Advertisement putting optimization system and method for commodity promotion

Abstract

The invention relates to the technical field of advertisement putting, in particular to an advertisement putting optimization system and method for commodity promotion, and the system comprises a data alignment unit which is used for obtaining the multi-source interaction data of a user, the multi-source interaction data comprises click stream data, transaction data and social behavior data, and the data alignment unit is used for obtaining the multi-source interaction data of the user; performing heterogeneous data space-time alignment on the multi-source interaction data to obtain aligned data corresponding to the multi-source interaction data; and the feature updating unit is used for carrying out cross-modal attention fusion on the alignment data so as to obtain attention features. According to the method, space-time alignment is performed through the multi-source interaction data, and the user characteristics are updated by using a cross-modal attention fusion technology, so that the multi-dimensional interest change and behavior pattern of the user can be accurately captured, the user demand is accurately matched in the advertisement putting process, and the advertisement effect is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of advertisement placement, and in particular to an advertisement placement optimization system and method for commodity promotion. Background Art

[0002] Traditional advertising systems often rely on static user portraits and a single data source (such as browsing history or purchase behavior) to match advertisements. This approach cannot accurately capture the dynamic changes and multi-dimensional characteristics of user interests, resulting in the inability to adapt to changes in user interests when advertising is delivered, which in turn affects the click-through rate and conversion rate of advertisements. In addition, traditional systems usually use simple tags or data based on historical behavior for interest matching, lack deep feature mapping methods, and do not use advanced technology to perform refined mathematical modeling of user interests. Therefore, the probability of mismatch between advertisements and user interests is high. In addition, traditional systems are generally based on simple budget allocation strategies, often allocating budgets according to preset rules, and lack a real-time dynamic adjustment mechanism for optimizing advertiser budgets. This method easily leads to waste of resources and unsatisfactory advertising effects, especially in the case of rapidly changing market environments. The budget allocation of traditional systems may not be able to adapt to rapidly changing market demands. In addition, traditional advertising systems often lack sufficient flexibility and dynamic adjustment capabilities. Advertising delivery strategies are generally difficult to adjust in a timely manner after being set in the initial stage of delivery, and cannot be optimized based on real-time feedback from the market and users. As a result, advertising effects may decline as the market environment changes. Summary of the invention

[0003] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide an advertisement placement optimization system and method for commodity promotion.

[0004] The technical solution adopted to solve the above technical problems is: an advertisement delivery optimization system for product promotion, comprising: A data alignment unit, the data alignment unit is used to obtain multi-source interaction data of the user, wherein the multi-source interaction data includes click stream data, transaction data and social behavior data, and perform heterogeneous data spatiotemporal alignment on the multi-source interaction data to obtain aligned data corresponding to the multi-source interaction data; A feature updating unit, the feature updating unit is used to perform cross-modal attention fusion on the aligned data to obtain attention features, and perform feature update on the user through the attention features according to a dual-gating mechanism to update user features; A benefit modeling unit, the benefit modeling unit is used to perform Hilbert space mapping on the user characteristics to obtain an orthogonal interest basis vector, obtain an ad slot feature vector and an environment feature vector, and construct a user-ad slot benefit matrix according to the orthogonal interest basis vector, the environment feature vector and the ad slot feature vector; A competition modeling unit, wherein the competition modeling unit is used to define an advertiser strategy space, wherein the advertiser strategy space includes an advertisement slot budget allocation, a creative aggressiveness coefficient, and a time period preference phase, and an indirect competition effect coefficient is calculated based on the advertiser strategy space and the user-ad slot benefit matrix.

[0005] Preferably, the system further comprises: A budget allocation unit, the budget allocation unit is used to construct a budget allocation objective function according to the indirect competition effect coefficient and the user-advertising slot benefit matrix, and solve the budget allocation objective function according to the reverse induction method to obtain an optimal budget allocation matrix for the advertiser; An advertisement delivery unit, the advertisement delivery unit is used to obtain the spatiotemporal feature vector when the advertisement is delivered, define a Q function approximator according to the advertiser's optimal budget allocation matrix and the spatiotemporal feature vector, and optimize the advertisement delivery strategy by maximizing the Q function approximator according to deep reinforcement learning to obtain the best bidding strategy.

[0006] Preferably, performing spatiotemporal alignment of heterogeneous data on the multi-source interactive data to obtain aligned data corresponding to the multi-source interactive data includes: The transaction data is interpolated by cubic spline to obtain an interpolation result of the transaction data, wherein the calculation formula of the interpolation result of the transaction data is as follows: ; in, Indicates at time The interpolation result of the transaction data at Indicates the total number of transaction data. represents the basis function based on Lagrange interpolation, and , Indicates at time Number of transactions at A social behavior heat decay coefficient is defined to perform heat decay processing on the social behavior data to obtain a heat decay result of the social behavior data, wherein the calculation formula of the heat decay result of the social behavior data is as follows: ; in, Indicates at time The heat decay results of the social behavior data at Indicates at time Social behavior data represents the social behavior heat decay coefficient, and , Represents the seconds of the minute; The interpolation results of the transaction data and the heat decay results of the social behavior data are aligned to a unified time axis according to the click stream data to obtain the aligned data corresponding to the multi-source interaction data, wherein the calculation formula of the aligned data is as follows: ; in, Indicates alignment data, Indicates at time Clickstream data from .

[0007] Preferably, cross-modal attention fusion is performed on the aligned data to obtain attention features, including: The attention head of each modality in the aligned data is calculated to obtain the attention corresponding to each modality in the aligned data, wherein the calculation formula for the attention corresponding to each modality in the aligned data is as follows: ; in, Indicates the alignment data The attention corresponding to each modality, , and represents the query matrix, key matrix and value matrix, , and Indicates the alignment data The projection matrix corresponding to each mode is represents the dimension of the feature, represents a learnable modality mask matrix; The attention corresponding to each modality in the aligned data is concatenated and linearly transformed to obtain the attention feature, where the calculation formula of the attention feature is as follows: ; in, Represents the attention feature, Represents a splicing operation, Represents a linear transformation matrix.

[0008] Preferably, updating the user's features through the attention features according to the dual-gating mechanism to obtain user features includes: According to the first gating mechanism, the importance score is performed through the attention feature in combination with the historical user features to obtain an importance score, wherein the calculation formula of the importance score is as follows: ; in, represents the importance score, represents the activation function, represents the first trainable weight matrix, represents the user features of the previous time step, Represents the attention feature of the current time step; According to the second gating mechanism, the user's features are fused by the importance score to obtain fusion information, wherein the calculation formula of the fusion information is as follows: ; in, represents fusion information, represents the second trainable weight matrix; The user characteristics are updated according to the importance score and the fusion information to update the user characteristics, wherein the updating formula of the user characteristics is as follows: ; in, represents the updated user features, represents the hyperbolic tangent function, represents the third trainable weight matrix; When the importance score is greater than a preset importance score threshold, the updated user feature corresponding to the importance score is stored as an event to obtain a memory library, wherein the expression of the memory library is as follows: ; in, represents the updated user feature sequence stored in the memory bank, Indicates the size of the memory bank.

[0009] Preferably, the calculation formula of the user-advertising space benefit matrix is ​​as follows: ; in, Indicates user and advertising space The benefits between represents the orthogonal interest basis vector, represents the ad slot feature vector, represents the environmental feature vector, Represents the time decay factor, which is used to control the timeliness of the ad slot. Indicates ad space The timeliness of Represents the historical click-through rate mixing coefficient, balancing the effects of new and old ad slots. Indicates the historical click-through rate of the ad slot.

[0010] Preferably, the calculation formula of the indirect competition effect coefficient is as follows: ; in, Indicates advertiser With advertisers The indirect interference strength between Indicates advertiser In the advertising space budget allocation on Indicates user and advertising space The benefits between Represents the differences in advertiser strategy space; The budget allocation objective function is as follows: ; in, Indicates ad space The benefits represents the budget competition factor, represents the regularization coefficient, represents the indirect interference coefficient, Indicates advertiser total budget.

[0011] Preferably, the budget allocation objective function is solved according to the reverse induction method to obtain the advertiser's optimal budget allocation matrix, including: Initialize the budget allocation strategy matrix for each advertiser; Iteratively update the budget allocation strategy of each advertiser, wherein the update formula of the advertiser's budget allocation strategy is as follows: ; in, Indicates at time Advertiser budget allocation strategy, Indicates that all advertisers at the time The sum of budget allocation strategies; A convergence condition is defined to control the accuracy of updating the budget allocation strategy, wherein the convergence condition is as follows: ; in, Indicates at time The budget allocation strategy matrix is Indicates at time The budget allocation strategy matrix is Indicates the preset accuracy threshold; When the update of the advertiser's budget allocation strategy meets a preset convergence condition, the iterative update of the advertiser's budget allocation strategy is stopped to obtain the advertiser's optimal budget allocation moment.

[0012] Preferably, a Q function approximator is defined according to the advertiser's optimal budget allocation matrix and the spatiotemporal feature vector, and an advertising delivery strategy is optimized by maximizing the Q function approximator according to deep reinforcement learning to obtain an optimal bidding strategy, including: A Q-function approximator is defined according to the advertiser's optimal budget allocation matrix and the spatiotemporal feature vector, wherein the Q-function approximator is as follows: ; in, represents the Q-function myopic device, represents the weight parameter of the Q function, and represents the network weight, represents the spatiotemporal feature vector, and the spatiotemporal feature vector is obtained by extracting the geographic location information, time information and device information through MLP. Indicates the ad slot and bid embedding corresponding to the action. Indicates the user's status information; Generate actions based on states. That is, given a state, the policy network outputs the optimal ad slot selection and bid combination. The parameters of the Q-function approximator are optimized by back-propagation, and the policy network is updated so that the Q-function approximator can maximize the reward in the environment space.

[0013] The technical solution adopted to solve the above technical problem is: an advertisement delivery optimization method for product promotion, which is applicable to the advertisement delivery optimization system for product promotion, comprising: Acquire multi-source interaction data of the user, wherein the multi-source interaction data includes click stream data, transaction data and social behavior data, and perform spatiotemporal alignment of heterogeneous data on the multi-source interaction data to obtain aligned data corresponding to the multi-source interaction data; Performing cross-modal attention fusion on the aligned data to obtain attention features, and updating features of the user through the attention features according to a dual-gating mechanism to update user features; Performing Hilbert space mapping on the user features to obtain orthogonal interest basis vectors, obtaining an advertisement slot feature vector and an environment feature vector, and constructing a user-ad slot benefit matrix according to the orthogonal interest basis vectors, the environment feature vectors, and the advertisement slot feature vectors; Define an advertiser strategy space, wherein the advertiser strategy space includes an advertisement slot budget allocation, a creative aggressiveness coefficient, and a time period preference phase, and calculate an indirect competition effect coefficient based on the advertiser strategy space; Constructing a budget allocation objective function according to the indirect competition effect coefficient and the user-advertising slot benefit matrix, and solving the budget allocation objective function according to the reverse induction method to obtain an optimal budget allocation matrix for the advertiser; The spatiotemporal feature vector of the advertisement is obtained, a Q-function approximator is defined according to the advertiser's optimal budget allocation matrix and the spatiotemporal feature vector, and the advertisement delivery strategy is optimized by maximizing the Q-function approximator according to deep reinforcement learning to obtain the best bidding strategy.

[0014] The beneficial effects of the present invention are as follows: (1) The present invention performs spatiotemporal alignment through multi-source interactive data and utilizes cross-modal attention fusion technology to update user features, which can accurately capture the user's multi-dimensional interest changes and behavior patterns, thereby accurately matching user needs during the advertising process and improving advertising effects. In addition, by mapping user features to Hilbert space and obtaining orthogonal interest basis vectors, the different dimensions of user potential interests can be more accurately captured. This mapping method helps to accurately match user interests with advertising positions, reduce the mismatch between advertising and user interests, and thus improve advertising click-through rates and conversion rates. (2) The present invention further introduces the calculation of the indirect competition effect coefficient through the user-advertising position benefit matrix and combines it with the advertiser's strategy space. This design can effectively optimize advertising. The budget allocation of advertisers can maximize the return on advertising investment, while avoiding resource waste or reduced advertising effect due to unreasonable budget allocation. Through deep reinforcement learning technology, the Q function approximator is used to continuously optimize the advertising delivery strategy. Reinforcement learning can adjust the strategy based on real-time feedback, so that advertising delivery can continuously adapt to market changes and obtain the best bidding strategy. This dynamic optimization process can help advertisers continuously improve the advertising delivery effect and adjust budget allocation in real time to cope with different market environments. (3) The present invention can optimize the advertising delivery strategy while considering market competition by introducing the indirect competition effect coefficient. This can not only help advertisers better allocate budgets, but also avoid excessive competition and market saturation, improve advertising effects, and ensure that each advertiser can obtain the best delivery return in a highly competitive environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram of the system architecture of the overall system in an embodiment of the present invention; Figure 2 A schematic flow chart of the steps of an overall method in an embodiment of the present invention.

[0016] Figure numerals: 1. Data alignment unit; 2. Feature update unit; 3. Benefit modeling unit; 4. Competition modeling unit; 5. Budget allocation unit; 6. Advertising delivery unit. DETAILED DESCRIPTION

[0017] Embodiment 1, as Figure 1 As shown, the present invention proposes an advertisement delivery optimization system for commodity promotion, comprising: The data alignment unit 1 is used to obtain multi-source interaction data of the user, wherein the multi-source interaction data includes click stream data, transaction data and social behavior data, and perform spatiotemporal alignment of heterogeneous data on the multi-source interaction data to obtain aligned data corresponding to the multi-source interaction data; Feature updating unit 2, which is used to perform cross-modal attention fusion on the aligned data to obtain attention features, and update user features through the attention features according to the dual-gating mechanism to update user features; The benefit modeling unit 3 is used to perform Hilbert space mapping on user characteristics to obtain orthogonal interest basis vectors, obtain ad slot feature vectors and environment feature vectors, and construct a user-ad slot benefit matrix based on the orthogonal interest basis vectors, the environment feature vectors and the ad slot feature vectors; Competition modeling unit 4, competition modeling unit 4 is used to define the advertiser strategy space, wherein the advertiser strategy space includes advertising space budget allocation, creative aggressiveness coefficient and time period preference phase, and calculates the indirect competition effect coefficient based on the advertiser strategy space and the user-advertising space benefit matrix.

[0018] In the present invention, multi-source interactive data includes interactive data between users and goods or advertisements. Usually, clickstream data reflects the user's click behavior on advertisements or goods, transaction data reflects whether the user has made a purchase, and social behavior data reflects the user's interaction on social platforms (such as likes, comments, etc.); heterogeneous data spatiotemporal alignment is due to the different spatiotemporal characteristics of these data sources, and timestamp alignment is required to make them reflect the user's real behavior at the same time; cross-modal attention fusion is due to the different feature distributions and importance of different data sources (such as clickstreams, transactions, and social data), and attention mechanisms (especially cross-modal attention mechanisms) are required to perform weighted fusion on them to highlight the importance of each data in the current task; Hilbert space mapping refers to mapping user features into Hilbert space, which can better process high-dimensional data, and through the selection of orthogonal bases, each interest dimension can be made They are independent and unrelated to each other to avoid interference between different interests; the orthogonal interest basis vectors represent the user's independent interest points. Through orthogonalization, the model can more clearly distinguish the user's interest tendencies; the benefit matrix refers to the user-advertising slot benefit matrix constructed by combining the user's interests with the advertising slot and environmental characteristics. The matrix describes the matching degree between the user and each advertising slot and can be used to evaluate the effectiveness of advertising; the advertiser's strategy space includes budget allocation, creative aggressiveness coefficient (i.e. the promotion strength of the advertisement) and time period preference phase (i.e. in what time period the advertiser wants to place the advertisement). These factors affect the choice of advertising slot and the effectiveness of advertising; advertising is not only affected by direct competitors, but also by other indirect competitive factors. By analyzing the competitive relationship between advertising slots and advertisers, the model can estimate these indirect effects and incorporate them into the delivery decision to avoid repeated or ineffective delivery.

[0019] In an optional embodiment, the system further includes: Budget allocation unit 5, budget allocation unit 5 is used to construct a budget allocation objective function according to the indirect competition effect coefficient and the user-advertising space benefit matrix, and solve the budget allocation objective function according to the reverse induction method to obtain the advertiser's optimal budget allocation matrix; Advertisement delivery unit 6, which is used to obtain the spatiotemporal feature vector during advertisement delivery, define a Q-function approximator according to the advertiser's optimal budget allocation matrix and the spatiotemporal feature vector, and optimize the advertisement delivery strategy by maximizing the Q-function approximator according to deep reinforcement learning to obtain the best bidding strategy.

[0020] It should be noted that the indirect competition effect coefficient represents the competition effect between advertising slots or advertisers during the advertising process. These effects include the impact of other advertisers’ advertising, the supply and demand relationship of advertising slots and other factors. The introduction of indirect competition effect makes the budget allocation not only consider the direct competition between advertisers, but also consider how these external factors affect the advertising effect. The construction of the objective function is usually to maximize the total benefit of the advertiser. The objective function will combine the indirect competition effect and the benefit matrix to determine the budget allocation of different advertising slots. The reverse induction method plays an important role in solving the budget allocation problem. Usually, this method starts from the final result and reversely calculates the optimal strategy in the decision-making process step by step. The reverse induction method can help Starting from the effect of advertising delivery, the budget allocation of each advertising position is gradually derived; the spatiotemporal feature vector in advertising delivery reflects the time period, location, user activities and other information of advertising delivery, which have a significant impact on the effect of advertising delivery, because different time periods, locations and user behavior characteristics will affect the effect indicators such as the click-through rate and conversion rate of advertising; Q function is a core concept in reinforcement learning, which is used to represent the expected benefits of taking an action under a given state. In advertising delivery, Q function can quantify the benefits that can be obtained by delivering advertisements in a specific spatiotemporal environment. By defining the Q function approximator, the relationship between decision-making and benefits in the advertising delivery process can be simulated, thereby providing advertisers with the best delivery strategy; deep reinforcement learning is used here to optimize advertising delivery strategies. By constantly interacting with the environment (that is, constantly adjusting the advertising delivery strategy and observing the effect), the deep reinforcement learning model will gradually learn how to maximize the Q function, thereby optimizing the advertising bidding strategy.

[0021] Embodiment 2, an advertisement placement optimization system for product promotion proposed by the present invention, compared with embodiment 1, this embodiment further includes: performing spatiotemporal alignment of heterogeneous data on multi-source interactive data to obtain aligned data corresponding to the multi-source interactive data, including: The transaction data is interpolated by cubic spline to obtain the interpolation result of the transaction data, wherein the calculation formula of the interpolation result of the transaction data is as follows: ; in, Indicates at time The interpolation result of the transaction data at Indicates the total number of transaction data. represents the basis function based on Lagrange interpolation, and , Indicates at time Number of transactions at The social behavior heat decay coefficient is defined to perform heat decay processing on the social behavior data to obtain the heat decay result of the social behavior data. The calculation formula of the heat decay result of the social behavior data is as follows: ; in, Indicates at time The heat decay results of the social behavior data at Indicates at time Social behavior data represents the social behavior heat decay coefficient, and , Represents the seconds of the minute; According to the clickstream data, the interpolation results of the transaction data and the heat decay results of the social behavior data are aligned to the same time axis to obtain the aligned data corresponding to the multi-source interaction data. The calculation formula of the aligned data is as follows: ; in, Indicates alignment data, Indicates at time Clickstream data from .

[0022] In this embodiment, cubic spline interpolation is a method of interpolation by connecting multiple segments of cubic polynomials (each segment of the polynomial has a continuous and smooth derivative at each interpolation point). It is very common in practical applications, especially for problems that require maintaining data smoothness (such as curve fitting); Lagrange interpolation is a method of interpolation by calculating a polynomial of a given set of data points; social behavior data usually has a heat decay effect, which means that the influence of social behavior will gradually weaken over time.

[0023] In an optional embodiment, cross-modal attention fusion is performed on the aligned data to obtain attention features, including: The attention head of each modality in the aligned data is calculated to obtain the attention corresponding to each modality in the aligned data. The calculation formula for the attention corresponding to each modality in the aligned data is as follows: ; in, Indicates the alignment data The attention corresponding to each modality, , and represents the query matrix, key matrix and value matrix, , and Indicates the alignment data The projection matrix corresponding to each mode is represents the dimension of the feature, represents a learnable modality mask matrix; The attention corresponding to each modality in the aligned data is concatenated and linearly transformed to obtain the attention feature, where the calculation formula of the attention feature is as follows: ; in, Represents the attention feature, Represents a splicing operation, Represents a linear transformation matrix.

[0024] In an optional embodiment, the user feature is updated through the attention feature according to the double gating mechanism to obtain the user feature, including: According to the first gating mechanism, the importance score is scored by combining the historical user characteristics with the attention feature to obtain the importance score, where the calculation formula of the importance score is as follows: ; in, represents the importance score, represents the activation function, represents the first trainable weight matrix, represents the user features of the previous time step, Represents the attention feature of the current time step; According to the second gating mechanism, user features are fused through importance scores to obtain fusion information, where the calculation formula of the fusion information is as follows: ; in, represents fusion information, represents the second trainable weight matrix; The user features are updated according to the importance score and the fusion information to update the user features. The update formula of the user features is as follows: ; in, represents the updated user features, represents the hyperbolic tangent function, represents the third trainable weight matrix; When the importance score is greater than the preset importance score threshold, the updated user features corresponding to the importance score are stored as events to obtain a memory library, where the expression of the memory library is as follows: ; in, represents the updated user feature sequence stored in the memory bank, Indicates the size of the memory bank.

[0025] It should be noted that the memory bank is used to save important user feature sequences. Whenever the user's behavior or features undergo significant changes at a certain time step, the model will store the feature sequence in the memory bank for future reference. The size of the memory bank may be limited, and when storing new features, there may be strategies to decide whether to delete old features.

[0026] In an optional embodiment, the calculation formula of the user-advertising slot benefit matrix is ​​as follows: ; in, Indicates user and advertising space The benefits between represents the orthogonal interest basis vector, represents the ad slot feature vector, represents the environmental feature vector, Represents the time decay factor, which is used to control the timeliness of the ad slot. Indicates ad space The timeliness of Represents the historical click-through rate mixing coefficient, balancing the effects of new and old ad slots. Indicates the historical click-through rate of the ad slot.

[0027] In an optional embodiment, the calculation formula of the indirect competition effect coefficient is as follows: ; in, Indicates advertiser With advertisers The indirect interference strength between Indicates advertiser In the advertising space budget allocation on Indicates user and advertising space The benefits between Represents the differences in advertiser strategy space; The budget allocation objective function is as follows: ; in, Indicates ad space The benefits represents the budget competition factor, represents the regularization coefficient, represents the indirect interference coefficient, Indicates advertiser total budget.

[0028] In an optional embodiment, the budget allocation objective function is solved according to the reverse induction method to obtain the advertiser's optimal budget allocation matrix, including: Initialize the budget allocation strategy matrix for each advertiser; Iteratively update the budget allocation strategy of each advertiser, where the update formula of the advertiser's budget allocation strategy is as follows: ; in, Indicates at time Advertiser budget allocation strategy, Indicates that all advertisers at the time The sum of budget allocation strategies; Define the convergence condition to control the accuracy of the budget allocation strategy update, where the convergence condition is as follows: ; in, Indicates at time The budget allocation strategy matrix is Indicates at time The budget allocation strategy matrix is Indicates the preset accuracy threshold; When the update of the advertiser's budget allocation strategy meets the preset convergence condition, the iterative update of the advertiser's budget allocation strategy is stopped to obtain the advertiser's optimal budget allocation matrix.

[0029] It should be noted that each advertiser has a budget allocation strategy matrix, which is usually a two-dimensional matrix, in which each row represents an advertiser's budget allocation strategy, and each column represents the budget allocation at different times (or different advertising positions, etc.). The initialization matrix can be randomly initialized or set based on some prior information. Initialization is usually the starting point for solving nonlinear optimization problems.

[0030] In an optional embodiment, a Q-function approximator is defined according to the advertiser's optimal budget allocation matrix and the spatiotemporal feature vector, and the advertising delivery strategy is optimized by maximizing the Q-function approximator according to deep reinforcement learning to obtain the best bidding strategy, including: The Q-function approximator is defined according to the advertiser's optimal budget allocation matrix and the spatiotemporal eigenvector, where the Q-function approximator is as follows: ; in, represents the Q-function myopic device, represents the weight parameter of the Q function, and represents the network weight, represents the spatiotemporal feature vector, and the spatiotemporal feature vector is obtained by extracting the geographic location information, time information and device information through MLP. Indicates the ad slot and bid embedding corresponding to the action. Indicates the user's status information; Generate actions based on states. That is, given a state, the policy network outputs the optimal ad slot selection and bid combination. Through back propagation, the parameters of the Q-function approximator are optimized and the policy network is updated so that the Q-function approximator can maximize the reward in the environment space.

[0031] It should be noted that geographic location information usually indicates the user's location characteristics when the advertiser wants to display advertisements at a specific location. The geographic location can be represented by longitude and latitude, city information, etc.; time information indicates that the advertiser wants to run advertisements within a certain period of time. Time information is usually represented by different granularities such as hours, days, weeks, etc., or directly encoded using a timestamp; device information reflects the type of device used by the user, such as mobile phones, tablets, computers, etc.

[0032] Embodiment three, as Figure 2 As shown, the present invention proposes an advertisement delivery optimization method for commodity promotion, which is applicable to the advertisement delivery optimization system for commodity promotion, and includes: S1. Acquire multi-source interaction data of users, wherein the multi-source interaction data includes click stream data, transaction data, and social behavior data, and perform spatiotemporal alignment of heterogeneous data on the multi-source interaction data to obtain aligned data corresponding to the multi-source interaction data; S2. Perform cross-modal attention fusion on the aligned data to obtain attention features, and update the user features through the attention features according to the double-gating mechanism to update the user features; S3, performing Hilbert space mapping on user features to obtain orthogonal interest basis vectors, obtaining ad slot feature vectors and environment feature vectors, and constructing a user-ad slot benefit matrix based on the orthogonal interest basis vectors, environment feature vectors, and ad slot feature vectors; S4. define an advertiser strategy space, wherein the advertiser strategy space includes the allocation of advertising budget, creative aggressiveness coefficient and time period preference phase, and calculate the indirect competition effect coefficient according to the advertiser strategy space; S5. construct a budget allocation objective function according to the indirect competition effect coefficient and the user-advertising slot benefit matrix, and solve the budget allocation objective function according to the reverse induction method to obtain the advertiser's optimal budget allocation matrix; S6. Obtain the spatiotemporal feature vector when the advertisement is delivered, define a Q-function approximator according to the advertiser's optimal budget allocation matrix and the spatiotemporal feature vector, and optimize the advertisement delivery strategy by maximizing the Q-function approximator according to deep reinforcement learning to obtain the best bidding strategy.

[0033] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge scope of technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. An advertising placement optimization system for commodity promotion, characterized in that: include: A data alignment unit (1), the data alignment unit (1) is used to obtain multi-source interaction data of a user, wherein the multi-source interaction data includes click stream data, transaction data and social behavior data, and perform spatiotemporal alignment of heterogeneous data on the multi-source interaction data to obtain aligned data corresponding to the multi-source interaction data; A feature updating unit (2), the feature updating unit (2) being used to perform cross-modal attention fusion on the aligned data to obtain attention features, and to perform feature update on the user through the attention features according to a dual-gating mechanism to update user features; A benefit modeling unit (3), the benefit modeling unit (3) is used to perform Hilbert space mapping on the user characteristics to obtain an orthogonal interest basis vector, obtain an advertisement position feature vector and an environment feature vector, and construct a user-advertising position benefit matrix based on the orthogonal interest basis vector, the environment feature vector and the advertisement position feature vector; A competition modeling unit (4), the competition modeling unit (4) is used to define an advertiser strategy space, wherein the advertiser strategy space includes an advertisement slot budget allocation, a creative aggressiveness coefficient and a time period preference phase, and an indirect competition effect coefficient is calculated based on the advertiser strategy space and the user-ad slot benefit matrix.

2. The advertisement placement optimization system for commodity promotion according to claim 1, characterized in that: The system further comprises: A budget allocation unit (5), the budget allocation unit (5) being used to construct a budget allocation objective function according to the indirect competition effect coefficient and the user-advertising space benefit matrix, and to solve the budget allocation objective function according to the reverse induction method to obtain an optimal budget allocation matrix for the advertiser; An advertisement delivery unit (6), the advertisement delivery unit (6) is used to obtain a spatiotemporal feature vector when an advertisement is delivered, define a Q function approximator according to the advertiser's optimal budget allocation matrix and the spatiotemporal feature vector, and optimize the advertisement delivery strategy by maximizing the Q function approximator according to deep reinforcement learning to obtain an optimal bidding strategy.

3. The advertisement placement optimization system for commodity promotion according to claim 2, characterized in that: Performing spatiotemporal alignment of heterogeneous data on the multi-source interactive data to obtain aligned data corresponding to the multi-source interactive data includes: The transaction data is interpolated by cubic spline to obtain an interpolation result of the transaction data, wherein the calculation formula of the interpolation result of the transaction data is as follows: ； in, Indicates at time The interpolation result of the transaction data at Indicates the total number of transaction data. represents the basis function based on Lagrange interpolation, and , Indicates at time Number of transactions at A social behavior heat decay coefficient is defined to perform heat decay processing on the social behavior data to obtain a heat decay result of the social behavior data, wherein the calculation formula of the heat decay result of the social behavior data is as follows: ； in, Indicates at time The heat decay results of the social behavior data at Indicates at time Social behavior data represents the social behavior heat decay coefficient, and , Represents the seconds of the minute; The interpolation results of the transaction data and the heat decay results of the social behavior data are aligned to a unified time axis according to the click stream data to obtain the aligned data corresponding to the multi-source interaction data, wherein the calculation formula of the aligned data is as follows: ； in, Indicates alignment data, Indicates at time Clickstream data from .

4. The advertisement placement optimization system for commodity promotion according to claim 3, characterized in that: Perform cross-modal attention fusion on the aligned data to obtain attention features, including: The attention head of each modality in the aligned data is calculated to obtain the attention corresponding to each modality in the aligned data, wherein the calculation formula for the attention corresponding to each modality in the aligned data is as follows: ； in, Indicates the alignment data The attention corresponding to each modality, , and represents the query matrix, key matrix and value matrix, , and Indicates the alignment data The projection matrix corresponding to each mode is Represents the dimension of the feature, represents a learnable modality mask matrix; The attention corresponding to each modality in the aligned data is concatenated and linearly transformed to obtain the attention feature, where the calculation formula of the attention feature is as follows: ； in, Represents the attention feature, Represents a splicing operation, Represents a linear transformation matrix.

5. The advertisement placement optimization system for commodity promotion according to claim 4, characterized in that: The user feature is updated by the attention feature according to the double-gating mechanism to obtain the user feature, including: According to the first gating mechanism, the importance score is performed through the attention feature in combination with the historical user features to obtain an importance score, wherein the calculation formula of the importance score is as follows: ； in, represents the importance score, represents the activation function, represents the first trainable weight matrix, represents the user features of the previous time step, Represents the attention feature of the current time step; According to the second gating mechanism, the user's features are fused by the importance score to obtain fusion information, wherein the calculation formula of the fusion information is as follows: ； in, represents fusion information, represents the second trainable weight matrix; The user characteristics are updated according to the importance score and the fusion information to update the user characteristics, wherein the updating formula of the user characteristics is as follows: ； in, represents the updated user features, represents the hyperbolic tangent function, represents the third trainable weight matrix; When the importance score is greater than a preset importance score threshold, the updated user feature corresponding to the importance score is stored as an event to obtain a memory library, wherein the expression of the memory library is as follows: ； in, represents the updated user feature sequence stored in the memory bank, Indicates the size of the memory bank.

6. The advertisement placement optimization system for commodity promotion according to claim 5, characterized in that: The calculation formula of the user-advertising space benefit matrix is ​​as follows: ； in, Indicates user and advertising space The benefits between represents the orthogonal interest basis vector, represents the ad slot feature vector, represents the environmental feature vector, Represents the time decay factor, which is used to control the timeliness of the ad slot. Indicates ad space The timeliness of Represents the historical click-through rate mixing coefficient, balancing the effects of new and old ad slots. Indicates the historical click-through rate of the ad slot.

7. The advertisement placement optimization system for commodity promotion according to claim 6, characterized in that: The calculation formula of the indirect competition effect coefficient is as follows: ； in, Indicates advertiser With advertisers The indirect interference strength between Indicates advertiser In the advertising space budget allocation on Indicates user and advertising space The benefits between Represents the differences in advertiser strategy space; The budget allocation objective function is as follows: ； in, Indicates ad space The benefits represents the budget competition factor, represents the regularization coefficient, represents the indirect interference coefficient, Indicates advertiser total budget.

8. The advertisement placement optimization system for commodity promotion according to claim 7, characterized in that: The budget allocation objective function is solved according to the reverse induction method to obtain the advertiser's optimal budget allocation matrix, including: Initialize the budget allocation strategy matrix for each advertiser; Iteratively update the budget allocation strategy of each advertiser, wherein the update formula of the advertiser's budget allocation strategy is as follows: ； in, Indicates at time Advertiser budget allocation strategy, Indicates that all advertisers at the time The sum of budget allocation strategies; A convergence condition is defined to control the accuracy of updating the budget allocation strategy, wherein the convergence condition is as follows: ； in, Indicates at time The budget allocation strategy matrix is Indicates at time The budget allocation strategy matrix is Indicates the preset accuracy threshold; When the update of the advertiser's budget allocation strategy meets a preset convergence condition, the iterative update of the advertiser's budget allocation strategy is stopped to obtain the advertiser's optimal budget allocation matrix.

9. The advertisement placement optimization system for commodity promotion according to claim 8, characterized in that: A Q-function approximator is defined according to the advertiser's optimal budget allocation matrix and the spatiotemporal feature vector, and an advertising delivery strategy is optimized by maximizing the Q-function approximator according to deep reinforcement learning to obtain an optimal bidding strategy, including: A Q-function approximator is defined according to the advertiser's optimal budget allocation matrix and the spatiotemporal feature vector, wherein the Q-function approximator is as follows: ； in, represents the Q-function myopic device, represents the weight parameter of the Q function, and represents the network weight, represents the spatiotemporal feature vector, and the spatiotemporal feature vector is obtained by extracting the geographic location information, time information and device information through MLP. Indicates the ad slot and bid embedding corresponding to the action. Indicates the user's status information; Generate actions based on states. That is, given a state, the policy network outputs the optimal ad slot selection and bid combination. The parameters of the Q-function approximator are optimized by back-propagation, and the policy network is updated so that the Q-function approximator can maximize the reward in the environment space.

10. An advertisement placement optimization method for product promotion, which is applicable to an advertisement placement optimization system for product promotion as claimed in any one of claims 9, characterized in that: include: Acquire multi-source interaction data of the user, wherein the multi-source interaction data includes click stream data, transaction data, and social behavior data, and perform spatiotemporal alignment of heterogeneous data on the multi-source interaction data to obtain aligned data corresponding to the multi-source interaction data; Performing cross-modal attention fusion on the aligned data to obtain attention features, and updating features of the user through the attention features according to a dual-gating mechanism to update user features; Performing Hilbert space mapping on the user features to obtain orthogonal interest basis vectors, obtaining an advertisement slot feature vector and an environment feature vector, and constructing a user-ad slot benefit matrix according to the orthogonal interest basis vectors, the environment feature vectors, and the advertisement slot feature vectors; Define an advertiser strategy space, wherein the advertiser strategy space includes an advertisement slot budget allocation, a creative aggressiveness coefficient, and a time period preference phase, and calculate an indirect competition effect coefficient based on the advertiser strategy space; Constructing a budget allocation objective function according to the indirect competition effect coefficient and the user-advertising slot benefit matrix, and solving the budget allocation objective function according to the reverse induction method to obtain an optimal budget allocation matrix for the advertiser; The spatiotemporal feature vector of the advertisement is obtained, a Q-function approximator is defined according to the advertiser's optimal budget allocation matrix and the spatiotemporal feature vector, and the advertisement delivery strategy is optimized by maximizing the Q-function approximator according to deep reinforcement learning to obtain the best bidding strategy.

Images