisement. i,j .

[0011] The short-term interest evaluation value can capture the user's interest changes in a timely manner, improving the timeliness of recommendations. This is crucial for the recommendation system because users' interests often change dynamically, and capturing these changes in a timely manner can significantly improve the recommendation effect. The normalization process makes the evaluation result more intuitive and easier to compare, facilitating subsequent processing by the system.

[0012] Furthermore, step S500 includes: Step S501: Obtain the first evaluation vector r1 composed of the short-term interest evaluation values of the i-th user for all advertisements, and obtain the second evaluation vector r2 composed of the long-term interest evaluation values of the i-th user for all advertisements; Step S502: Concatenate r1 and r2 to obtain the feature vector f of the i-th user. i , f i = (r1 || r2), where || represents the concatenation operation. Take the i-th user as node i in the graph structure, and the feature vector fi is the node feature vector of node i; Step S503: Obtain the feature vector f of the i-th user. i and the feature vector f of the m-th user. m , calculate the feature vector f. iand the feature vector f m The similarity sim(f i , f m ), set the similarity threshold θ. When sim(f i , f m ) > θ, establish an undirected edge between node i and node j, where sim represents the similarity calculation function; Step S504: Calculate the joint encoding p i,k , of node i and node k, where LeakyReLU represents the linear activation function, a represents the attention vector, W represents the feature transformation matrix, and || represents the concatenation operation; Step S505: Obtain all nodes that have an undirected edge connection with node i, and jointly form a neighbor node set U including itself with all nodes and node i. When there is an undirected edge between node m and node i, normalize the joint encoding p i,j of node i and node m to obtain the attention coefficient e i,j of node i to node j, ; Step S506: Perform feature aggregation on node features through the graph attention network, , where σ represents the Sigmoid function, represents the feature aggregation result of the (l + 1)-th propagation layer in the graph attention network. When the feature fusion is completed in the L-th propagation layer, take the output result of the L-th propagation layer as the behavior trend vector vi of the i-th user.

[0013] Use the graph structure to represent the interest relationship between users, and mine the potential connections between users through the graph attention network. It combines the short-term and long-term interest evaluation values of users to form an evaluation vector, constructs nodes in the graph structure, and calculates the similarity and attention coefficients between nodes, and finally generates the user behavior trend vector; The establishment of the interest graph can deeply mine the interest relationship between users and improve the personalization degree of recommendations. The generation of the trend vector enables the recommendation system to better understand the interest changes of users, so as to provide more accurate recommendations. The use of the graph attention network further improves the accuracy of feature aggregation, making the recommendation results more in line with the actual needs of users.

[0014] Furthermore, step S600 includes: Step S601: Obtain the time interval between the current time and the time when the i-th user last clicked on the j-th advertisement, and record the time interval as the relevant time interval tr i,j ; Step S602: Calculate the recommendation weight E i,j, where η is the balance coefficient, 0 < η < 1, and β is the time decay coefficient, 0 < β < 1; Obtain the recommendation weights of all advertisements for the i-th user, and record the advertisement with the largest recommendation weight as the target advertisement for the i-th user; Step S603: Denote a user to be pushed at the current time as the current user, obtain the historical interaction record of the current user, and calculate the behavior trend vector q of the current user; Step S604: When the similarity between the behavior trend vector q and the behavior trend vector vi is greater than the threshold, push the target advertisement to the current user.

[0015] The calculation of the recommendation weight can push advertisements that better meet the user's needs based on the user's real-time interests and behavioral characteristics. The comparison of the similarity of the behavior trend vectors improves the accuracy and personalization of the recommendation, making the recommendation results closer to the actual needs of the user.

[0016] Through the recommendation weight formula, the advertisement recommendation weight can be dynamically adjusted to ensure a reasonable combination of real-time behavior and historical behavior; when the user's historical interaction weight matrix frequently clicks on a certain type of advertisement recently, the system will assign a higher short-term interest score to this type of advertisement, thereby increasing its recommendation weight; while when the user has not interacted with a certain type of advertisement for a long time, its recommendation weight will be reduced to avoid over-recommendation.

[0017] To better implement the above method, a user behavior prediction system for Internet advertisement pushing is also proposed. The system includes: a historical data management module, a historical interaction weight matrix management module, a long-term interest evaluation module, a short-term interest evaluation module, a trend vector management module, and an advertisement recommendation module; The historical data management module is used to manage the historical data of user and advertisement interactions. The historical interaction weight matrix management module is used to collect the interaction data between users and advertisements and manage the historical interaction weight matrix. The long-term interest evaluation module is used to manage the long-term interest evaluation value of users. The short-term interest evaluation module is used to manage the short-term interest evaluation value of users. The trend vector management module is used to construct the interest map of users and manage the behavior trend vectors of users. The advertisement recommendation module is used to calculate the recommendation weights of advertisements in combination with historical data and push the advertisements that meet the conditions.

[0018] Further, the historical data management module includes: a basic weight calculation unit, an interaction weight calculation unit, and a historical interaction weight matrix generation unit. Among them, the basic weight calculation unit is used to collect the number of clicks of the user on the advertisement from the historical records and calculate the basic weight. The interaction weight calculation unit is used to calculate the interaction weight by obtaining the average time interval between the user's clicks on the advertisement and the advertisement conversion rate. The historical interaction weight matrix generation unit is used to perform normalization processing on the basic weight and the interaction weight to generate a historical interaction weight matrix; Further, the long-term interest evaluation module includes: a time interval management unit and a long-term interest evaluation value calculation unit. Among them, the time interval management unit is used to manage the maximum and minimum time intervals of the user's clicks on the advertisement within a unit period. The long-term interest evaluation value calculation unit is used to calculate the long-term interest evaluation value of the user; Further, the short-term interest evaluation module includes: an operation behavior management unit and a short-term interest evaluation value calculation unit. Among them, the operation behavior management unit is used to manage the user's operation behaviors and manage the operation scores of different operation behaviors. The short-term interest evaluation value calculation unit is used to normalize the original interaction score to obtain the short-term interest evaluation value; Further, the trend vector management module includes: a node evaluation unit, an interest graph management unit, an attention coefficient management unit, and a trend vector management unit. Among them, the node evaluation unit is used to obtain the short-term and long-term interest evaluation values of the user to obtain the node evaluation vector of the node. The interest graph management unit is used to evaluate the similarity of the node evaluation vectors and manage the graph structure of the interest graph. The attention coefficient management unit is used to calculate the joint encoding between user nodes, normalize the joint encoding to obtain the attention coefficient. The trend vector management unit is used to perform feature aggregation on the node features through a graph attention network to generate the behavior trend vector of the user; Further, the advertisement recommendation module includes: an advertisement recommendation weight calculation unit, a user comparison unit, and an advertisement push decision unit. Among them, the advertisement recommendation weight calculation unit is used to obtain the time interval between the current time and the time when the user last performed an operation behavior on the advertisement, and calculate the recommendation weight of the user for the advertisement based on the user's historical interaction data and the current operation time interval. The user comparison unit is used to compare the behavior trend vectors of the users. The advertisement push decision unit is used to push the target advertisement to the relevant users according to the comparison result between the similarity of the user behavior trend vectors and a preset threshold.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: Through the multi-level feature fusion technology, the present invention comprehensively captures the dynamic changes of user behavior. At the same time, by introducing the user behavior evolution graph and the adaptive weight allocation algorithm, the accuracy and real-time response ability of advertisement pushing are significantly improved. In actual application scenarios, this method can be widely applied to fields such as e-commerce advertisements, social media advertisements, and search engine advertisements, providing more efficient placement strategies for advertisers, while improving the user experience and the advertising placement effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic structural diagram of a user behavior prediction system for Internet advertisement pushing according to the present invention; Figure 2 It is a schematic flow diagram of a user behavior prediction method for Internet advertisement pushing according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0022] Embodiment: As Figure 1 - Figure 2 shown, the present invention provides a technical solution, a user behavior prediction method for Internet advertisement pushing, and the method includes: Step 100: Collect historical data of user interactions with advertisements in a unit cycle, and establish a historical record database composed of historical records of user interactions with advertisements; Among them, step S100 includes: Step 101: Set a unit cycle with a time length of T0, and obtain historical records of a number of users interacting with advertisements in unit time; Step 102: Collect all advertisement contents in the historical records to form an advertisement candidate set.

[0023] Step S200: Collect time features of each user's interaction with advertisements from the historical record database, and establish a historical interaction weight matrix; Among them, step S200 includes: Step S201: Collect the historical click count c i,j of the i-th user on the j-th advertisement, and calculate the basic weight of the i-th user and the j-th advertisement; Step S202: Obtain the average time interval ∆t ij, the advertisement conversion rate d of the j-th advertisement for the i-th user ij , calculate the interaction weight of the i-th user for the j-th advertisement , where γ represents the interaction decay coefficient, satisfying the condition 0 < γ < 1; The usual calculation method of the advertisement conversion rate is: advertisement conversion rate = (advertisement click volume / number of conversions) × 100%; ‌Number of conversions‌: The number of times a user completes a specific target behavior after clicking an advertisement, such as placing an order, registering, etc.; ‌Advertisement click volume‌: The total number of times an advertisement is clicked by users; Step S203: Normalize the basic weight and the interaction weight to obtain the historical interaction weight matrix Hi,j of the i-th user clicking on the j-th advertisement, , where n is the total number of advertisements in the advertisement candidate set.

[0024] In the embodiment, the interaction records of user 1, user 2, and user 3 regarding advertisement 1, advertisement 2, and advertisement 3 in a unit cycle are collected, where c11 = 5, c12 = 10, c13 = 2, c21 = 1, c22 = 20, c23 = 0, c31 = 3, c32 = 7, c33 = 15; Δt11 = 2, Δt12 = 1, Δt13 = 5, Δt21 = 10, Δt22 = 0.5, Δt23 = 100, Δt31 = 3, Δt32 = 25, Δt33 = 1; d11 = 0.3, d12 = 0.5, d13 = 0.1, d21 = 0.2, d22 = 0.6, d23 = 0.01, d31 = 0.4, d32 = 0.3, d33 = 0.7; Since the 2nd user did not click on the 3rd advertisement, the preset parameters are used to supplement Δt23 and d23; Calculate the basic weight and the interaction weight of each user regarding each advertisement respectively according to the formulas in step S201 and step S202; Normalize the basic weight and the interaction weight to generate the historical interaction weight matrix H, .

[0025] Step S300: Collect the time intervals of user-advertisement interactions in a unit cycle and calculate the long-term interest evaluation value of the user; Among them, step S300 includes: Step S301: Obtain the maximum time interval t and the minimum time interval t of the i-th user clicking on the j-th advertisement in a unit cycle max and the minimum time interval t min ; Step S302: Calculate the long-term interest evaluation value , where α is the balance factor, 0 < α < 1, λ1 represents the first rate control parameter, λ2 represents the second rate control parameter, and 0 < λ2 < λ1 < 1.

[0026] In the embodiment, t min and t max have the dimension of unit time, such as days or hours, and the units of λ1 and λ2 are 1 / unit time; In different scenarios, the value ranges of λ1 and λ2 are also different. For example, in the scenarios of short videos and news, the iteration speed of the content needs to be relatively fast, so the values of the rate control parameters are relatively large. In the scenarios of e-commerce or content communities, the content iteration speed is relatively slow, so the values of the rate control parameters are relatively small; Preferably, when the unit time is in days, in the scenarios of short videos and news, the value range of λ1 is (0.2, 0.5), and the value range of λ2 is (0.05, 0.1). In the scenarios of e-commerce or content communities, the value range of λ1 is (0.05, 0.1), and the value range of λ2 is (0.01, 0.03).

[0027] Step S400: Collect the operation behaviors of the user with the advertisement each time in a unit period, and calculate the short-term interest evaluation value of the user; Among them, step S400 includes: Step S401: Set the threshold score of the operation behavior, collect the operation behaviors of the i-th user with the j-th advertisement in a unit period, and sum the operation scores of the corresponding operation behaviors to obtain the original interaction score S of the i-th user with the j-th advertisement i,j raw ; The operation behaviors include: click, browse, and slide. Among them, the click operation score b1, the browse operation score b2, and the slide operation score b3 are correspondingly set, satisfying the condition b1 > b2 > 0 > b3; Step S403: Obtain the maximum value of the original interaction scores from the original interaction scores of all users with each advertisement, denoted as max(S raw ), and the minimum value of the original interaction scores is denoted as min(S raw ). The original interaction score of the i-th user with the j-th advertisement is Min-Max normalized to obtain the short-term interest evaluation value S of the i-th user with the j-th advertisement i,j .

[0028] In the embodiment, the click operation score b1 = 1, the browsing operation score b2 = 0.5, and the sliding operation score b3 = -0.3 are set. To further accurately distinguish between the browsing operation and the sliding operation, a time threshold discrimination mechanism can be introduced in the actual discrimination process. For example, if the stay time on the page exceeds a certain threshold, it is determined as a browsing operation, and if the stay time on the page is less than a certain threshold, it is determined as a sliding operation; In the embodiment, the original interaction score S of the i-th user with respect to the j-th advertisement is obtained i,j raw , and the short-term interest evaluation value S of the i-th user with respect to the j-th advertisement is calculated i,j , .

[0029] Step S500: Based on the historical interaction records between the user and the advertisement and the behavioral characteristics of the user's interaction with the advertisement, an interest graph of the user is established, and a trend vector of the user is generated through a graph attention network; Among them, step S500 includes: Step S501: Obtain the first evaluation vector r1 composed of the short-term interest evaluation values of the i-th user for all advertisements, and obtain the second evaluation vector r2 composed of the long-term interest evaluation values of the i-th user for all advertisements; Step S502: Concatenate r1 and r2 to obtain the feature vector f of the i-th user i , f i = (r1 || r2), where || represents the concatenation operation. The i-th user is used as node i in the graph structure, and the feature vector fi is the node feature vector of node i; Step S503: Obtain the feature vector f of the i-th user i and the feature vector f of the m-th user m , calculate the similarity sim(f i and the feature vector f m ), set the similarity threshold θ. When sim(f i , f m ) > θ, an undirected edge between node i and node j is established, where sim represents the similarity calculation function; i , f m ) > θ, an undirected edge between node i and node j is established, where sim represents the similarity calculation function; Step S504: Calculate the joint encoding p of node i and node k i,k , , where LeakyReLU represents the linear activation function, a represents the attention vector, W represents the feature transformation matrix, and || represents the concatenation operation; Step S505: Obtain all the nodes that are connected to node i by undirected edges. Together with node i, all these nodes form a set U of neighbor nodes that includes itself. When there is an undirected edge between node m and node i, perform normalization on the joint encoding p of node i and node m to obtain the attention coefficient e of node i to node j i,j ; i,j , ; Step S506: Perform feature aggregation on node features through a graph attention network , where σ represents the Sigmoid function represents the feature aggregation result of the (l + 1)-th propagation layer in the graph attention network. When the feature fusion is completed in the L-th propagation layer, take the output result of the L-th propagation layer as the behavior trend vector vi of the i-th user

[0030] In the embodiment, one input layer and two GAT propagation layers are used for feature fusion. At this time, L = 2 Obtain the initial feature vector of node i through vector splicing and denote it as f i 0 , f i 0 = (r1 || r2); Calculate the first-layer joint encoding p of node i and node j i,k 0 , , where W 0 represents the first-layer feature transformation matrix, and a0 represents the first-layer attention vector Perform normalization on the first-layer joint encoding of node i and node j to obtain the first-layer attention coefficient e i,j 0 , ; Perform feature aggregation on the initial feature vector through the first GAT propagation layer ; Perform transformation on the output result of the first GAT propagation layer through the second-layer feature transformation matrix W 1 and calculate the second-layer joint encoding p of node i and node j i,k 1 , , where a1 represents the second-layer attention vector Perform normalization on the second-layer joint encoding of node i and node j to obtain the second-layer attention coefficient e i,j 1 , ; Perform feature aggregation on the initial feature vector through the second GAT propagation layer ; Take the output result of the second propagation layer as the behavior trend vector vi of the i-th user, where vi = f i 2 .

[0031] Step S600: Collect the current interaction records between the user and the advertisement, calculate the recommendation weight of the user for the advertisement by combining the historical records of the user's interaction with the advertisement, and recommend the advertisement with the largest recommendation weight to another user with the same user behavior characteristics; Among them, step S600 includes: Step S601: Obtain the time interval between the current time and the time when the i-th user last clicked on the j-th advertisement, and record the time interval as the relevant time interval tr i,j ; Step S602: Calculate the recommendation weight E of the i-th user for the j-th advertisement i,j , , where η is a balance coefficient, 0 < η < 1, and β is a time decay coefficient, 0 < β < 1; Obtain the recommendation weights of the i-th user for all advertisements, and record the advertisement with the largest recommendation weight as the target advertisement of the i-th user; Step S603: Denote a user to be pushed at the current time as the current user, obtain the historical interaction records of the current user, and calculate the behavior trend vector q of the current user; Step S604: When the similarity between the behavior trend vector q and the behavior trend vector vi is greater than the threshold, push the target advertisement to the current user.

[0032] The system includes: a historical data management module, a historical interaction weight matrix management module, a long-term interest evaluation module, a short-term interest evaluation module, a trend vector management module, and an advertisement recommendation module; The historical data management module is used to manage the historical data of the interaction between the user and the advertisement. Among them, the historical data management module includes: a basic weight calculation unit, an interaction weight calculation unit, and a historical interaction weight matrix generation unit. The basic weight calculation unit is used to collect the number of times the user clicks on the advertisement from the historical records and calculate the basic weight. The interaction weight calculation unit is used to calculate the interaction weight by obtaining the average time interval between the user's clicks on the advertisement and the advertisement conversion rate. The historical interaction weight matrix generation unit is used to perform normalization processing on the basic weight and the interaction weight to generate a historical interaction weight matrix; The historical interaction weight matrix management module is used to collect the interaction data between users and advertisements. Among them, the long-term interest evaluation module includes: a time interval management unit and a long-term interest evaluation value calculation unit. Among them, the time interval management unit is used to manage the maximum and minimum time intervals for users to click on advertisements within a unit cycle, and the long-term interest evaluation value calculation unit is used to calculate the long-term interest evaluation value of users and manage the historical interaction weight matrix. The long-term interest evaluation module is used to manage the long-term interest evaluation value of users. Among them, the long-term interest evaluation module includes: a time interval management unit and a long-term interest evaluation value calculation unit. Among them, the time interval management unit is used to manage the maximum and minimum time intervals for users to click on advertisements within a unit cycle, and the long-term interest evaluation value calculation unit is used to calculate the long-term interest evaluation value of users. The short-term interest evaluation module is used to manage the short-term interest evaluation value of users. Among them, the short-term interest evaluation module includes: an operation behavior management unit and a short-term interest evaluation value calculation unit. Among them, the operation behavior management unit is used to manage the operation behavior of users and manage the operation scores of different operation behaviors, and the short-term interest evaluation value calculation unit is used to normalize the original interaction score to obtain the short-term interest evaluation value. The trend vector management module is used to construct the interest graph of users and manage the behavior trend vector of users. Among them, the trend vector management module includes: a node evaluation unit, an interest graph management unit, an attention coefficient management unit, and a trend vector management unit. Among them, the node evaluation unit is used to obtain the short-term and long-term interest evaluation values of users to obtain the node evaluation vector of the node, the interest graph management unit is used to evaluate the similarity of the node evaluation vector and manage the graph structure of the interest graph, the attention coefficient management unit is used to calculate the joint coding between user nodes, normalize the joint coding to obtain the attention coefficient, and the trend vector management unit is used to perform feature aggregation on the node features through the graph attention network to generate the behavior trend vector of users. The advertisement recommendation module is used to calculate the recommendation weight of advertisements in combination with historical data and push the advertisements that meet the conditions. Among them, the advertisement recommendation module includes: an advertisement recommendation weight calculation unit, a user comparison unit, and an advertisement push decision unit. Among them, the advertisement recommendation weight calculation unit is used to obtain the time interval between the current time and the time when the user last performed an operation on the advertisement, and calculate the recommendation weight of the user for the advertisement based on the historical interaction data of the user and the current operation time interval. The user comparison unit is used to compare the behavior trend vectors of users, and the advertisement push decision unit is used to push the target advertisement to relevant users according to the comparison result between the similarity of the user behavior trend vector and the preset threshold.

[0033] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above-described exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, in any regard, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Accordingly, all changes that fall within the meaning and scope of the equivalent elements of the claims are intended to be embraced within the present invention. Any reference signs in the claims should not be construed as limiting the claims involved.

Claims

1. A user behavior prediction method for Internet advertising push, characterized in that: The method includes: Step 100: Collect the historical data of the interaction between users and advertisements in a unit cycle, and establish a historical record database composed of the historical records of the interaction between users and advertisements; Step S200: Collect the time characteristics of the interaction between each user and the advertisement from the historical record database, and establish a historical interaction weight matrix; Step S300: Collect the time interval of the interaction between users and advertisements in a unit cycle, and calculate the long-term interest evaluation value of the user; Step S400: Collect the operation behavior of the user with the advertisement each time in a unit cycle, and calculate the short-term interest evaluation value of the user; Step S500: Based on the historical interaction records between the user and the advertisement and the behavioral characteristics of the interaction between the user and the advertisement, establish an interest map of the user, and generate a trend vector of the user through a graph attention network; Step S600: Collect the current interaction record of the user with the advertisement, calculate the recommendation weight of the user for the advertisement in combination with the historical record of the interaction between the user and the advertisement, and recommend the advertisement with the largest recommendation weight to another user with the same user behavior characteristics as the user.

2. The user behavior prediction method for Internet advertising push according to claim 1, wherein: Step S100 includes: Step 101: Set a unit cycle with a time length of T0, and obtain the historical records of the interaction between several users and advertisements in a unit time; Step 102: Collect all the advertisement contents in the historical records to form an advertisement candidate set.

3. The user behavior prediction method for Internet advertisement push according to claim 2, wherein: Step S200 includes: Step S200: Collect the time characteristics of the interaction between each user and the advertisement from the historical record database, and establish a historical interaction weight matrix; Step S200 includes: Step S201: Collect the historical click count c of the i-th user on the j-th advertisement i,j , and calculate the basic weight of the i-th user and the j-th advertisement ; Step S202: Obtain the average time interval ∆t for the i-th user to click on the j-th advertisement ij , and the advertisement conversion rate d of the j-th advertisement for the i-th user ij , and calculate the interaction weight of the i-th user for the j-th advertisement , where γ represents the interaction decay coefficient, satisfying the condition 0 < γ < 1; Step S203: Normalize the base weight and the interaction weight to obtain the historical interaction weight matrix H of the i-th user clicking on the j-th advertisement i,j , , where n is the total number of advertisements in the advertisement candidate set.

4. A user behavior prediction method for Internet advertising push according to claim 3, characterized in that: Step S300 includes: Step S301: Obtain the maximum time interval tmax and the minimum time interval tmin for the i-th user to click on the j-th advertisement in a unit period max and the minimum time interval t min ; Step S302: Calculate the long-term interest evaluation value , where α is a balance factor, 0 < α < 1, λ1 represents the first rate control parameter, λ2 represents the second rate control parameter, and 0 < λ2 < λ1 < 1.

5. The user behavior prediction method for Internet advertisement push according to claim 4, characterized in that: Step S400 includes: Step S401: Set the threshold score for the operation behavior. Collect the operation behavior of the i-th user on the j-th advertisement in a unit period, and sum up the operation scores corresponding to the operation behavior in the unit period to obtain the original interaction score S of the i-th user on the j-th advertisement i,j raw ; Step S402: The operation behaviors include: click, browse, and slide. Among them, the click operation score b1, the browse operation score b2, and the slide operation score b3 are set correspondingly, and b1>b2>0>b3 is satisfied; Step S403: Obtain the maximum value of the original interaction scores of all users for each advertisement, denoted as max(S raw ), and the minimum value of the original interaction scores, denoted as min(S raw ). Perform Min-Max normalization on the original interaction score of the i-th user for the j-th advertisement to obtain the short-term interest evaluation value S i,j of the i-th user for the j-th advertisement.

6. The user behavior prediction method for Internet advertising push according to claim 5, characterized in that: Step S500 includes: Step S501: Obtain the first evaluation vector r1 composed of the short-term interest evaluation values of the i-th user for all advertisements, and obtain the second evaluation vector r2 composed of the long-term interest evaluation values of the i-th user for all advertisements; Step S502: Concatenate r1 and r2 to obtain the feature vector f of the i-th user i , f i = (r1 || r2), where || represents the concatenation operation. Take the i-th user as node i in the graph structure, and the feature vector fi is the node feature vector of node i; Step S503: Obtain the feature vector f of the i-th user i and the feature vector f of the mth user m , calculate the eigenvector f i and the eigenvector f m The similarity sim(f i , f m ), set the similarity threshold θ, when sim(f i , f m )>θ, an undirected edge between node i and node j is established, where sim represents the similarity calculation function; Step S504: Calculate the joint encoding p of node i and node k i,k , , where LeakyReLU represents the linear activation function, a represents the attention vector, W represents the feature transformation matrix, and || represents the concatenation operation; Step S505: Obtain all the nodes that are connected to node i by undirected edges, and jointly form a set U of neighbor nodes including itself with all the nodes and node i. When there is an undirected edge between node m and node i, perform normalization on the joint encoding p i,j of node i and node m to obtain the attention coefficient e i,j of node i to node j ; Step S506: Aggregate node features through a graph attention network, , where σ represents the Sigmoid function, represents the feature aggregation result of the (l + 1)-th propagation layer in the graph attention network. When the feature fusion is completed in the L-th propagation layer, the output result of the L-th propagation layer is used as the behavior trend vector vi of the i-th user.

7. A user behavior prediction method for Internet advertisement push according to claim 6, characterized in that: Step S600 includes: Step S601: Obtain the time interval between the current time and the time when the i-th user last clicked on the j-th advertisement, and denote this time interval as the relevant time interval tr i,j ; Step S602: Calculate the recommendation weight E of the i-th user for the j-th advertisement i,j , , where η is a balance coefficient, 0 < η < 1, and β is a time decay coefficient, 0 < β < 1; Obtain the recommendation weight of the i-th user for all advertisements, and record the advertisement with the largest recommendation weight as the target advertisement of the i-th user; Step S603: Denote a user to be pushed at the current time as the current user, obtain the historical interaction record of the current user, and calculate the behavior trend vector q of the current user; Step S604: When the similarity between the behavior trend vector q and the behavior trend vector vi is greater than the threshold, push the target advertisement to the current user.

8. A user behavior prediction system for Internet advertising push, which is used to execute a user behavior prediction method for Internet advertising push according to any one of claims 1-6, characterized in that: The system includes: A historical data management module, a historical interaction weight matrix management module, a long-term interest evaluation module, a short-term interest evaluation module, a trend vector management module, and an advertisement recommendation module; The historical data management module is used to manage the historical data of user and advertisement interactions. The historical interaction weight matrix management module is used to collect the interaction data between users and advertisements and manage the historical interaction weight matrix. The long-term interest evaluation module is used to manage the long-term interest evaluation values of users. The short-term interest evaluation module is used to manage the short-term interest evaluation values of users. The trend vector management module is used to construct the interest graph of users and manage the behavior trend vectors of users. The advertisement recommendation module is used to calculate the recommendation weights of advertisements in combination with historical data and push the advertisements that meet the conditions.

9. A user behavior prediction system for Internet advertisement push according to claim 8, wherein: The historical data management module includes: a basic weight calculation unit, an interaction weight calculation unit, and a historical interaction weight matrix generation unit. Among them, the basic weight calculation unit is used to collect the number of clicks of users on advertisements from historical records and calculate the basic weight. The interaction weight calculation unit is used to calculate the interaction weight by obtaining the average time interval between user clicks on advertisements and the advertisement conversion rate. The historical interaction weight matrix generation unit is used to perform normalization processing on the basic weight and the interaction weight to generate a historical interaction weight matrix; The long-term interest evaluation module includes: a time interval management unit and a long-term interest evaluation value calculation unit. Among them, the time interval management unit is used to manage the maximum and minimum time intervals of user clicks on advertisements within a unit cycle. The long-term interest evaluation value calculation unit is used to calculate the long-term interest evaluation value of users; The short-term interest evaluation module includes: an operation behavior management unit and a short-term interest evaluation value calculation unit. Among them, the operation behavior management unit is used to manage the operation behaviors of users and manage the operation scores of different operation behaviors. The short-term interest evaluation value calculation unit is used to normalize the original interaction score to obtain the short-term interest evaluation value.

10. A user behavior prediction system for Internet advertisement push according to claim 8, wherein: The trend vector management module includes: a node evaluation unit, an interest graph management unit, an attention coefficient management unit, and a trend vector management unit. Among them, the node evaluation unit is used to obtain the short-term and long-term interest evaluation values of users to obtain the node evaluation vector of the node. The interest graph management unit is used to evaluate the similarity of the node evaluation vectors and manage the graph structure of the interest graph. The attention coefficient management unit is used to calculate the joint encoding between user nodes, normalize the joint encoding to obtain the attention coefficient. The trend vector management unit is used to perform feature aggregation on the node features through a graph attention network to generate the behavior trend vector of the user; The advertisement recommendation module includes: an advertisement recommendation weight calculation unit, a user comparison unit, and an advertisement push decision unit. Among them, the advertisement recommendation weight calculation unit is used to obtain the time interval between the current time and the time when the user last operated on an advertisement, and calculate the recommendation weight of the user for the advertisement based on the user's historical interaction data and the current operation time interval. The user comparison unit is used to compare the behavior trend vectors of users, and the advertisement push decision unit is used to push the target advertisement to relevant users according to the comparison result between the similarity of the user behavior trend vectors and a preset threshold.

Images