High-dimensional user portrait construction method based on multi-source data fusion and dynamic update

Abstract

This invention relates to the fields of artificial intelligence and big data processing, specifically disclosing a method for constructing high-dimensional user profiles based on multi-source data fusion and dynamic updates. This method constructs an initial high-dimensional feature vector, introduces an immune mechanism to identify "self" and "non-self" behaviors, establishes a dynamic antibody library with a lifecycle, and combines clonal selection, high-frequency mutation, and natural apoptosis mechanisms to achieve adaptive evolution of the profile. It also constructs a cross-user immune network, utilizes common group behaviors to generate regulatory factors to correct individual biases, and integrates adversarial attack detection, version management, and viability assessment modules. Through the above technical solutions, this invention improves the accuracy, timeliness, and anti-interference capability of user profiles, providing highly reliable data support for precision marketing, intelligent recommendation, and risk prevention and control.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and big data processing, specifically involving a method for constructing high-dimensional user profiles based on multi-source data fusion and dynamic updates. Background Technology

[0002] With the deep integration of big data technology and artificial intelligence, user profiling, as a core technology for characterizing individual attributes and behavioral preferences, is playing an increasingly crucial role in areas such as precision marketing, personalized recommendations, and risk control. Through the mining and modeling of massive amounts of heterogeneous data, the system can construct a high-dimensional digital feature matrix, enabling accurate prediction of complex user behaviors. This technology not only improves the utilization efficiency of data assets but also provides indispensable underlying support for building intelligent decision support systems and achieving refined operations.

[0003] User profile construction based on multi-source data fusion and dynamic updates is an important research topic in the field of digital governance and information processing. This technology aims to integrate data streams distributed across different platforms with different structures, and construct a dynamic model reflecting the real-time status of users through feature mapping and weight allocation. In a rapidly changing information environment, the system needs to ensure that profile features can spontaneously adjust as user behavior evolves to maintain the freshness and accuracy of the profile. This requires the system to possess data parsing capabilities, feature evolution rate, and adaptability to complex environments.

[0004] Existing user profiling solutions largely rely on static weight allocation and historical trajectory fitting, making it difficult to eliminate noise and malicious injection from multi-source data. Traditional models lack mechanisms for identifying and filtering fake traffic, adversarial attacks, and dirty data, leading to semantic biases in the generated profiles and an inability to maintain robustness in contaminated data environments. Because feature update mechanisms often lag behind real-world scenarios, systems struggle to capture instantaneous shifts in user interests and changes in social attributes, causing profile models to become rigid. Existing solutions generally lack the ability to collaboratively correct common group characteristics and individual biases, making it difficult to achieve self-healing and self-adaptation in complex and dynamic interactions.

[0005] There is an urgent need for a method to construct high-dimensional user profiles based on multi-source data fusion and dynamic updates. Summary of the Invention

[0006] The purpose of this invention is to provide a method for constructing high-dimensional user profiles based on multi-source data fusion and dynamic updates, which can solve the problems in the background art mentioned above.

[0007] To achieve the above objectives, the technical solution adopted by this invention is as follows: a method for constructing high-dimensional user profiles based on multi-source data fusion and dynamic updating, comprising the following specific steps: Step 1: Construct an initial user profile model. Standardize user behavior data from multiple heterogeneous data sources and map it to a unified high-dimensional feature space to form an initial profile feature vector. Step 2: Define normal behavior patterns as the self-set. Establish a dynamic boundary for the self-set through historical behavior trajectories and common group characteristics. Identify abnormal behaviors deviating from this boundary as non-self signals. Step 3: Construct a dynamic profile antibody library. The antibody library consists of multiple profile feature subsets with lifecycles. Each subset corresponds to a type of user behavior pattern and has an affinity assessment mechanism. Step 4: Perform immune recognition on newly injected user behavior data. Determine whether the data matches the self-set. Whether to trigger antibody activation; if the matching degree is lower than a preset threshold, initiate a clonal selection mechanism to generate a high-affinity antibody variant; Step 5: Execute high-frequency mutation and natural apoptosis mechanisms. After the new antibody is activated, perform local structural mutations on the relevant feature subset to adapt to behavioral drift, and perform elimination operations on antibodies that have not been activated for a long time and whose affinity is consistently lower than a predetermined level; Step 6: Construct a cross-user immune network, generate immune regulatory factors based on the common behavioral characteristics of group users in a specific scenario, and use these factors to reverse-correct bias features in individual profiles; Step 7: Output a high-dimensional user profile optimized by dynamic evolution and cross-defense mechanisms for real-time decision support of downstream application systems.

[0008] Preferably, in step 1, the multi-source data includes, but is not limited to, social platform interaction logs, e-commerce transaction records, location service trajectories, and device usage behavior. All raw data undergoes semantic alignment and timestamp normalization before entering the feature space to ensure that behavioral events from different sources are comparable on a unified timeline.

[0009] Preferably, in step 2, the dynamic boundary of the self-set is determined by the behavior frequency distribution within the sliding time window and the stability of the cluster center. When a certain behavior remains stable in multiple consecutive window periods and is more than a predetermined distance from other behavior clusters, it is included in the self-set.

[0010] Preferably, in step 3, each subset of portrait features contains several basic attribute dimensions and their associated weights. The weights decay over time and are adjusted by affinity feedback. Affinity is calculated by the similarity between the current behavior and the pattern represented by the subset.

[0011] Preferably, in step 4, the immune recognition process adopts a multi-layer filtering strategy, which eliminates violations based on a rule engine, then filters potential non-self signals through a density-based anomaly detection algorithm, and finally verifies them by matching existing antibodies in the antibody library.

[0012] Preferably, in step 5, the high-frequency mutation mechanism is achieved by randomly perturbing and recombinating the characteristic dimensions of the activated antibody, with the perturbation amplitude dynamically adjusted according to the rate of behavioral change, while the natural apoptosis mechanism determines whether to remove the antibody based on the comprehensive score of antibody survival time and cumulative affinity.

[0013] Preferably, in step 6, the construction of the immune network depends on the user grouping results. The system periodically performs unsupervised clustering on all users and extracts the core behavioral patterns of each group as immune regulatory factors. When the difference between an individual profile and the pattern of the group to which it belongs exceeds a certain range, a correction process is triggered.

[0014] Preferably, in step 7, the output high-dimensional user profile not only includes static demographic attributes, but also integrates dynamic interest tags, behavioral intention predictions, and risk propensity indicators. All dimensions are marked with confidence level to indicate the reliability level of the feature at the current moment.

[0015] Preferably, the method further includes an adversarial attack detection module, which continuously monitors behavioral mutation patterns in the data stream. When a large number of similar behaviors are detected to occur in a short period of time and do not conform to the normal distribution pattern, the relevant data source is automatically isolated and a profile rollback mechanism is initiated.

[0016] Preferably, the method also includes a profile version management mechanism, in which the system generates a snapshot for each major update and records the reason for the update, and supports reverting to any historical version when necessary, ensuring the auditability and controllability of profile evolution.

[0017] Preferably, the method further includes a multi-granularity profile fusion strategy. For different application scenarios, the system can dynamically combine coarse-grained general profiles and fine-grained vertical domain profiles to form a task-oriented composite profile view.

[0018] Preferably, the method further includes a portrait freshness assessment unit, which calculates the overall portrait freshness score based on the time interval of the most recent behavior, the behavior diversity index, and the feature update frequency, and adjusts the trust weight of the portrait by downstream applications accordingly.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention introduces the core principle of biological immune mechanisms to construct a high-dimensional user profile system with self-recognition, self-adaptation and self-repair capabilities.

[0020] 2. The system can distinguish between normal behavior and noise interference, and demonstrates robustness when faced with malicious data injection such as order brushing and bot traffic.

[0021] 3. By assigning lifecycles to profile features and introducing a high-frequency mutation mechanism, the system can quickly respond to substantial changes in user interests, improving the timeliness and accuracy of profiles.

[0022] 4. By leveraging the corrective effect of herd immunity networks on individual biases, the stability and generalization ability of the profile are further enhanced in complex and dynamic environments.

[0023] 5. This invention realizes a paradigm shift in user profiling from static description to dynamic life form, providing a highly reliable data foundation for key applications such as precision marketing, intelligent recommendation, and risk prevention and control. Attached Figure Description

[0024] Figure 1 The flowchart is based on the present invention; Figure 2 This is a schematic diagram of the overall technical architecture and data flow of the high-dimensional user profile construction method based on multi-source data fusion and dynamic updating according to the present invention; Figure 3 This is a flowchart illustrating the construction, immune recognition, and clonal selection evolution of the profiling antibody library based on biological immune mechanisms according to the present invention. Figure 4 This is a flowchart illustrating the immune network construction and individual profile bias correction based on common behavioral characteristics of group users according to the present invention. Figure 5 This is a flowchart illustrating the multi-source heterogeneous data standardization processing, high-dimensional space mapping, and self-set dynamic boundary definition according to the present invention. Detailed Implementation

[0025] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0026] A method for constructing high-dimensional user profiles based on multi-source data fusion and dynamic updates is proposed. In the specific system execution logic, this method is implemented through the collaboration of a distributed computing platform and a real-time data processing engine.

[0027] Step 1 involves constructing an initial user profile model. User behavior data from multiple heterogeneous data sources is standardized and mapped to a unified high-dimensional feature space to form an initial profile feature vector. During this step, the system obtains raw behavior data from multiple heterogeneous data sources through a data access gateway. This multi-source data includes social media platform interaction logs, e-commerce transaction records, location service trajectories, and device usage behavior. For social media platform interaction logs, the system extracts the topic distribution of user posts, the hierarchical relationship of interactive objects, the sentiment polarity of comments, and the temporal density of forwarding behavior. For e-commerce transaction records, the system extracts features such as order amount, product category hierarchy, brand preference weight, promotional sensitivity, and return frequency. For location service trajectories, the system identifies the user's residence, workplace, and frequently visited points of interest types through spatiotemporal clustering of latitude and longitude sequences, and calculates the transition probability between different points of interest. For device usage behavior, the system records device hardware parameters, operating system version, the category distribution of installed applications, and daily active duration.

[0028] During the standardization process, all raw data undergoes semantic alignment and timestamp normalization before entering the feature space. Semantic alignment is achieved by constructing a globally unified ontology mapping table, mapping behavioral tags with the same meaning but different expressions from different platforms to the same standardized terms. For example, "add to cart" on e-commerce platforms and "favorite items" on social platforms are aligned in terms of intent. Timestamp normalization converts behavioral events from all sources into millisecond-level values ​​in Coordinated Universal Time (UTC), ensuring that behavioral events from different sources are comparable on a unified timeline. When mapping to a unified high-dimensional feature space, the system uses feature engineering techniques to transform unstructured data into fixed-length floating-point vectors. For categorical features, one-hot encoding is used for expansion, transforming features with N categories into N-dimensional binary vectors; for numerical features, a max-min scaling method is used to map them to the range between 0 and 1.

[0029] Step 2 defines normal behavior patterns as a self-set. A dynamic boundary for this self-set is established using historical behavior trajectories and common group characteristics. Abnormal behaviors deviating from this boundary are identified as non-self signals. In specific implementation, the dynamic boundary of the self-set is determined jointly by the behavior frequency distribution within a sliding time window and the stability of the cluster centers. The system sets a sliding time window of 30 natural days with a step size of 1 natural day. Within each window period, the system calculates the user's behavior frequency across various feature dimensions and uses a density clustering algorithm to find the center of the behavior distribution. When a certain type of behavior remains stable for more than 5 consecutive window periods, and the movement distance of its corresponding cluster center in the feature space is less than a preset stability threshold, and the distance to other known abnormal behavior clusters is greater than a predetermined safety distance, this type of behavior is included in the self-set. The system monitors newly generated behavior trajectories in real time, calculates the spatial distance between the new behavior point and the self-set cluster center, and if this distance exceeds the predefined dynamic boundary range, the behavior is marked as a non-self signal, indicating that it may be noise data, fraudulent behavior, or bot traffic.

[0030] Step 3 involves constructing a dynamic profile antibody library. This library consists of multiple profile feature subsets with lifecycles, each corresponding to a user behavior pattern and equipped with an affinity assessment mechanism. Each profile feature subset contains several basic attribute dimensions and their associated weights. Each antibody in the library represents a specific area of ​​interest or behavioral habit. To simulate the characteristics of a biological immune system, the system assigns an initial lifespan value to each antibody. During each update cycle, the system checks the antibody's activity; if the antibody is successfully activated, its lifespan increases; otherwise, it decays over time. The weight decay logic follows a linear or exponential model, meaning the feature weight gradually decreases over time. The affinity assessment mechanism calculates the cosine similarity between the current behavior vector and the vectors of each feature subset in the antibody library. The sum of the component products of the corresponding dimensions of the two vectors is calculated, and then divided by the product of the magnitudes of the two vectors; the resulting value is the affinity score. The affinity score reflects the degree of matching between the current behavior and the pattern represented by the antibody.

[0031] Step 4 involves performing immune recognition on the newly injected user behavior data. The system determines whether antibody activation is triggered based on the matching degree between the data and the user's self-assessment. If the matching degree is lower than a preset threshold, a clonal selection mechanism is initiated to generate high-affinity antibody variants. The immune recognition process employs a multi-layered filtering strategy. The first layer is hard filtering based on a rule engine. The system removes attack behaviors or dirty data according to preset blacklists, violation frequency thresholds, and logical contradiction judgment rules. The second layer is density-based anomaly detection algorithm filtering. The system calculates the point density of new data points in the local space. If the density is lower than the surrounding area, it is marked as a potential non-self signal. The third layer involves matching and verification with existing antibodies in the antibody library. The system calculates the maximum affinity score between the new data and all antibodies in the library. If this maximum affinity score is lower than a preset activation threshold, it indicates that the existing profiling model cannot accurately describe the new behavior. The system determines that the user may have experienced interest drift or exhibited a new behavioral pattern. The system then initiates a clonal selection mechanism, selecting several antibodies with relatively high affinity to the current behavior as parent antibodies, replicating and mutating them to generate antibody variants that accurately cover the new behavioral characteristics.

[0032] Step 5 involves implementing high-frequency mutation and natural apoptosis mechanisms. After new antibodies are activated, local structural mutations are performed on relevant feature subsets to adapt to behavioral drift. Antibodies that have not been activated for a long time and whose affinity remains below a predetermined level are eliminated. The high-frequency mutation mechanism is achieved by randomly perturbing and recombinating the feature dimensions of activated antibodies. During the mutation process, the system dynamically adjusts the perturbation amplitude according to the rate of change in user behavior. If a drastic change in user behavior is detected in a short period of time, such as a shift from workplace-related searches to maternal and infant-related purchases, the system increases the mutation probability and perturbation amount so that the profile model can quickly converge to the new state. Specific mutation operations include adding small deviation values ​​that conform to a normal distribution to specific components in the feature vector, or swapping the weight allocation of two related feature dimensions. The natural apoptosis mechanism is determined based on the antibody's survival time and cumulative affinity score. The system periodically traverses the antibody library. For antibodies that have not been activated for 90 consecutive natural days and whose cumulative affinity score moving average is below a predetermined threshold, the system determines that the behavioral pattern represented by the antibody is outdated or invalid, performs a physical deletion operation, and releases system storage resources.

[0033] Step 6 involves constructing a cross-user immune network. An immune regulatory factor is generated based on the common behavioral characteristics of a user group in a specific scenario, and this factor is used to correct biased features in individual profiles. The construction of the immune network relies on large-scale user grouping results. The system periodically analyzes the profile features of all users using unsupervised clustering algorithms (such as Expectation-Maximization clustering) to identify highly similar user groups. For each group, the system extracts the mean vector of its core behavioral patterns as the immune regulatory factor for that group. When the deviation of an individual's profile feature value in a specific dimension from the mean vector of its group exceeds three standard deviations, the system triggers a correction process. The immune regulatory factor acts on the individual feature vector with a decay coefficient, pulling it towards the group center. This cross-defense mechanism can identify and correct profile biases caused by individual abnormal transactions or misoperations, leveraging collective intelligence to enhance the robustness of individual profiles.

[0034] Step 7 outputs a high-dimensional user profile optimized by dynamic evolution and cross-defense mechanisms, used for real-time decision support in downstream application systems. The output high-dimensional user profile has a hierarchical structure, including not only static demographic attributes (such as age group, gender prediction, and occupational classification), but also dynamic interest tags (such as long-term and short-term interest distributions), behavioral intent predictions (such as the probability of the next purchase), and risk propensity indicators (such as account theft risk scores). All profile dimensions are labeled with confidence levels. Confidence levels are calculated based on the reliability of the data source, the activation frequency of antibodies, the time decay factor of behavior, and the correction strength of the immune network. When downstream application systems (such as personalized recommendation engines or risk control systems) access the profile data, they adjust the weights of different feature dimensions based on the confidence levels. If the confidence level of a certain interest tag is low, the recommendation system will reduce the exposure of related content to ensure the accuracy and reliability of decision support.

[0035] The method also includes an adversarial attack detection module. This module continuously monitors behavioral mutation patterns in the input data stream. When it detects a large number of behaviors with similar characteristics or originating from the same IP segment occurring within a short time step (e.g., within 10 minutes), and the statistical distribution characteristics of these behaviors do not conform to the normal probability distribution of the user or their group, the system determines that an adversarial attack has occurred. The system automatically isolates the relevant data sources, stops updating the current profile, and initiates a profile rollback mechanism to restore the user profile state to the previous security snapshot version before the attack.

[0036] The system also integrates a profile version management mechanism. Whenever a profile undergoes a significant evolutionary update (such as clonal selection producing a new core antibody or cross-population migration), the system automatically generates a snapshot containing complete feature vectors, antibody library status, and environmental parameters. The snapshot records the triggering reason for the update, the associated behavioral data ID, and a timestamp. Through the version management interface, administrators can trace the profile status at any historical moment, audit and analyze the evolution path, and ensure the controllability of the profile evolution process.

[0037] To meet the differentiated needs of various business scenarios, the method also includes a multi-granularity profile fusion strategy. The system stores complete fine-grained feature vectors at the underlying level, but can dynamically combine them at the output based on task requirements. In general global search scenarios, the system outputs a coarse-grained general profile view, retaining only the top-level interest category tags; in vertical domain precision marketing scenarios, the system calls upon fine-grained vertical domain profiles to display refined behavioral tendencies within specific categories. This strategy ensures data processing efficiency while enabling the construction of task-oriented composite profile views.

[0038] This method also includes a profile freshness assessment unit. This unit calculates the overall freshness score of the profile, taking into account the time interval between the most recent behavior, the diversity index of behavior types, and the update frequency of feature vectors. If no new data is injected into a profile for a long period, its freshness score will gradually decrease over time. Downstream applications will adjust the trust weight of the profile accordingly. When the freshness score falls below the minimum safety value, the system will trigger a dormant user activation strategy or mark the profile as dormant.

[0039] Example 2: In another specific application scenario, the present invention is applied to real-time risk control and precise user segmentation in a large e-commerce platform.

[0040] In step 1, the system accesses real-time transaction data, app clickstream data, search history, and user device fingerprint information from the e-commerce platform. For clickstream data, the system uses a distributed stream processing framework for windowed processing, extracting the user's path depth, page dwell time distribution, and bounce rate within a 30-minute session. For device fingerprints, it extracts records of unique device identifier changes, whether the device is jailbroken or rooted, and changes in geographical location. After standardization, all data is transformed into an initial profile feature vector with 1024 dimensions, each dimension representing a specific business characteristic or implicit representation.

[0041] In step 2, the dynamic boundary construction of the self-set takes into account the holiday effect. When calculating the frequency distribution of behavior, the system introduces a time period correction factor. For explosive purchasing behavior during major promotional events such as Singles' Day, the system identifies this as normal "promotion-sensitive self" behavior, rather than an abnormal "bot-driven order-faking" signal, by comparing it with common characteristics of the same period in history. The dynamic boundary automatically expands and contracts according to the start and end of the promotional period.

[0042] In step 3, the constructed profile antibody library was optimized for e-commerce business. For example, it specifically constructed "high-frequency, low-average-order-value purchase antibodies," "cyclical fast-moving consumer goods demand antibodies," and "digital product enthusiast antibodies." The weight allocation in each antibody subset reflects the user's engagement depth in that dimension. The antibody affinity calculation not only considers vector similarity but also introduces confidence interval determination.

[0043] In step 4, for a newly injected abnormally high-value order, the immune recognition module uses a first-layer filter to detect that the order's delivery address is too far from the user's historically frequently used address (over 500 kilometers). Then, a second-layer density detection is performed, revealing that this type of high-value order is an outlier in the global distribution for that time period. A third-layer antibody matching reveals that the affinity of this behavior with any existing subset of the user's antibodies is below 0.15. The system determines that cloning selection is triggered, but due to the high risk level, the system temporarily suspends the formal update of this profile and directs it to the variant library for verification.

[0044] In step 5, the system performs high-frequency mutation on the validated new behavioral patterns. For example, if a user transitions from a "single youth" to a "wedding preparation" state, their search keywords will drastically shift from "single-player games" to "wedding photography and wedding supplies." After a week of continuous activation and validation, the new antibodies generated by clonal selection show a sustained increase in affinity scores, triggering the inhibition and apoptosis process of the old "single-player games" antibody.

[0045] In step 6, the system identifies the user as belonging to the "preparing-for-wedding" group through a cross-user immune network. Utilizing common characteristics of this group (such as a shared demand for furniture and home appliances), the system generates immune regulatory factors to correct biases in the "home appliance preference" dimension of the individual profile caused by data gaps. If the individual profile has a weight of 0 for the "refrigerator" feature, but group characteristics show that users at this stage have an 80% probability of paying attention to refrigerators, the system automatically adjusts the individual's weight for that dimension upwards, achieving proactive prediction.

[0046] In step 7, the output high-dimensional user profile is pushed to the recommendation engine in real time. Based on the profile's "Wedding Preparation Confidence Level" of 0.92, the recommendation engine completely switches the personalized displays on the homepage to wedding-related themes. The risk control system, based on the profile's "Account Security Confidence Level" of 0.98, determines that the previous abnormally large orders were genuine significant family expenditures, unlocks the account, and completes the entire closed-loop process.

[0047] This embodiment further illustrates that, through multi-source data fusion and dynamic updates, the system can achieve accurate capture of user intent and real-time risk prevention and control in complex business environments.

[0048] Example 3: In another specific implementation scenario, the present invention is applied to interest evolution tracking and anti-fraud in mobile information applications.

[0049] In step 1, the collected data focuses on content consumption characteristics, including keyword tags for reading articles, percentage of time spent on the page, scrolling speed, and interactive behavior at the bottom of the article. During the standardization process, natural language processing techniques are used to transform the article title and body text into topic vectors, which are then mapped to a 512-dimensional high-dimensional space.

[0050] In step 2, the normal behavior boundary defines the user's "reading comfort zone." If a user suddenly clicks on 50 unrelated articles within 1 minute, and the duration of each click is less than 2 seconds, the system calculates that this behavior trajectory completely deviates from the dynamic boundary of the self-set, identifies it as a "non-self signal," and determines it to be crawled or maliciously inflated.

[0051] In step 3, the profile antibody library reflects the user's knowledge structure. For example, "deep technology reading antibody" has a longer lifespan and slower weight decay, while "breaking hot news antibody" has a short lifespan; if a hot event is not clicked again within 3 days, the relevant antibody enters its death countdown.

[0052] In the immune recognition step 4, when a user begins reading about a completely new field (such as switching from "literature" to "finance" due to a job change), the system generates a new "financial knowledge antibody" through cloning selection. In the early stages of evolution, this antibody is in an observed state, and its confidence level increases with the depth of reading.

[0053] In step 5, the high-frequency mutation mechanism enables the profile to capture subtle refinements of user interests, ranging from broad variations like "sports" to more specific ones like "Premier League" and "tactical analysis." For older interests such as "childhood animation," which users have completely lost interest in, the system uses a natural attrition mechanism to reduce their feature weights to extremely low levels, preventing them from interfering with current accurate recommendations.

[0054] In step 6, using an immune network, the system discovers that the user's behavioral patterns highly match those of a group of "professional investors." Through immune regulatory factors, the system can identify emerging financial topics that are popular within this group but not yet reflected in the individual's profile, thus achieving "discovery-based recommendations."

[0055] In step 7, the output high-dimensional profile supports millisecond-level sorting of the news feed. Each candidate news item is compared with the antibody clusters in the profile for affinity, and only the candidate with the highest affinity score-confidence product is displayed. The "bot probability index" in the profile is used in the downstream advertising billing system to automatically eliminate invalid clicks generated by fraudulent traffic.

[0056] This embodiment demonstrates the superiority of the present invention in handling high-frequency, fine-grained interest drift, and achieves adaptive evolution of the profile through a biological immune mechanism.

[0057] Example 4: The present invention also has application value in the scenario of building user profiles in the context of IoT devices.

[0058] In step 1, the data source is expanded to include data collected from smart home sensors, such as the opening and closing times of smart door locks, brightness adjustment habits of smart lighting, and temperature preferences of smart air conditioners. Standardization processing transforms the sensor values ​​from the physical world into feature vectors representing user lifestyle habits.

[0059] In step 2, the self set establishes a dynamic boundary for the "stay-at-home mode" by analyzing the family life trajectory over the past three months. For example, the peak water and electricity consumption period from 7:00 to 8:00 AM on weekdays is defined as the core self set. If there is continuous electrical appliance turning on at 3:00 AM on a certain day, the system identifies this non-self signal through dynamic boundary detection.

[0060] In step 3, the antibodies in the profiling antibody library correspond to different life scenarios, such as "weekday early rising scenario" and "weekend home leisure scenario". The affinity of the antibody is calculated by the pattern matching degree of the sensor sequence.

[0061] In step 4, if the family structure changes (e.g., an elderly person moves in), the newly generated behavioral data shows a lower match with the existing antibody database of "young couples." The system then initiates cloning selection to generate new antibody variants adapted to the "lifestyle habits of the elderly."

[0062] In step 5, through high-frequency mutation, the system continuously fine-tunes its automatic adjustment preferences for indoor temperature and brightness. The original low-temperature preference antibody targeting young people naturally dies due to decreased affinity, and is replaced by a moderate-temperature preference antibody targeting the elderly.

[0063] In step 6, the immune network generates energy-saving adjustment factors by comparing the energy usage patterns of thousands of similar households. If an individual profile shows that its energy consumption is higher than the common characteristics of similar households, the system uses a reverse correction mechanism to mark the profile with a "high energy-saving potential" label and triggers the optimization operation strategy of smart devices.

[0064] In step 7, the output high-dimensional profile provides decision support for the smart home control system, realizing the transformation from passively responding to commands to actively sensing and adaptively adjusting.

[0065] Example 5: This example focuses on describing the defense process of this method when facing large-scale adversarial attacks.

[0066] In step 1, a large amount of manually simulated click data was mixed into the data stream received by the system. This data statistically attempts to mimic the click rate and dwell time of normal users, but it exhibits mechanical characteristics in the distribution of micro-time intervals.

[0067] In step 2, the self-set boundary measures the randomness of behavior by introducing the concept of "entropy." Normal user behavior often exhibits a degree of randomness and unpredictability, while the behavior generated by the attack script displays a high degree of determinism. The system detects that the entropy value of newly entered data is significantly lower than the average level of the self-set, identifying it as a non-self signal.

[0068] In step 3, the system constructs specialized "defensive antibodies" that are specifically designed to capture known attack patterns.

[0069] In step 4, the immune recognition module initiates adversarial attack detection. When a high structural similarity is detected among a large number of newly activated antibody variants within a short period of time, the system determines that this is not a genuine user interest drift, but a cooperative attack.

[0070] In step 5, the system limits the high-frequency mutation rate for such suspicious behavior to prevent the profiling model from being "poisoned" by malicious data.

[0071] In step 6, the cross-user immune network comes into play. Since attacks are often conducted across multiple accounts, the system analyzes group characteristics and identifies multiple different accounts exhibiting completely consistent anomalous deviations within the same timeframe. The immune network generates a strong negative feedback modulator, blocking profile updates for these accounts.

[0072] In step 7, the risk propensity index of the relevant account is marked as the highest level in the output profile results, and the downstream system immediately triggers the verification code verification or manual review process.

[0073] In summary, the embodiments of this invention construct a dynamic, robust, and insightful high-dimensional user profiling system by simulating the self-recognition, self-evolution, and network regulation mechanisms of the biological immune system. This system can accurately filter noise, adapt to substantial changes in user behavior in real time, and maintain a high degree of realism and accuracy in complex environments involving multi-source data fusion.

[0074] In this invention, all numerical judgment logic is not expressed using mathematical formulas, but rather described through specific logical steps. For example, when calculating feature weight decay, the system performs the following operations: obtaining the time difference between the current moment and the last feature update moment, dividing this time difference by a preset decay period to obtain the decay factor; then multiplying the original weight by the negative power of the decay coefficient (achieved through iterative multiplication) to obtain the updated weight value. When calculating spatial distance, the system performs the following operations: extracting the component values ​​of the two vectors in each dimension, calculating the difference between the corresponding dimension component values; summing the squares of all differences, and taking the square root of the sum, the result is the Euclidean distance between the two vectors in the feature space. When performing affinity threshold judgment, the system compares the calculated similarity value with a floating-point number pre-stored in the configuration table. If the similarity value is greater than or equal to the floating-point number, the activation branch logic is executed; otherwise, the ignore or mutation branch logic is executed.

[0075] The unsupervised clustering process in the system is implemented through the following logic: a fixed number of initial centroids are randomly selected in the feature space; the textual description distance from each data point to each centroid is calculated, and the data points are classified into the clusters to which the nearest centroids belong; for each cluster, the arithmetic mean of the components of all data points in each dimension is calculated, and the mean vector is used as the new centroid of the cluster; the above classification and centroid update process is repeated until the moving distance of the centroids is less than the set minimum value or the preset iteration limit is reached, thus completing the automatic division of the user group.

[0076] In the specific operation of the clone selection mechanism, the system sorts the existing antibodies from high to low according to their affinity scores; selects the top 10% of antibodies as seed antibodies; determines the number of copies of each seed antibody based on its affinity, with higher affinity resulting in more clones; performs random bit flipping or numerical fine-tuning on the feature vector of each clone to introduce mutations; and puts the mutated clones into the next round of immune recognition testing, retaining the best-performing variants for inclusion in the formal antibody library.

[0077] In profile version management, snapshot storage employs incremental storage technology. The system only records the feature dimensions and values ​​that have changed from the previous version. When backtracking is needed, the system starts from the latest full snapshot and applies these incremental change records in reverse or forward order, reconstructing the profile state at any historical moment in memory. This approach optimizes storage space usage while ensuring auditability.

[0078] The execution logic of the multi-granularity profile fusion strategy is as follows: The system predefines a tree structure containing multi-level categories, with each low-level feature mapped to a leaf node of the tree. When an application requests a coarse-grained profile, the system performs an upward aggregation operation, weighting or summing the feature weights of leaf nodes belonging to the same parent node to generate a higher-level category score. When a fine-grained profile is requested, the system transmits the original feature values ​​of the leaf nodes.

[0079] The calculation process for the profile freshness assessment unit is as follows: Obtain the difference between the current system time and the most recent behavior timestamp (referred to as the stagnation duration); calculate the proportion of unique antibodies accessed by the user in the past 7 calendar days to the total antibody library (referred to as the diversity ratio); obtain the total number of dimensions of the feature vector that changed in the past 24 hours (referred to as the update frequency). The freshness score equals the product of the diversity ratio and the update frequency, divided by the logarithmic value of the stagnation duration. The system maps this score to a range of 0 to 100, serving as an important reference for profile trustworthiness.

[0080] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing high-dimensional user profiles based on multi-source data fusion and dynamic updating, characterized in that, Includes the following steps: Step 1: Construct an initial user profile model by standardizing user behavior data from multiple heterogeneous data sources and mapping the user behavior data to a unified high-dimensional feature space to form an initial profile feature vector. Step 2: Define normal behavior patterns as a set of self-identities, establish dynamic boundaries of self-identities through historical behavior trajectories and common characteristics of the group, and identify abnormal behaviors that deviate from the dynamic boundaries as non-self signals; Step 3: Construct a dynamic profile antibody library. The dynamic profile antibody library consists of multiple profile feature subsets with lifecycles. Each profile feature subset corresponds to a type of user behavior pattern and has an affinity assessment mechanism. Step 4: Perform immune recognition on the newly injected user behavior data, and determine whether to trigger antibody activation based on its matching degree with the self set. If the matching degree is lower than a preset threshold, start the clonal selection mechanism to generate a high-affinity antibody variant. Step 5: Execute high-frequency mutation and natural apoptosis mechanism. After the high-affinity antibody variant is activated, local structural mutations are performed on the feature subset to adapt to behavioral drift, and elimination operations are performed on antibodies that have not been activated for a long time and whose affinity is consistently below the predetermined level. Step 6: Construct a cross-user immune network, generate immune regulatory factors based on the common behavioral characteristics of group users in the scenario, and use the immune regulatory factors to reverse correct the bias features in individual profiles. Step 7: Output a high-dimensional user profile optimized by dynamic evolution and cross-defense mechanisms for real-time decision support of downstream application systems.

2. The method for constructing a high-dimensional user profile based on multi-source data fusion and dynamic updating according to claim 1, characterized in that, Step 1 specifically includes: Raw behavioral data is obtained from the multiple heterogeneous data sources through a data access gateway. The multi-source data includes social platform interaction logs, e-commerce transaction records, location service trajectories, and device usage behavior. Based on the social platform interaction logs, extract the topic distribution of user posts, the hierarchical relationship of interactive objects, the sentiment polarity of comments, and the temporal density of forwarding behavior; For the e-commerce transaction records, extract features such as order amount, product category level, brand preference weight, promotion sensitivity, and return frequency; For the location service trajectory, spatiotemporal clustering of latitude and longitude sequences is performed to identify the user's residence, workplace, and frequently accessed point of interest types, and the transition probability between different points of interest is calculated; In the standardization process, a globally unified ontology mapping table is constructed to map behavioral tags with the same meaning but different expressions from different platforms to the same standardized term, and semantic alignment is performed. All behavioral events from all sources are uniformly converted to millisecond-level values ​​in Coordinated Universal Time (UTC), and timestamp normalization is performed. Feature engineering techniques are used to transform unstructured data into fixed-length floating-point vectors. Categorical features are expanded using one-hot encoding, while numerical features are mapped to the range between zero and one by calculating the difference between the maximum and minimum values ​​and scaling them.

3. The method for constructing a high-dimensional user profile based on multi-source data fusion and dynamic updating according to claim 2, characterized in that, Step 2 specifically includes: Set the length of the sliding time window and the step size, calculate the frequency of user behavior on each feature dimension within each window period, and use density clustering algorithm to determine the center of behavior distribution; When a certain type of behavior remains stable for multiple consecutive window periods, and the movement distance of its corresponding cluster center in the feature space is less than a preset stability threshold, while the distance to other known abnormal behavior clusters is greater than a predetermined safety distance, the behavior is included in the self set. Real-time monitoring of newly generated behavioral trajectories, calculation of the spatial distance between the new behavioral point and each cluster center in the self set, the calculation logic of the spatial distance is as follows: extract the component values ​​of two vectors in each dimension and calculate the component difference in the corresponding dimension, sum the squares of all component differences, and finally perform a square root operation on the sum. If the spatial distance exceeds the predefined dynamic boundary range, the behavior is marked as the non-self signal.

4. The method for constructing a high-dimensional user profile based on multi-source data fusion and dynamic updating according to claim 3, characterized in that, In step 3, the subset of portrait features includes several basic attribute dimensions and their associated weights. The execution steps of the affinity assessment mechanism include: Each antibody in the antibody library is assigned an initial life value. The activity of the antibody is checked in each update cycle. If the antibody is successfully activated, its life value is increased. If it is not activated, its life value is decreased over time. Perform a weight decay operation, obtain the time difference between the current time and the last time the feature was updated, divide the time difference by the preset decay period to obtain the decay factor, and multiply the original weight by the negative power of the decay coefficient to obtain the updated weight value. The cosine similarity between the current behavior vector and the feature subset vectors in the antibody library is calculated to obtain the affinity score. The specific calculation process is as follows: calculate the sum of the products of the corresponding dimensions of the two vectors, and then divide by the product of the magnitudes of the two vectors.

5. The method for constructing a high-dimensional user profile based on multi-source data fusion and dynamic updating according to claim 4, characterized in that, In step 4, the immune recognition process employs a multi-layer filtering strategy, specifically including: The first layer of filtering is performed, based on the rule engine, which removes abnormal behavior or dirty data according to the preset blacklist, violation frequency threshold and logical contradiction judgment rules. The second layer of filtering is performed by using a density-based anomaly detection algorithm to calculate the point density of new data points in the local space. If the point density is lower than the average density of the surrounding area, it is marked as a potential non-self signal. The third layer of filtering is performed to calculate the maximum affinity score between the new data and all antibodies in the antibody library. If the highest affinity score is lower than the preset activation threshold, it is determined that the user has experienced interest drift or a new behavioral pattern has emerged. When the clonal selection mechanism is activated, the existing antibodies are sorted from high to low according to their affinity scores, and the top-ranked antibodies are selected as seed antibodies. The number of copies of each seed antibody is determined by the affinity; the higher the affinity, the more clones are generated.

6. The method for constructing a high-dimensional user profile based on multi-source data fusion and dynamic updating according to claim 5, characterized in that, Step 5 specifically includes: The high-frequency mutation mechanism is achieved by randomly perturbing and recombinating the characteristic dimensions of the activated antibody. The perturbation amplitude is dynamically adjusted according to the rate of change of user behavior. If the amount of change in user behavior within a preset time exceeds the change threshold, the mutation probability and perturbation amount are increased. Mutation operations include adding a small, normally distributed bias value to the components of a feature vector, or swapping the weights of two feature dimensions. The natural apoptosis mechanism periodically traverses the antibody library. For antibodies that have not been activated for a predetermined number of consecutive days and whose cumulative affinity score moving average is lower than a predetermined elimination threshold, the behavioral pattern represented by the antibody is determined to be invalid, and a physical deletion operation is performed to release storage resources.

7. The method for constructing a high-dimensional user profile based on multi-source data fusion and dynamic updating according to claim 6, characterized in that, Step 6 specifically includes: The profile features of all users are analyzed using an unsupervised clustering algorithm to identify user groups with high similarity. The unsupervised clustering process is implemented through the following logic: In the feature space, a fixed number of initial center points are randomly selected. The distance from each data point to each center point is calculated, and the data points are classified into the cluster to which the nearest center point belongs. For each cluster, the arithmetic mean of the components of all data points in each dimension is calculated. The average vector is used as the new center point of the cluster. The classification and center point update process is repeated until the moving distance of the center point is less than the preset minimum value. Extract the mean vector of the core behavioral patterns of each user group as the immune regulatory factor for that user group; When the deviation of an individual's profile feature value in a dimension from the mean vector of its group exceeds a preset multiple of the standard deviation, a correction process is triggered. The immune regulatory factor, combined with the attenuation coefficient, is applied to the individual feature vector to pull it toward the center of the group, thereby correcting the profile deviation caused by abnormal transactions or misoperations.

8. The method for constructing a high-dimensional user profile based on multi-source data fusion and dynamic updating according to claim 7, characterized in that, In step 7, the output high-dimensional user profile has a hierarchical structure, specifically including: integrating static demographic attributes, dynamic interest tags, behavioral intention prediction and risk propensity indicators; Confidence markers are added to all profile dimensions. These confidence markers are calculated by comprehensively considering the reliability score of the data source, the activation frequency of the antibody, the time decay factor of the behavior, and the correction strength of the immune network. When downstream application systems call the profile data, they adjust the weights of different feature dimensions according to the confidence level marker. If the confidence level of a certain feature dimension is lower than the preset security threshold, the influence of that feature dimension is reduced during the decision support process.

9. The method for constructing a high-dimensional user profile based on multi-source data fusion and dynamic updating according to claim 8, characterized in that, The high-dimensional user profile construction method also includes adversarial attack detection and version management, specifically including: The adversarial attack detection module continuously monitors the behavioral mutation patterns in the input data stream and calculates the entropy value of the behavioral trajectory to measure the randomness of the behavior. If the entropy value of the newly entered data is lower than the average entropy value of the self set, it is identified as a non-self signal. When a large number of behaviors with similar characteristics or from the same address range are detected to occur within a time step, and the statistical distribution of the behaviors does not conform to the normal probability distribution, it is determined that an adversarial attack has been launched, the data source is automatically isolated and the profile rollback mechanism is initiated. A profile version management mechanism is implemented to generate a snapshot containing complete feature vectors, antibody library status, and environmental parameters when the profile undergoes a major evolution update. Incremental storage technology is used to record only the feature dimensions and values ​​that have changed in the current version relative to the previous version. When backtracking is needed, starting from the latest full snapshot, incremental change records are applied in chronological order to reconstruct the profile state at any historical moment in memory.

10. The method for constructing a high-dimensional user profile based on multi-source data fusion and dynamic updating according to claim 9, characterized in that, The high-dimensional user profile construction method also includes multi-granularity profile fusion and vividness evaluation, specifically including: A tree structure containing multi-level categories is predefined, and each low-level feature is mapped to a leaf node of the tree structure; When an application requests a coarse-grained profile, an upward aggregation operation is performed, which weights the feature weights of leaf nodes belonging to the same parent node to average or sum them, generating a higher-level category score. When a fine-grained profile is requested, the original feature values ​​of the leaf nodes are directly passed through. The overall freshness score of the profile is calculated through the profile freshness evaluation unit. The calculation logic is as follows: the difference between the current system time and the most recent behavior timestamp is obtained as the stagnation time; the proportion of the number of independent antibodies reached by the user within the preset time to the total antibody library is calculated as the diversity ratio; and the total number of dimensions of the feature vector that have changed within the preset period is obtained as the update frequency. Multiply the diversity ratio by the update frequency, then divide by the logarithmic mapping value of the stagnation time, and map the calculation result to a preset score range as a reference for the profile trust level. When the freshness score is lower than the minimum safety value, trigger the user activation strategy or mark the profile as dormant.

Images