A method, device and computing equipment for intelligent push based on big data

Abstract

The present application provides a method, device and computing equipment for intelligent push based on big data. Among them, a multi-dimensional data portrait is constructed based on the collected user data; interest tags and group tags of the target user are extracted; a personalized recommendation model is trained based on the user interest model and the hybrid recommendation algorithm; a context-aware push strategy is generated based on the current geographic location, device status, time node and social activity status of the target user obtained, combined with the prediction results of the personalized recommendation model; based on the context-aware push strategy, the corresponding content is pushed to the target user. The technical solution provided by the present application can improve the accuracy and efficiency of the pushed content.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of big data analysis, and in particular to a big data-based intelligent push method, device and computing equipment. Background Art

[0002] In the Internet era, with the explosive growth of information and the increasing demand for personalization, how to effectively and accurately push content that users are interested in to users has become an important issue. Intelligent content push technology aims to achieve efficient distribution of personalized content by analyzing user behavior habits, interest preferences and other information. This technology is widely used in news information, e-commerce, social media and other fields, greatly improving user experience and platform activity and stickiness.

[0003] The current intelligent push system mainly relies on several recommendation algorithms, including but not limited to content-based recommendations, collaborative filtering recommendations, and knowledge graph-based recommendations. These algorithms can infer users' interests based on data such as their browsing history, purchase history, and search behavior, and recommend relevant content accordingly. For example, collaborative filtering algorithms analyze the similarities between users or items to make recommendations, while content-based recommendations focus more on the specific content features that users liked in the past.

[0004] However, existing intelligent push solutions still have some limitations and shortcomings. First, most recommendation algorithms do not work well when dealing with the user cold start problem, that is, it is difficult to make accurate recommendations for newly registered users or users with less behavioral data. Secondly, the real-time and accuracy of recommendation systems are often challenged, especially when faced with massive and constantly changing data. How to quickly update the model and maintain the relevance of the recommendations is a major problem. Finally, traditional recommendation algorithms often ignore the specific scenarios of users (such as geographic location, time point, etc.), resulting in a mismatch between the recommended content and the user's current needs, affecting user experience and satisfaction. Therefore, it is particularly important to develop an intelligent push method that can comprehensively consider multi-dimensional user information and adapt to dynamic environmental changes. Summary of the invention

[0005] The embodiments of the present application provide a big data-based intelligent push method, device and computing equipment to solve the problems of low accuracy and poor push efficiency of pushed content in the prior art.

[0006] In a first aspect, an embodiment of the present application provides an intelligent push method based on big data, including:

[0007] Collect user data of target users from multiple heterogeneous data sources;

[0008] Extracting features from the user data to construct a multi-dimensional data portrait of the target user;

[0009] By analyzing the multidimensional data portrait of the target user, interest tags and group tags of the target user are extracted and stored in a pre-established user interest model, wherein the interest tags are obtained by identifying multidimensional feature vectors in the multidimensional data portrait;

[0010] Based on the user interest model, a hybrid recommendation algorithm combining deep neural network, collaborative filtering and graph neural network is used to train a personalized recommendation model, wherein the personalized recommendation model is used to generate prediction results according to the interest tags and group tags in the user interest model, and the prediction results include the content of interest to the target user;

[0011] Generate a context-aware push strategy based on the acquired current geographic location, device status, time node, and social activity status of the target user, combined with the prediction result of the personalized recommendation model, wherein the context-aware push strategy includes at least: push timing and push content;

[0012] Based on the context-aware push strategy, corresponding content is pushed to the target user.

[0013] Optionally, the user data includes different types of user data, and the different types of user data include at least: user historical behavior data, device usage data, geographic location information, social network interaction data, and user environment data;

[0014] The extracting features of the user data to construct a multi-dimensional data portrait of the target user includes:

[0015] For different types of user data, corresponding feature extraction methods are used to extract key features that reflect the user behavior and interests of the target user to construct a feature set;

[0016] The feature set is integrated into a multi-dimensional feature vector to construct a multi-dimensional data portrait of the user, and the multi-dimensional feature vector is used to indicate the feature vector of the target user in multiple dimensions.

[0017] Optionally, for different types of user data, corresponding feature extraction methods are used to extract key features for reflecting user behaviors and interests to construct a feature set, including:

[0018] Extracting user behavior characteristics from the user historical behavior data through time series analysis, wherein the user behavior characteristics at least include access frequency, access duration, preferred content, and behavior trend;

[0019] Extracting device preference features from the device usage data, the device preference features at least including device type, device usage frequency, and commonly used applications;

[0020] Analyzing the geographic location information using a clustering algorithm to extract the user's geographic location features, the geographic location features at least including the user's permanent location, travel patterns, and areas of interest;

[0021] Analyzing the social network interaction data by natural language processing technology to extract the user's social tendency characteristics, wherein the social tendency characteristics at least include emotional tendency, topic preference and social activity;

[0022] In combination with the user environment data, analyzing the differences in user behavior under different environmental conditions to extract environmental behavior features related to the environment;

[0023] A feature set is constructed based on the user behavior feature, the device preference feature, the geographic location feature, the social tendency feature, and the environmental behavior feature.

[0024] Optionally, the extracting the interest tags and group tags of the target user by analyzing the multi-dimensional data portrait of the target user includes:

[0025] Performing feature selection on the multidimensional feature vector in the multidimensional data portrait of the target user, screening out representative user behavior features, device preference features, geographic location features, social tendency features, and environmental behavior features as user features;

[0026] Constructing a feature matrix based on the feature vectors corresponding to the user features, and optimizing the initial cluster center points according to the feature matrix to determine the target cluster center points;

[0027] The target cluster center point is used as an input of a hierarchical clustering algorithm, so as to perform multi-level interest grouping on the target user in combination with the hierarchical clustering algorithm to obtain a multi-level interest grouping result;

[0028] A user interest model is generated based on the multi-level interest clustering results. The user interest model records the interest tags and group tags of the target user. The interest tags are generated based on the characteristics and preferences shown by the target user in the clustering process. The group tags refer to the group identifiers to which the target user belongs in different clustering levels.

[0029] Optionally, constructing a feature matrix based on the feature vector corresponding to the user feature, and optimizing the initial cluster center point according to the feature matrix to determine the target cluster center point includes:

[0030] Constructing a feature matrix based on the feature vectors corresponding to the user features, wherein the rows of the feature matrix contain the feature values ​​of the target user on the user features, and the columns of the feature matrix contain the specific performance of the target user on representative user behavior features, device preference features, geographic location features, social tendency features, and environmental behavior features;

[0031] In the feature matrix, randomly select a feature vector corresponding to any user feature of the target user as the first initial cluster center point, calculate the distance from the feature vectors corresponding to the remaining user features in the feature matrix to the first initial cluster center point, and construct a probability distribution according to the square value of the distance, and then randomly select the next initial cluster center point until a preset number of initial cluster center points are selected;

[0032] Each eigenvector in the feature matrix is ​​assigned to the nearest initial cluster center point, and the sum of squared distances between each eigenvector and the corresponding initial cluster center point is calculated, and the sum of squared distances is minimized through iterative optimization until a preset number of initial cluster center points converge to form a target cluster center point.

[0033] Optionally, the target cluster center point is used as an input of a hierarchical clustering algorithm to perform multi-level interest grouping on the target user in combination with the hierarchical clustering algorithm to obtain a multi-level interest grouping result, including:

[0034] Using the target cluster center point, the feature vectors corresponding to the user features are assigned to a plurality of initial groups to generate preliminary interest grouping results;

[0035] Based on the preliminary interest grouping results, construct a similarity matrix of the user characteristics, and use a hierarchical clustering algorithm to gradually merge the groups with the highest similarity to form a hierarchical clustering tree;

[0036] In the hierarchical clustering tree, by setting different similarity thresholds, a multi-level interest group structure is divided to identify the clustering relationship of the target user at different interest levels;

[0037] By comparing the interest group structures at different levels, the rationality of the interest group structures at different levels is evaluated to select the best interest group structure as the multi-level interest grouping result.

[0038] Optionally, the training of a personalized recommendation model based on the user interest model using a hybrid recommendation algorithm combining a deep neural network, a collaborative filtering algorithm and a graph neural network includes:

[0039] Extracting interaction data between the user and the content from the different types of user data;

[0040] The interest tags and group tags in the user interest model are converted into interest feature vectors and group feature vectors, and the interaction data between the user and the content is converted into user content feature vectors, and the result is input into a deep neural network. The interest feature vectors, the group feature vectors and the user content feature vectors are represented in high dimensions through an embedding layer to obtain a nonlinear relationship, wherein the nonlinear relationship refers to the relationship between the user interest feature, the group feature vector and the user content feature vector;

[0041] Using the interaction data between the user and the content to construct an interaction matrix between the user and the content, calculating the similarity between the content of the target user through a collaborative filtering algorithm, and generating a matching score between the user and the content, wherein the similarity between the content refers to the degree of similarity between two or more content items, and the similarity is calculated based on the interaction data between the user and the content;

[0042] Building a relationship graph between the user and the content based on the interaction data between the user and the content, and applying a graph neural network to perform feature propagation and aggregation to obtain an interest propagation path between the user and the content;

[0043] The personalized recommendation model is trained by combining the nonlinear relationship, the matching score between the user and the content, and the interest propagation path between the user and the content.

[0044] Optionally, generating a context-aware push strategy based on the acquired current geographic location, device status, time node, and social activity status of the target user in combination with the prediction result of the personalized recommendation model includes:

[0045] The geographical location, device status, current time node and social activity status of the target user acquired in real time are used as elements to integrate into a context information set;

[0046] Determining the weight and influence of each element in the context information set;

[0047] Generate a preliminary push content list according to the prediction result of the personalized recommendation model and the weight and influence of each element in the context information set;

[0048] Optimizing the preliminary push content list by preset optimization goals to generate a context-aware push strategy;

[0049] The preset optimization goals include at least one or any combination of the following: optimization of push timing, improvement of content relevance, optimization of push coverage, optimization of conversion rate, and maximization of user engagement.

[0050] In a second aspect, the embodiment of the present application provides an intelligent push device based on big data, including:

[0051] A collection module, used to collect user data of target users from multiple heterogeneous data sources;

[0052] A construction module, used for extracting features from the user data to construct a multi-dimensional data portrait of the target user;

[0053] An extraction module, configured to extract interest tags and group tags of the target user by analyzing the multidimensional data portrait of the target user, and store them in a pre-established user interest model, wherein the interest tags are obtained by identifying multidimensional feature vectors in the multidimensional data portrait;

[0054] A training module, for training a personalized recommendation model based on the user interest model using a hybrid recommendation algorithm combining deep neural network, collaborative filtering and graph neural network, wherein the personalized recommendation model is used to generate prediction results according to interest tags and group tags in the user interest model, and the prediction results include the content of interest to the target user;

[0055] A generation module, configured to generate a context-aware push strategy based on the acquired current geographic location, device status, time node, and social activity status of the target user and the prediction result of the personalized recommendation model, wherein the context-aware push strategy includes at least: push timing and push content;

[0056] A push module is used to push corresponding content to the target user based on the context-aware push strategy.

[0057] In a third aspect, an embodiment of the present application provides a computing device, comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement the big data-based intelligent push method as described in the first aspect above.

[0058] In an embodiment of the present application, user data of a target user is collected from multiple heterogeneous data sources; feature extraction is performed on the user data to construct a multidimensional data portrait of the target user; interest tags and group tags of the target user are extracted by analyzing the multidimensional data portrait of the target user, and stored in a pre-established user interest model, wherein the interest tags are obtained by identifying multidimensional feature vectors in the multidimensional data portrait; based on the user interest model, a personalized recommendation model is trained using a hybrid recommendation algorithm combining deep neural networks, collaborative filtering, and graph neural networks, the personalized recommendation model being used to generate prediction results based on the interest tags and group tags in the user interest model, the prediction results including the content of interest to the target user; a context-aware push strategy is generated based on the current geographic location, device status, time node, and social activity status of the target user obtained, combined with the prediction results of the personalized recommendation model, the context-aware push strategy comprising at least: push timing and push content; based on the context-aware push strategy, the corresponding content is pushed to the target user.

[0059] These and other aspects of the present application will become more clearly understood in the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0061] Figure 1 A flowchart of an intelligent push method based on big data provided in an embodiment of the present application;

[0062] Figure 2 A schematic diagram of the structure of an intelligent push device based on big data provided in an embodiment of the present application;

[0063] Figure 3 A schematic diagram of the structure of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0064] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0065] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to different types.

[0066] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0067] Figure 1 A flowchart of an intelligent push method based on big data is provided for an embodiment of the present application, such as Figure 1 As shown, the method includes:

[0068] 101. Collect user data of target users from multiple heterogeneous data sources;

[0069] In this step, the target user may refer to a single user, and targeted push content is generated by obtaining the user data of the single user, thereby ensuring the accuracy of the push content. The user data includes different types of user data, and the different types of user data include at least: user historical behavior data, device usage data, geographic location information, social network interaction data, and user environment data;

[0070] Among them, user historical behavior data refers to the user's behavior records when using an application, website or other services in the past, such as browsing history, search history, purchase history, etc.

[0071] Device usage data: including the type of device used by the user (such as mobile phone, tablet computer, PC, etc.), operating system version, screen size, hardware configuration and other information.

[0072] Geographic location information: refers to the user's location information, which can be obtained through GPS, Wi-Fi positioning, base station positioning and other methods.

[0073] Social network interaction data: refers to the user's interactive behavior on social media, such as likes, comments, and shares.

[0074] User environmental data: Generally refers to information about the user's environment, such as indoor or outdoor, weather conditions, surrounding sound environment, etc.

[0075] In the embodiments of the present application, the process of collecting user data of the target user from multiple heterogeneous data sources may include:

[0076] User historical behavior data: Information such as the user's browsing history, search keywords, clicked advertisements, purchased goods, etc. within a certain period in the past can be collected. For example, an e-commerce platform can record which product pages the user has viewed, which products have been purchased, and the evaluations and feedback on the products in the past month.

[0077] Device usage data: The system can collect information such as the device model used by the user, the operating system version, and the screen resolution. For example, for a mobile application, the system can record that the user is using an iPhone 14 Pro Max, running the iOS 17 system, and the screen resolution is 1284x2778 pixels.

[0078] Geographical location information: The system can obtain the user's current location information through the user's mobile device. For example, if the user is using a map navigation software, the system can obtain the user's GPS coordinates in real time (such as latitude 39.9042°, longitude 116.4074°) to provide recommendation services for nearby locations.

[0079] Social network interaction data: The system can collect interaction information from the user's social media accounts. For example, if the user often shares food photos on Weibo, then the system can record the time, location of these shares, and related topic tags (such as #Food#, #Beijing Food#).

[0080] User environmental data: The system can collect information about the user's environment, such as indoor / outdoor, weather conditions, etc. For example, if the user's mobile phone is equipped with corresponding sensors, the system can know whether the user is currently indoor or outdoor, and the current weather conditions (sunny, rainy, etc.).

[0081] By comprehensively integrating these different types of user data, the system can more comprehensively understand the user's behavior habits, preferences, and the situation they are in, so as to provide more personalized services and recommendations.

[0082] 102. Extract features from the user data to construct a multi-dimensional data portrait of the target user;

[0083] Optionally, in the embodiments of the present application, the process of "extracting features from the user data to construct a multi-dimensional data portrait of the target user" in step 102 may include:

[0084] 1021. For different types of user data, using corresponding feature extraction methods, extracting key features for reflecting user behaviors and interests of the target user to construct a feature set;

[0085] 1022. Integrate the feature set into a multi-dimensional feature vector to construct a multi-dimensional data portrait of the user, where the multi-dimensional feature vector is used to indicate the feature vector of the target user in multiple dimensions.

[0086] The process of "using corresponding feature extraction methods for different types of user data to extract key features for reflecting user behaviors and interests to construct a feature set" in step 1021 may include:

[0087] Through time series analysis, user behavior features are extracted from the user historical behavior data, and the user behavior features include at least access frequency, access duration, preferred content and behavior trends; device preference features are extracted from the device usage data, and the device preference features include at least device type, device usage frequency and commonly used applications; the geographic location information is analyzed using a clustering algorithm to extract the user's geographic location features, and the geographic location features include at least the user's permanent location, travel patterns and areas of interest; the social network interaction data is analyzed through natural language processing technology to extract the user's social tendency features, and the social tendency features include at least emotional tendencies, topic preferences and social activity; combined with the user environment data, the user's behavior differences under different environmental conditions are analyzed to extract environmental behavior features related to the environment; a feature set is constructed based on the user behavior features, the device preference features, the geographic location features, the social tendency features and the environmental behavior features.

[0088] In the above steps 1021 to 1022, user behavior characteristics: extract user behavior patterns from user historical behavior data through time series analysis and other methods, including access frequency (such as the number of visits per day), access duration (such as the duration of each visit), preferred content (such as frequently visited content types) and behavior trends (such as behavior patterns that change over time).

[0089] Device preference characteristics: Understand users' device preferences by analyzing the types of devices they use (such as smartphones, tablets), device usage frequency (such as daily usage time), and common applications (such as the top five most commonly used apps).

[0090] Geographic location characteristics: Use clustering algorithms to analyze the user's geographic location information to identify the user's permanent location (such as home, office), travel patterns (such as commuting routes) and areas of interest (such as frequently visited shopping malls).

[0091] Social tendency features: Analyze the user's interactive content on social networks through natural language processing technology, extract the user's emotional tendency (such as positive, negative), topic preference (such as liking to discuss topics such as technology and travel) and social activity (such as the number of daily status posts).

[0092] Environmental behavior characteristics: Consider the differences in user behavior under different environmental conditions, such as the difference in user behavior indoors and outdoors, or activity preferences under different weather conditions.

[0093] In the embodiment of the present application, for different types of user data, corresponding feature extraction methods are used to extract key features for reflecting user behaviors and interests to construct a feature set, and the process may include:

[0094] User behavior characteristics: Assume that we have collected a user's browsing history showing that the user visits an e-commerce website 5 times a week, stays on the website for an average of 10 minutes each time, prefers to buy sports equipment, and has started to frequently search for outdoor activity related products in recent months. We can convert this information into specific numerical features, such as a visit frequency of 5 times / week, a visit duration of 10 minutes / time, a preferred content of sports equipment, and a behavioral trend of increased searches for outdoor activity products.

[0095] Device preference features: By analyzing the user's device usage data, it is found that the user mainly uses the iPhone 13Pro, which is used for about 5 hours a day, and the most frequently used applications are WeChat, Douyin, Taobao, Zhihu and Didi Chuxing. This information can be converted into a feature vector, such as the device type is iPhone 13Pro, the device usage frequency is 5 hours / day, and the commonly used applications are WeChat, Douyin, Taobao, Zhihu and Didi Chuxing.

[0096] Geographical location characteristics: By analyzing the user's geographic location information through a clustering algorithm, it can be found that the user's main activity area is concentrated in the city center. Every morning between 7 and 9 am, he leaves from his residence in Dongcheng District to his workplace in Xicheng District and returns between 6 and 8 pm. In addition, he often appears in several large shopping malls in Chaoyang District on weekends. This information can be summarized as the permanent location is Dongcheng District, the travel pattern is to travel between Dongcheng District and Xicheng District during peak hours in the morning and evening, and the area of ​​interest is shopping malls in Chaoyang District.

[0097] Social tendency characteristics: By analyzing the user's social media interaction content through natural language processing technology, it can be learned that the user tends to post and share photos and articles about travel and food, the emotions expressed are mostly positive and optimistic, and the active time on the social platform every day is from 7 to 10 pm. These can be extracted as positive emotional tendency, travel and food theme preference, and social activity of 3 hours / day.

[0098] Environmental behavior characteristics: Considering the differences in user behavior in different environments, for example, when users are outdoors, they may use navigation apps and health tracking apps more frequently; while when users are indoors, they are more likely to use entertainment apps such as video players and music players. In addition, users are more inclined to go out in sunny weather, but are more willing to stay at home and use indoor entertainment apps on rainy days. The differences in behavior under these environmental conditions can be recorded as environmental behavior characteristics.

[0099] Through the above feature extraction process, we can construct a feature set containing multiple features, further integrate them into a multi-dimensional feature vector, and thus construct a multi-dimensional data portrait of the user. Such a portrait can help enterprises and platforms better understand the needs and preferences of users and provide more accurate services and recommendations.

[0100] 103. Analyze the multi-dimensional data portrait of the target user to extract the interest tags and group tags of the target user, and store them in a pre-established user interest model;

[0101] Wherein, the interest tag is obtained by identifying a multidimensional feature vector in the multidimensional data portrait;

[0102] Optionally, in the embodiment of the present application, the process of “extracting interest tags and group tags of the target user by analyzing the multi-dimensional data portrait of the target user” in step 103 may include:

[0103] 1031. Perform feature selection on the multidimensional feature vector in the multidimensional data portrait of the target user, screen out representative user behavior features, device preference features, geographic location features, social tendency features, and environmental behavior features, and use them as user features;

[0104] 1032. Construct a feature matrix based on the feature vectors corresponding to the user features, and optimize the initial cluster center points according to the feature matrix to determine the target cluster center points;

[0105] In step 1032, "constructing a feature matrix based on the feature vectors corresponding to the user features, and optimizing the initial cluster center points according to the feature matrix to determine the target cluster center points" may include: constructing a feature matrix based on the feature vectors corresponding to the user features, wherein the rows of the feature matrix contain the feature values ​​of the target user on the user features, and the columns of the feature matrix contain the specific performances of the target user on representative user behavior features, device preference features, geographic location features, social tendency features, and environmental behavior features; randomly selecting any feature vector corresponding to the user features of the target user in the feature matrix as the first initial cluster center point, calculating the distance from the feature vectors corresponding to the remaining user features in the feature matrix to the first initial cluster center point, and constructing a probability distribution according to the square value of the distance, and then randomly selecting the next initial cluster center point until a preset number of initial cluster center points are selected; assigning each feature vector in the feature matrix to the initial cluster center point with the closest distance, and calculating the sum of the squares of the distances between each feature vector and the corresponding initial cluster center point, and minimizing the sum of the squares of the distances through iterative optimization until the preset number of initial cluster center points converge to form the target cluster center point.

[0106] 1033. Using the target cluster center point as an input of a hierarchical clustering algorithm, and combining the hierarchical clustering algorithm to perform multi-level interest grouping on the target user, thereby obtaining a multi-level interest grouping result;

[0107] In step 1033, the process of "using the target cluster center point as the input of the hierarchical clustering algorithm to perform multi-level interest grouping on the target user in combination with the hierarchical clustering algorithm to obtain multi-level interest grouping results" may include: using the target cluster center point to assign the feature vectors corresponding to the user characteristics to multiple initial groups to generate preliminary interest grouping results; based on the preliminary interest grouping results, constructing a similarity matrix of the user characteristics, and using a hierarchical clustering algorithm to gradually merge the groups with the highest similarity to form a hierarchical clustering tree; in the hierarchical clustering tree, by setting different similarity thresholds, dividing a multi-level interest group structure to identify the clustering relationship of the target user at different interest levels; by comparing the interest group structures at different levels, evaluating the rationality of the interest group structures at different levels, so as to select the best interest group structure as the multi-level interest grouping result.

[0108] 1034. Generate a user interest model according to the multi-level interest grouping result, wherein the user interest model records the interest tags and group tags of the target user;

[0109] The interest tags are generated according to the characteristics and preferences shown by the target users in the clustering process, and the group tags refer to the group identifiers to which the target users belong in different clustering levels.

[0110] In the above steps 1031 to 1034, feature selection: the most representative and discriminative features are selected from the multi-dimensional feature vector to reduce redundant information and improve the efficiency and accuracy of subsequent analysis.

[0111] Feature matrix: A two-dimensional array in which each row represents a sample (here, the target user) and each column represents a feature. Feature matrix is ​​a common data structure in machine learning, used to store and process feature data.

[0112] Initial cluster center point optimization: In cluster analysis, choosing the right initial cluster center points is crucial to the final clustering results. Optimizing this process usually involves randomly selecting initial center points and using specific strategies (such as k-means++) to select other initial center points to ensure that they are as dispersed as possible in the data space.

[0113] Hierarchical clustering algorithm: A clustering method that produces a tree-like structure (called a hierarchical cluster tree or dendrogram) that shows the similarities between data points. Hierarchical clustering can be agglomerative (bottom-up) or divisive (top-down).

[0114] Multi-level interest clustering: By performing multi-level clustering analysis on user data, users’ interest preferences can be revealed from different levels from coarse-grained to fine-grained.

[0115] Interest tags: tags generated based on the user's characteristics and preferences, used to describe the user's areas of interest.

[0116] Group label: An identifier that indicates the group to which a user belongs in different clustering levels. It can be used to describe the group category to which the user belongs.

[0117] In the embodiment of the present application, an example scenario is provided as follows:

[0118] Assume that we have built a multidimensional data portrait of a user, and now we need to further analyze the data to extract the user's interest tags and group tags.

[0119] We first selected the most representative features from the user data, such as users preferring to read technology articles (user behavior features), mainly using Android phones (device preference features), living in Beijing (geographic location features), frequently participating in outdoor activities (social tendency features), and being more active in sunny weather (environmental behavior features).

[0120] Based on these selected features, we construct a feature matrix. For example, a row of the matrix may represent the feature value of a user, while the columns represent the specific features mentioned above. The value in each cell reflects the user's specific performance on a certain feature, such as the frequency of preferring to read science and technology articles.

[0121] To determine the initial cluster center points, we first randomly selected a user feature vector as the first center point. Then, we calculated the distance from all other user feature vectors to this center point and constructed a probability distribution based on the squared value of the distance. Next, we randomly selected the next center point, making sure it is far enough from the selected center point. This process is repeated until we have selected a predetermined number of initial center points.

[0122] We use a hierarchical clustering algorithm to generate a hierarchical clustering tree. In this example, we first assign each user to the initial cluster center point that is closest to it. Then, we construct a similarity matrix and gradually merge the groups with the highest similarity to form a hierarchical clustering tree. By setting different similarity thresholds, we can divide different interest group structures from the hierarchical clustering tree, thereby identifying the clustering relationship of users at different interest levels.

[0123] Finally, we generated a user interest model based on the results of hierarchical clustering. For example, a user's interest tags may include "technology enthusiasts" and "outdoor sports enthusiasts", while group tags may include "Beijing technology enthusiasts group" and "outdoor sports enthusiasts group". These tags are stored in the user's interest model for subsequent use.

[0124] Through the above steps, we can not only accurately identify users’ areas of interest, but also understand their sense of belonging in different groups, so as to provide more personalized services and recommendations.

[0125] 104. Based on the user interest model, a hybrid recommendation algorithm combining deep neural network, collaborative filtering and graph neural network is used to train a personalized recommendation model, wherein the personalized recommendation model is used to generate prediction results based on interest tags and group tags in the user interest model, and the prediction results include the content of interest to the target user.

[0126] Optionally, in the embodiment of the present application, step 104 may include:

[0127] 1041. Extracting interaction data between the user and the content from the different types of user data;

[0128] 1042. Convert the interest tags and group tags in the user interest model into interest feature vectors and group feature vectors, convert the interaction data between the user and the content into user content feature vectors, and input them into a deep neural network. Perform a high-dimensional representation of the interest feature vectors, the group feature vectors, and the user content feature vectors through an embedding layer to obtain a nonlinear relationship, wherein the nonlinear relationship refers to the relationship between the user interest feature, the group feature vector, and the user content feature vector.

[0129] 1043. constructing a user-content interaction matrix using the user-content interaction data, calculating the similarity between the target user's content using a collaborative filtering algorithm, and generating a user-content matching score, wherein the similarity between the content refers to the degree of similarity between two or more content items, and the similarity is calculated using the user-content interaction data;

[0130] 1044. Construct a relationship graph between the user and the content based on the interaction data between the user and the content, and apply a graph neural network to perform feature propagation and aggregation to obtain an interest propagation path between the user and the content;

[0131] 1045. Train a personalized recommendation model by combining the nonlinear relationship, the matching score between the user and the content, and the interest propagation path between the user and the content.

[0132] In the above steps 1041 to 1045, the interaction data between the user and the content refers to the interaction records between the user and various contents (such as articles, videos, etc.), such as clicks, favorites, comments, etc.

[0133] Deep Neural Network (DNN): A neural network model composed of multiple layers of nodes, used to process complex nonlinear relationships. In recommendation systems, DNN can be used to capture the complex associations between user interests and content features.

[0134] Embedding layer: A layer in a deep learning model that converts discrete categorical variables into continuous vector representations to facilitate model learning.

[0135] Non-linear relationship: refers to the relationship between variables that cannot be simply represented by a straight line or a plane, but rather exists in a more complex pattern. In a recommendation system, this relationship may exist between user interests, group characteristics, and content characteristics.

[0136] Collaborative filtering algorithm: A recommendation algorithm that recommends content to users by finding similar users or similar items based on the user-item rating matrix.

[0137] User-content interaction matrix: a two-dimensional matrix in which rows represent users, columns represent content items, and the element values ​​in the matrix reflect the user's rating or interaction intensity on the content.

[0138] Graph Neural Network (GNN): A special type of neural network architecture that is suitable for processing graph-structured data, capable of capturing relationships between nodes and performing feature propagation and aggregation.

[0139] Feature propagation and aggregation: In graph neural networks, each node updates its own feature representation based on the features of its neighboring nodes, a process called feature propagation. Afterwards, the nodes aggregate these updated features to form a higher-level representation.

[0140] Interest propagation path: In the user-content relationship graph, the path from one user to another or from a user to content, which reflects how user interests spread through social networks or common interests.

[0141] In the embodiments of the present application, the following example scenarios are provided:

[0142] Assuming that we have completed the previous steps, we now want to train a personalized recommendation model based on the user's interest model.

[0143] We collected user browsing, liking, sharing, commenting and other behavioral data, which formed the basis of user-content interaction. We mapped user interest tags (such as "technology enthusiasts") and group tags (such as "Beijing technology enthusiasts group") into feature vectors. For example, "technology enthusiasts" may be mapped to a vector with higher values ​​for dimensions related to technology. We created a user-content interaction matrix, where each row represents a user, each column represents a content item, and the values ​​in the matrix reflect the user's rating of the content or the intensity of interaction. Based on the user-content interaction data, we constructed a graph structure in which nodes represent users or content, and edges represent the interaction relationship between users and content. We combined deep neural networks, collaborative filtering, and graph neural networks to train recommendation models.

[0144] It should be noted that the above embodiment includes the following details:

[0145] Extracting user-content interaction data: We extracted user behavior records from the database, such as user A clicking on article B, user C collecting movie D, etc.

[0146] Convert interest tags and group tags into feature vectors: We use an embedding layer to map “technology enthusiasts” into a vector [0.8, 0.1, 0.1, ...], and “Beijing technology enthusiasts group” into [0.9, 0.05, 0.05, ...], and convert the user-content interaction data into user-content feature vectors.

[0147] Constructing an interaction matrix: We created a user-content interaction matrix. For example, if user A clicks on article B, then the value in the matrix may be 1 in the row corresponding to user A and the column corresponding to article B, indicating that an interaction has occurred.

[0148] Build a graph of user-content relationships: We build a graph where user nodes are connected to the article nodes they interacted with, and article nodes may also be connected due to similarity. For example, user A and user B both clicked on article B, so in the graph, there is a connection between user A, user B, and article B.

[0149] Training personalized recommendation model: Use deep neural network to capture the nonlinear relationship between user interest features, group features and content features. For example, we can convert user A's interest tags and group tags into vectors through the embedding layer, and input them into the deep neural network together with the interaction data between user A and article B to learn the nonlinear relationship between them. Use collaborative filtering algorithm to calculate the matching score between users and content. We use the user-content interaction matrix to calculate the similarity between user A and other users or article B, and recommend content that user A may be interested in. Apply graph neural network to propagate and aggregate interest information between users and content. We use the constructed user-content relationship graph to let the graph neural network learn the association between nodes and propagate the user's interest features to related content nodes. Combining the results of the above three parts, we trained a comprehensive model that can generate prediction results based on user interest features, group features and content features, that is, recommend article B to user A, etc.

[0150] Evaluating recommendation results: We use test set data to evaluate the effectiveness of the recommendation model. For example, we adjust model parameters by calculating indicators such as the click-through rate of recommended content and user satisfaction.

[0151] Through the above steps, we can effectively train a personalized recommendation model that can recommend content of interest to users based on their interests and behavioral habits.

[0152] 105. Generate a context-aware push strategy based on the acquired current geographic location, device status, time node, and social activity status of the target user and the prediction result of the personalized recommendation model;

[0153] The context-aware push strategy includes at least: push timing and push content;

[0154] Optionally, in the embodiment of the present application, step 105 may specifically include:

[0155] 1051. Using the real-time acquired geographic location, device status, current time node, and social activity status of the target user as elements to integrate into a context information set;

[0156] 1052. Determine the weight and influence of each element in the context information set;

[0157] Optionally, the characteristics of the target user's geographical location (such as a city, business district, residential area, etc.) can be identified to evaluate the relevance of the location to the pushed content, so as to determine the weight and influence of the geographical location; the relevance of the device status to the pushed content can be determined by the target user's device status (such as connection status, power level, and usage scenario), so as to determine the weight and influence of the device status; the suitability of the push timing can be evaluated by the current time node (such as weekdays, weekends, holidays, morning and evening rush hours), so as to determine the weight and influence of the current time node; the target user's social activity status (such as social activity and current activity status (such as social network interaction, gathering, travel)) can be used to determine the target user's acceptance of the pushed content, so as to determine the weight and influence of the social activity status;

[0158] 1053. Generate a preliminary push content list according to the prediction result of the personalized recommendation model and the weight and influence of each element in the context information set;

[0159] 1054. Optimizing the preliminary push content list according to a preset optimization target to generate a context-aware push strategy;

[0160] The preset optimization goals include at least one or any combination of the following: optimization of push timing, improvement of content relevance, optimization of push coverage, optimization of conversion rate, and maximization of user engagement.

[0161] In this step, context-aware push strategy: A push strategy that is customized according to the user's current environment and status, aiming to improve the relevance of the pushed content and the receiver's satisfaction.

[0162] Geographic location: refers to the user's current physical location, which can be a specific place (such as a restaurant, shopping mall, etc.) or a wider area (such as a city, country, etc.).

[0163] Device Status: refers to the operating status of the user's device, including but not limited to battery level, network connection type (such as Wi-Fi or mobile data), whether the screen is on, etc.

[0164] Time node: refers to a specific time period or date, such as weekdays, weekends, holidays, etc., as well as a time period in a day (such as morning and evening).

[0165] Social activity status: refers to the user's current social behavior and activities, such as attending a party, traveling, or interacting on a social network.

[0166] Weight and influence: refers to the importance of each element in the context information set in determining the push strategy. Factors with higher weights have a greater impact on the final decision.

[0167] Initial push content list: The initial recommended content list generated based on the user interest model and personalized recommendation model.

[0168] Preset optimization goals: refers to specific goals set to improve push effects, such as improving content relevance, optimizing push time, etc.

[0169] In the embodiment of the present application, an example scenario is provided as follows:

[0170] Suppose we are developing an application that can recommend relevant content to users based on factors such as the user's current location, device status, time node, and social activity status. Here are the specific steps to achieve this function.

[0171] First, we need to obtain information such as the user's geographic location, device status, current time node, and social activity status in real time, and integrate this information into a contextual information set.

[0172] Next, we need to determine the weight and influence of each element in the context information set.

[0173] For example, if the user is in a shopping mall, the location might be weighted higher because it indicates that the user might be more interested in shopping-related content.

[0174] If the user's device is low on battery, the device state might be weighted higher because it could affect the user's ability to view content.

[0175] If the current time is the weekend, the weight of the time node may be higher because users may have more free time to view push content on weekends.

[0176] If the user is engaging in a social activity, such as attending a party or traveling, the weight of the social activity status will also be higher because the user may be more interested in content related to the activity.

[0177] Using the prediction results of the personalized recommendation model and combining the weight and influence of each element in the context information set, a preliminary push content list is generated.

[0178] Further optimize the initial push content list based on preset optimization goals to ensure that the push content is more in line with the user's interests and current status.

[0179] It should be noted that the above embodiment includes the following embodiment details:

[0180] Integration of context information collection: User location: Assume that the user is currently in Taikoo Li, Sanlitun, Chaoyang District, Beijing, which is a bustling commercial district. Device status: The user's phone battery is 70%, connected to Wi-Fi, and unlocked. Time node: Today is 8 o'clock on Saturday night. Social activity status: The user recently posted some party photos on social media, indicating that the user may be attending a party.

[0181] Determine the weight and influence of each element: Geographic location: Since the user is in a commercial area, the weight of the geographic location is set to 0.4, because this indicates that the user may be interested in nearby dining and entertainment content. Device status: The user's device is in good condition, so the weight of the device status is set to 0.1, because the user has enough power and network connection to view the content. Time node: Since it is a weekend night, the weight of the time node is set to 0.3, because users usually have more leisure time on weekend nights. Social activity status: The user is attending a party, and the weight of the social activity status is set to 0.2, because this indicates that the user may be interested in activities related to the party.

[0182] Generate a preliminary push content list: Based on the prediction results of the personalized recommendation model and the weights of each element in the context information set, we generate a preliminary push content list.

[0183] For example:

[0184] Recommend popular restaurants and bars nearby.

[0185] Recommend upcoming concerts and art exhibitions.

[0186] Provides information on some weekend specials.

[0187] Optimize the initial push content list: Optimize push timing: Considering the user's current time node, we decided to push content around 8:30 pm, which is the time when the user is most likely to check his phone. Improve content relevance: Considering that the user is attending a party, we added party-related content, such as nearby bars with live music performances. Optimize push coverage: We made sure that the pushed content covered multiple categories to meet the different interests of users. Optimize conversion rate: We analyzed historical data and found that providing discount coupons can increase user engagement and conversion rate, so we included coupons in the push content. Maximize user engagement: We also took into account that users may share this information with friends, so we selected some content suitable for sharing with multiple people, such as team games and group activities.

[0188] Through the above steps, we can develop a more personalized push strategy based on the user's contextual information to improve user engagement and satisfaction.

[0189] 106. Push corresponding content to the target user based on the context-aware push strategy.

[0190] In this step, context-aware push strategy: This is a push strategy based on factors such as the user's current geographic location, device status, time node, and social activity status, aiming to improve the relevance of push content and the recipient's satisfaction.

[0191] Target users: refers to users selected based on the user interest model and personalized recommendation model determined in the previous step.

[0192] Push content: The content that is most suitable for the current context, as determined by the context-aware push strategy. This content should be highly relevant to the user’s interests and current status.

[0193] In the embodiments of the present application, the following example scenarios are provided:

[0194] Assume that we have completed the previous steps, including analyzing user data, building user interest models, training personalized recommendation models, and generating context-aware push strategies. Now we need to push corresponding personalized content to users based on these strategies.

[0195] It should be noted that the above embodiment includes the following embodiment details:

[0196] Determine the target user: We have determined a target user, such as user A, through the previous steps.

[0197] Integrate contextual information collection: We have obtained user A’s geographic location (located in Taikoo Li Sanlitun, Chaoyang District, Beijing), device status (the phone’s battery is 70%, it is connected to Wi-Fi, and it is unlocked), time node (today is 8 o’clock on Saturday night), and social activity status (the user recently posted some party photos on social media, indicating that the user may be attending a party).

[0198] Determine the weight and influence of each element: We have determined the weight and influence of each element in the context information set, for example, the weight of geographic location is 0.4, the weight of device status is 0.1, the weight of time node is 0.3, and the weight of social activity status is 0.2.

[0199] Generate a preliminary push content list: We use the prediction results of the personalized recommendation model and the weight and influence of each element in the context information set to generate a preliminary push content list, for example:

[0200] Recommend popular restaurants and bars nearby.

[0201] Recommend upcoming concerts and art exhibitions.

[0202] Provides information on some weekend specials.

[0203] Optimize the initial push content list: We further optimized the initial push content list based on the preset optimization goals, such as:

[0204] Push timing optimization: Decided to push content around 8:30 pm, which is the time when users are most likely to check their phones.

[0205] Improved relevance: Added content related to gatherings, such as nearby bars with live music.

[0206] Optimize push coverage: Ensure that the pushed content covers multiple categories to meet the different interests of users.

[0207] Conversion rate optimization: coupons were added to push content.

[0208] Maximize user engagement: Select content that is suitable for multiple people to share, such as team games and group activities.

[0209] Push content based on context-aware push strategies:

[0210] According to the optimized push strategy, we decided to push the following content to user A:

[0211] Offers for a nearby bar with live music.

[0212] Discounted tickets for an upcoming concert in Sanlitun.

[0213] Some special offers for team games and group activities.

[0214] Execute push:

[0215] We sent a push notification to User A at 8:30 PM with a link and a brief description of the above content.

[0216] The title of the push notification might be: "Enjoy special offers in Sanlitun tonight!"

[0217] The push content includes specific discount information, event time, location, and instructions on how to book or participate.

[0218] Through the above steps, we can effectively push personalized content to users based on their interests and current context, thereby improving user engagement and satisfaction.

[0219] Figure 2 A schematic diagram of the structure of a smart push device based on big data is provided for the embodiment of the present application, such as Figure 2 As shown, the device comprises:

[0220] A collection module 21 is used to collect user data of target users from multiple heterogeneous data sources;

[0221] A construction module 22, configured to extract features from the user data to construct a multi-dimensional data portrait of the target user;

[0222] An extraction module 23 is used to extract interest tags and group tags of the target user by analyzing the multidimensional data portrait of the target user, and store them in a pre-established user interest model, wherein the interest tags are obtained by identifying the multidimensional feature vectors in the multidimensional data portrait;

[0223] The training module 24 is used to train a personalized recommendation model based on the user interest model using a hybrid recommendation algorithm that combines deep neural networks, collaborative filtering, and graph neural networks. The personalized recommendation model is used to generate prediction results based on the interest tags and group tags in the user interest model, and the prediction results include the content of interest to the target user.

[0224] A generating module 25 is used to generate a context-aware push strategy based on the current geographical location, device status, time node and social activity status of the target user obtained, combined with the prediction result of the personalized recommendation model, wherein the context-aware push strategy at least includes: push timing and push content;

[0225] The push module 26 is used to push corresponding content to the target user based on the context-aware push strategy.

[0226] Optionally, in an embodiment of the present application, the user data includes different types of user data, and the different types of user data include at least: user historical behavior data, device usage data, geographic location information, social network interaction data, and user environment data;

[0227] The construction module 22 is specifically used to use corresponding feature extraction methods for different types of user data to extract key features that reflect the user behavior and interests of the target user to construct a feature set; integrate the feature set into a multidimensional feature vector to construct a multidimensional data portrait of the user, and the multidimensional feature vector is used to indicate the feature vector of the target user in multiple dimensions.

[0228] Optionally, in an embodiment of the present application, the construction module 22 is specifically used to extract user behavior characteristics from the user historical behavior data through time series analysis, and the user behavior characteristics include at least access frequency, access duration, preferred content and behavior trends; extract device preference characteristics from the device usage data, and the device preference characteristics include at least device type, device usage frequency and commonly used applications; use a clustering algorithm to analyze the geographic location information to extract the user's geographic location characteristics, and the geographic location characteristics include at least the user's permanent location, travel patterns and areas of interest; analyze the social network interaction data through natural language processing technology to extract the user's social tendency characteristics, and the social tendency characteristics include at least emotional tendency, topic preference and social activity; combine the user environment data to analyze the user's behavior differences under different environmental conditions to extract environmental behavior characteristics related to the environment; construct a feature set based on the user behavior characteristics, the device preference characteristics, the geographic location characteristics, the social tendency characteristics and the environmental behavior characteristics.

[0229] Optionally, in an embodiment of the present application, the extraction module 23 is specifically used to perform feature selection on the multidimensional feature vector in the multidimensional data portrait of the target user, screen out representative user behavior characteristics, device preference characteristics, geographic location characteristics, social tendency characteristics and environmental behavior characteristics, and use them as user characteristics; construct a feature matrix based on the feature vectors corresponding to the user characteristics, optimize the initial clustering center points according to the feature matrix to determine the target clustering center points; use the target clustering center points as the input of the hierarchical clustering algorithm, and perform multi-level interest grouping on the target user in combination with the hierarchical clustering algorithm to obtain multi-level interest grouping results; generate a user interest model based on the multi-level interest grouping results, and the user interest model records the interest tags and group tags of the target user, and the interest tags are generated according to the characteristics and preferences shown by the target user in the clustering process, and the group tags refer to the group identifiers to which the target user belongs in different clustering levels.

[0230] Optionally, in an embodiment of the present application, the extraction module 23 is also used to construct a feature matrix based on the feature vectors corresponding to the user features, the rows of the feature matrix contain the feature values ​​of the target user on the user features, and the columns of the feature matrix contain the specific performance of the target user in representative user behavior characteristics, device preference characteristics, geographic location characteristics, social tendency characteristics and environmental behavior characteristics; in the feature matrix, randomly select any user feature corresponding feature vector of the target user as the first initial cluster center point, calculate the distance from the feature vectors corresponding to the remaining user features in the feature matrix to the first initial cluster center point, and construct a probability distribution based on the square value of the distance, and then randomly select the next initial cluster center point until a preset number of initial cluster centers are selected; assign each feature vector in the feature matrix to the nearest initial cluster center point, and calculate the sum of the squares of the distances between each feature vector and the corresponding initial cluster center point, and minimize the sum of the squares of the distances through iterative optimization until the preset number of initial cluster centers converge to form a target cluster center point.

[0231] Optionally, in an embodiment of the present application, the extraction module 23 is also used to use the target cluster center point to assign the feature vectors corresponding to the user features to multiple initial groups to generate a preliminary interest grouping result; based on the preliminary interest grouping result, a similarity matrix of the user features is constructed, and a hierarchical clustering algorithm is used to gradually merge the groups with the highest similarity to form a hierarchical clustering tree; in the hierarchical clustering tree, a multi-level interest group structure is divided by setting different similarity thresholds to identify the clustering relationship of the target user at different interest levels; by comparing the interest group structures at different levels, the rationality of the interest group structures at different levels is evaluated to select the best interest group structure as the multi-level interest grouping result.

[0232] Optionally, in an embodiment of the present application, the training module 24 is specifically used to extract user-content interaction data from the different types of user data; convert the interest tags and group tags in the user interest model into interest feature vectors and group feature vectors, convert the user-content interaction data into user content feature vectors, and input them into a deep neural network, and perform high-dimensional representation of the interest feature vectors, the group feature vectors, and the user content feature vectors through an embedding layer to obtain a nonlinear relationship, wherein the nonlinear relationship refers to the relationship between the user interest feature, the group feature vector, and the user content feature vector; construct a user-content interaction matrix using the user-content interaction data, calculate the similarity between the contents of the target user through a collaborative filtering algorithm, and generate a matching score between the user and the content, wherein the similarity between the contents refers to the degree of similarity between two or more content items, and the similarity is calculated through the user-content interaction data; construct a user-content relationship graph based on the user-content interaction data, and apply a graph neural network to perform feature propagation and aggregation to obtain an interest propagation path between the user and the content; and train a personalized recommendation model in combination with the nonlinear relationship, the user-content matching score, and the user-content interest propagation path.

[0233] Optionally, in the embodiment of the present application, the generating module 25 is specifically used to use the real-time acquired geographic location, device status, current time node and social activity status of the target user as elements to integrate into a context information set; determine the weight and influence of each element in the context information set; generate a preliminary push content list according to the prediction result of the personalized recommendation model and the weight and influence of each element in the context information set; optimize the preliminary push content list by preset optimization goals to generate a context-aware push strategy;

[0234] The preset optimization goals include at least one or any combination of the following: optimization of push timing, improvement of content relevance, optimization of push coverage, optimization of conversion rate, and maximization of user engagement.

[0235] Figure 2 The intelligent push device based on big data can execute Figure 1 The implementation principle and technical effect of the intelligent push method based on big data described in the illustrated embodiment will not be described in detail. The specific manner in which each module and unit performs operations in the intelligent push device based on big data in the above embodiment has been described in detail in the embodiment of the method, and will not be elaborated here.

[0236] In one possible design, Figure 2 The big data-based intelligent push device of the embodiment shown can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0237] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0238] The processing component 32 is used to: collect user data of the target user from multiple heterogeneous data sources; perform feature extraction on the user data to construct a multidimensional data portrait of the target user; extract interest tags and group tags of the target user by analyzing the multidimensional data portrait of the target user, and store them in a pre-established user interest model, wherein the interest tags are obtained by identifying multidimensional feature vectors in the multidimensional data portrait; based on the user interest model, a personalized recommendation model is trained using a hybrid recommendation algorithm combining deep neural networks, collaborative filtering and graph neural networks, the personalized recommendation model is used to generate prediction results based on the interest tags and group tags in the user interest model, the prediction results include the content of interest to the target user; based on the current geographic location, device status, time node and social activity status of the target user obtained, combined with the prediction results of the personalized recommendation model, a context-aware push strategy is generated, the context-aware push strategy includes at least: push timing and push content; based on the context-aware push strategy, the corresponding content is pushed to the target user.

[0239] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to perform the above method.

[0240] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0241] Of course, the computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0242] The input / output interface provides an interface between the processing component and the peripheral interface module, which may be an output device, an input device, etc.

[0243] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0244] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0245] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 The illustrated embodiment is an intelligent push method based on big data.

[0246] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0247] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0248] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0249] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An intelligent push method based on big data, characterized in that: include: Collect user data of target users from multiple heterogeneous data sources; Extracting features from the user data to construct a multi-dimensional data portrait of the target user; By analyzing the multidimensional data portrait of the target user, interest tags and group tags of the target user are extracted and stored in a pre-established user interest model, wherein the interest tags are obtained by identifying multidimensional feature vectors in the multidimensional data portrait; Based on the user interest model, a hybrid recommendation algorithm combining deep neural network, collaborative filtering and graph neural network is used to train a personalized recommendation model, wherein the personalized recommendation model is used to generate prediction results according to the interest tags and group tags in the user interest model, and the prediction results include the content of interest to the target user; Generate a context-aware push strategy based on the acquired current geographic location, device status, time node, and social activity status of the target user, combined with the prediction result of the personalized recommendation model, wherein the context-aware push strategy includes at least: push timing and push content; Pushing corresponding content to the target user based on the context-aware push strategy; The step of analyzing the multi-dimensional data portrait of the target user to extract the interest tags and group tags of the target user includes: Performing feature selection on the multidimensional feature vector in the multidimensional data portrait of the target user, screening out representative user behavior features, device preference features, geographic location features, social tendency features, and environmental behavior features as user features; Constructing a feature matrix based on the feature vectors corresponding to the user features, and optimizing the initial cluster center points according to the feature matrix to determine the target cluster center points; The target cluster center point is used as an input of a hierarchical clustering algorithm, so as to perform multi-level interest grouping on the target user in combination with the hierarchical clustering algorithm to obtain a multi-level interest grouping result; A user interest model is generated based on the multi-level interest clustering results. The user interest model records the interest tags and group tags of the target user. The interest tags are generated based on the characteristics and preferences shown by the target user in the clustering process. The group tags refer to the group identifiers to which the target user belongs in different clustering levels.

2. The method according to claim 1, characterized in that The user data includes different types of user data, and the different types of user data include at least: user historical behavior data, device usage data, geographic location information, social network interaction data, and user environment data; The extracting features of the user data to construct a multi-dimensional data portrait of the target user includes: For different types of user data, corresponding feature extraction methods are used to extract key features that reflect the user behavior and interests of the target user to construct a feature set; The feature set is integrated into a multi-dimensional feature vector to construct a multi-dimensional data portrait of the user, and the multi-dimensional feature vector is used to indicate the feature vector of the target user in multiple dimensions.

3. The method according to claim 2, characterized in that For different types of user data, corresponding feature extraction methods are used to extract key features that reflect user behaviors and interests to construct feature sets, including: Extracting user behavior characteristics from the user historical behavior data through time series analysis, wherein the user behavior characteristics at least include access frequency, access duration, preferred content, and behavior trend; Extracting device preference features from the device usage data, the device preference features including at least device type, device usage frequency, and commonly used applications; Analyzing the geographic location information using a clustering algorithm to extract the user's geographic location features, the geographic location features at least including the user's permanent location, travel patterns, and areas of interest; Analyzing the social network interaction data by natural language processing technology to extract the user's social tendency characteristics, wherein the social tendency characteristics at least include emotional tendency, topic preference and social activity; In combination with the user environment data, analyzing the differences in user behavior under different environmental conditions to extract environmental behavior features related to the environment; A feature set is constructed based on the user behavior feature, the device preference feature, the geographic location feature, the social tendency feature, and the environmental behavior feature.

4. The method according to claim 1, characterized in that: The constructing a feature matrix based on the feature vector corresponding to the user feature, and optimizing the initial cluster center point according to the feature matrix to determine the target cluster center point includes: Constructing a feature matrix based on the feature vectors corresponding to the user features, wherein the rows of the feature matrix contain the feature values ​​of the target user on the user features, and the columns of the feature matrix contain the specific performance of the target user on representative user behavior features, device preference features, geographic location features, social tendency features, and environmental behavior features; In the feature matrix, randomly select a feature vector corresponding to any user feature of the target user as the first initial cluster center point, calculate the distance from the feature vectors corresponding to the remaining user features in the feature matrix to the first initial cluster center point, and construct a probability distribution according to the square value of the distance, and then randomly select the next initial cluster center point until a preset number of initial cluster center points are selected; Each eigenvector in the feature matrix is ​​assigned to the nearest initial cluster center point, and the sum of squared distances between each eigenvector and the corresponding initial cluster center point is calculated, and the sum of squared distances is minimized through iterative optimization until a preset number of initial cluster center points converge to form a target cluster center point.

5. The method according to claim 1, characterized in that The step of using the target cluster center point as the input of the hierarchical clustering algorithm, and combining the hierarchical clustering algorithm to perform multi-level interest grouping on the target user to obtain a multi-level interest grouping result includes: Using the target cluster center point, the feature vectors corresponding to the user features are assigned to a plurality of initial groups to generate preliminary interest grouping results; Based on the preliminary interest grouping results, construct a similarity matrix of the user characteristics, and use a hierarchical clustering algorithm to gradually merge the groups with the highest similarity to form a hierarchical clustering tree; In the hierarchical clustering tree, by setting different similarity thresholds, a multi-level interest group structure is divided to identify the clustering relationship of the target user at different interest levels; By comparing the interest group structures at different levels, the rationality of the interest group structures at different levels is evaluated to select the best interest group structure as the multi-level interest grouping result.

6. The method according to claim 2, characterized in that The method of training a personalized recommendation model based on the user interest model using a hybrid recommendation algorithm combining a deep neural network, a collaborative filtering algorithm and a graph neural network includes: Extracting interaction data between the user and the content from the different types of user data; The interest tags and group tags in the user interest model are converted into interest feature vectors and group feature vectors, and the interaction data between the user and the content is converted into user content feature vectors, and the result is input into a deep neural network. The interest feature vectors, the group feature vectors and the user content feature vectors are represented in a high dimension through an embedding layer to obtain a nonlinear relationship, wherein the nonlinear relationship refers to the relationship between the interest feature vectors, the group feature vectors and the user content feature vectors; Using the interaction data between the user and the content to construct an interaction matrix between the user and the content, calculating the similarity between the content of the target user through a collaborative filtering algorithm, and generating a matching score between the user and the content, wherein the similarity between the content refers to the degree of similarity between two or more content items, and the similarity is calculated based on the interaction data between the user and the content; Building a relationship graph between the user and the content based on the interaction data between the user and the content, and applying a graph neural network to perform feature propagation and aggregation to obtain an interest propagation path between the user and the content; The personalized recommendation model is trained by combining the nonlinear relationship, the matching score between the user and the content, and the interest propagation path between the user and the content.

7. The method according to claim 1, characterized in that The step of generating a context-aware push strategy based on the acquired current geographic location, device status, time node, and social activity status of the target user and the prediction result of the personalized recommendation model includes: The geographical location, device status, current time node and social activity status of the target user acquired in real time are used as elements to integrate into a context information set; Determining the weight and influence of each element in the context information set; Generate a preliminary push content list according to the prediction result of the personalized recommendation model and the weight and influence of each element in the context information set; Optimizing the preliminary push content list by preset optimization goals to generate a context-aware push strategy; The preset optimization goals include at least one or any combination of the following: optimization of push timing, improvement of content relevance, optimization of push coverage, optimization of conversion rate, and maximization of user engagement.

8. An intelligent push device based on big data, characterized in that: include: A collection module, used to collect user data of target users from multiple heterogeneous data sources; A construction module, used for extracting features from the user data to construct a multi-dimensional data portrait of the target user; An extraction module, configured to extract interest tags and group tags of the target user by analyzing the multidimensional data portrait of the target user, and store them in a pre-established user interest model, wherein the interest tags are obtained by identifying multidimensional feature vectors in the multidimensional data portrait; A training module, for training a personalized recommendation model based on the user interest model using a hybrid recommendation algorithm combining deep neural network, collaborative filtering and graph neural network, wherein the personalized recommendation model is used to generate prediction results according to interest tags and group tags in the user interest model, and the prediction results include the content of interest to the target user; A generation module, configured to generate a context-aware push strategy based on the acquired current geographic location, device status, time node, and social activity status of the target user and the prediction result of the personalized recommendation model, wherein the context-aware push strategy includes at least: push timing and push content; A push module, used for pushing corresponding content to the target user based on the context-aware push strategy; The extraction module is specifically used to perform feature selection on the multidimensional feature vector in the multidimensional data portrait of the target user, screen out representative user behavior characteristics, device preference characteristics, geographic location characteristics, social tendency characteristics and environmental behavior characteristics, and use them as user characteristics; construct a feature matrix based on the feature vectors corresponding to the user characteristics, optimize the initial cluster center points according to the feature matrix to determine the target cluster center points; use the target cluster center points as the input of the hierarchical clustering algorithm, and perform multi-level interest grouping on the target user in combination with the hierarchical clustering algorithm to obtain multi-level interest grouping results; generate a user interest model based on the multi-level interest grouping results, the user interest model records the interest tags and group tags of the target user, the interest tags are generated according to the characteristics and preferences shown by the target user in the clustering process, and the group tags refer to the group identifiers to which the target user belongs in different clustering levels.

9. A computing device, characterized in that It comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement the big data-based intelligent push method as described in any one of claims 1 to 7.

Images