Data delivery optimization method and system based on multi-modal data

Abstract

The invention discloses a data delivery optimization method based on multi-modal data, and the method comprises the steps: obtaining user identity information and associated multi-modal data, and inputting the user identity information and the associated multi-modal data into a portrait building model; the portrait establishment model adjusts the analysis weight of the multi-modal data based on the platform traffic type information, and establishes user portrait information by analyzing the multi-modal information; generating a basic delivery strategy based on the user portrait information; determining a propagation relation chain including a sharer user and a receiver user based on the user identity information; acquiring historical sharing information and social network structure information in the propagation relation chain, and constructing a social influence model; and optimizing the basic putting strategy based on the bidirectional association features output by the social influence model to generate an optimized putting strategy. According to the method, the problems of low precision and low conversion rate in a traditional data delivery method are solved through a mode of adjusting the analysis weight of the multi-modal data and a mode of optimizing the basic delivery strategy based on the bidirectional association features.

Description

Technical Field

[0001] The present invention relates to the technical field of data delivery, and specifically refers to a data delivery optimization method and system based on multimodal data. Background Art

[0002] Data delivery refers to the process of presenting advertising content to potential consumers through various channels and platforms to achieve business goals such as brand promotion, product marketing, and sales growth. In the current data delivery process, in order to improve the accuracy of delivery, a personal profile (i.e., a multi-dimensional information model) is usually established for users, and the most attractive data is selected for delivery based on the personal profile, thereby enhancing the delivery effect and conversion rate.

[0003] However, the existing data delivery methods still have the following deficiencies:

[0004] 1. Traditional data delivery methods rely on the collection of multimodal data such as user behavior data and social attribute data (such as text and video), analyze and process the multimodal data based on equal weights, and construct a multi-dimensional information model of users based on the analysis results. However, the core content forms of different platforms may vary significantly. For example, 70% of the interest characteristics of users on short video platforms may be contained in video frames, but the existing data analysis methods usually adopt the mean distribution method, resulting in unreasonable allocation of analysis capacity, thus affecting the accuracy of user profile information.

[0005] 2. When processing the social network data between users, the current data delivery methods generally adopt a one-way propagation model (i.e., from the sharer to the recipient). For example, only adjust the personal profile of the recipient based on the sharing information of the sharer, ignoring the two-way conduction effect between users in the social network.

[0006] Therefore, there is an urgent need for a data delivery optimization method and system that can adaptively adjust the processing weights of multimodal data according to the mainstream data types of the platform and can activate the value of the social network between users. Summary of the Invention

[0007] In order to optimize the uneven allocation of data analysis capacity in traditional data delivery and activate the two-way conduction effect between users, and improve the accuracy of data delivery, the present invention provides a data delivery optimization method and system based on multimodal data.

[0008] The first invention object of the present application is achieved through the following technical solutions:

[0009] A data delivery optimization method based on multimodal data, comprising the steps:

[0010] Obtain user identity information and obtain multimodal data associated with the user identity information as the input of the portrait building model for pre-training. The multimodal data includes historical user behavior information, corresponding content feature information, and platform traffic type information;

[0011] The portrait building model adjusts the analysis weights of the multimodal data based on the platform traffic type information, and establishes user portrait information by analyzing the historical user behavior information and the content feature information;

[0012] When receiving the user portrait information output by the portrait building model, generate a basic placement strategy based on the user portrait information;

[0013] Determine the propagation relationship chain based on the user identity information. The propagation relationship chain includes the sharing user and the receiving user;

[0014] Obtain the historical sharing information and social network structure information between the sharing user and the receiving user in the propagation relationship chain, and construct a social influence model;

[0015] Optimize the basic placement strategy based on the two-way association features output by the social influence model to generate an optimized placement strategy.

[0016] By adopting the above technical solution, obtain the user identity information and its associated multimodal data, and input them into the portrait building model. The portrait building model adjusts the analysis weights of the multimodal data based on the platform browsing type information in the multimodal data, and establishes user portrait information by analyzing the historical user behavior information and content feature information in the multimodal data, so as to generate a basic placement strategy based on the user portrait information; determine the propagation relationship chain including the sharing user and the receiving user based on the user identity information, obtain the historical sharing information and social network structure information between the sharing user and the receiving user in the propagation relationship chain, so as to establish a social influence model, and finally optimize the basic placement strategy based on the two-way association features output by the social influence model to generate an optimized placement strategy; in this application, by adjusting the analysis weights of the multimodal data based on the platform traffic type information, the analysis capacity allocation of the multimodal data is optimized, and by establishing a propagation relationship chain, a social influence model is constructed to optimize the basic placement strategy, activate the social network value between users, and thus improve the accuracy and conversion rate of data placement between users.

[0017] In a preferred example of this application, it can be further configured that: the portrait building model includes a data classification layer, a weight adjustment layer, and a feature fusion layer. The steps of the portrait building model adjusting the analysis weights of the multimodal data based on the platform traffic type information and establishing user portrait information by analyzing the historical user behavior information and the content feature information include the steps:

[0018] The data classification layer classifies the historical user behavior information and its corresponding content feature information into data modalities to obtain feature tensors of different modalities;

[0019] The weight adjustment layer analyzes the weight information of the feature tensors of different modalities based on the platform traffic type information and adjusts the weights;

[0020] The feature fusion layer extracts and fuses features from the historical user behavior information and the content feature information based on the weight adjustment result and outputs user portrait information.

[0021] By adopting the above technical solutions, classifying the historical user behavior information and the corresponding content feature information into data modalities, analyzing and adjusting the weight information of the feature tensors based on the platform traffic type information, and processing the historical user behavior information and the corresponding content feature information based on the weight adjustment result, a user portrait with higher accuracy can be obtained more efficiently.

[0022] In a preferred example of the present application, it can be further configured as: the step of obtaining the historical sharing information and the social network structure information between the sharer user and the recipient user in the propagation relationship chain and constructing a social influence model includes the steps of:

[0023] Obtain the historical sharing information between the sharer user and the recipient user, and generate low-dimensional vectors of user nodes with each user as a node, so as to generate social network structure information;

[0024] Construct a social influence model based on the historical sharing information and the social network structure information.

[0025] By adopting the above technical solutions, obtain the historical sharing information between the sharer user and the recipient user, generate corresponding low-dimensional vectors with users as nodes, and generate social network structure information based on this, so as to construct a social influence model based on the historical sharing information and the social network structure information, and realize the two-way conduction effect of users in the social network in the way of a two-way propagation model between users.

[0026] In a preferred example of the present application, it can be further configured as: before the step of optimizing the basic placement strategy based on the two-way association features output by the social influence model to generate an optimized placement strategy, the following steps are further included:

[0027] Obtain the historical consumption information of the user, use it as the input of the pre-trained consumption prediction model, and perform matching based on the generated predicted consumption information to obtain the corresponding predicted consumption level information.

[0028] By adopting the above technical solution, the historical consumption information of the user is obtained and input into the consumption prediction model, and the generated predicted consumption information is matched to obtain the corresponding predicted consumption level information, which is convenient for subsequent optimization of the delivered content based on the predicted consumption level information.

[0029] In a preferred example of the present application, it can be further configured that: the step of obtaining the historical consumption information of the user, as the input quantity of the pre-trained consumption prediction model, and matching based on the generated predicted consumption information to obtain the corresponding predicted consumption level information includes the steps:

[0030] Obtain the historical consumption information of the user, and extract the consumption ability characteristics and consumption preference characteristics;

[0031] Input the consumption ability characteristics and consumption preference characteristics into the pre-trained consumption prediction model to generate predicted consumption information and predicted consumption preference information;

[0032] Compare the predicted consumption information based on the preset level matching rule to obtain the corresponding predicted consumption level information.

[0033] By adopting the above technical solution, the historical consumption information of the user is obtained for feature extraction, the extracted consumption ability characteristics and consumption preference characteristics are input into the consumption prediction model to generate predicted consumption information and predicted consumption preference information, and the predicted consumption information is compared based on the level matching rule to obtain the predicted consumption level information, realizing the prediction of the user's consumption ability and consumption preference based on the historical consumption information, improving the accuracy of data delivery, and indirectly improving the conversion rate at the same time.

[0034] In a preferred example of the present application, it can be further configured that: the step of optimizing the basic delivery strategy based on the two-way association features output by the social influence model to generate an optimized delivery strategy includes the steps:

[0035] The social influence model analyzes the social graph based on the user identity information, determines the interaction frequency of the propagation relationship chain, and mines the second-degree propagation relationship chain to generate two-way association features;

[0036] Based on the predicted consumption level information corresponding to the user, preliminarily optimize the basic delivery strategy to generate a preliminary delivery strategy;

[0037] Based on the two-way association features, perform secondary optimization on the preliminary delivery strategy to generate an optimized delivery strategy.

[0038] By adopting the above technical solution, the social influence model performs social graph analysis on the propagation relationship chain, determines its interaction frequency, and mines the second-degree propagation relationship chain, thereby generating corresponding bidirectional association features. Based on the predicted consumption level information corresponding to the user, the basic placement strategy is optimized to obtain a preliminary placement strategy. Then, based on the bidirectional association features, the preliminary placement strategy is optimized to generate an optimized placement strategy, realizing the optimization of the placement content of the basic placement strategy based on the predicted consumption level information to obtain a preliminary placement strategy, and then optimizing the placement method of the preliminary placement strategy based on the bidirectional association features to obtain the final optimized placement strategy.

[0039] In a preferred example of the present application, it can be further configured that: the step of the social influence model performing social graph analysis based on user identity information, determining the interaction frequency of the propagation relationship chain, and mining the second-degree propagation relationship chain to generate bidirectional association features includes the steps of:

[0040] Obtain the historical interaction information between the sharer user and the recipient user, perform interaction frequency judgment, and respectively generate bidirectional association features corresponding to high-frequency interaction relationships and bidirectional association features corresponding to low-frequency interaction relationships;

[0041] Mine the second-degree propagation relationship chain formed based on indirect interaction, and generate bidirectional association features corresponding to the indirect interaction relationship.

[0042] By adopting the above technical solution, the historical interaction information between the users of the propagation relationship chain is analyzed to perform interaction frequency judgment, and corresponding bidirectional association features are generated therefrom, and the second-degree propagation relationship chain is mined, thereby generating bidirectional association features corresponding to the indirect interaction relationship, realizing the frequency judgment and analysis of the propagation relationship chain, and thus performing differential processing.

[0043] In a preferred example of the present application, it can be further configured that: the basic placement strategy includes placement content and preferential strategies. The step of initially optimizing the basic placement strategy based on the predicted consumption level information corresponding to the user to generate a preliminary placement strategy includes the steps of:

[0044] Optimize the placement content of the basic placement strategy based on the predicted consumption level information;

[0045] Optimize the preferential strategy of the basic placement strategy based on the predicted consumption preference information;

[0046] Integrate the optimized placement content and preferential strategies to generate a preliminary placement strategy.

[0047] By adopting the above technical solutions, the content of the basic delivery strategy is optimized based on the predicted consumption level information, and the preferential strategy of the basic delivery strategy is optimized based on the predicted consumption preference information. The preliminary delivery strategy is generated by integrating the optimized content of the delivery and the preferential strategy, so as to match the corresponding delivery products based on the predicted consumption level information as the content of the delivery, and generate the corresponding preferential strategy based on the predicted consumption preference information. Thus, the preliminary delivery strategy is generated by integrating the content of the delivery and the preferential strategy, ensuring that both the content of the delivery and the preferential strategy are within the acceptable range of users, and improving the actual conversion rate of data delivery.

[0048] In a preferred example of the present application, it can be further configured that: the step of further optimizing the preliminary delivery strategy based on the bidirectional association feature to generate an optimized delivery strategy includes the steps of:

[0049] Grouping based on the interaction frequency corresponding to the bidirectional association feature, and the groups include a priority push group and an interaction incentive group;

[0050] Classify the propagation relationship chains associated with the interaction frequencies with high-frequency interaction relationships into the priority push group, and perform priority push optimization based on the corresponding preliminary delivery strategy to generate the corresponding optimized delivery strategy;

[0051] Classify the propagation relationship chains associated with the interaction frequencies with low-frequency interaction relationships and indirect interaction relationships into the interaction incentive group, and perform incentive optimization based on the corresponding preliminary delivery strategy to generate the corresponding optimized delivery strategy.

[0052] By adopting the above technical solutions, grouping is performed based on the interaction frequency corresponding to the bidirectional association feature. The propagation relationship chains associated with the interaction frequencies with high-frequency interaction relationships are classified into the priority push group, and the corresponding preliminary delivery strategy is used for priority push optimization to generate the corresponding optimized delivery strategy; the propagation relationship chains associated with the interaction frequencies with low-frequency interaction relationships and indirect interaction relationships are classified into the interaction incentive group, and the corresponding preliminary delivery strategy is used for incentive optimization to generate the corresponding optimized delivery strategy; by grouping the propagation relationship chains with different interaction frequencies and adopting different optimization schemes for the preliminary delivery strategies corresponding to different groups, the propagation relationship chains with high-frequency interaction relationships can continue to maintain high-frequency interaction, and the remaining propagation relationship chains are interactively incentivized, thereby activating the two-way conduction effect between users.

[0053] The second above-mentioned invention object of the present application is achieved by the following technical solutions:

[0054] A data delivery optimization system based on multi-modal data, including:

[0055] A data information acquisition module, which is used to acquire user identity information and acquire multimodal data associated with the user identity information as the input of a pre-trained portrait building model. The multimodal data includes historical user behavior information, corresponding content feature information, and platform traffic type information;

[0056] A portrait building module, which is used to adjust the analysis weight of multimodal data based on the platform traffic type information, and establish user portrait information by analyzing the historical user behavior information and the content feature information;

[0057] A strategy generation module, which is used to generate a basic placement strategy based on the user portrait information when receiving the user portrait information output by the portrait building model;

[0058] A social influence module, which is used to determine a propagation relationship chain based on the user identity information. The propagation relationship chain includes a sharer user and a recipient user;

[0059] A social analysis module, which is used to acquire the historical sharing information and social network structure information between the sharer user and the recipient user in the propagation relationship chain, and construct a social influence model;

[0060] A strategy optimization module, which is used to optimize the basic placement strategy based on the two-way association features output by the social influence model to generate an optimized placement strategy.

[0061] By adopting the above technical solutions, a data information acquisition module, which is used to acquire user identity information and acquire multimodal data associated with the user identity information as the input of a pre-trained portrait building model. The multimodal data includes historical user behavior information, corresponding content feature information, and platform traffic type information; a portrait building module, which is used to adjust the analysis weight of multimodal data based on the platform traffic type information, and establish user portrait information by analyzing the historical user behavior information and the content feature information; a strategy generation module, which is used to generate a basic placement strategy based on the user portrait information when receiving the user portrait information output by the portrait building model; a social influence module, which is used to determine a propagation relationship chain based on the user identity information. The propagation relationship chain includes a sharer user and a recipient user; a social analysis module, which is used to acquire the historical sharing information and social network structure information between the sharer user and the recipient user in the propagation relationship chain, and construct a social influence model; a strategy optimization module, which is used to optimize the basic placement strategy based on the two-way association features output by the social influence model to generate an optimized placement strategy.

[0062] The beneficial effects of a data placement optimization method and system based on multimodal data of the present invention are as follows:

[0063] 1. By obtaining and analyzing platform traffic type information, the portrait building model adjusts the analysis weights of multimodal data based on the analysis results, and uses the portrait building model with adjusted analysis weights to analyze the multimodal data, thereby establishing user portrait information. This method realizes the adjustment of the analysis weights of multimodal data according to the mainstream types of the platform, and solves the problems of unreasonable transport capacity allocation caused by the average distribution method in traditional data placement methods, resulting in low accuracy of user portraits and low operation efficiency.

[0064] 2. By determining the propagation relationship chain based on user identity information and outputting two-way association features by the social influence model constructed based on the propagation relationship chain, and then optimizing the basic placement strategy based on the two-way association features, this method realizes the effect of optimizing the basic placement strategy based on the two-way conduction relationship between users, solves the problems such as low accuracy of user portraits caused by traditional data placement methods using one-way propagation models to process social network data between users, and has the advantage of improving the conversion rate of data placement. Description of the Drawings

[0065] Figure 1 It is a flowchart of an embodiment of a data placement optimization method based on multimodal data according to the present application;

[0066] Figure 2 It is a flowchart of an implementation of step S20 in an embodiment of a data placement optimization method based on multimodal data according to the present application;

[0067] Figure 3 It is a flowchart of an implementation of step S01 in an embodiment of a data placement optimization method based on multimodal data according to the present application;

[0068] Figure 4 It is a flowchart of an implementation of step S60 in an embodiment of a data placement optimization method based on multimodal data according to the present application;

[0069] Figure 5 It is a flowchart of an implementation of step S602 in an embodiment of a data placement optimization method based on multimodal data according to the present application;

[0070] Figure 6 It is a flowchart of an implementation of step S603 in an embodiment of a data placement optimization method based on multimodal data according to the present application. Detailed Embodiments

[0071] The following is a further detailed description of the present application in conjunction with the attached Figure 1-6 drawings.

[0072] In one embodiment, as Figure 1 shown, the present application discloses a data placement optimization method based on multimodal data, which specifically includes the following steps:

[0073] S10: Obtain user identity information and obtain multimodal data associated with the user identity information as the input for the portrait building model. The multimodal data includes historical user behavior information, corresponding content feature information, and platform traffic type information.

[0074] In this embodiment, the user identity information is the basic data that uniquely identifies the user identity within the platform. The multimodal data is a composite information that integrates various data forms such as text, image, audio, and video. The portrait building model is a model that generates dimensions such as interest preferences by integrating user identity information and multimodal behavior data. The historical user behavior information is the interaction trajectory of the user on the platform, including clicks, browsing, dwell time, purchase records, etc. The content feature information is the information that characterizes the relevance between the content corresponding to the historical user behavior information and the user's interests. The platform traffic type information is the traffic proportion information of the platform corresponding to each modal information such as text, pictures, videos, etc.

[0075] Specifically, obtain the user identity information, capture the corresponding historical user behavior information, the content feature information corresponding to the historical user behavior information, and the traffic proportion information of the platform corresponding to each modal information type, and input them into the portrait building model.

[0076] S20: The portrait building model adjusts the analysis weight of the multimodal data based on the platform traffic type information, and establishes user portrait information by analyzing the historical user behavior information and the content feature information.

[0077] In this embodiment, the analysis weight is the different importance coefficients assigned to data such as text, images, and behaviors. The user portrait information is information that includes dimensions such as the user's basic attributes, interest tags, and behavior characteristics.

[0078] Specifically, adjust the analysis weight of the portrait building model for the multimodal data based on the platform traffic type information. For example, for a platform where short videos are the mainstream, increase the analysis weight of the short video data in the corresponding multimodal data, and analyze the historical user behavior information and content feature information through the portrait building model according to the adjusted analysis weight, thereby establishing user portrait information.

[0079] S30: When receiving the user portrait information output by the portrait building model, generate a basic placement strategy based on the user portrait information.

[0080] In this embodiment, the basic placement strategy is the initial execution plan for data placement.

[0081] Specifically, extract effective tags from the user portrait information, select placement content that meets the requirements of the matching degree with the effective tags, and generate the corresponding basic placement strategy.

[0082] S40: Determine the dissemination relationship chain based on the user identity information, where the dissemination relationship chain includes the sharing user and the receiving user;

[0083] In this embodiment, the dissemination relationship chain is the path structure of information flowing in the user network, which is a chain or network relationship formed by the sharing user (information sending node) and the receiving user (information receiving node) through social interaction. The sharing user is the starting node of information dissemination, and transmits the information to other users through active behaviors such as forwarding. The receiving user is the end or intermediate node of information dissemination, and participates in the dissemination through passive reception such as browsing, liking or active interaction such as commenting, secondary forwarding, etc.

[0084] Specifically, perform node identification based on the user identity information, including the sharing user and the receiving user, and perform topological expansion on the relationship chain corresponding to the nodes, so as to establish the corresponding dissemination relationship chain.

[0085] S50: Obtain the historical sharing information and social network structure information between the sharing user and the receiving user in the dissemination relationship chain, and construct a social influence model;

[0086] In this embodiment, the historical sharing information is the record of the user's past dissemination behavior, including data such as sharing time, content type, etc. The social network structure information is the topological feature describing the user's social relationship, including indicators such as node degree, community clustering, etc. The social influence model is a model for quantifying the user's dissemination ability, and predicts the diffusion effect of information in the social network through the node influence weight.

[0087] Specifically, obtain data such as the sharing time and sharing content type between the sharing user and the receiving user, and obtain the social network structure information describing the user's social relationship, and construct a social influence model based on this.

[0088] S60: Optimize the basic placement strategy based on the two-way association features output by the social influence model to generate an optimized placement strategy.

[0089] In this embodiment, the two-way association feature is a dynamic relationship index of mutual influence between users, which is divided into a positive association feature, that is, the dissemination effect of the sharing user on the receiving user, such as the click-through rate of the receiving user after forwarding, and a negative association feature, that is, the feedback influence of the receiving user on the sharing user, such as the like behavior improving the subsequent activity of the sharing user. The optimized placement strategy is an upgraded solution optimized by superimposing data such as the social influence analysis results on the basic strategy.

[0090] Specifically, obtain the two-way association features output by the social influence model, and perform optimization steps on the basic placement strategy for priority placement or incentive optimization, so as to generate the final optimized placement strategy.

[0091] In one embodiment, the portrait building model includes a data classification layer, a weight adjustment layer, and a feature fusion layer. As Figure 2 shown, step S20 includes the steps of:

[0092] S201: The data classification layer performs data modality classification on historical user behavior information and its corresponding content feature information to obtain feature tensors of different modalities;

[0093] S202: The weight adjustment layer analyzes the weight information of feature tensors of different modalities based on the platform traffic type information and performs weight adjustment;

[0094] S203: The feature fusion layer performs feature extraction and feature fusion on the historical user behavior information and the content feature information based on the weight adjustment result and outputs user portrait information.

[0095] In this embodiment, the data classification layer is a technical module that divides multi-modal data into different modalities according to its manifestation forms, such as text, image, user behavior sequence, etc. The weight adjustment layer is a technical module that dynamically adjusts the importance weights of different modality features, and the feature fusion layer is the core module that integrates multi-modal features and generates a unified user portrait.

[0096] Specifically, the data classification layer of the portrait building model performs data modality classification on multi-modal data to obtain feature tensors of different modalities. The weight adjustment layer analyzes and adjusts the weights of feature tensors of different modalities based on the platform traffic type information. The feature fusion layer performs feature extraction and feature fusion on the historical user behavior information and the content feature information based on the weight adjustment result, thereby outputting user portrait information.

[0097] In one embodiment, step S50 includes the steps of:

[0098] S501: Obtain the historical sharing information between the sharer user and the recipient user, and generate a low-dimensional vector of the user node with each user as a node, thereby generating social network structure information;

[0099] S502: Construct a social influence model based on the historical sharing information and the social network structure information.

[0100] In this embodiment, the low-dimensional vector is a fixed-dimensional dense vector obtained by compressing high-dimensional user behavior data such as historical sharing records and social relationships through an embedding technique.

[0101] Specifically, obtain the historical sharing information between the sharer user and the recipient user, extract the user interaction behaviors from the historical sharing information, and compress the high-dimensional behavior data of each user into low-dimensional vectors with each user as a node, so as to generate the social network structure information, and fuse the low-dimensional vectors of the users and the structural features, thereby constructing a social influence model.

[0102] In one embodiment, before step S60, the following steps are further included:

[0103] S01: Obtain the historical consumption information of the user as the input of the pre-trained consumption prediction model, and perform matching based on the generated predicted consumption information to obtain the corresponding predicted consumption level information.

[0104] In this embodiment, the historical consumption information is the user's past consumption behavior data, including purchase time, commodity category, consumption amount, etc. The consumption prediction model is a model for predicting future consumption trends or amounts, and the predicted consumption information is the quantitative result output by the model, including the future consumption amount range, consumption frequency, etc. The predicted consumption level information is the consumption level divided according to the prediction result, including the mass consumption group, the middle class, the high net worth, etc.

[0105] Specifically, extract the historical consumption information corresponding to the user from the transaction records, input the preprocessed data into the pre-trained consumption prediction model, generate the predicted consumption information for the future period, and perform level matching based on this to obtain the corresponding predicted consumption level information.

[0106] In one embodiment, as Figure 3 shown, step S01 includes the steps:

[0107] S011: Obtain the historical consumption information of the user, and extract the consumption ability characteristics and consumption preference characteristics;

[0108] S012: Input the consumption ability characteristics and consumption preference characteristics into the pre-trained consumption prediction model to make it generate the predicted consumption information and predicted consumption preference information;

[0109] S013: Compare the predicted consumption information based on the preset level matching rules to obtain the corresponding predicted consumption level information.

[0110] In this embodiment, the consumption ability characteristics are indicators for quantifying the user's economic strength, such as the monthly average consumption amount, the upper limit of a single consumption, etc. The consumption preference characteristics are the user's selection tendencies for promotion plans such as consumption discounts and promotions. The predicted consumption preference information is the promotion plan information that the user may be willing to accept in the future, such as promotion plans like full reduction for the total amount of goods, second piece of goods at half price, etc.

[0111] Specifically, obtain the user's historical consumption information, extract the consumption ability characteristics and consumption preference characteristics therefrom, and input them into the consumption prediction model, so as to generate predicted consumption information, such as the remaining disposable consumption amount within the user's consumption limit or the remaining consumption amount of the average monthly consumption amount within a specific period of time, etc., and predicted consumption preference information, such as the consumption preference that the user prefers to make up the full reduction when purchasing goods or prefers to hoard large quantities of goods with offers such as buy one get one free, etc., and then compare the predicted consumption information based on the hierarchical matching rule to obtain the corresponding predicted consumption level information.

[0112] In one embodiment, as Figure 4 shown, step S60 includes the steps:

[0113] S601: The social influence model performs social graph analysis based on the user identity information, determines the interaction frequency of the propagation relationship chain, and mines the second-degree propagation relationship chain to generate bidirectional association features;

[0114] S602: Based on the predicted consumption level information corresponding to the user, preliminarily optimize the basic placement strategy to generate a preliminary placement strategy;

[0115] S603: Based on the bidirectional association features, perform secondary optimization on the preliminary placement strategy to generate an optimized placement strategy.

[0116] In this embodiment, the social graph is a networked data structure describing the social relationships between users, the interaction frequency is the number of social interactions between users within a specific period of time, and the second-degree propagation relationship chain is the path through which information spreads through two or more jumps.

[0117] Specifically, the social influence model takes users as nodes and interaction behaviors as edges to construct a directed graph, performs social graph analysis, determines the interaction frequency of the propagation relationship chain, and mines the second-degree propagation relationship chain, thereby generating bidirectional association features, optimizes the basic placement strategy based on the predicted consumption level information to obtain a preliminary placement strategy, and then optimizes the preliminary placement strategy based on the bidirectional association features to obtain an optimized placement strategy.

[0118] In one embodiment, step S601 includes the steps:

[0119] S6011: Obtain the historical interaction information between the sharer user and the recipient user, perform interaction frequency judgment, and generate bidirectional association features corresponding to the high-frequency interaction relationship and bidirectional association features corresponding to the low-frequency interaction relationship respectively;

[0120] S6012: Mine the second-degree propagation relationship chain formed based on indirect interaction to generate bidirectional association features corresponding to the indirect interaction relationship.

[0121] In this embodiment, the interaction frequency is judged by counting the number of interactions between users within a specific time period, and is divided into high-frequency or low-frequency interaction relationships based on a threshold.

[0122] Specifically, historical interaction information of the propagation relationship chain is obtained for interaction frequency judgment, and two-way association features corresponding to high-frequency interaction relationships and two-way association features corresponding to low-frequency interaction relationships are respectively generated. The second-degree propagation relationship chain formed based on indirect interaction is mined, and two-way association features corresponding to the indirect interaction relationship are generated.

[0123] In one embodiment, as Figure 5 shown, step S602 includes the steps:

[0124] S6021: Optimize the delivery content of the basic delivery strategy based on the predicted consumption level information;

[0125] S6022: Optimize the preferential strategy of the basic delivery strategy based on the predicted consumption preference information;

[0126] S6023: Generate a preliminary delivery strategy by integrating the optimized delivery content and preferential strategy.

[0127] In this embodiment, the delivery content is an associated product whose price matches the predicted consumption level information, and the preferential strategy is a promotional method adjusted according to the user's consumption level and preference.

[0128] Specifically, based on the predicted consumption level information, the corresponding associated products are matched to optimize the delivery content of the basic delivery strategy. The preferential strategy of the basic delivery strategy is optimized based on the predicted consumption preference information, and a preliminary delivery strategy is generated by integrating the optimized delivery content and preferential strategy.

[0129] In one embodiment, as Figure 6 shown, step S603 includes the steps:

[0130] S6031: Group based on the interaction frequency corresponding to the two-way association feature. The groups include a priority push group and an interaction incentive group;

[0131] S6032: Classify the propagation relationship chains associated with the interaction frequencies with high-frequency interaction relationships into the priority push group, and perform priority push optimization based on the corresponding preliminary delivery strategy to generate the corresponding optimized delivery strategy;

[0132] S6033: Classify the propagation relationship chains associated with the interaction frequencies with low-frequency interaction relationships and indirect interaction relationships into the interaction incentive group, and perform incentive optimization based on the corresponding preliminary delivery strategy to generate the corresponding optimized delivery strategy.

[0133] In this embodiment, the priority push group is the user group corresponding to the high-frequency interaction relationship. Because of its strong social influence, it is preferentially reached to maximize the dissemination efficiency and improve the conversion rate. The interaction incentive group is the user group with low-frequency or indirect interactions. Their participation and dissemination willingness are enhanced through targeted incentives.

[0134] Specifically, grouping is performed according to the interaction frequency corresponding to the two-way association feature. The dissemination relationship chain associated with the interaction frequency with a high-frequency interaction relationship is classified into the priority push group, and priority push optimization is performed based on the corresponding preliminary placement strategy to generate the corresponding optimized placement strategy. The dissemination relationship chain associated with the interaction frequency with a low-frequency interaction relationship and an indirect interaction relationship is classified into the interaction incentive group, and incentive optimization is performed based on the corresponding preliminary placement strategy to generate the corresponding optimized placement strategy.

[0135] The data placement optimization system based on multimodal data includes:

[0136] The data information acquisition module is used to acquire user identity information and acquire the multimodal data associated with the user identity information as the input of the pre-trained portrait establishment model. The multimodal data includes historical user behavior information, corresponding content feature information, and platform traffic type information;

[0137] The portrait establishment module is used to adjust the analysis weight of the multimodal data based on the platform traffic type information, and establish user portrait information by analyzing the historical user behavior information and the content feature information;

[0138] The strategy generation module is used to generate a basic placement strategy based on the user portrait information when receiving the user portrait information output by the portrait establishment model;

[0139] The social influence module is used to determine the dissemination relationship chain based on the user identity information. The dissemination relationship chain includes the sharing user and the receiving user;

[0140] The social analysis module is used to obtain the historical sharing information and social network structure information between the sharing user and the receiving user in the dissemination relationship chain, and construct a social influence model;

[0141] The strategy optimization module is used to optimize the basic placement strategy based on the two-way association feature output by the social influence model to generate an optimized placement strategy.

[0142] For the specific limitations of a data placement optimization system based on multimodal data, reference can be made to the limitations of a data placement optimization method based on multimodal data in the above text, which will not be elaborated here. Each module in the above data placement optimization system based on multimodal data can be implemented in whole or in part by software, hardware, or a combination thereof. The above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0143] The present invention and its embodiments have been described above. Such description is not restrictive. What is shown in the drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. In general, if those of ordinary skill in the art are inspired by it and, without departing from the spirit of the present invention, design similar structural forms and embodiments to this technical solution without creative efforts, they should all fall within the protection scope of the present invention.

Claims

1. A data placement optimization method based on multi-modal data, characterized in that: Including the steps: Obtain user identity information, and obtain multimodal data associated with the user identity information as the input of the portrait building model for pre-training. The multimodal data includes historical user behavior information, corresponding content feature information, and platform traffic type information; The portrait building model adjusts the analysis weights of the multimodal data based on the platform traffic type information, and establishes user portrait information by analyzing the historical user behavior information and the content feature information; When receiving the user portrait information output by the portrait building model, generate a basic placement strategy based on the user portrait information; Determine the propagation relationship chain based on the user identity information. The propagation relationship chain includes the sharing user and the receiving user; Obtain the historical sharing information and social network structure information between the sharing user and the receiving user in the propagation relationship chain, and construct a social influence model; Optimize the basic placement strategy based on the two-way association features output by the social influence model to generate an optimized placement strategy.

2. The data placement optimization method based on multimodal data according to claim 1, wherein: The portrait building model includes a data classification layer, a weight adjustment layer, and a feature fusion layer. The step of the portrait building model adjusting the analysis weights of the multimodal data based on the platform traffic type information and establishing user portrait information by analyzing the historical user behavior information and the content feature information includes the steps: The data classification layer classifies the historical user behavior information and its corresponding content feature information into data modalities to obtain feature tensors of different modalities; The weight adjustment layer analyzes the weight information of the feature tensors of different modalities based on the platform traffic type information and performs weight adjustment; The feature fusion layer extracts and fuses the features of the historical user behavior information and the content feature information based on the weight adjustment result, and outputs user portrait information.

3. The data placement optimization method based on multimodal data according to claim 1, wherein: The step of obtaining the historical sharing information and social network structure information between the sharing user and the receiving user in the propagation relationship chain and constructing a social influence model includes the steps: Obtain the historical sharing information between the sharing user and the receiving user, and generate low-dimensional vectors of user nodes with each user as a node, so as to generate social network structure information; Construct a social influence model based on the historical sharing information and social network structure information.

4. A data placement optimization method based on multimodal data according to claim 1, characterized in that: Before the step of optimizing the basic placement strategy based on the two-way association features output by the social influence model to generate an optimized placement strategy, the following steps are also included: Obtain the historical consumption information of the user as the input of the pre-trained consumption prediction model, and perform matching based on the generated predicted consumption information to obtain the corresponding predicted consumption level information.

5. The data placement optimization method based on multimodal data according to claim 4, wherein: The step of obtaining the historical consumption information of the user as the input of the pre-trained consumption prediction model and performing matching based on the generated predicted consumption information to obtain the corresponding predicted consumption level information includes the steps: Obtain the historical consumption information of the user, and extract the consumption ability feature and the consumption preference feature; Input the consumption ability feature and the consumption preference feature into the pre-trained consumption prediction model to generate predicted consumption information and predicted consumption preference information; Compare the predicted consumption information based on the preset hierarchical matching rules to obtain the corresponding predicted consumption level information.

6. The data placement optimization method based on multimodal data according to claim 5, wherein: The steps of optimizing the basic placement strategy based on the bidirectional association features output by the social influence model to generate an optimized placement strategy include the steps of: The social influence model analyzes the social graph based on the user identity information, determines the interaction frequency of the propagation relationship chain, and mines the second-degree propagation relationship chain to generate bidirectional association features; Preliminarily optimize the basic placement strategy based on the predicted consumption level information corresponding to the user to generate a preliminary placement strategy; Perform secondary optimization on the preliminary placement strategy based on the bidirectional association features to generate an optimized placement strategy.

7. The data placement optimization method based on multimodal data according to claim 6, wherein: The steps of the social influence model analyzing the social graph based on the user identity information, determining the interaction frequency of the propagation relationship chain, and mining the second-degree propagation relationship chain to generate bidirectional association features include the steps of: Obtain the historical interaction information between the sharer user and the recipient user, judge the interaction frequency, and generate the bidirectional association features corresponding to the high-frequency interaction relationship and the bidirectional association features corresponding to the low-frequency interaction relationship respectively; Mine the second-degree propagation relationship chain formed by indirect interaction to generate the bidirectional association features corresponding to the indirect interaction relationship.

8. The data placement optimization method based on multimodal data according to claim 6, characterized in that: The basic placement strategy includes the placement content and the preferential strategy. The steps of preliminarily optimizing the basic placement strategy based on the predicted consumption level information corresponding to the user to generate a preliminary placement strategy include the steps of: Optimize the placement content of the basic placement strategy based on the predicted consumption level information; Optimize the preferential strategy of the basic placement strategy based on the predicted consumption preference information; Generate a preliminary placement strategy by comprehensively optimizing the placement content and the preferential strategy.

9. The data placement optimization method based on multimodal data according to claim 6, wherein: The steps of performing secondary optimization on the preliminary placement strategy based on the bidirectional association features to generate an optimized placement strategy include the steps of: Group based on the interaction frequency corresponding to the bidirectional association features. The groups include the priority push group and the interaction incentive group; Classify the propagation relationship chains associated with the interaction frequencies with high-frequency interaction relationships into the priority push group, and perform priority push optimization based on the corresponding preliminary placement strategy to generate the corresponding optimized placement strategy; Classify the propagation relationship chains associated with the interaction frequencies with low-frequency interaction relationships and indirect interaction relationships into the interaction incentive group, and perform incentive optimization based on the corresponding preliminary placement strategy to generate the corresponding optimized placement strategy.

10. A data delivery optimization system based on multimodal data, characterized in that: Include: A data information acquisition module for acquiring user identity information and acquiring multi-modal data associated with the user identity information as the input of a pre-trained portrait building model. The multi-modal data includes historical user behavior information, corresponding content feature information, and platform traffic type information; A portrait building module for adjusting the analysis weight of the multi-modal data based on the platform traffic type information, and establishing user portrait information by analyzing the historical user behavior information and the content feature information; A strategy generation module for generating a basic placement strategy based on the user portrait information when receiving the user portrait information output by the portrait building model; A social influence module for determining a propagation relationship chain based on the user identity information. The propagation relationship chain includes a sharer user and a recipient user; A social analysis module, which is used to obtain historical sharing information and social network structure information between sharer users and recipient users in a dissemination relationship chain, and construct a social influence model; A strategy optimization module, which is used to optimize a basic placement strategy based on two-way association features output by the social influence model to generate an optimized placement strategy.

Images