A system for matching and accelerating the generation of user portraits

Abstract

The present invention proposes a system for accelerating the generation of user portraits through matching, which belongs to the field of user portrait generation and includes a data collection module, a behavior query module, a behavior analysis module, a behavior prediction module and a result display module; the data collection module is used to collect the behavior data and transaction data of TOB-end users, the behavior data includes the user's browsing behavior, search behavior and purchase behavior, and the transaction data includes the user's purchase record and transaction amount; the behavior query module sorts the data collected in the data collection module according to keywords, industries and time ranges, and filters the behavior data and transaction data collected by the data collection module according to limited conditions when querying; the system realizes efficient processing, analysis and prediction of TOB-end user behavior and transaction data in an automated and intelligent manner, generates accurate user portraits, and displays the prediction results in an intuitive manner, thereby improving the decision-making efficiency and effectiveness of the enterprise.

Description

Technical Field

[0001] The present invention belongs to the field of user portrait generation, and in particular relates to a system for matching and accelerating the generation of user portraits. Background Art

[0002] In today's digital age, competition among enterprises is becoming increasingly fierce, especially in the TOB (business-to-business) market. Understanding customer needs and providing personalized services have become the key to gaining competitive advantage. User portraits are an important tool for enterprises to obtain users' deep needs, habits and preferences. They build users' virtual images by analyzing users' behavioral data, helping enterprises to accurately locate user characteristics and realize personalized product recommendations and services.

[0003] With the development of big data and artificial intelligence technology, the generation and application of user portraits have become more intelligent and automated. On the TOB side, user behavior data is usually more complex and rich, including but not limited to users' browsing behavior, search behavior, transaction behavior, and social interaction behavior. The effective analysis and use of these data are crucial to building high-quality user portraits.

[0004] Traditional user portrait generation methods usually rely on manual data collection and analysis, which is not only time-consuming and labor-intensive, but also difficult to process large-scale data and capture real-time changes in user behavior, thus affecting the accuracy and timeliness of user portraits. Therefore, how to use advanced technology to achieve automatic generation and real-time updating of user portraits has become an urgent problem to be solved in the TOB market.

[0005] The TOB user portrait automatic generation system relies on a large amount of user behavior data to generate user portraits. However, the quality of the data is a challenge. For example, the data may contain errors, duplications, or missing information, which will affect the accuracy of the system. Because the amount of data on the TOB side is often very large, and the types and formats of the data may also vary, this places high demands on the ability to process and analyze data. Traditional data processing and analysis techniques may not be able to meet this demand. The market environment and user behavior are constantly changing, so user portraits also need to be able to be updated in real time to reflect the latest user behavior. However, real-time processing and analysis of large amounts of data to update user portraits is a technical challenge. In the process of collecting and processing user behavior data, user privacy needs to be strictly protected. At the same time, because user portraits contain a lot of sensitive information, data security also needs to be considered in the process of storing and using user portraits. The user portrait automatic generation system usually includes a prediction module to predict the user's future behavior. However, the accuracy of the prediction is affected by many factors, such as model selection, feature selection, data quality, etc. Although the user portrait automatic generation system can generate detailed user portraits, how to interpret and use these portraits is a challenge. For example, how to interpret the characteristics in the user portrait, how to formulate marketing strategies based on the user portrait, etc. Each user is unique, and their needs and behaviors are also different. Therefore, how to generate a user portrait that can reflect the uniqueness of the user is a challenge.

[0006] The key point of current development is that the current user portrait generation system has performance bottlenecks in the big data environment. Traditional methods are often unable to effectively handle large and complex data sets, resulting in slow generation speed. This affects the feasibility of real-time applications, especially in scenarios that require instant response, such as online advertising and personalized recommendation systems.

[0007] Existing technologies have limitations in extracting user features. Traditional algorithms have difficulty capturing complex user behaviors and preference patterns, resulting in a lack of accuracy and depth in the generated user portraits. This affects the system’s understanding of user needs and reduces the quality of personalized services.

[0008] Many current systems have deficiencies in matching algorithms. Due to the lack of effective matching strategies, the system often has difficulty accurately matching the generated user profile with the actual user behavior. This results in inaccurate recommendations of products or services, which reduces user experience and satisfaction.

[0009] To address these issues, the new generation of user portrait generation system uses advanced big data processing technology to improve data processing speed and efficiency. At the same time, deep learning algorithms are introduced for user feature extraction, enabling the system to more accurately capture user behavior patterns and generate more accurate and comprehensive user portraits.

[0010] In terms of matching, the new system uses advanced matching algorithms, combining user historical behavior, points of interest and real-time dynamic information to achieve more accurate user portrait matching. This not only improves the accuracy of recommendations, but also enhances the system's understanding of user needs, thereby improving the level of personalized services.

[0011] The new generation of user portrait generation system effectively solves the bottlenecks and shortcomings of existing technologies by integrating advanced big data processing, deep learning and matching algorithms. This will bring more powerful and efficient solutions to application areas such as personalized recommendations and advertising. Summary of the invention

[0012] The present invention proposes a system for accelerating the generation of user portraits. The system solves the problem of slow data query and matching speed in traditional user portrait generation systems through GPU-accelerated K-nearest neighbor query, embedding model vectorization and other technologies, and realizes efficient user behavior analysis and prediction.

[0013] The technical solution of the present invention is implemented as follows: a system for matching and accelerating the generation of user portraits, comprising a data collection module, a behavior query module, a behavior analysis module, a behavior prediction module and a result display module;

[0014] The data collection module is used to collect the behavior data and transaction data of TOB end users, the behavior data includes the user's browsing behavior, search behavior and purchase behavior, and the transaction data includes the user's purchase record and transaction amount;

[0015] The behavior query module sorts the data collected in the data collection module according to keywords, industries and time ranges, and filters the behavior data and transaction data collected by the data collection module according to the limiting conditions when querying;

[0016] The behavior analysis module extracts the information content of the user behavior defined in the behavior query module, marks the frequency of occurrence of the content, establishes a behavior data relevance weight value, sorts the user behavior information after composite weighting according to the frequency and relevance weight value, and sends the sorting information to the behavior prediction module;

[0017] The behavior prediction module calls the sorting information, establishes a user O2O behavior map according to the sorting information, establishes a user portrait according to the user O2O behavior map, and stores the user portrait in the backend server;

[0018] When building a user portrait through the user O2O behavior graph, all data are vectorized using the embedding model, and the norm of the vector is uniformly scaled to 1 to obtain n vectors; an index containing the data and vectors is created, and the data and the uniformly scaled vector are saved to the index;

[0019] When the behavior query module performs a query, it first performs a gpu-accelerated K-nearest-neighbor query on the content and the standard vector of each data group. If the query result is greater than the set threshold, it is considered similar, and the most similar i similar groups are selected. The gpu-accelerated K-nearest-neighbor query is performed on the content to be queried and all the data in the similar groups. If it is greater than the set threshold, it is concluded that there are j most matching data in i similar groups;

[0020] The best matching data is imported into the behavior prediction module, and a probability table is formed according to the probability of subsequent user behavior of the best matching data, and the probability table is sent to the result display module; after receiving the probability table, the result display module synchronously forms pie charts, bar charts and curve charts according to the probability table, and visualizes the pie charts, bar charts and curve charts through the display panel, and regularly captures the pie charts, bar charts and curve charts and uploads them to the backend server, and associates the captured charts with the user portraits.

[0021] Compared with traditional data collection, the data collection module of this application document not only covers the basic behavior data of users (such as browsing, searching, and purchasing behaviors), but also includes detailed transaction data (such as purchase records and transaction amounts). This comprehensive data collection can provide richer data support for generating more accurate user portraits. The prior art adopts a relatively simple query method, while the behavior query module in this system can perform highly customized data sorting and screening according to a variety of conditions (keywords, industries, time ranges), thereby providing more accurate and relevant data input to subsequent modules. This system directly uses the output of the behavior analysis module (the frequency and correlation weight value of the behavior information) for the behavior prediction module to establish a user O2O behavior map and generate a user portrait. This deep combination makes the model for predicting the probability of subsequent user behavior more accurate and optimizes the entire prediction process. Unlike the static or manually updated visualization in the prior art, this system provides dynamic visualization results, including pie charts, bar charts, and curve charts, and can automatically and regularly capture charts and upload them to the backend server, and associate them with user portraits. This allows users to obtain the latest data analysis results and make it easier to understand and apply these results.

[0022] Traditional user profile generation systems have slow data query and matching problems. Typically, the query algorithms used by traditional systems are inefficient and cannot effectively process large-scale data sets, resulting in slow query speeds. In contrast, the new system introduces GPU-accelerated K-nearest neighbor query technology, which takes advantage of GPU parallel computing to significantly increase query speeds. By leveraging the powerful computing power of the GPU, the system can complete data query and matching in a shorter time, thereby achieving rapid analysis and prediction of user behavior.

[0023] The new system uses an embedding model to vectorize user behavior. In traditional systems, feature extraction and vectorization of user behavior often use rule-based or simple statistical methods, which makes it difficult to capture the potential associations and patterns of the data. The embedding model can map complex user behavior data into a low-dimensional vector space, retaining the inherent structure and semantic information of the data, thereby better expressing the characteristics of user behavior. This enables the system to more accurately capture user interests and preferences when matching and recommending, improving the accuracy and personalization of recommendations.

[0024] The new system also introduces a behavior prediction module, which predicts users' possible future behaviors and preferences by analyzing users' historical behavior data and real-time dynamic information. Traditional systems often lack this prediction function and can only perform simple matching and recommendations based on existing data, and cannot flexibly respond to changes and dynamics in user behavior. The new system, by combining historical data and real-time information, can more accurately predict user behavior trends and provide more effective support for personalized recommendations.

[0025] In summary, the new generation of user portrait generation system solves the problems of slow data processing, low feature extraction accuracy and low recommendation accuracy in traditional systems by introducing advanced technologies such as GPU-accelerated query algorithms, embedding model vectorization and behavior prediction modules, and realizes rapid analysis of user behavior and personalized recommendations. The introduction and application of these technical solutions have significantly improved the system's performance and efficiency, providing users with a better and more personalized service experience.

[0026] The entire technical solution provides more comprehensive data support, enhancing the dimension and depth of user portraits, thereby improving the pertinence and effectiveness of marketing and service strategies. The highly customized query function makes data screening more accurate and ensures the data quality of the analysis and prediction modules. The close integration of behavioral analysis and prediction makes the generation of user portraits and behavioral predictions more coherent, improving the intelligence level of the system. Dynamic visualization and automatically updated result display improve the readability and practicality of information, allowing decision makers to respond quickly to market changes.

[0027] As a preferred implementation, when the data and norm-uniformly scaled vectors are saved to the index, the cosine similarity between the two vectors is calculated in parallel using the GPU. If the similarity is greater than a set threshold, the corresponding two vectors are considered to be associated. All vectors are plotted and the graph is compressed using the Louvain algorithm to obtain m data groups, and the average vector of all data in each data group is calculated, and the average vector is used as the standard vector of the group.

[0028] As a preferred implementation, when establishing the user O2O behavior map, the behavior prediction module reconstructs the sorting information data model through at least one of time series analysis, regression analysis, classification algorithm, clustering algorithm or deep learning algorithm, and refines the user behavior data through data reconstruction to improve the prediction accuracy of the user O2O behavior map.

[0029] As a preferred implementation, when extracting user behavior information, the behavior analysis module divides user types according to the user behavior information to form different categories of user behavior information, and makes major distinctions when establishing user portraits based on user behavior characteristics and consumption characteristics.

[0030] As a preferred implementation, when the behavior prediction module calls the sorting information, a user behavior event sequence is generated based on the sorting information. Each event represents a behavior of the user at a certain point in time. The user behavior event sequence is analyzed through a sequence analysis algorithm to establish the rules of user behavior, and a user O2O behavior map is established based on the rules of user behavior.

[0031] After adopting the above technical solution, the beneficial effects of the present invention are: through automated data collection, query and analysis processes, the efficiency of user behavior data processing is greatly improved, and the complexity and error rate of manual operations are reduced. The system can generate more accurate and comprehensive user portraits through in-depth analysis of user behavior data and transaction data, and establish user O2O behavior maps using behavior prediction modules, providing companies with more targeted user insights. The introduction of GPU-accelerated query algorithms and other advanced technologies has greatly improved the speed and efficiency of data processing, enabling the system to complete user behavior analysis and prediction in a shorter time. Through methods such as embedding model vectorization, the system can more accurately capture user interests and preferences, thereby improving the accuracy of personalized recommendations and user satisfaction.

[0032] By introducing the behavior prediction module, the system can more accurately predict the user's behavior trends by analyzing the user's historical behavior data and real-time dynamic information, thereby enhancing the understanding and prediction of user behavior. The behavior prediction module can predict the probability of the user's subsequent behavior based on the user's O2O behavior map and user portrait, helping enterprises to more accurately predict and understand user needs and behavior trends. The result display module generates pie charts, bar charts and curve charts to make the prediction results and user portraits more intuitive and easier to understand. At the same time, the regular chart capture and upload function allows the user portrait and prediction results to be updated in real time, providing the latest market information.

[0033] Through in-depth analysis and prediction of user behavior and transaction data, it can provide important support for the company's marketing, product development and service optimization decisions, help improve the company's competitiveness and increase market share. The system realizes efficient processing, analysis and prediction of TOB user behavior and transaction data in an automated and intelligent way, generates accurate user portraits, and displays prediction results in an intuitive way, thereby improving the company's decision-making efficiency and effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0035] Figure 1 It is a flow chart of the system of the present invention. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0037] Example:

[0038] like Figure 1 As shown, a system for matching and accelerating the generation of user portraits includes a data collection module, a behavior query module, a behavior analysis module, a behavior prediction module and a result display module;

[0039] The data collection module is used to collect the behavior data and transaction data of TOB end users, the behavior data includes the user's browsing behavior, search behavior and purchase behavior, and the transaction data includes the user's purchase record and transaction amount;

[0040] The behavior query module sorts the data collected in the data collection module according to keywords, industries and time ranges, and filters the behavior data and transaction data collected by the data collection module according to the limiting conditions when querying;

[0041] The behavior analysis module extracts the information content of the user behavior defined in the behavior query module, marks the frequency of occurrence of the content, establishes a behavior data relevance weight value, sorts the user behavior information after composite weighting according to the frequency and relevance weight value, and sends the sorting information to the behavior prediction module;

[0042] The behavior prediction module calls the sorting information, establishes a user O2O behavior map according to the sorting information, establishes a user portrait according to the user O2O behavior map, and stores the user portrait in the backend server;

[0043] When building a user portrait through the user O2O behavior graph, all data are vectorized using the embedding model, and the norm of the vector is uniformly scaled to 1 to obtain n vectors; an index containing the data and vectors is created, and the data and the uniformly scaled vector are saved to the index;

[0044] When the behavior query module performs a query, it first performs a gpu-accelerated K-nearest-neighbor query on the content and the standard vector of each data group. If the query result is greater than the set threshold, it is considered similar, and the most similar i similar groups are selected. The gpu-accelerated K-nearest-neighbor query is performed on the content to be queried and all the data in the similar groups. If it is greater than the set threshold, it is concluded that there are j most matching data in i similar groups;

[0045] The best matching data is imported into the behavior prediction module, and a probability table is formed according to the probability of subsequent user behavior of the best matching data, and the probability table is sent to the result display module; after receiving the probability table, the result display module synchronously forms pie charts, bar charts and curve charts according to the probability table, and visualizes the pie charts, bar charts and curve charts through the display panel, and regularly captures the pie charts, bar charts and curve charts and uploads them to the backend server, and associates the captured charts with the user portraits.

[0046] The working principle and workflow of the system for matching and accelerating the generation of user portraits in this application document are as follows: Data collection module: This module is responsible for collecting the behavior data and transaction data of TOB users. The behavior data includes the user's browsing, searching and purchasing behavior, while the transaction data includes the purchase record and transaction amount.

[0047] Behavior query module: The data collected by the data collection module is sorted according to keywords, industries and time ranges. The behavior query module filters data according to limited conditions when querying, including behavior data and transaction data.

[0048] Behavior analysis module: This module extracts the user behavior information defined in the query module and marks its frequency of occurrence. Then it establishes the behavior data relevance weight value, performs compound weighted sorting according to the frequency and relevance weight value, and sends the sorting information to the behavior prediction module.

[0049] Behavior prediction module: Receives the ranking information sent by the behavior analysis module, builds the user O2O behavior map based on the ranking information, and then builds the user portrait, and stores the user portrait in the background server. When building the user portrait, the embedding model is used for vectorization processing, and the norm of the data vector is uniformly scaled to 1 to obtain n vectors.

[0050] The behavior query module queries again: When the behavior query module queries, it first performs GPU-accelerated K-nearest neighbor queries on the content and the standard vectors of each data group. If the query result is greater than the set threshold, it is considered similar, and the most similar i similar groups are selected. Then the content to be queried and all the data in the similar groups are subjected to GPU-accelerated K-nearest neighbor queries. If it is greater than the set threshold, it is concluded that there are j most matching data in i similar groups.

[0051] Result display module: Receives the most matching data sent by the behavior prediction module, and forms a probability table based on the probability of subsequent user behavior based on the most matching data. The result display module synchronously forms pie charts, bar charts, and curve charts based on the probability table, and visualizes these charts through the display panel. The charts are periodically captured and uploaded to the backend server to be associated with the user portrait.

[0052] By introducing technologies such as GPU-accelerated K-nearest neighbor query and embedding model vectorization, the data processing speed and efficiency are improved, enabling the system to complete user behavior analysis and prediction more quickly. Through embedding model vectorization processing, users' interests and preferences can be captured more accurately, thereby improving the accuracy of personalized recommendations. The introduction of a behavior prediction module can more accurately predict user behavior trends by analyzing user historical behavior and real-time dynamic information, thereby enhancing the ability to understand and predict user behavior.

[0053] The operation of vectorizing all data using the embedding model and uniformly scaling it to unit norm plays an important role in the system. The main purpose of this step is to convert the original user behavior data into a vector representation for subsequent calculation and analysis. Specifically, the embedding model is used to map each user's behavior data into a high-dimensional vector space, where each dimension represents a specific behavior or feature. Through this vectorized representation, the similarity and correlation between user behaviors can be better captured, thereby achieving more accurate user portrait generation and behavior prediction.

[0054] The process of scaling vectors to unit norm is to unify the scale of data so that the values ​​in different dimensions have the same importance. The advantage of doing so is that it can eliminate the impact of scale differences between different dimensions on the calculation results and ensure that the contribution of each dimension to the final result is balanced. At the same time, scaling vectors to unit norm can also simplify the complexity of subsequent calculations, speed up calculations, and improve the stability and generalization ability of the model.

[0055] Creating an index containing data and vectors and saving the scaled vectors to the index is to facilitate subsequent similarity queries and matching operations. By storing vectors in the index, the data items most similar to a given query vector can be quickly retrieved, thereby achieving efficient K-nearest neighbor queries. This index structure can greatly reduce query time and save storage space, enabling the system to process large-scale user behavior data sets.

[0056] The query speed and response time of the system are further accelerated by performing GPU-accelerated K-nearest neighbor queries on the content and the standard vectors of each data group. Taking advantage of GPU parallel computing, multiple query requests can be processed simultaneously, thereby achieving efficient similarity calculation and matching operations. K-nearest neighbor query is a commonly used similarity search algorithm that calculates the distance between the query vector and all vectors in the database, and then selects the K closest vectors as the query result. GPU acceleration can significantly increase the calculation speed of this process, allowing the system to find data items similar to user behavior more quickly, thereby improving the accuracy and personalization of recommendations.

[0057] In summary, using the embedding model to vectorize the data and uniformly scale it to the unit norm, as well as using GPU acceleration for K-nearest neighbor queries, are all aimed at improving the performance and efficiency of the system, thereby achieving fast and accurate user behavior analysis and personalized recommendations. The introduction and application of these technical means enable the system to better understand the user's behavioral characteristics, more accurately capture the user's interests and preferences, and provide users with a better quality and more personalized service experience.

[0058] When the data and norm-scaled vectors are saved to the index, the cosine similarity between the two vectors is calculated in parallel using the GPU. If the similarity is greater than the set threshold, the corresponding two vectors are considered to be related; all vectors are plotted, and the graph is compressed using the Louvain algorithm to obtain m data groups, and the average vector of all data in each data group is calculated, and the average vector is used as the standard vector of the group. By using the GPU to calculate the cosine similarity of vectors in parallel, compressing the graph using the Louvain algorithm, and calculating the average vector in the data group as the standard vector, an innovative strategy is presented for the processing and representation of vector data. Compared with the existing technology, this method has significant differences and advantages in terms of computational efficiency, accuracy of data representation, and intelligent data processing.

[0059] Traditional vector similarity calculation usually uses CPU for serial calculation, while this method greatly improves the calculation efficiency by using GPU to calculate cosine similarity in parallel. Compared with CPU, GPU has more parallel processing units, which is particularly suitable for processing large-scale vector calculation tasks. Through GPU parallel calculation, more vectors can be processed in the same time, improving the calculation efficiency, which is particularly suitable for big data processing. This method improves the accuracy of data representation by calculating the average vector of all data in each data group as the standard vector of the group. Traditional methods may directly select a certain data as the standard, which may ignore the information of other data in the group, resulting in a decrease in the accuracy of representation. However, this method makes full use of the information of all data in the group by calculating the average vector, thereby improving the accuracy of data representation. This method uses the Louvain algorithm to compress the graph, obtains m data groups, and intelligently processes a large amount of vector data. The Louvain algorithm is a community discovery algorithm that can effectively identify and divide the community structure in the data. By using the Louvain algorithm, this method can intelligently process and organize a large amount of vector data, improving the efficiency and effect of data processing.

[0060] When establishing the user O2O behavior map, the behavior prediction module reconstructs the sorting information data model through at least one of time series analysis, regression analysis, classification algorithm, clustering algorithm or deep learning algorithm, and refines the user behavior data through data reconstruction to improve the prediction accuracy of the user O2O behavior map. In this process, the behavior prediction module reconstructs the sorting information data model through at least one of time series analysis, regression analysis, classification algorithm, clustering algorithm or deep learning algorithm. The time series analysis is used to analyze a series of data points observed over a period of time. This can help predict future behaviors and trends. The regression analysis is used to estimate the relationship between two or more variables. This can help understand which factors affect the user's behavior. The classification algorithm is used to predict the category label of the data. This can help predict the behavior that the user may take. The clustering algorithm is used to group the data. This can help understand the user's behavior pattern. The deep learning algorithm can process a large amount of complex data. This can help improve the accuracy of the prediction. Through these methods, the behavior prediction module can refine the user behavior data and improve the accuracy of the user O2O behavior map prediction. In this way, enterprises can more accurately predict the future behavior of users and formulate more effective strategies.

[0061] When extracting user behavior information, the behavior analysis module divides the user types according to the user behavior information to form different categories of user behavior information, and makes a major distinction when establishing a user portrait based on the user's behavior characteristics and consumption characteristics. When processing user behavior information, the behavior analysis module will classify users according to the user's behavior information to form different categories of user behavior information. This classification may be based on the user's browsing behavior, search behavior, purchase behavior, etc. For example, some users may often browse a certain type of product, so they can be classified as users interested in this product; other users may often purchase in a certain time period, so they can be classified as users active in this time period. When establishing a user portrait, the behavior analysis module will make a major distinction based on the user's behavior characteristics and consumption characteristics. This is a higher-level classification, which may be based on the user's value, potential, needs, etc. For example, some users may buy a lot of high-value products, so they can be classified as high-value users; other users may often browse but rarely purchase, so they can be classified as potential users.

[0062] Such classification and distinction can not only help enterprises better understand users, but also help them formulate more targeted marketing strategies. For example, for high-value users, enterprises can provide more discounts and services; for potential users, enterprises can provide more guidance and incentives.

[0063] When the behavior prediction module calls the sorting information, it generates a user's behavior event sequence based on the sorting information. Each event represents a behavior of the user at a certain point in time. The user's behavior event sequence is analyzed by the sequence analysis algorithm to establish the law of user behavior, and the user's O2O behavior map is established based on the law of user behavior. The behavior prediction module first calls the sorting information provided by the behavior analysis module. The sorting information may contain indicators such as the frequency and correlation of user behavior. This information helps determine which behaviors are the most typical and most likely to be repeated by the user. The behavior prediction module uses the sorting information to generate the user's behavior event sequence. This sequence is arranged in chronological order and records the user's behavior at various time points, such as searching for a keyword, browsing a product, or making a purchase at a certain time point. This serialized data format provides a time dimension for analyzing user behavior and helps to understand the evolution of user behavior. Next, the module uses a sequence analysis algorithm to analyze the user's behavior event sequence. The sequence analysis algorithm may include methods such as time series analysis, Markov chain, and sequence pattern mining. Its purpose is to identify laws and patterns from time series data. Through the application of sequence analysis algorithms, the system can identify regularities in user behavior, such as users may make specific searches or purchases at certain times of the week, or typically view related products after viewing a certain product.

[0064] According to the rules of user behavior, the system builds the user's O2O (online to offline) behavior map. The O2O behavior map is a visual representation of user behavior patterns, which shows the relationship and conversion path between users' online behaviors (such as online searches, browsing and purchases) and offline behaviors (such as in-store purchases and service use). Through this series of steps, the behavior prediction module can build a dynamic and personalized user behavior model based on the user's historical behavior data. This behavior model can help companies more accurately predict users' future behavior, thereby providing more personalized services and product recommendations, optimizing user experience, and improving marketing effectiveness.

[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A system for accelerating the generation of user portraits, characterized in that: It includes data collection module, behavior query module, behavior analysis module, behavior prediction module and result display module; The data collection module is used to collect the behavior data and transaction data of TOB end users, the behavior data includes the user's browsing behavior, search behavior and purchase behavior, and the transaction data includes the user's purchase record and transaction amount; The behavior query module sorts the data collected in the data collection module according to keywords, industries and time ranges, and filters the behavior data and transaction data collected by the data collection module according to the limiting conditions when querying; The behavior analysis module extracts the information content of the user behavior defined in the behavior query module, marks the frequency of occurrence of the content, establishes a behavior data relevance weight value, sorts the user behavior information after composite weighting according to the frequency and relevance weight value, and sends the sorting information to the behavior prediction module; The behavior prediction module calls the sorting information, establishes a user O2O behavior map according to the sorting information, establishes a user portrait according to the user O2O behavior map, and stores the user portrait in the backend server; When building a user portrait through the user O2O behavior graph, all data are vectorized using the embedding model, and the norm of the vector is uniformly scaled to 1 to obtain n vectors; an index containing the data and vectors is created, and the data and the uniformly scaled vector are saved to the index; When the behavior query module performs a query, it performs a gpu-accelerated K-nearest-neighbor query on the content and the standard vector of each data group. If the query result is greater than the set threshold, it is considered similar, and the most similar i similar groups are selected. The gpu-accelerated K-nearest-neighbor query is performed on the content to be queried and all the data in the similar groups. If it is greater than the set threshold, it is concluded that there are j most matching data in i similar groups; Import the best matching data into the behavior prediction module, form a probability occurrence table based on the probability of subsequent behavior of the user of the best matching data, and send the probability occurrence table to the result display module; After receiving the probability table, the result display module synchronously forms a pie chart, a bar chart and a curve chart according to the probability table, and visualizes the pie chart, the bar chart and the curve chart through the display panel, and regularly captures the pie chart, the bar chart and the curve chart and uploads them to the backend server, and associates the captured chart with the user portrait; when the vector after the data and norm are uniformly scaled is saved to the index, the cosine similarity between the two vectors is calculated in parallel by using the GPU, and if the similarity is greater than the set threshold, the corresponding two vectors are considered to be associated; all vectors are plotted, and the graph is compressed using the Louvain algorithm to obtain m data groups, and the average vector of all data in each data group is calculated, and the average vector is used as the standard vector of the group; when establishing the user O2O behavior map, the behavior prediction module reconstructs the sorting information data model through at least one of time series analysis, regression analysis, classification algorithm, clustering algorithm or deep learning algorithm, and refines the user behavior data through data reconstruction to improve the prediction accuracy of the user O2O behavior map.

2. A system for accelerating the generation of user portraits according to claim 1, characterized in that: When extracting user behavior information, the behavior analysis module divides user types according to the user behavior information to form different categories of user behavior information, and makes major classifications when establishing user portraits based on the user's behavior characteristics and consumption characteristics.

3. A system for accelerating the generation of user portraits according to claim 1, characterized in that: When the behavior prediction module calls the sorting information, a user behavior event sequence is generated based on the sorting information. Each event represents a behavior of the user at a certain point in time. The user behavior event sequence is analyzed through a sequence analysis algorithm to establish the rules of user behavior. Based on the rules of user behavior, a user O2O behavior map is established.

Images