Large Model Information Push Method and System Based on Industry Corpus

Abstract

The present invention relates to the field of machine learning technology, and discloses a large model information push method and system based on industry corpus. The method first obtains the positive and negative behavior record sequences of the target user, extracts behavior features and generates behavior weights to construct a multi-dimensional user portrait vector; then obtains a semantic embedding vector based on a pre-trained industry language model, generates positive and negative preference embedding vectors through similarity threshold adjustment, and obtains an adjusted push content set accordingly; finally, intelligently sorts the candidate push content to generate a precise push content sequence. The present invention realizes precise and dynamic personalized recommendation by accurately capturing the immediate needs and emotional fluctuations of users, effectively improving the user experience and reducing information fatigue.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and more specifically, to a large model information push method and system based on industry corpora. Background Art

[0002] With the popularization of the Internet and mobile devices, the application of information push technology in various industries has received extensive attention. Especially in industries such as finance and e-commerce, personalized push technology has become an important means to enhance user experience, optimize user services, and improve business conversion rates.

[0003] The patent application with the publication number CN116489216A discloses an information push method, device, electronic device, storage medium, and program product, which determines whether to push a specific product based on the customer's preference information and push permission information. This method decides whether to push a product to a target user based on the user's first preference information and push permission information. Although this method can reduce the blindness of pushing to a certain extent and avoid the push of irrelevant information, there are still limitations: First, this method relies on preset push permission information and cannot dynamically adjust the recommended content according to the user's real-time behavior or emotional fluctuations; Second, the determination of product push is limited to the matching with the customer's preference information, lacking a comprehensive analysis of more dimensions such as the user's emotion and behavior state. Therefore, how to more accurately capture the dynamic needs of users and provide real-time personalized push through more fine-grained sentiment analysis and semantic processing remains an unsolved problem.

[0004] The patent application with the publication number CN110460674A discloses an information push method, device, and system, which ensures the security and integrity of information by signing and verifying the push information. This method uses a push gateway to generate signature information and verifies the security of information transmission through a public key and private key mechanism. Although this method is of great significance in the security of information transmission, it mainly focuses on the security issues during the push process and does not involve how to improve the personalization and accuracy of push content. This method fails to make full use of information such as the user's behavior data and sentiment analysis to optimize push content, nor does it consider how to improve the relevance of push content through intelligent algorithms. Therefore, although the security of information transmission is guaranteed, the quality and personalization of push content are still the shortcomings of the current technology.

[0005] The push methods in the prior art often limit to static data analysis or simple rule-based determination, lacking dynamic response to the user's real-time behavior and emotional fluctuations, resulting in problems such as over-simplification and singleness of push content. Summary of the Invention

[0006] To overcome the above-mentioned defects of the prior art, the present invention provides a large model information push method and system based on industry corpora, which deeply excavates user behavior preferences, constructs a domain language model by integrating industry corpora, and dynamically optimizes the pushed content to improve the accuracy and personalization of information push.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A large model information push method based on industry corpora, comprising:

[0009] Obtaining a positive behavior record sequence and a negative behavior record sequence of a target user; extracting first behavior features from the positive behavior record sequence and second behavior features from the negative behavior record sequence; generating a first behavior weight for the first behavior features and a second behavior weight for the second behavior features; generating a multi-dimensional user portrait vector according to the first behavior features, the first behavior weight, the second behavior features and the second behavior weight;

[0010] Constructing a pre-trained industry language model, obtaining a first semantic embedding vector and a second semantic embedding vector according to the multi-dimensional user portrait vector and the pre-trained industry language model; adjusting the first semantic embedding vector and the second semantic embedding vector according to a preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; obtaining an adjusted push content set according to the positive preference embedding vector and the negative preference embedding vector;

[0011] Obtaining a third candidate push content set based on the adjusted push content set; performing intelligent sorting on the content in the third candidate push content set to generate an accurate push content sequence.

[0012] Further, the obtaining the positive behavior record sequence and the negative behavior record sequence of the target user includes:

[0013] Constructing an industry corpus, and obtaining first historical behavior data of the target user in the industry field from the industry corpus; the first historical behavior data includes n1 behavior types of the target user, and the n1 behavior types include positive behaviors and negative behaviors; n1 is a positive integer;

[0014] Obtaining a positive behavior record sequence and a negative behavior record sequence according to the first historical behavior data.

[0015] Further, the extracting the first behavior features from the positive behavior record sequence includes:

[0016] Performing word segmentation on the text content in the positive behavior record sequence to segment the text content into a word sequence of the smallest semantic units;

[0017] Performing part-of-speech tagging on the segmented word sequence;

[0018] Based on part-of-speech tagging, identify the named entities in the text content;

[0019] Extract keywords and topic words from the word sequence;

[0020] Use keywords, topic words, and named entity words as the first row as features.

[0021] Furthermore, the first-row weights for generating the first-row features and the second-row weights for generating the second-row features include:

[0022] Set the basic weights for different behavior types in n1 behavior types, and construct an initial behavior weight matrix;

[0023] According to the first historical behavior data, statistically analyze the behavior intensity distribution of different behavior types;

[0024] Construct a behavior intensity quantization function, and according to the behavior intensity quantization function, map the behavior intensity distribution of different behavior types to the corresponding weight adjustment coefficients;

[0025] Multiply the basic weights in the initial behavior weight matrix by the corresponding weight adjustment coefficients to generate an adjusted behavior weight matrix;

[0026] In the adjusted behavior weight matrix, use the weights corresponding to positive behaviors as the first-row weights for the first-row features, and use the weights corresponding to negative behaviors as the second-row weights for the second-row features.

[0027] Furthermore, the generation of the multi-dimensional user portrait vector includes:

[0028] Construct a positive preference dimension according to the first-row features and the first-row weights;

[0029] Construct a negative preference dimension according to the second-row features and the second-row weights;

[0030] Fuse the positive preference dimension and the negative preference dimension to generate a multi-dimensional user portrait vector.

[0031] Furthermore, the generation of the positive preference embedding vector and the negative preference embedding vector includes:

[0032] Calculate the similarity between the first semantic embedding vector and the second semantic embedding vector to obtain the positive and negative preference similarity score D1;

[0033] If the positive and negative preference similarity score D1 is greater than θ1, trigger the embedding vector adjustment process to obtain the positive preference embedding vector and the negative preference embedding vector.

[0034] Furthermore, the embedding vector adjustment process includes:

[0035] Step S2221: Randomly sample a noise vector, and adjust the first semantic embedding vector according to the noise vector to obtain an adjusted first semantic embedding vector, which is marked as a positive preference embedding vector;

[0036] Step S2222: Adjust the second semantic embedding vector according to the noise vector to obtain an adjusted second semantic embedding vector, which is marked as a negative preference embedding vector;

[0037] Step S2223: Calculate the similarity D2 between the positive preference embedding vector and the negative preference embedding vector, and determine whether D2 is less than or equal to θ1. If so, complete the embedding vector adjustment process; otherwise, return to Step S2221, randomly sample the noise vector again, and repeat Steps S2221 - S2223 until D2 ≤ θ1.

[0038] Furthermore, the obtaining of the adjusted push content set includes:

[0039] Perform semantic similarity retrieval on the text content in the industry corpus using the positive preference embedding vector to generate a first candidate push content set;

[0040] Perform semantic similarity retrieval on the text content in the industry corpus using the negative preference embedding vector to generate a second candidate push content set;

[0041] Based on the first candidate push content set and the second candidate push content set, obtain the adjusted push content set.

[0042] The generating of the first candidate push content set includes:

[0043] Convert the text content in the industry corpus into corresponding semantic embedding vectors;

[0044] Calculate the similarity between the positive preference embedding vector in the multi-dimensional user profile vector and the semantic embedding vector of the text content in the industry corpus to obtain a first similarity score;

[0045] Select the top K text contents with the highest first similarity scores in the industry corpus to generate a first candidate push content set.

[0046] Furthermore, the obtaining of the adjusted push content set based on the first candidate push content set and the second candidate push content set includes:

[0047] Obtain the semantic embedding vectors of the text contents in the first candidate push content set and the second candidate push content set;

[0048] According to the semantic embedding vectors of the text contents in the first candidate push content set and the second candidate push content set, calculate the semantic similarity S2 between each text content in the first candidate push content set and all text contents in the second candidate push content set, and form a similarity matrix;

[0049] In the similarity matrix, screen out the content pairs with a semantic similarity S2 higher than the second similarity threshold θ2, and mark them as highly overlapping contents;

[0050] Remove the marked highly overlapping contents from the first candidate push content set to obtain an adjusted push content set.

[0051] An information push system based on a large model of industry corpus, which is used to implement the above-mentioned information push method based on a large model of industry corpus. The system includes:

[0052] Portrait construction module: used to obtain the positive behavior record sequence and negative behavior record sequence of the target user; extract the first behavior feature from the positive behavior record sequence, and extract the second behavior feature from the negative behavior record sequence; generate the first behavior weight of the first behavior feature and the second behavior weight of the second behavior feature; generate a multi-dimensional user portrait vector according to the first behavior feature, the first behavior weight, the second behavior feature and the second behavior weight;

[0053] Push content generation module: used to construct a pre-trained industry language model, and obtain the first semantic embedding vector and the second semantic embedding vector according to the multi-dimensional user portrait vector and the pre-trained industry language model; adjust the first semantic embedding vector and the second semantic embedding vector according to the preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; obtain an adjusted push content set according to the positive preference embedding vector and the negative preference embedding vector;

[0054] Push content sorting module: based on the adjusted push content set, obtain a third candidate push content set; perform intelligent sorting on the contents in the third candidate push content set to generate an accurate push content sequence.

[0055] An electronic device includes a memory, a central processing unit, and a computer program stored on the memory and executable on the central processing unit. When the central processing unit executes the computer program, it implements the above-mentioned information push method based on a large model of industry corpus.

[0056] A computer-readable storage medium stores a computer program, and when the computer program is executed, it implements the above-mentioned information push method based on a large model of industry corpus.

[0057] Compared with the prior art, the beneficial effects of the present invention are:

[0058] By integrating the user's real-time behavior data, sentiment analysis, and semantic embedding technology, the present invention can capture the changing needs of users in real time, dynamically adjust the recommended content, and ensure that each push highly matches the user's immediate interests and emotional state. This process avoids the static nature of traditional push systems and the problem of over-reliance on historical data, enabling the recommended content to not only conform to the user's long-term interests but also quickly respond to the changing current needs of users, thus greatly enhancing the relevance and accuracy of the recommendation.

[0059] In addition, by adopting semantic similarity calculation and deduplication mechanisms, the present invention effectively avoids information redundancy and information fatigue, ensuring that the pushed content is both rich and diverse without repetition. By optimizing the push strategy through reinforcement learning, the pre-trained industry language model can continuously adjust the ranking of the recommended content based on the user's real-time feedback, making the pushed content not only conform to the user's interests but also provide the most suitable content at the best time, further improving the user's engagement and conversion rate. The introduction of reinforcement learning enables the push strategy to self-optimize, adapt to market changes and the dynamic changes in user behavior, thereby continuously enhancing the recommendation effect and user satisfaction.

[0060] Through multi-dimensional data analysis and intelligent optimization, the present invention not only improves the quality and personalization of the pushed content but also enhances the ability to respond to market and user demand changes, capable of providing more accurate, efficient, and personalized services. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0062] Figure 1 It is the principle flowchart of the large model information push method based on industry corpus in the present invention;

[0063] Figure 2 It is the method flowchart for extracting the first behavior feature from the positive behavior record sequence in the large model information push method based on industry corpus in the present invention;

[0064] Figure 3 It is the method flowchart for generating the first behavior weight of the first behavior feature and the second behavior weight of the second behavior feature in the large model information push method based on industry corpus in the present invention;

[0065] Figure 4 It is the method flowchart for generating the first candidate push content set in the large model information push method based on industry corpus in the present invention;

[0066] Figure 5 This is the flowchart of the method for obtaining the adjusted push content set in the large model information push method based on industry corpus of the present invention;

[0067] Figure 6 This is the functional module diagram of the large model information push system based on industry corpus in the present invention. Detailed implementation manners

[0068] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0069] Embodiment 1

[0070] Please refer to Figure 1 As shown, this embodiment provides a large model information push method based on industry corpus, including:

[0071] Step S1000, obtaining the positive behavior record sequence and negative behavior record sequence of the target user; extracting the first behavior feature from the positive behavior record sequence, and extracting the second behavior feature from the negative behavior record sequence; generating the first behavior weight of the first behavior feature and the second behavior weight of the second behavior feature; generating a multi-dimensional user portrait vector according to the first behavior feature, the first behavior weight, the second behavior feature and the second behavior weight;

[0072] Further, step S1000 includes:

[0073] Step S1100, obtaining the positive behavior record sequence and negative behavior record sequence of the target user;

[0074] Further, step S1100 includes:

[0075] Step S1110, constructing an industry corpus, and obtaining the first historical behavior data of the target user in the industry field from the industry corpus; the first historical behavior data includes n1 behavior types of the target user, and the n1 behavior types include positive behaviors and negative behaviors; n1 is a positive integer;

[0076] Step S1120, obtaining the positive behavior record sequence and negative behavior record sequence according to the first historical behavior data.

[0077] Specifically, the purpose of step S1100 is to identify and extract positive and negative behavior records by analyzing the historical behavior of the target user. These behavior records will provide data support for subsequent product push and service recommendation, whether it is to push real-time industry analysis and exclusive financial products for investors in the financial field or to recommend suitable products for customers in other consumer fields. In this way, more accurate personalized services can be achieved.

[0078] An industry corpus refers to a large dataset of user behaviors collected in a certain industry (such as finance, consumption, healthcare, etc.), which can represent the behavior characteristics and preferences of most users in the industry. This corpus is the basic data source for analysis, recommendation, and prediction. Constructing an industry corpus refers to the process of collecting and organizing all relevant user behavior data within a certain industry (such as the financial field, the consumer field, etc.). The industry corpus not only contains the specific behavior records of the target user but also includes the behavior data of all users in the industry, having broad representativeness. By analyzing the industry corpus, the specific behavior patterns of the target user can be mined, and then suitable products or services can be recommended for them.

[0079] In the financial field, the industry corpus includes the following data:

[0080] Transaction behavior data: including investors' stock trading, fund purchases, bond trading, etc.

[0081] History of financial product purchases: financial products, insurance products, etc. purchased by users in the past.

[0082] Investment preferences: investors' risk tolerance, preferred investment fields, etc.

[0083] Industry information interaction: industry analysis reports, investment strategy articles, etc. read and interacted with by users.

[0084] In other consumer fields, the industry corpus includes:

[0085] Shopping behavior data: types of goods purchased by users, purchase frequency, consumption amount, etc.

[0086] Search and browsing records: users' search and browsing records of goods or services, as well as their interest in specific products.

[0087] Social interaction records: interaction records of users on social platforms or with brand customer service, such as likes, comments, consultations, etc.

[0088] The first historical behavior data refers to the first-hand behavior data of the target user within a certain industry field, usually a detailed record of all the user's behaviors in the past period of time. It includes the user's interaction records with the industry, reflecting the user's preferences, needs, and potential interests.

[0089] By constructing an industry corpus, service providers in the financial or consumer fields can comprehensively understand the historical behaviors of target users, and then identify users' preferences, needs, and pain points. In the financial field, this helps to accurately recommend financial products and investment opportunities based on users' investment history and preferences; in the consumer field, it can recommend suitable goods or services to customers, avoid over-promotion, and improve the accuracy of recommendations and user satisfaction.

[0090] For example, assume that user A has purchased equity funds and bond funds through a bank in the past year and has regular monthly deposits. These behavior records are marked as "conservative investors" in the industry corpus. Similarly, user B has frequently engaged in high-risk stock trading and often adjusted the investment portfolio in the past year, and these behaviors are marked as "aggressive investors". Based on these labels, low-risk financial products can be recommended for user A, and high-risk equity funds can be recommended for user B.

[0091] Step S1120 divides positive and negative behavior records based on the first historical behavior data. The purpose of this process is to classify users' behaviors to accurately judge which behaviors meet the needs of target users and which behaviors may bring risks or adverse effects. Positive behaviors are usually those that demonstrate users' good credit, preferences, or needs, while negative behaviors are those related to users' disadvantages or risks.

[0092] Positive behavior refers to user behavior that meets expectations or is beneficial. For example, in the financial field, positive behaviors include users' continuous deposits, on-time repayments, and investment in financial products with moderate risks; in the consumer field, positive behaviors include users' frequent purchases of healthy foods, high loyalty, and good product evaluations.

[0093] Negative behavior refers to user behavior that is contrary to expectations or may trigger risks. For example, in the financial field, negative behaviors include users' overdue repayments, excessive high-risk investments, and credit card debts; in the consumer field, negative behaviors include users' frequent returns, poor product ratings, and frequent complaint records.

[0094] By identifying and recording positive and negative behaviors, it can help accurately classify and evaluate users, and then provide personalized services for them. For users in the financial field, identifying positive behaviors can help recommend financial products suitable for their risk tolerance, while negative behaviors can prompt the system to warn of potential default or high-risk investment behaviors. In the consumer field, positive behaviors can help recommend goods or services that match users' needs, while negative behaviors can help the system avoid recommending too many goods that may cause user dissatisfaction.

[0095] Step S1200: Extract the first behavioral feature from the positive behavioral record sequence and the second behavioral feature from the negative behavioral record sequence; generate the first behavioral weight for the first behavioral feature and the second behavioral weight for the second behavioral feature.

[0096] Further, step S1200 includes:

[0097] Step S1210: Extract the first behavioral feature from the positive behavioral record sequence.

[0098] The purpose of step S1210 is to extract useful features from the positive behavioral record sequence to provide data support for subsequent intelligent marketing decisions. The process of feature extraction involves in-depth analysis of the text content of user behavior, extracting keywords, topic words, and named entities, thereby establishing the first behavioral feature that can reflect user needs, interests, and preferences. These features will be used to recommend personalized financial products or services to users.

[0099] Further, as Figure 2 shown, step S1210 includes:

[0100] Step S1211: Perform word segmentation on the text content in the positive behavioral record sequence to segment the text content into a sequence of words as the smallest semantic units.

[0101] Specifically, first perform word segmentation on the text content in the positive behavioral record sequence. Word segmentation refers to dividing continuous text into independent smallest semantic units, that is, words or phrases. Since words in natural language can have different combinations and meanings, the process of word segmentation can extract the independent meaning of each word or phrase in the context. This is the first step in processing user behavior data and analyzing their intentions. Word segmentation is a technique in text processing that divides coherent sentences in the text into lexical units. The quality of word segmentation directly affects subsequent semantic analysis and information extraction. Semantic units in language processing refer to the smallest units with independent meanings, which are usually a word or a phrase in Chinese.

[0102] Key steps of word segmentation:

[0103] Text cleaning: Clean up punctuation marks, meaningless characters (such as carriage returns, line breaks), etc.

[0104] Segmentation: Divide the sentences in the text into individual words according to spaces, punctuation, etc.

[0105] Standardization: Normalize some synonyms or variant words into a unified form.

[0106] For example, user A enters "I want to know the differences between stock funds and bond funds" in a financial management APP. During the word segmentation process, this text is divided into the following sequence of words: "I / want / to know / stock funds / and / bond funds / of / differences".

[0107] Word segmentation is a basic step in natural language processing. It converts the user's natural language behavior data into a format that machines can understand and analyze. By splitting the user's text content into individual sequences of words, it is possible to more precisely analyze the user's needs and intentions, laying a foundation for subsequent feature extraction and push systems.

[0108] Step S1212, perform part-of-speech tagging on the segmented sequence of words;

[0109] Specifically, when performing part-of-speech tagging on the segmented sequence of words, the goal is to assign a part-of-speech tag (such as noun, verb, adjective, etc.) to each word. Part-of-speech tagging helps to understand the grammatical role and meaning of words in a sentence, thus enabling more accurate information extraction and semantic analysis. Part-of-speech tags are identifiers assigned to each word, indicating the grammatical role of the word in the sentence, such as noun (NN), verb (VB), adjective (JJ), etc.

[0110] For example, in the sentence "I want to know the differences between stock funds and bond funds":

[0111] "I" is a pronoun (PRP);

[0112] "want" is a verb (VB);

[0113] "know" is a verb (VB);

[0114] "stock funds" is a noun phrase (NP);

[0115] "and" is a conjunction (CC);

[0116] "bond funds" is a noun phrase (NP);

[0117] "of" is a particle (PART);

[0118] "differences" is a noun (NN);

[0119] Part-of-speech tagging can help accurately understand the role of each word in a sentence, facilitating the extraction of keywords and themes in subsequent steps. Without part-of-speech tagging, it may not be possible to accurately distinguish whether "fund" in "bond funds" is a noun or a word with other meanings. Through part-of-speech tagging, it is possible to identify key nouns (such as types of financial products) and verbs (such as investment, purchase, etc.) from the text, laying a foundation for precise recommendation of financial products.

[0120] For example, user C inputs: "I am interested in low-risk funds". After part-of-speech tagging:

[0121] "I" is a pronoun (PRP);

[0122] "To" is a preposition (IN);

[0123] "Low-risk" is an adjective (JJ);

[0124] "Fund" is a noun (NN);

[0125] "Be interested in" is a verb phrase (VP);

[0126] Through these tags, the system can further extract keywords related to "fund" and "low-risk".

[0127] Step S1213, based on part-of-speech tagging, identify the named entities in the text content;

[0128] Specifically, based on the part-of-speech tagging results, identify the named entities in the text. Named entities usually refer to specific and independently meaningful entities, such as company names, place names, dates, product names, etc. In the financial field, the identification of named entities can help the system identify key information such as the names of investment products and company names.

[0129] The identification of named entities helps to accurately extract the key information in the text. Especially in the financial field, identifying specific entities such as funds, stocks, and bank names can help the system better understand the user's needs and provide more accurate product recommendations.

[0130] Step S1214, extract keywords and topic words from the word sequence;

[0131] Specifically, extract the keywords and topic words related to the user's needs from the word sequence. Keywords are the core words that can represent the user's needs. For example, in the investment field, keywords may include "stocks", "funds", "bonds", etc. Topic words are the main topics or directions refined from the entire behavior record, and are words or phrases that can summarize the user's need theme, such as "low-risk wealth management products", "stock fund investment". By extracting keywords and topic words, the core intention of the user can be understood, and appropriate products can be recommended based on this.

[0132] Extracting keywords and subject terms helps to accurately identify the user's main needs and provide personalized recommendations based on this information. If the user frequently mentions "bond funds" and "low risk", it can be inferred that the user prefers low-risk investment products, and relevant wealth management products can be recommended accordingly. This not only improves the accuracy of recommendations but also greatly enhances the user experience. For example, in the behavior records of user E on the financial platform, the terms "low-risk funds", "bond funds", and "long-term investment" are often entered. These words will be extracted as keywords and subject terms, and based on this information, low-risk bond funds with long-term holding can be recommended to the user.

[0133] Step S1215 takes the keywords, subject terms, and named entity terms as the first behavioral features.

[0134] Specifically, step S1215 combines the keywords, subject terms, and named entity terms extracted in step S1214 as the first behavioral features. These features will be used to describe the user's investment preferences, needs, and interests, providing an important basis for subsequent personalized recommendations and intelligent marketing decisions. By taking the keywords, subject terms, and named entity terms as the first behavioral features, the user's core needs and preferences can be accurately captured. This feature processing enables subsequent intelligent marketing to more precisely meet the user's needs, improving the recommendation matching degree and user satisfaction. For example, assume that user F enters "low-risk bond funds", and the keywords extracted by the system are "low risk" and "bond funds". These words, as the first feature of the user's behavior, can be used to push low-risk bond fund products that match their preferences.

[0135] Step S1220 extracts the second behavioral features from the negative behavior record sequence; the second behavioral features include irrelevant words, negative sentiment words, and ignored entity words.

[0136] Specifically, in the financial field, especially in the banking industry, the user's negative behavior records can reveal the potential risks of user dissatisfaction, complaints, and churn. Therefore, step S1220 extracts the second behavioral features from the negative behavior record sequence, aiming to accurately construct the user's negative profile by analyzing the user's dissatisfaction and negative behaviors, helping to predict user churn, optimize the service experience, and improve user relationship management.

[0137] Negative behavior records usually include user complaints, negative feedback, bad reviews, service cancellations, or unsatisfactory interactions, etc. In this step, the extracted second behavioral features include the following three types:

[0138] Irrelevant words: In negative behavior records, users may mention content that has nothing to do with the business itself. Irrelevant words can be descriptions of situations, backgrounds, or unrelated information that are not related to the service. For example, a user may mention "It rained heavily yesterday" or "There was a serious traffic jam on the road" when making a complaint. Although these words appear in the record, they are not relevant to the analysis of negative behavior and therefore need to be filtered out to avoid interfering with sentiment analysis or behavior prediction.

[0139] Negative sentiment words: Negative sentiment words are words that directly express negative emotions in users' feedback. For example, "disappointed", "angry", and "dissatisfied" are typical negative sentiment words. Extracting negative sentiment words helps to accurately capture users' emotional states and quantify their attitudes. By identifying negative sentiment words, banks can effectively evaluate users' emotional fluctuations and thus take corresponding measures in advance to avoid user churn.

[0140] Ignored entity words: In some cases, users may mention specific services or products, but these entities do not play a dominant role in users' dissatisfaction. Ignored entity words usually refer to specific service or product entities mentioned in negative feedback that are not related to users' emotions. For example, a user may mention "The application is very bad" but does not clearly explain why the application is bad, or only mentions the specific product name when complaining. In this case, these entity information should be ignored because they do not directly affect users' negative emotions.

[0141] By accurately extracting irrelevant words, negative sentiment words, and ignored entity words, it is possible to remove irrelevant information when analyzing users' negative feedback, thus ensuring more accurate sentiment analysis and behavior prediction. This processing helps to provide an effective basis for user relationship management, reduce the risk of user churn, and make timely optimization decisions.

[0142] Step S1230, generate the first behavior weight of the first behavior feature and the second behavior weight of the second behavior feature.

[0143] Further, as Figure 3 shown, step S1230 includes:

[0144] Step S1231, set the basic weights of different behavior types in n1 behavior types and construct an initial behavior weight matrix;

[0145] Step S1232, according to the first historical behavior data, count the behavior intensity distributions of different behavior types;

[0146] Step S1233, construct a behavior intensity quantization function, and according to the behavior intensity quantization function, map the behavior intensity distributions of different behavior types to the corresponding weight adjustment coefficients;

[0147] Step S1234: Multiply the basic weights in the initial behavior weight matrix by the corresponding weight adjustment coefficients to generate an adjusted behavior weight matrix.

[0148] Step S1235: In the adjusted behavior weight matrix, use the weights corresponding to positive behaviors as the first behavior weights of the first behavior feature, and use the weights corresponding to negative behaviors as the second behavior weights of the second behavior feature.

[0149] Specifically, by generating an adjusted behavior weight matrix in step S1230, based on the user's behavior data, corresponding weights are assigned to the first behavior feature (positive behavior) and the second behavior feature (negative behavior). These weights will affect the personalized recommendation strategy, which is used to optimize the recommendation of financial services. For example, in the financial field such as banks, AI robots replace traditional customer service staff and salespersons for intelligent marketing.

[0150] The behavior types are classified according to the user's historical behavior data, usually including positive behaviors (such as purchasing financial products, repaying on time, etc.) and negative behaviors (such as overdue repayment, frequent consultations, etc.). The basic weights reflect the initial importance of each behavior type. In the initial stage, positive behaviors are usually given higher weights, while negative behaviors are given lower weights. For example: Positive behaviors (such as repaying on time, frequently purchasing low-risk products) may obtain higher basic weights (such as 0.8 or 1.0), indicating that these behaviors have a greater impact on subsequent recommendations. Negative behaviors (such as overdue repayment, frequent returns) may obtain lower basic weights (such as 0.2 or 0.3), indicating that these behaviors have a smaller impact on the push system and should be suppressed. Based on these basic weights, an initial behavior weight matrix is constructed, which assigns a preliminary weight value to each behavior type.

[0151] By setting basic weights for each behavior type, the importance of each behavior type can be reasonably defined, ensuring that positive behaviors are given priority and reducing the interference of negative behaviors on recommendations. For example, positive behaviors (such as repaying on time) directly reflect the user's credit status, while negative behaviors (such as overdue repayment) may indicate the user's credit risk and reduce the probability of recommending certain financial products.

[0152] Behavior intensity refers to the activity frequency, duration, or participation degree of the user under a certain behavior type, reflecting the degree of the user's investment in a specific behavior. Behaviors with high frequency, long duration, and high participation degree will be considered "high-intensity" behaviors, while behaviors with low frequency, short time, and low participation degree are "low-intensity". The behavior intensity distribution describes the distribution of the user's investment intensity in different types of behaviors, which helps to identify user preferences and activity levels.

[0153] For example:

[0154] High-intensity behavior: The user frequently purchases financial products and continuously invests a large amount every month.

[0155] Low-intensity behavior: The user occasionally browses the bank's financial product page but never makes an actual purchase.

[0156] By counting the intensity of these behaviors, it is possible to identify which behavior types are the main behaviors of the user and which behaviors are only occasional secondary behaviors. Counting the distribution of behavior intensity can help the system identify the user's activity level and preferences. For a financial product push system, the system should give priority to behaviors with higher user intensity (such as frequently purchasing financial products or repaying on time) in order to provide more personalized and accurate recommendations for users. On the contrary, behaviors with lower intensity can be regarded as secondary information with less impact. For example, the behavior data of User B is as follows:

[0157] Purchase of financial products: The frequency is once a month and the investment amount is large;

[0158] Consultation on loan products: The frequency is once a quarter and no actual application is made;

[0159] By counting the intensity of these behaviors, it can be concluded that User B has a higher intensity in purchasing financial products and a lower intensity in loan consultation, and the recommendation strategy can be adjusted according to these intensity distributions.

[0160] The behavior intensity quantization function determines how to adjust the basic weight according to the user's behavior intensity. For example, the following simple linear function can be used to quantify the behavior intensity:

[0161] Adjustment coefficient ;

[0162] Among them, the behavior intensity is the actual performance of the user in a certain behavior type, and the maximum behavior intensity is the maximum value of all user behaviors in this behavior type. This function can convert the behavior intensity into an adjustment coefficient, thereby adjusting the weights in the initial weight matrix. For example, assume that a user's behavior intensity of purchasing financial products is 8 (assuming the maximum intensity is 10), then the adjustment coefficient for this behavior type will be 0.8, indicating that this behavior has a greater impact on the recommendation.

[0163] By constructing a behavior intensity quantization function, the system can flexibly adjust the weights according to the user's actual behavior intensity, so that behaviors with higher intensity can have a greater influence on the push system. For example, for users who frequently purchase financial products, their behavior intensity is high and the adjustment coefficient is large. The system should assign a higher weight to this behavior, so that when making recommendations, it is more inclined to recommend financial products. For example, the behavior intensity of User C in purchasing financial products is 7 and the maximum intensity is 10, so the adjustment coefficient is 0.7, and the basic weight of purchasing financial products can be adjusted according to this adjustment coefficient.

[0164] Multiply the base weights in the initial behavior weight matrix by the corresponding weight adjustment coefficients. This process makes the weights of each behavior type more accurately reflect the user's behavior intensity and participation, thereby ensuring a higher degree of personalization in subsequent recommendations. Through this operation, the weights of different behavior types can be dynamically adjusted, enabling high-intensity user behaviors to obtain higher recommendation priorities. For example, if user A has a high intensity of behavior in purchasing financial products, the adjusted weight will be larger than that of other behavior types, and relevant products will be recommended first.

[0165] Step S1235 finally generates the weights of two behavior features, providing support for personalized recommendations. By assigning the weights of positive and negative behaviors to the two features respectively, it is possible to accurately judge the user's preferences and potential risks based on the user's behavior data, and adjust the recommendation strategy accordingly. For example, financial products related to positive behaviors (such as financial management products, credit cards) can be recommended first, while negative behaviors (such as high-risk investments) can be appropriately avoided. For example, if user A shows a strong positive behavior in purchasing financial products and the adjusted weight is 0.8, this will become the first behavior weight of the first behavior feature; if user A has a low record of overdue repayment and the adjusted weight of the negative behavior is 0.3, this will become the second behavior weight of the second behavior feature.

[0166] In step S1300, generate a multi-dimensional user portrait vector based on the first behavior feature, the first behavior weight, the second behavior feature, and the second behavior weight.

[0167] Furthermore, step S1300 includes:

[0168] In step S1310, construct a positive preference dimension according to the first behavior feature and the first behavior weight;

[0169] In step S1320, construct a negative preference dimension according to the second behavior feature and the second behavior weight;

[0170] In step S1330, fuse the positive preference dimension and the negative preference dimension to generate a multi-dimensional user portrait vector.

[0171] Specifically, the goal of step S1300 is to generate a multi-dimensional user portrait vector based on the user's behavior data, which comprehensively combines the user's behavior characteristics in financial services and the weights of these behaviors. Financial institutions, such as banks, usually collect the user's behavior data, such as deposits, loans, investments, payment behaviors, etc., and perform weighted processing on these data to understand the user's preferences, needs, risk preferences and other characteristics. This process not only helps banks identify the user's needs, but also enables them to provide personalized services or products and make more targeted marketing decisions in the future.

[0172] The first type of behavioral characteristics generally refers to the positive behavioral data of users, such as users' deposits, loan applications, investment purchases, credit card consumption, etc. Positive behaviors reflect users' preferences or interests in financial products, showing their financial needs and activity levels. For example, users frequently making deposits, investing in wealth management products, or actively participating in specific promotional activities are all positive behaviors. In practical applications, the contributions of different behaviors to the user portrait may vary, so it is necessary to assign weights to these behaviors. The assignment of behavioral weights depends on the frequency, amount, duration of the behavior, and the importance of the user. For example, a large deposit may reflect a user's dependence on bank services more than an occasional small transaction, so its weight is higher; while frequent small-scale consumption or short-term loans may represent the user's general needs and have a lower weight. By combining the first type of behavioral characteristics with the first type of behavioral weights, a positive preference dimension can be constructed, that is, the user's positive behavioral preferences. This dimension can reflect the user's core needs and interests. For example, users who frequently invest in funds indicate a strong interest in investment and wealth management, and banks can infer the user's demand for wealth management products based on this. Constructing the positive preference dimension helps to accurately identify users' interests and needs in specific financial products, thereby achieving personalized marketing and improving user satisfaction and loyalty. For example, banks can recommend relevant wealth management products and loan offers to users based on their deposit and investment behaviors, enhancing the user experience and product sales effectiveness.

[0173] The second type of behavioral characteristics generally refers to the negative behavioral data of users, such as users' complaints, service cancellations, rejected loan applications, negative reviews, etc. Negative behaviors usually reflect users' dissatisfaction with financial services or the potential risk of churn. The weights of negative behaviors also depend on the frequency, severity of the behavior, and its potential impact on user churn. For example, when a user makes multiple complaints or threatens to close their account in a short period, the weight should be relatively large; while an occasional minor complaint has a smaller weight. By combining the second type of behavioral characteristics with the second type of behavioral weights, a negative preference dimension can be constructed, that is, the user's behavioral tendency in negative situations. For example, if a user frequently complains about the bank's services or has posted negative reviews on the bank platform, it indicates that the user has strong negative emotions and may lack trust in some of the bank's services. Constructing the negative preference dimension helps to promptly identify users' dissatisfaction and churn risks, thereby preventing the further spread of negative emotions. Intervention can be carried out at the initial stage of users' dissatisfaction, by improving services and providing compensation to mitigate the negative impact and prevent user churn. In addition, the negative preference dimension can also help analyze and optimize service processes, enhance the user experience, and reduce user complaints.

[0174] In step S1330, the positive preference dimension and the negative preference dimension are fused to form a comprehensive multi-dimensional user profile vector. This fusion process combines the positive and negative behaviors of users with certain weights through weighted summation to obtain a comprehensive user profile. For example, the positive preference dimension and the negative preference dimension can occupy different weight ratios respectively, or be dynamically adjusted according to their impact on the user churn risk. The fused user profile vector contains the positive preferences, negative emotions of users, and the comprehensive performance of both. This multi-dimensional user profile vector not only shows the needs and interests of users, but also reveals their emotional attitudes and behavior patterns. This comprehensive vector provides strong data support for subsequent product recommendations, personalized services, and marketing decisions. By fusing the positive and negative preference dimensions, the behaviors and emotional attitudes of users can be comprehensively understood, and a more accurate user profile can be constructed. For example, the positive preference dimension reflects the user's interest in financial products, while the negative preference dimension reveals the user's potential dissatisfaction with services. The fused profile can help banks formulate more effective user management strategies, avoid blindly promoting products that users are not interested in, and timely adjust service quality to avoid user churn.

[0175] Step S2000: Construct a pre-trained industry language model. According to the multi-dimensional user profile vector and the pre-trained industry language model, obtain the first semantic embedding vector and the second semantic embedding vector; adjust the first semantic embedding vector and the second semantic embedding vector according to a preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; obtain an adjusted push content set according to the positive preference embedding vector and the negative preference embedding vector;

[0176] Furthermore, step S2000 includes:

[0177] Step S2100: Construct a pre-trained industry language model. According to the multi-dimensional user profile vector and the pre-trained industry language model, obtain the first semantic embedding vector and the second semantic embedding vector;

[0178] Furthermore, step S2100 includes:

[0179] Step S2110: Construct a pre-trained industry language model;

[0180] Specifically, in the financial field, especially in the banking industry, the pre-trained industry language model (Industry-Specific Pre-trained Language Model) can help automate tasks such as user services, sales promotion, and user sentiment analysis. By learning the professional terms, knowledge background, and user behavior data of the financial industry, this model can better understand and generate natural language content related to financial operations.

[0181] The architectures of pre-trained industry language models are usually based on the Transformer structure, such as BERT, GPT, and T5. The Transformer architecture can effectively handle long-distance dependencies through the Self-Attention Mechanism, so it performs particularly well when dealing with complex language patterns in the financial field. Different Transformer variants are suitable for different task scenarios: the Encoder-Decoder architecture (such as T5) is good at generative tasks, such as generating bank recommendations or automatic responses; the Encoder-only architecture (such as BERT) is more suitable for understanding tasks, such as sentiment analysis, user question classification, etc.; while the Decoder-only architecture (such as GPT) is widely used in generative tasks and dialogue generation.

[0182] Pre-trained industry language models need to process text inputs closely related to the industry, such as user consultation records, complaint content, bank product descriptions, and user transaction data. During the preprocessing and embedding of the input data, the model can gradually learn the professional language features and expressions in the financial industry. The pre-training of the model depends on a large amount of financial domain data, including bank service conversations, user feedback, industry reports, and investment analysis articles. These data need to be cleaned and annotated to remove irrelevant content in order to retain financial industry-related vocabulary and sentiment expressions. During the training process, the goal of the model is usually a language modeling task (such as Masked Language Modeling, MLM) or an autoregressive generation task (such as Autoregressive Modeling), and it gradually learns context information by predicting masked words or generating the next part of the text.

[0183] In model training, the input data is usually preprocessed text, which may be presented in single-sentence or multi-sentence form. For example, the format "User consultation: question, Bank reply: reply content" can be used as a typical input. Depending on the task, the true label may be a "positive" or "negative" label in a sentiment analysis task, or the correct reply corresponding to the input text in a text generation task. After training, the results output by the model mainly include two forms: one is semantic embedding vectors, which can capture the deep information of the text content and help the model understand the sentiment and intention in the financial field; the other is generated content. For generative tasks, the model can directly generate automated responses, such as generating accurate answers to user consultations in a user service scenario.

[0184] Building a pre-trained industry language model can enable financial institutions such as banks to better understand users' needs, emotional states, and intentions, thereby providing more accurate services or recommendations. For the automation of customer service and salesman replacement in the financial industry, this model can automatically generate personalized responses, reduce the pressure on human customer service, and improve response speed and accuracy.

[0185] Step S2120: Input the positive preference dimensions in the multi-dimensional user profile vector into the pre-trained industry language model to obtain a first semantic embedding vector.

[0186] Specifically, input the positive preference dimensions in the user profile into the pre-trained industry language model to obtain a first semantic embedding vector. The positive preference dimensions usually include the needs and interests demonstrated by users in their past behaviors. For example, if a user frequently purchases a certain financial product (such as a fund, loan, etc.), it indicates that they have a high interest in these products. Convert this information into text descriptions, such as "The user has a strong interest in fund investment" or "The user often applies for personal loans". Input these texts into the pre-trained industry language model, and the model generates a first semantic embedding vector by processing these texts and applying its built-in semantic knowledge. The semantic embedding vector is a high-dimensional vector that represents the deep semantic information about the user's preferences. It not only contains the user's behavioral characteristics but also information on industry terms and emotional tendencies. By inputting the positive preference dimensions into the pre-trained industry language model, it is possible to understand the user's needs and interests more deeply and provide personalized financial service recommendations. The first semantic embedding vector can provide accurate data support for subsequent intelligent customer service, marketing pushes, etc., thereby enhancing the user experience and the service efficiency of the bank.

[0187] Step S2130: Input the negative preference dimensions of the multi-dimensional user profile vector into the pre-trained industry language model to obtain a second semantic embedding vector.

[0188] Specifically, the negative preference dimensions usually include the user's negative behavior data, such as complaint records, dissatisfied feedback, rejected services, etc. Convert this information into text descriptions, for example: "The user is dissatisfied with the bank's service" or "The user has complained about the high handling fees". After inputting these texts into the pre-trained industry language model, the model will generate a second semantic embedding vector according to its inherent semantic understanding ability. This vector contains the semantic information of negative emotions, such as the user's aversion to certain bank services or distrust of certain products. By inputting the negative preference dimensions into the pre-trained industry language model, it is possible to better identify the user's dissatisfaction and negative emotions. The second semantic embedding vector generated by the model provides a basis for the bank to predict user churn and helps the bank to optimize services and retain users in a timely manner for negative emotions.

[0189] Step S2200: Adjust the first semantic embedding vector and the second semantic embedding vector according to a preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector;

[0190] Further, step S2200 includes:

[0191] Step S2210: Calculate the similarity between the first semantic embedding vector and the second semantic embedding vector to obtain a positive and negative preference similarity score D1;

[0192] The calculation of the similarity between the first semantic embedding vector and the second semantic embedding vector includes:

[0193] ;

[0194] Where:

[0195] The first part (Gaussian weight term):

[0196] : The squared Euclidean distance between the first semantic embedding vector and the second semantic embedding vector , which is used to measure the overall geometric distance between the two vectors in the semantic space. The greater the geometric distance of the embedding vectors, the more significant the semantic difference between them.

[0197] : The Gaussian weight term, which decays using the exponential function of the negative squared distance, emphasizes similar cases with smaller distances, and at the same time rapidly decays cases with larger distances. It can assign higher similarity weights to semantic embedding points with closer distances and reduce the influence of points with larger distances.

[0198] : The exponential function with the base e, the base of the natural logarithm.

[0199] The second part (weight correction term):

[0200] : The dynamic weight correction term, which combines the dynamic weights of each dimension to perform a weighted sum of the local distances of the embedding vectors.

[0201] : The dimension size of the embedding vector (usually 300 - 768).

[0202] : The th dimension's dynamic weight, indicating the importance of this dimension in the semantic similarity calculation. By analyzing industry corpora, the semantic distribution variance corresponding to each dimension is statistically calculated: is the Variance of dimensions). It is used to reduce the impact of high-variance dimensions on similarity calculation and highlight the stable contribution of low-variance dimensions.

[0203] : The squared local distance between two vectors on the -th dimension, which is used to quantify local differences.

[0204] This term can combine the dynamic weights of each dimension and endow the overall similarity with a more refined semantic characterization ability.

[0205] The third part (angle similarity term):

[0206] : The included angle between two vectors (expressed in radians), which is obtained through the inverse cosine function of cosine similarity. is the dot product of two vectors, and are their Euclidean norms. The smaller the included angle, the more similar the two vectors are; the larger the included angle, the less similar the two vectors are.

[0207] : Normalize the included angle to so that it can be directly involved in weighting.

[0208] : The reverse form of the angle similarity, the closer to 1 indicates the more similar, and the closer to 0 indicates the less similar.

[0209] Capture the semantic consistency of two vectors in direction through the angle correction of cosine similarity.

[0210] The fourth part (semantic shift correction term):

[0211] : The semantic shift correction term, which combines the dynamic ratio of the first-order norm and the second-order norm, and is used to correct the overall semantic shift of the vector.

[0212] : The first-order norm (Manhattan distance) of the first semantic embedding vector and the second semantic embedding vector, which is used to measure the semantic difference between two vectors.

[0213] : The sum of the Euclidean norms of two vectors, which is used to measure the overall semantic concentration of two vectors.

[0214] : The semantic shift correction coefficient, a hyperparameter, which controls the contribution of the semantic shift term to the final similarity. Value range: . Through tuning the parameters on the validation set, select the optimal value that can optimize the recommendation performance.

[0215] : Normalization adjustment coefficient, a hyperparameter used to balance the normalization effect of the offset term. Value range: . Obtained by tuning parameters through the validation set.

[0216] The semantic offset correction term can capture the overall semantic offset trend of the embedding vectors and optimize the similarity distortion problem caused by directionality and local weights.

[0217] When increases, the Gaussian weight term decays rapidly, making decrease. When the Manhattan distance increases, the semantic offset correction term increases, further reducing . When the included angle between the two vectors increases, the cosine similarity decreases, resulting in the angle correction term decreasing, thereby reducing . The weights of high-variance dimensions are lower, weakening the influence on similarity calculation; the weights of low-variance dimensions are higher, making the influence of local distance on increase.

[0218] This formula combines geometric distance, dynamic weights, directionality, and semantic offset correction, and can measure the similarity of semantic embedding vectors more comprehensively. It not only considers the overall distance of the vectors, but also combines the importance of local dimensions and direction consistency, and has the ability to describe similarity from multiple angles. The dynamic weight distribution and the normalization adjustment coefficient are introduced to enable the formula to adapt to different scenarios and the semantic characteristics of embedding vectors. Through the hyperparameters and regulation, the sensitivity of similarity can be adjusted in different tasks (such as recommendation or filtering). The Gaussian weight term can rapidly decay abnormal vectors (too far away), reducing the influence of outliers on similarity calculation. The semantic offset correction term can correct the distortion of directionality and local deviation on similarity.

[0219] Step S2220, set the similarity threshold θ1. If the positive and negative preference similarity scores D1 are greater than θ1, trigger the embedding vector adjustment process to obtain the positive preference embedding vector and the negative preference embedding vector.

[0220] The embedding vector adjustment process includes:

[0221] Step S2221, randomly sample noise vectors, adjust the first semantic embedding vector according to the noise vectors to obtain the adjusted first semantic embedding vector, marked as the positive preference embedding vector;

[0222] Step S2222: Adjust the second semantic embedding vector according to the noise vector to obtain the adjusted second semantic embedding vector, denoted as the negative preference embedding vector.

[0223] Step S2223: Calculate the similarity D2 between the positive preference embedding vector and the negative preference embedding vector, and determine whether D2 is less than or equal to θ1. If so, complete the embedding vector adjustment process; otherwise, return to Step S2221, randomly sample the noise vector again, and repeat Steps S2221 - S2223 until D2 ≤ θ1.

[0224] Specifically, a user profile usually consists of multiple dimensions, including the user's positive preferences (such as the preference for certain financial products) and negative preferences (such as dissatisfaction with certain services). The purpose of Step S2220 is to ensure that there is a clear distinction between the semantic embedding vectors of positive and negative preferences when making personalized recommendations for users, so as to accurately reflect the user's needs and emotional state.

[0225] The similarity threshold θ1 is a preset value used to determine whether the similarity between the positive and negative preference embedding vectors exceeds the acceptable range. This value is usually adjusted based on experience or through cross - validation, etc., to ensure the balance of the model. For example, if θ1 is set to 0.8, when the similarity between positive and negative preferences is greater than 0.8, it means that these two vectors are too close and need to be adjusted. By setting the threshold θ1, it is possible to avoid excessive overlap of semantic information in the positive and negative preference dimensions, ensuring that the model can identify the multi - dimensional behavior characteristics of users, rather than merging positive and negative behaviors into the same preference.

[0226] In Step S2221, the first semantic embedding vector is adjusted by sampling the noise vector. The purpose is to increase the difference between the positive and negative preference embedding vectors and make them more distinguishable. The noise vector is a random vector, usually sampled from the standard normal distribution, that is, a normal distribution with a mean of 0 and a variance of 1. The introduction of noise can help adjust the spatial distribution of the vector, thus avoiding the embedding vectors being too similar. The noise vector and the first semantic embedding vector are weighted and synthesized to obtain a new second semantic embedding vector. This process is similar to perturbing the original vector, thereby adjusting its position in the vector space, so that the adjusted first semantic embedding vector and the second semantic embedding vector have a greater difference. By introducing the noise vector and adjusting the first semantic embedding vector, it is possible to effectively avoid overly similar embedding vectors and ensure sufficient semantic differences between positive and negative preferences. This is crucial for accurately modeling the user's positive and negative emotions. Especially in a personalized push system, it can help the model more precisely identify the user's needs and emotions, thereby optimizing services and recommendations.

[0227] Similar to step S2221, the goal of step S2222 is to adjust the second semantic embedding vector to further increase the distance between it and the first semantic embedding vector, thereby ensuring that they are significantly distinguishable in the semantic space. The noise vector used here is the same as that in step S2221. By adjusting the second semantic embedding vector, the distinction between positive and negative preferences is further enhanced. The adjustment of the second semantic embedding vector helps to more clearly distinguish the user's negative emotional information, such as dissatisfaction with certain financial products or services, thus improving the model's ability to understand user needs.

[0228] By continuously adjusting the first semantic embedding vector and the second semantic embedding vector in step S2223 until the similarity between positive and negative preferences meets a preset threshold, the final user profile can be ensured to be more accurate. In this way, it can help provide more precise personalized services and product recommendations, avoid the confusion between positive and negative preferences, and improve user satisfaction and loyalty.

[0229] In step S2300, an adjusted set of push content is obtained based on the positive preference embedding vector and the negative preference embedding vector.

[0230] Furthermore, step S2300 includes:

[0231] In step S2310, an industry corpus is loaded, and semantic similarity retrieval is performed on the text content in the industry corpus using the positive preference embedding vector to generate a first candidate set of push content;

[0232] Furthermore, as Figure 4 shown, step S2310 includes:

[0233] In step S2311, the text content in the industry corpus is converted into corresponding semantic embedding vectors;

[0234] In step S2312, the similarity between the positive preference embedding vector in the multi-dimensional user profile vector and the semantic embedding vector of the text content in the industry corpus is calculated to obtain a first similarity score;

[0235] In step S2313, the text content in the industry corpus with the first similarity score ranked among the top K is selected to generate a first candidate set of push content.

[0236] Specifically, the goal of step S2310 is to screen out the text content that best matches the user's needs by calculating the semantic similarity between the user's positive preference embedding vector and the relevant text in the industry corpus, thereby generating a candidate set of push content. This process mainly relies on semantic embedding technology to convert text into vector representations and perform matching by calculating similarity. In this way, personalized recommended content can be provided according to the user's preferences, thereby improving user satisfaction.

[0237] The industry corpus contains industry-related text data, such as product introductions, terms of service, frequently asked questions, market reports, investment advice, etc. These texts are the basis for communicating with users and represent the diversity and professionalism of the products. The process of loading the industry corpus usually includes reading text files, database queries, etc., to ensure that the text content in the corpus is complete and can perform efficient semantic retrieval. By loading the industry corpus, rich industry knowledge and user interaction content can be obtained, providing basic support for subsequent personalized push. The quality of the corpus directly affects the accuracy of the push content and the user experience.

[0238] Step S2311 converts the text content in the industry corpus into corresponding semantic embedding vectors. Semantic embedding vectors are a method of text representation that can map text into a high-dimensional vector space, making texts with similar semantics closer in the vector space. Common semantic embedding models include Word2Vec, GloVe, BERT, etc. Through semantic embedding, texts of different lengths and different expression methods can be converted into a unified vector representation, facilitating subsequent similarity calculation.

[0239] Step S2312 calculates the similarity between the positive preference embedding vector in the multi-dimensional user profile vector and the semantic embedding vector of the text content in the industry corpus to obtain the first similarity score. Similarity calculation measures the semantic proximity of two vectors. Common similarity metrics include cosine similarity, Euclidean distance, etc. Cosine similarity calculates the cosine value of the angle between two vectors, with a numerical range between [-1, 1]. The larger the value, the higher the similarity. Euclidean distance calculates the straight-line distance between two vectors in space, and the smaller the distance, the higher the similarity. By calculating the similarity score, the text content that best matches the user's needs can be effectively screened according to the user's positive preferences. The core advantage of this process is that it can capture the deep semantic connection between the user's needs and the text content, thereby realizing personalized recommendation and enhancing the user's experience and engagement.

[0240] Step S2313 sorts the text content in the corpus according to the similarity calculation result obtained in step S2312, and selects the top K texts with the highest similarity as candidate push content. The selection of the K value needs to balance the quality and diversity of the push content: a too small K value may lead to overly limited push content, while a too large K value may introduce irrelevant content. Usually, the optimal K value can be determined through methods such as A / B testing based on user feedback and business requirements. By screening out the content most relevant to the user's needs, it can be ensured that the pushed text information is more in line with the user's interests and needs, thereby improving the accuracy of the recommendation and the user's satisfaction. In addition, this process can also optimize the push efficiency and reduce the interference of irrelevant content.

[0241] Exemplarily, assume that user A has shown a preference for low-risk investment products in past financial activities and is relatively concerned about the quality of the bank's user services. In step S2310, the positive preference embedding vector of user A has been generated through their behavioral data. Next, the industry corpus contains text descriptions of multiple investment products and services.

[0242] Step S2311: The texts in the industry corpus (such as "low-risk fund product Q", "user service optimization", etc.) are converted into semantic embedding vectors.

[0243] Step S2312: Calculate the similarity between the positive preference embedding vector of user A and the semantic embedding vector of each text. Assume that the text "low-risk fund product Q" has a relatively high similarity with the positive preference of user A, while "user service optimization" has a relatively low similarity.

[0244] Step S2313: Screen out the top K texts with the highest similarity scores. Assume the top two texts are "low-risk fund product Q" and "user service optimization", then these two texts will form the first candidate push content set.

[0245] Through this method, the bank can accurately push personalized content that meets the user's needs, improve user engagement, and enhance the user's satisfaction and loyalty to the bank's services.

[0246] Step S2320, use the negative preference embedding vector to perform semantic similarity retrieval on the text content in the industry corpus to generate a second candidate push content set;

[0247] Furthermore, step S2320 includes:

[0248] Step S2321, calculate the similarity between the negative preference embedding vector in the multi-dimensional user profile vector and the semantic embedding vector of the text content in the industry corpus to obtain a second similarity score;

[0249] Step S2322, screen out the top M text contents in the industry corpus with the second similarity scores to generate a second candidate push content set.

[0250] Specifically, the purpose of step S2320 is to screen out the text content related to the user's dissatisfaction, negative emotions or rejections from the industry corpus according to the user's negative preference embedding vector. By calculating the similarity between the negative preference embedding vector and each text in the corpus, a text set that matches the user's emotions and dissatisfaction is generated. The goal of this process is to avoid pushing content that may trigger negative emotions in users and help the bank better identify the pain points and needs of users, thereby improving service quality.

[0251] In step S2321, the similarity between the user's negative preference embedding vector and the semantic embedding vector of each text in the industry corpus is first calculated. Similar to step S2312, the similarity between the negative preference embedding vector and the text content is calculated here, with the aim of finding content related to user dissatisfaction and negative emotions. The similarity calculation is completed by comparing the negative preference embedding vector with the semantic embedding vector of each text in the industry corpus. Commonly used similarity measurement methods include cosine similarity, Euclidean distance, etc. By calculating the similarity between the negative preference embedding vector and the text, the text content that matches the user's negative emotions can be accurately identified. This process is crucial for the push system because it can effectively prevent the push of content that may cause user dissatisfaction or distrust. In this way, banks can optimize push content and better manage user emotions.

[0252] In step S2322, according to the second similarity score calculated in step S2321, the top M text contents that best match the user's negative preference are screened out, and these texts are used as the second candidate push content set. These text contents are usually related to the user's negative emotions or needs, and are intended to avoid pushing content that may exacerbate negative emotions or dissatisfaction. By screening out the content that best matches the user's negative preference, it is possible to avoid pushing content that arouses user disgust, thereby improving user satisfaction and loyalty. This step helps to optimize the user experience, especially in terms of service quality and product push, and helps banks avoid the risk of user churn.

[0253] Step S2330: obtaining an adjusted push content set based on the first candidate push content set and the second candidate push content set.

[0254] Furthermore, if Figure 5 As shown, step S2330 includes:

[0255] Step S2331, obtaining semantic embedding vectors of text contents in the first candidate push content set and the second candidate push content set;

[0256] Step S2332, calculating the semantic similarity S2 between each text content in the first candidate push content set and all text content in the second candidate push content set based on the semantic embedding vectors of the text content in the first candidate push content set and the second candidate push content set, to form a similarity matrix;

[0257] Step S2333, in the similarity matrix, filter out content pairs whose semantic similarity S2 is higher than a second similarity threshold θ2, and mark them as highly overlapping content;

[0258] Step S2334: remove the marked highly overlapping contents from the first candidate push content set to obtain an adjusted push content set.

[0259] Specifically, the goal of step S2330 is to optimize the final push content by combining the first candidate push content set and the second candidate push content set, removing duplicate or highly overlapping content. This step mainly calculates the semantic similarity between contents, filters out highly duplicate content, and performs deduplication to ensure that the content finally pushed to users is diverse and personalized, avoiding overly repetitive pushes.

[0260] The semantic embedding vector mentioned in step S2331 is a technique that maps text to a high-dimensional space. Through semantic embedding, text content can be represented as a real vector with a fixed dimension, where texts with similar semantics are closer in the vector space. Common semantic embedding models include Word2Vec, GloVe, and BERT, etc. These models learn the semantic representations of words or sentences through unsupervised or semi-supervised training on large-scale corpora. By obtaining the semantic embedding vectors of candidate push content, the similarity between them can be calculated in the vector space, providing a basis for subsequent detection of overlapping content.

[0261] Step S2332 calculates the semantic similarity between each text content in the first candidate push content set and all text contents in the second candidate push content set to form a similarity matrix. The semantic similarity can be obtained by calculating metrics such as the cosine similarity or Euclidean distance between two semantic embedding vectors. The cosine similarity calculates the cosine value of the angle between vectors, with a value range of [-1, 1]. The larger the value, the closer the vector directions are, that is, the more similar the text semantics are. The Euclidean distance calculates the straight-line distance between vectors, and the smaller the distance, the more similar the text semantics are. By constructing the similarity matrix, the semantic overlap degree between the two groups of content can be understood more clearly. This step helps to accurately identify the relationship between the first candidate push content and the second candidate push content, thus making a more reasonable screening decision and avoiding overly repetitive or highly relevant content.

[0262] Step S2333 filters out content pairs with semantic similarity exceeding the threshold in the similarity matrix according to the preset second similarity threshold θ2 and marks them as highly overlapping content. The value of the second similarity threshold θ2 needs to be adjusted according to the specific application scenario and the characteristics of the push content. If the threshold is set too high, it may lead to ineffective identification of overlapping content; if the threshold is set too low, it may filter out semantically related but not duplicate content. Therefore, it is necessary to tune parameters and conduct tests on actual data to select an appropriate threshold to balance the filtering precision and recall rate. By removing highly overlapping content, the repetition of pushed information can be avoided, thereby improving the diversity and personalization of the push and avoiding excessive pushing of similar content. This can not only enhance the user experience but also increase the accuracy and effectiveness of the push system.

[0263] In step S2334, based on the height-overlapping content marked in step S2333, these contents are excluded from the first candidate push content set. The exclusion operation ensures that the final push content set does not contain duplicate or highly similar texts, thus guaranteeing that the content pushed to users is both diverse and targeted. Through this step, over-pushing to users can be avoided, reducing the sense of information fatigue, while enhancing user engagement and loyalty.

[0264] Exemplarily, assume that user C shows strong interest in low-risk financial products while expressing dissatisfaction with excessive bank handling fees.

[0265] Step S2331: The texts in the first candidate push set include "Recommendation of Low-Risk Financial Products" and "Fund Investment Product Y", and the texts in the second candidate push set include "New Policy on Handling Fees" and "Service Quality Improvement Plan".

[0266] Step S2332: Calculate the similarity between each first candidate push text and all second candidate push texts to obtain a similarity matrix. Assume that the similarity between "Recommendation of Low-Risk Financial Products" and "New Policy on Handling Fees" is 0.85, exceeding the second similarity threshold θ2.

[0267] Step S2333: Screen out the content pairs with similarity higher than the second similarity threshold θ2. For example, "Recommendation of Low-Risk Financial Products" and "New Policy on Handling Fees" are marked as "highly overlapping content".

[0268] Step S2334: Exclude "Recommendation of Low-Risk Financial Products" from the first candidate push set to generate an adjusted push content set, ensuring that the remaining content does not contain duplicate or highly similar texts.

[0269] By removing content with too high similarity, the bank ensures that each pushed content contains more unique options, avoiding duplication and information overload. In this way, users will receive more different content that meets their needs and interests, thus enhancing their experience. By excluding duplicate push content, users will not be troubled by the same information. Through semantic similarity screening of push content, it can be ensured that each piece of information recommended to users has high relevance and pertinence. Whether it is the recommendation of financial products or the notice of service quality improvement, it can better match the preferences and needs of users, enhancing the personalization and effectiveness of push content. In a financial environment with a large amount of information, it is crucial to avoid pushing the same recommendation content repeatedly. Information fatigue will cause users to lose interest in recommendations and even generate aversion. Through this deduplication process, the bank can improve users' attention and satisfaction with push content.

[0270] Step S3000: Obtain a third candidate push content set based on the adjusted push content set; perform intelligent sorting on the content in the third candidate push content set to generate an accurate push content sequence.

[0271] Further, step S3000 includes:

[0272] Step S3100: Build and train a semantic adversarial generative network based on the adjusted push content set;

[0273] Further, step S3100 includes:

[0274] Step S3110: Mark the adjusted push content set as positive samples and mark the second candidate push content set as negative samples to construct a training data set for the semantic adversarial generative network;

[0275] Step S3120: Train the semantic adversarial generative network according to the training data set of the semantic adversarial generative network.

[0276] Specifically, the goal of step S3100 is to build and train a semantic adversarial generative network (SAGN) through the adjusted push content set to generate more personalized recommendation content that better meets user needs and emotions. The design inspiration of the semantic adversarial generative network comes from generative adversarial networks (GANs). It uses the way of adversarial training and through the competition of two networks (generator and discriminator) to generate more natural and highly personalized push content. The challenge of personalized recommendation lies in how to dynamically adjust the recommended content according to the user's preferences, needs and emotional states. The semantic adversarial generative network can help the bank automatically generate recommended content that meets its needs according to the positive and negative emotional preferences of users, thereby improving the accuracy of recommendations and user satisfaction.

[0277] Prepare the training data set for the semantic adversarial generative network in step S3110. This data set consists of positive samples and negative samples:

[0278] Positive samples: From the adjusted push content set. These contents are push texts that have been deduplicated and optimized, representing recommended content that meets user needs and preferences. When generating training data, the positive samples should include all the content finally pushed to users, because they represent the content that the system considers to be the most relevant and in line with user needs.

[0279] Negative samples: These come from the second candidate push content set, which usually contains push information that does not match the user's needs or may trigger negative emotions. The second candidate push content set typically includes content related to the user's negative preferences, such as excessive fees, poor user service quality, and other information that does not meet the user's preferences.

[0280] Construct a training data set:

[0281] Input: The adjusted push content set (positive samples) and the second candidate push content set (negative samples).

[0282] Output: A data set labeled with positive and negative samples for training the semantic adversarial generation network.

[0283] Provide these labeled samples to the training network so that the generator learns to generate push content that meets the user's needs, and enables the discriminator to effectively distinguish between positive and negative samples, thereby guiding the generator to optimize its generation strategy.

[0284] By constructing a training data set of positive and negative samples, it is possible to provide sufficient training materials for the semantic adversarial generation network, ensuring that the generated recommended content not only meets the user's preferences but also avoids pushing irrelevant or user-repellent content. Such training can make the push system more accurate and personalized in practical applications.

[0285] In step S3120, the semantic adversarial generation network is trained using the training data set constructed in step S3110. The training process includes adversarial training of the generator and the discriminator:

[0286] Generator:

[0287] The task of the generator is to generate personalized recommended content based on the user's portrait (including positive and negative preferences). It learns from the training data set how to generate content that matches the user's needs and, by continuously optimizing the generation strategy, makes the generated content as close as possible to the positive samples at the semantic level.

[0288] Discriminator:

[0289] The task of the discriminator is to judge the content generated by the generator and determine whether it is "real" or "fake". Real content usually comes from positive samples (i.e., the adjusted push content set), while fake content comes from the content generated by the generator. Through continuous learning, the discriminator can better distinguish between positive and fake samples, thereby guiding the generator to optimize the generation strategy.

[0290] Training process:

[0291] Adversarial Training: The generator and the discriminator engage in an adversarial process during training. The generator tries to generate content similar to positive samples by improving its generation strategy, while the discriminator enhances its recognition ability by distinguishing between real and forged content. Through continuous iteration, both the generator and the discriminator make progress together, and the generator can ultimately generate high-quality recommended content.

[0292] Loss Function: The loss function in the training process usually includes the generator loss and the discriminator loss. The generator loss is used to measure the quality of the generated content, while the discriminator loss is used to measure its ability to distinguish between real and forged content.

[0293] By training the semantic adversarial generation network, higher personalization can be achieved in the pushed content, generating content that better suits the user's needs and emotional state. After training, the generator can freely generate personalized recommendations for different user preferences, while the discriminator helps ensure the high quality and high relevance of the recommended content. Through adversarial training, the generated content can reach a high level in both diversity and accuracy, greatly improving user satisfaction.

[0294] Step S3200: Obtain the real-time behavior data of the target user, and based on the real-time behavior data and the semantic adversarial generation network, obtain the third candidate push content set;

[0295] Specifically, in step S3200, based on the real-time behavior data of the target user and combined with the trained semantic adversarial generation network, new push content is generated. The core of this step is to use the user's latest behavior and preferences to dynamically adjust the recommended content to ensure that the pushed content as much as possible meets the user's current interests and needs.

[0296] Real-time behavior data refers to the user's immediate interaction information, usually including the user's transaction records, click behavior, browsing history, consultation records, etc. These data reflect the user's current interests, needs, and emotional state, and can help predict the products or services that the user will need soon. For example, if a user has recently frequently browsed the wealth management product page, then their interest in investment products may be relatively strong. Real-time behavior data is usually obtained through various data collection tools or technologies (such as Web log analysis, App behavior tracking, user interaction data analysis, etc.).

[0297] The Semantic Adversarial Generation Network (SAGN) updates the user profile based on real-time behavior data and generates content that best matches the user's interests. For example, if a user has recently invested in a low-risk fund, SAGN may generate a push notification recommending related products of that fund. In this way, SAGN ensures that the push notifications not only conform to the user's historical preferences but also adapt to their current behavior patterns. By combining real-time behavior data and the semantic adversarial generation network, instant and dynamic personalized recommendations can be provided. This behavior-based data-driven push can increase user engagement because it ensures that the recommended content is timely and relevant, thus enhancing the user experience and reducing the risk of user churn.

[0298] Step S3300: Input the third candidate push content set into a pre-constructed push strategy reinforcement learning model to obtain the optimal push strategy; perform intelligent sorting on the content in the third candidate push content set according to the optimal push strategy to generate a precise push content sequence.

[0299] Specifically, step S3300 inputs the generated third candidate push content set into the pre-constructed push strategy reinforcement learning model to generate the optimal push strategy. The reinforcement learning model simulates different push strategies and optimizes according to a preset goal to determine which strategy can maximize user satisfaction and push effect.

[0300] The push strategy reinforcement learning model is a machine learning model that uses reinforcement learning (RL) algorithms to automatically adjust push strategies. Reinforcement learning is a learning method based on reward feedback. The model tries different push strategies to obtain the optimal result. The core of the reinforcement learning model is the reward and punishment mechanism: the execution of each push strategy will bring a "reward" or "punishment" based on the user's reaction (such as click, purchase, browsing time, etc.). Eventually, the model will learn which push strategies are the most effective and utilize them in actual applications. The input of the push strategy reinforcement learning model is the third candidate push content set, and the output is the optimal push strategy.

[0301] Reinforcement learning process:

[0302] Reward and punishment mechanism: The reinforcement learning model is trained based on the user's reaction to the push content. For example, if a user clicks on a certain recommended content, the push strategy of that recommended content will be rewarded; if the user does not click, a punishment will be received. Through continuous feedback and adjustment, the model finally learns how to optimize the push strategy to maximize user engagement and satisfaction.

[0303] Optimization Goal: The goal of the push strategy is to maximize the long-term value of users, rather than just single interactions. For example, if the recommended financial products help users improve their investment returns or enhance users' trust in the bank, then the recommendation will be regarded as a successful push strategy.

[0304] Through the optimization of the reinforcement learning model for the push strategy, optimal decisions can be made in the selection and ranking of push content, ensuring that the recommended content can maximize users' interests and engagement. This adaptive and intelligent push strategy can significantly improve users' satisfaction, increase users' activity, and effectively reduce the pushing of ineffective or overly repetitive content.

[0305] Intelligently rank the third candidate push content set according to the optimal push strategy to generate the final accurate push content sequence. This ranking process is not simply sorting by relevance, but comprehensively considering multiple factors such as the timeliness of the content, users' interests, push history, etc., to ensure that each push content is sent to users at the best time and in the most appropriate context.

[0306] Key Factors for Intelligent Ranking:

[0307] Timeliness: Users' needs may change at any time, so the recommended financial products and services need to be updated in a timely manner. For example, if a user has recently viewed a certain financial product, then the relevant updated information should be displayed preferentially.

[0308] Users' Long-Term Interests: In addition to based on real-time behavior, users' long-term interests also need to be considered. For example, if a user has long invested in low-risk financial products, then preferentially push relevant low-risk investment products.

[0309] Users' Historical Interaction Data: Based on the historical data of users' interactions with the bank (such as purchase records, service feedback, etc.), the model can infer which content may be more in line with users' long-term needs.

[0310] Through intelligent ranking, an optimized push content sequence can be provided for users, so that each push message is delivered to users at the best time and in the best context. This optimization can significantly improve users' engagement, purchase conversion rate, and satisfaction, while reducing over-pushing and information fatigue.

[0311] Embodiment 2

[0312] Based on Embodiment 1, this embodiment provides a large model information push method based on industry corpus, including:

[0313] Step S1233, construct a behavior intensity quantization function, and according to the behavior intensity quantization function, map the behavior intensity distribution of different behavior types to the corresponding weight adjustment coefficients of the behavior types;

[0314] The described behavior intensity quantization function includes:

[0315]

[0316] Wherein:

[0317] represents the behavior type weight adjustment coefficient of.

[0318] represents the behavior type basic weight of, which is a preset constant reflecting the general importance of this behavior type. It can be set according to business experience and historical data analysis.

[0319] represents the behavior type behavior frequency of, that is, the number of times the user performs this behavior within a certain time range. It can be obtained by counting the user behavior logs.

[0320] represents the behavior type time span of, that is, the time range when the user performs this behavior. It can be obtained by calculating the time difference between the first and last times the user performs this behavior.

[0321] is the first scaling factor used to adjust the order of magnitude of the ratio of behavior frequency to time span. It can be set according to the distribution characteristics of the actual data.

[0322] represents the behavior type average duration of, reflecting the user's participation when performing this behavior. For example, for the browsing behavior, it can be measured by the average browsing duration; for the commenting behavior, it can be measured by the average number of comment words.

[0323] is the second scaling factor used to adjust the order of magnitude of the average duration. It can be set according to the distribution characteristics of the actual data.

[0324] represents the behavior type content relevance of, reflecting the matching degree between the content involved in this behavior and the user's interests. It can be measured by calculating the similarity between the content features and the user preference features.

[0325] is the third scaling factor used to adjust the order of magnitude of the content relevance. It can be set according to the distribution characteristics of the actual data.

[0326] This formula comprehensively considers multiple factors affecting the behavior intensity and generates a weight adjustment coefficient that can reflect the behavior intensity through non-linear combination. Specifically:

[0327] Logarithmic function Maps the ratio of behavior frequency to time span to a bounded interval, and increases as the ratio increases, reflecting the positive impact of the density of behavior on intensity.

[0328] Hyperbolic tangent function Maps the average duration to interval, and increases monotonically as the duration increases, reflecting the positive impact of user engagement on intensity.

[0329] Exponential function Maps the content relevance to interval, and increases monotonically as the relevance increases, reflecting the positive impact of content matching degree on intensity.

[0330] Base weight Plays a role in adjusting the overall level of the weight adjustment coefficient of different behavior types, reflecting the prior importance differences of different behavior types.

[0331] Generally speaking, this formula can calculate a comprehensive weight adjustment coefficient according to the multi-dimensional characteristics of user behavior. When a user shows high frequency, long time, high engagement and high relevance for a certain behavior type, the weight adjustment coefficient of this behavior type will increase accordingly, reflecting the strong interest of the user in this type of content. On the contrary, if the user's performance for a certain behavior type is inactive or the interest is not high, the weight adjustment coefficient will decrease accordingly. In this way, it is possible to more comprehensively and accurately characterize the preference intensity of users for different types of content, providing better weight information for subsequent personalized recommendations.

[0332] It should be noted that this formula contains multiple scaling factors , whose role is to normalize inputs of different orders of magnitude while maintaining the monotonicity of the function, so that the weight adjustment coefficients of different behavior types are in a similar numerical range, facilitating subsequent weight comparison and fusion. In practical applications, these scaling factors need to be appropriately adjusted according to the distribution characteristics of the data to achieve the best quantization effect. In addition, this formula is a possible form of the behavior intensity quantization function, and in practice, the formula can be further improved and optimized according to business requirements and data characteristics.

[0333] Example 3

[0334] Based on Example 1, this example provides a large model information push system based on industry corpus, such asFigure 6 As shown in the figure, it includes:

[0335] Image construction module: used to obtain the positive behavior record sequence and negative behavior record sequence of the target user; extract the first behavior feature from the positive behavior record sequence and the second behavior feature from the negative behavior record sequence; generate the first behavior weight of the first behavior feature and the second behavior weight of the second behavior feature; generate a multi-dimensional user image vector according to the first behavior feature, the first behavior weight, the second behavior feature and the second behavior weight;

[0336] Push content generation module: used to construct a pre-trained industry language model, and obtain the first semantic embedding vector and the second semantic embedding vector according to the multi-dimensional user image vector and the pre-trained industry language model; adjust the first semantic embedding vector and the second semantic embedding vector according to the preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; obtain an adjusted push content set according to the positive preference embedding vector and the negative preference embedding vector;

[0337] Push content sorting module: based on the adjusted push content set, obtain the third candidate push content set; perform intelligent sorting on the content in the third candidate push content set to generate an accurate push content sequence.

[0338] In the image construction module, the obtaining of the positive behavior record sequence and negative behavior record sequence of the target user includes:

[0339] Step S1110, construct an industry corpus, and obtain the first historical behavior data of the target user in the industry field from the industry corpus; the first historical behavior data includes n1 behavior types of the target user, and the n1 behavior types include positive behaviors and negative behaviors; n1 is a positive integer;

[0340] Step S1120, obtain the positive behavior record sequence and negative behavior record sequence according to the first historical behavior data.

[0341] In the image construction module, the extraction of the first behavior feature from the positive behavior record sequence includes:

[0342] Step S1211, perform word segmentation on the text content in the positive behavior record sequence, and segment the text content into a word sequence of the smallest semantic units;

[0343] Step S1212, perform part-of-speech tagging on the segmented word sequence;

[0344] Step S1213, identify named entities in the text content based on part-of-speech tagging;

[0345] Step S1214, extract keywords and topic words from the word sequence;

[0346] Step S1215, using keywords, subject words, and named entity words as the first row as features.

[0347] In the portrait construction module, the first row weight for generating the first row feature and the second row weight for generating the second row feature include:

[0348] Step S1231, set the basic weights of different behavior types among n1 behavior types, and construct an initial behavior weight matrix;

[0349] Step S1232, according to the first historical behavior data, count the behavior intensity distribution of different behavior types;

[0350] Step S1233, construct a behavior intensity quantization function, and according to the behavior intensity quantization function, map the behavior intensity distribution of different behavior types to the corresponding weight adjustment coefficients;

[0351] Step S1234, multiply the basic weights in the initial behavior weight matrix by the corresponding weight adjustment coefficients to generate an adjusted behavior weight matrix;

[0352] Step S1235, in the adjusted behavior weight matrix, use the weights corresponding to positive behaviors as the first row weights of the first row feature, and use the weights corresponding to negative behaviors as the second row weights of the second row feature.

[0353] In the portrait construction module, generating a multi-dimensional user portrait vector according to the first row feature, the first row weight, the second row feature, and the second row weight includes:

[0354] Step S1310, construct a positive preference dimension according to the first row feature and the first row weight;

[0355] Step S1320, construct a negative preference dimension according to the second row feature and the second row weight;

[0356] Step S1330, fuse the positive preference dimension and the negative preference dimension to generate a multi-dimensional user portrait vector.

[0357] In the push content generation module, obtaining the first semantic embedding vector and the second semantic embedding vector includes:

[0358] Step S2110, construct a pre-trained industry language model;

[0359] Step S2120, input the positive preference dimension in the multi-dimensional user portrait vector into the pre-trained industry language model to obtain the first semantic embedding vector;

[0360] Step S2130: Input the negative preference dimensions of the multi-dimensional user portrait vector into the pre-trained industry language model to obtain the second semantic embedding vector.

[0361] In the push content generation module, the generation of the positive preference embedding vector and the negative preference embedding vector includes:

[0362] Step S2210: Calculate the similarity between the first semantic embedding vector and the second semantic embedding vector to obtain the positive and negative preference similarity score D1.

[0363] Step S2220: Set the similarity threshold θ1. If the positive and negative preference similarity score D1 is greater than θ1, trigger the embedding vector adjustment process to obtain the positive preference embedding vector and the negative preference embedding vector.

[0364] The embedding vector adjustment process includes:

[0365] Step S2221: Randomly sample a noise vector, and adjust the first semantic embedding vector according to the noise vector to obtain the adjusted first semantic embedding vector, which is marked as the positive preference embedding vector.

[0366] Step S2222: Adjust the second semantic embedding vector according to the noise vector to obtain the adjusted second semantic embedding vector, which is marked as the negative preference embedding vector.

[0367] Step S2223: Calculate the similarity D2 between the positive preference embedding vector and the negative preference embedding vector, and determine whether D2 is less than or equal to θ1. If so, complete the embedding vector adjustment process; otherwise, return to Step S2221, randomly sample a new noise vector, and repeat Steps S2221 - S2223 until D2 ≤ θ1.

[0368] In the push content generation module, the obtaining of the adjusted push content set includes:

[0369] Step S2310: Load the industry corpus, and use the positive preference embedding vector to perform semantic similarity retrieval on the text content in the industry corpus to generate the first candidate push content set.

[0370] Step S2320: Use the negative preference embedding vector to perform semantic similarity retrieval on the text content in the industry corpus to generate the second candidate push content set.

[0371] Step S2330: Based on the first candidate push content set and the second candidate push content set, obtain the adjusted push content set.

[0372] The said Step S2310 includes:

[0373] Step S2311: Convert the text content in the industry corpus into the corresponding semantic embedding vector.

[0374] Step S2312: Calculate the similarity between the positive preference embedding vector in the multi-dimensional user profile vector and the semantic embedding vector of the industry corpus text content to obtain a first similarity score.

[0375] Step S2313: Screen the text content in the industry corpus with the top K first similarity scores to generate a first candidate push content set.

[0376] The said step S2320 includes:

[0377] Step S2321: Calculate the similarity between the negative preference embedding vector in the multi-dimensional user profile vector and the semantic embedding vector of the industry corpus text content to obtain a second similarity score.

[0378] Step S2322: Screen the text content in the industry corpus with the top M second similarity scores to generate a second candidate push content set.

[0379] The said step S2330 includes:

[0380] Step S2331: Obtain the semantic embedding vectors of the text content in the first candidate push content set and the second candidate push content set.

[0381] Step S2332: According to the semantic embedding vectors of the text content in the first candidate push content set and the second candidate push content set, calculate the semantic similarity S2 between each text content in the first candidate push content set and all text content in the second candidate push content set to form a similarity matrix.

[0382] Step S2333: In the similarity matrix, screen out the content pairs with a semantic similarity S2 higher than the second similarity threshold θ2 and mark them as highly overlapping content.

[0383] Step S2334: Remove the marked highly overlapping content from the first candidate push content set to obtain an adjusted push content set.

[0384] In the push content sorting module, the obtaining of the third candidate push content set includes:

[0385] Step S3100: Based on the adjusted push content set, construct and train a semantic adversarial generation network.

[0386] Step S3200: Obtain the real-time behavior data of the target user, and according to the real-time behavior data and the semantic adversarial generation network, obtain the third candidate push content set.

[0387] The said step S3100 includes:

[0388] Step S3110: Mark the adjusted push content set as positive samples, mark the second candidate push content set as negative samples, and construct a training data set for the semantic adversarial generation network.

[0389] Step S3120: Train the semantic adversarial generation network according to the training data set of the semantic adversarial generation network.

[0390] In the push content sorting module, the generation of the precise push content sequence includes:

[0391] Step S3300: Input the third candidate push content set into a pre-constructed push policy reinforcement learning model to obtain an optimal push policy; perform intelligent sorting on the content in the third candidate push content set according to the optimal push policy to generate a precise push content sequence.

[0392] Embodiment 4

[0393] This embodiment discloses an electronic device, which may include one or more processors and one or more memories. Among them, computer-readable code is stored in the memory, and when the computer-readable code is run by one or more processors, it can execute the above-mentioned large model information push method based on industry corpora.

[0394] The method or system according to the embodiments of the present application can also be implemented by means of the architecture of an electronic device. The electronic device may include a bus, one or more CPUs, a read-only memory (ROM), a random access memory (RAM), a communication port connected to a network, input / output components, a hard disk, etc. The storage device in the electronic device, such as ROM or hard disk, can store the large model information push method provided by the present application. The large model information push method based on industry corpora may include, for example: obtaining a positive behavior record sequence and a negative behavior record sequence of a target user; extracting a first behavior feature from the positive behavior record sequence and a second behavior feature from the negative behavior record sequence; generating a first behavior weight for the first behavior feature and a second behavior weight for the second behavior feature; generating a multi-dimensional user portrait vector according to the first behavior feature, the first behavior weight, the second behavior feature, and the second behavior weight; constructing a pre-trained industry language model, and obtaining a first semantic embedding vector and a second semantic embedding vector according to the multi-dimensional user portrait vector and the pre-trained industry language model; adjusting the first semantic embedding vector and the second semantic embedding vector according to a preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; obtaining an adjusted push content set according to the positive preference embedding vector and the negative preference embedding vector; obtaining a third candidate push content set based on the adjusted push content set; performing intelligent sorting on the content in the third candidate push content set to generate a precise push content sequence.

[0395] Furthermore, the electronic device may further include a user interface. Of course, the architecture disclosed in the present invention is only exemplary. When implementing different devices, one or more components of the electronic device disclosed in the present invention may be omitted according to actual needs.

[0396] Embodiment 5

[0397] This embodiment discloses a computer-readable storage medium with computer-readable instructions stored thereon. When the computer-readable instructions are run by a processor, the method for pushing large model information based on industry corpus according to the implementation manner of the present application can be executed. The storage medium includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0398] In addition, according to the implementation manner of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the present application provides a non-transitory machine-readable storage medium storing machine-readable instructions that can be run by a processor to execute instructions corresponding to the method steps provided by the present application, such as: obtaining a positive behavior record sequence and a negative behavior record sequence of a target user; extracting a first behavior feature from the positive behavior record sequence and a second behavior feature from the negative behavior record sequence; generating a first behavior weight for the first behavior feature and a second behavior weight for the second behavior feature; generating a multi-dimensional user portrait vector according to the first behavior feature, the first behavior weight, the second behavior feature, and the second behavior weight; constructing a pre-trained industry language model, and obtaining a first semantic embedding vector and a second semantic embedding vector according to the multi-dimensional user portrait vector and the pre-trained industry language model; adjusting the first semantic embedding vector and the second semantic embedding vector according to a preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; obtaining an adjusted push content set according to the positive preference embedding vector and the negative preference embedding vector; obtaining a third candidate push content set based on the adjusted push content set; and performing intelligent sorting on the content in the third candidate push content set to generate an accurate push content sequence. When this computer program is executed by a central processing unit (CPU), the above functions defined in the method of the present application are executed.

[0399] The methods, systems, and devices of the present application can be implemented in many ways. For example, the methods, systems, and devices of the present application can be implemented through software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is only for illustration purposes, and the steps of the method of the present application are not limited to the specific order described above, unless otherwise specifically stated. In addition, in some embodiments, the present application can also be implemented as a program recorded in a recording medium, and these programs include machine-readable instructions for implementing the method according to the present application. Therefore, the present application also covers a recording medium storing a program for executing the method according to the present application.

[0400] In addition, parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive elaboration.

[0401] As described above in the specific embodiments, the objectives, technical solutions, and beneficial effects of the present invention have been further described in detail. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A large model information push method based on industry corpus, characterized in that, The method includes: Obtaining a positive behavior record sequence and a negative behavior record sequence of a target user; extracting first behavior features from the positive behavior record sequence and second behavior features from the negative behavior record sequence; generating a first behavior weight for the first behavior features and a second behavior weight for the second behavior features; generating a multi-dimensional user profile vector based on the first behavior features, the first behavior weight, the second behavior features, and the second behavior weight. Constructing a pre-trained industry language model, and obtaining a first semantic embedding vector and a second semantic embedding vector based on the multi-dimensional user profile vector and the pre-trained industry language model; adjusting the first semantic embedding vector and the second semantic embedding vector according to a preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; using the negative preference embedding vector to perform semantic similarity retrieval on the text content in the industry corpus to generate a second candidate push content set; obtaining an adjusted push content set according to the positive preference embedding vector and the negative preference embedding vector. The generating of the positive preference embedding vector and the negative preference embedding vector includes: calculating the similarity between the first semantic embedding vector and the second semantic embedding vector to obtain a positive and negative preference similarity score D1; if the positive and negative preference similarity score D1 is greater than θ1, triggering an embedding vector adjustment process to obtain the positive preference embedding vector and the negative preference embedding vector. The embedding vector adjustment process includes: Step S2221: Randomly sample a noise vector, adjust the first semantic embedding vector according to the noise vector to obtain an adjusted first semantic embedding vector, marked as the positive preference embedding vector. Step S2222: Adjust the second semantic embedding vector according to the noise vector to obtain an adjusted second semantic embedding vector, marked as the negative preference embedding vector. Step S2223: Calculate the similarity D2 between the positive preference embedding vector and the negative preference embedding vector, and determine whether D2 is less than or equal to θ1. If so, complete the embedding vector adjustment process; otherwise, return to Step S2221, randomly sample a noise vector again, and repeat Steps S2221 - S2223 until D2 ≤ θ1. Based on the adjusted push content set, constructing and training a semantic adversarial generation network, where the adjusted push content set is marked as positive samples, and the second candidate push content set is marked as negative samples to construct a training data set; obtaining the real-time behavior data of the target user, and obtaining a third candidate push content set according to the real-time behavior data and the semantic adversarial generation network; performing intelligent sorting on the content in the third candidate push content set to generate a precise push content sequence.

2. The method for pushing large model information based on industry corpus according to claim 1, wherein The obtaining of the positive behavior record sequence and the negative behavior record sequence of the target user includes: Constructing an industry corpus, and obtaining the first historical behavior data of the target user in the industry field from the industry corpus; the first historical behavior data includes n1 behavior types of the target user, and the n1 behavior types include positive behaviors and negative behaviors; n1 is a positive integer. According to the first historical behavior data, obtaining the positive behavior record sequence and the negative behavior record sequence.

3. The method for pushing large model information based on industry corpora according to claim 1, wherein The extracting of the first behavior features from the positive behavior record sequence includes: Tokenize the text content in the positive behavior record sequence, and segment the text content into a sequence of words as the smallest semantic units; Perform part-of-speech tagging on the sequence of words after tokenization; Based on the part-of-speech tagging, identify the named entities in the text content; Extract keywords and topic words from the sequence of words; Use the keywords, topic words, and named entity words as the first behavior features.

4. The method for pushing large model information based on industry corpus according to claim 2, wherein, The first behavior weights for generating the first behavior features and the second behavior weights for generating the second behavior features include: Set the basic weights for different behavior types among n1 behavior types, and construct an initial behavior weight matrix; According to the first historical behavior data, count the behavior intensity distributions of different behavior types; Construct a behavior intensity quantization function, and according to the behavior intensity quantization function, map the behavior intensity distributions of different behavior types to the corresponding weight adjustment coefficients; Multiply the basic weights in the initial behavior weight matrix by the corresponding weight adjustment coefficients to generate an adjusted behavior weight matrix; In the adjusted behavior weight matrix, use the weights corresponding to positive behaviors as the first behavior weights for the first behavior features, and use the weights corresponding to negative behaviors as the second behavior weights for the second behavior features.

5. The large model information push method based on industry corpus according to claim 1, wherein, The generation of the multi-dimensional user profile vector includes: Construct a positive preference dimension according to the first behavior features and the first behavior weights; Construct a negative preference dimension according to the second behavior features and the second behavior weights; Fuse the positive preference dimension and the negative preference dimension to generate a multi-dimensional user profile vector.

6. The method for pushing large model information based on industry corpus according to claim 2, wherein The obtaining of the adjusted push content set includes: Use the positive preference embedding vector to perform semantic similarity retrieval on the text content in the industry corpus to generate a first candidate push content set; Based on the first candidate push content set and the second candidate push content set, obtain the adjusted push content set; The generation of the first candidate push content set includes: Convert the text content in the industry corpus into the corresponding semantic embedding vectors; Calculate the similarity between the positive preference embedding vector in the multi-dimensional user profile vector and the semantic embedding vectors of the text content in the industry corpus to obtain the first similarity score; Select the top K text contents in the industry corpus with the highest first similarity scores to generate a first candidate push content set.

7. The method for pushing large model information based on industry corpus according to claim 6, characterized in that The obtaining of the adjusted push content set based on the first candidate push content set and the second candidate push content set includes: Obtain the semantic embedding vectors of the text contents in the first candidate push content set and the second candidate push content set; According to the semantic embedding vectors of the text contents in the first candidate push content set and the second candidate push content set, calculate the semantic similarity S2 between each text content in the first candidate push content set and all text contents in the second candidate push content set to form a similarity matrix; In the similarity matrix, select the content pairs with the semantic similarity S2 higher than the second similarity threshold θ2 and mark them as highly overlapping contents; Remove the marked highly overlapping contents from the first candidate push content set to obtain the adjusted push content set.

8. An information push system of a large model based on industry corpus, which is used to implement the information push method of the large model based on industry corpus described in any one of claims 1-7, characterized in that, The system includes: Image construction module: used to obtain the positive behavior record sequence and negative behavior record sequence of the target user; extract the first behavior feature from the positive behavior record sequence and the second behavior feature from the negative behavior record sequence; generate the first behavior weight of the first behavior feature and the second behavior weight of the second behavior feature; generate a multi-dimensional user portrait vector according to the first behavior feature, the first behavior weight, the second behavior feature and the second behavior weight; Push content generation module: used to construct a pre-trained industry language model, obtain the first semantic embedding vector and the second semantic embedding vector according to the multi-dimensional user portrait vector and the pre-trained industry language model; adjust the first semantic embedding vector and the second semantic embedding vector according to the preset similarity threshold θ1 to generate a positive preference embedding vector and a negative preference embedding vector; obtain an adjusted push content set according to the positive preference embedding vector and the negative preference embedding vector; Push content sorting module: based on the adjusted push content set, obtain the third candidate push content set; perform intelligent sorting on the content in the third candidate push content set to generate an accurate push content sequence.

Images