Method and system for precision marketing based on multi-channel user portrait

Abstract

This invention discloses a precise marketing method and system based on multi-channel user profiles, belonging to the field of intelligent marketing technology. It achieves cross-channel identity alignment by binding de-identified dynamic association tags to behavioral fragments from various channels as a unified anchor point, while protecting privacy. Secondly, it employs a two-stage matching mechanism: when fragments are insufficient, the fuzzy inference level is dynamically adjusted based on data volume and dispersion, outputting a matching confidence score; once the score is met, an initial profile is generated. All matched fragments must undergo a verification and observation period before being merged with the current profile, triggering incremental retraining for continuous iteration and updates. Finally, when marketing is triggered, the latest profile is called, and push notifications are only executed after dual verification of tag matching and cross-channel compliance. This solution achieves adaptive and precise mapping and continuous iteration from multi-channel fragments to profiles.

Description

Technical Field

[0001] This invention relates to the field of intelligent marketing technology, specifically to a precise marketing method and system based on multi-channel user profiles. Background Technology

[0002] With the rapid development of mobile internet and digital marketing, businesses typically reach users through multiple channels such as mobile applications, WeChat mini-programs, offline stores, third-party e-commerce platforms, and social media. User behavior data from these different channels is scattered and stored in their respective independent systems, creating a typical data silo phenomenon. How to accurately link this fragmented data to the same real user, build a unified user profile, and implement targeted marketing accordingly is a core technical challenge currently facing the field of targeted marketing.

[0003] In existing technologies, cross-channel user identification mainly employs ID-Mapping technology, which links the same user across different channels using strong feature attributes such as phone number and device ID. However, ID-Mapping relies excessively on strong features like phone numbers, and when these features are missing, encrypted, or anonymized, the matching accuracy drops sharply. Furthermore, existing mechanisms struggle to effectively distinguish between ordinary users and members, leading to confusion in identity tags. Current technologies employ a match-and-merge strategy, directly incorporating matching fragments into the main profile. This approach has serious risks: if a fragment matches incorrectly, the erroneous information directly contaminates the user profile, and all subsequent training and marketing decisions based on that profile will be built on erroneous data, leading to continuous accumulation of errors that are difficult to trace and correct. Simultaneously, marketing content delivery lacks accuracy verification. Existing technologies typically deliver content based on match, without implementing dual verification mechanisms for profile tag matching accuracy and cross-channel compliance, resulting in insufficient marketing delivery accuracy and a poor user experience.

[0004] Therefore, in order to address the above issues, there is an urgent need for precise marketing methods and systems based on multi-channel user profiles. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a precise marketing method and system based on multi-channel user profiles. It solves the problem that directly merging fragmented user profiles from multiple channels leads to erroneous fragments contaminating the main profile and causing continuous accumulation of errors in subsequent training and marketing decisions.

[0006] To achieve the above objectives, this invention provides the following technical solution: a precision marketing method based on multi-channel user profiles, comprising the following steps: collecting fragments of user behavior from various channels, binding each fragment with a de-identified dynamic association tag as a unified anchor point for cross-channel identity alignment, and associating all fragments of the same user with the same tag; performing fragment matching in two stages with a preset fragment quantity threshold: the first stage is when the fragment quantity threshold is not reached, dynamically adjusting the fuzzy inference level based on the current collected fragment quantity and the dispersion of the inference results, processing fragments based on the adjusted level, outputting the first score of single fragment features, the second score of inter-fragment association, and fusing to obtain the third score of matching confidence, and entering a verification observation period for matched fragments without merging, entering the second stage after the fragment quantity threshold is reached, generating an initial profile, and matching new fragments with the latest version profile as a reference, entering a verification observation period after matching, merging with the current profile after verification, triggering the retraining and updating of the profile to the next version profile after each merging, and continuously iterating without end; when a marketing trigger request is received, calling the latest version profile, and sequentially performing profile tag matching degree verification and cross-channel delivery compliance verification on the candidate marketing content, and executing precise push for the corresponding channel if both verifications pass.

[0007] Furthermore, the de-identified dynamic association tag is generated based on the user's irreversible encrypted identifier in each channel. There is no reversible mapping relationship between the de-identified dynamic association tag and the user's identity information. It is dynamically updated according to a preset period as the user's cross-channel behavior data accumulates. The updated de-identified dynamic association tag inherits all the fragments associated with the original de-identified dynamic association tag.

[0008] Furthermore, the dynamic adjustment of the fuzzy inference level includes: presetting at least three inference levels, each corresponding to a different set of fuzzy inference parameters. The inference level is the sensitivity level of the membership function during fuzzy inference, with a high inference level corresponding to a narrowband high-sensitivity membership function, a medium inference level corresponding to a wideband medium-sensitivity membership function, and a low inference level corresponding to a wideband low-sensitivity membership function; determining the baseline inference level based on the interval where the currently collected fragments are located, and then adjusting the baseline inference level based on the dispersion of the inference results. The higher the dispersion, the more it shifts towards the higher inference level, and the lower the dispersion, the more it shifts towards the lower inference level.

[0009] Furthermore, the first score of the single fragment feature is calculated based on the membership degree of the channel identifier, behavior type, timestamp, and content attributes carried by the fragment and the de-identified dynamic association tag; the second score of the fragment association is calculated based on the degree of coordination between the fragment and other collected fragments in time series, behavior series, and spatial trajectory; the third score of the matching confidence is obtained by dynamically weighting and fusing the first score and the second score, and the dynamic weight tilts towards the second score as the number of currently collected fragments increases.

[0010] Furthermore, the duration of the verification observation period is dynamically determined based on the credibility of the source channel, the completeness of the attributes, and the third score of the matching confidence of the fragment. The condition for the verification to be merged with the current profile is that, during the verification observation period, the behavioral consistency index and attribute consistency index of the fragment and the current profile in at least two preset dimensions continuously reach or exceed the corresponding dimension thresholds. A virtual marketing verification mechanism is also set up during the verification observation period. Virtual marketing content is determined and pushed based on the historical behavior of the user corresponding to the fragment. The user's response to the virtual marketing content is collected, and the response result is compared with the user's historical behavior pattern. When the comparison result exceeds the preset deviation range, the fragment is determined to fail the verification. The historical behavior pattern is the behavior pattern in the historical behavior data collected before the user corresponding to the fragment enters the verification observation period.

[0011] Furthermore, the step of triggering retraining and updating the portrait to the next version portrait after each merge includes: using the merged fragment set as training samples, performing incremental updates on the label weights and feature vectors of the current version portrait to generate the next version portrait. Each version portrait carries a version number and a timestamp. When a marketing trigger request is received, the portrait corresponding to the latest version number is called. When the core feature change of the portrait is lower than a preset change threshold after a preset number of consecutive merges, full retraining is triggered to replace the incremental update.

[0012] Furthermore, the profile tag matching degree verification specifically includes: extracting the target tags carried by the candidate marketing content, matching each target tag with the tags in the latest version of the profile, calculating the semantic similarity and weight matching degree between each target tag and the tags in the latest version of the profile, and comprehensively obtaining the matching degree score between the candidate marketing content and the current user profile. When the matching degree score reaches the preset accurate delivery threshold, the candidate marketing content passes the profile tag matching degree verification.

[0013] Furthermore, the cross-channel delivery compliance verification includes: obtaining the user's marketing authorization status and frequency limit on each channel; verifying, based on the target delivery channel corresponding to the candidate marketing content, whether the number of marketing messages the user has received on that channel within the current time period has not exceeded the frequency limit of that channel, and whether the user has valid authorization to receive that type of marketing content on that channel. The valid authorization indicates that the user has authorization for that type of marketing content on that channel and that the authorization is valid. When the number of marketing messages the user has received on that channel within the current time period has not exceeded the frequency limit of that channel, and the user has valid authorization to receive that type of marketing content on that channel, the candidate marketing content passes the cross-channel delivery compliance verification.

[0014] Furthermore, a feedback collection mechanism is also set up during the verification observation period: when the fragment enters the verification observation period, the subsequent behavior feedback of the user corresponding to the fragment on at least one channel is monitored. When the subsequent behavior feedback is consistent with the predicted behavior model of the current profile, the remaining duration of the verification observation period is shortened according to a preset shortening rule. When the subsequent behavior feedback is inconsistent with the predicted behavior model of the current profile, the remaining duration of the verification observation period is extended according to a preset extension rule.

[0015] The precision marketing system based on multi-channel user profiles applies the aforementioned precision marketing method based on multi-channel user profiles, including: a multi-channel fragment collection module, used to collect fragments of user behavior from various channels, binding each fragment with a de-identified dynamic association tag as a unified anchor point for cross-channel identity alignment, with all fragments of the same user associated with the same tag; and a fragment matching module, which performs fragment matching in two stages based on a preset fragment quantity threshold: the first stage occurs when the fragment quantity threshold is not reached, dynamically adjusting the fuzzy inference level based on the currently collected fragment quantity and the dispersion of the inference results, processing fragments based on the adjusted level, and outputting the first score of single fragment features and fragment... The second score is obtained by correlation, and the third score is obtained by fusion. The matched fragments enter the verification and observation period and are not merged. After the fragment quantity threshold is reached, the second stage is entered to generate the initial profile. Subsequent new fragments are matched with the latest version profile as a reference. After matching, they enter the verification and observation period. If the verification is successful, they are merged with the current profile. After each merge, the profile is triggered to be retrained and updated to the next version profile. The iteration continues without end. The precise marketing matching module is used to call the latest version profile when a marketing trigger request is received. It performs profile tag matching degree verification and cross-channel delivery compliance verification on the candidate marketing content in sequence. If both verifications are successful, the precise push of the corresponding channel is executed.

[0016] The present invention has the following beneficial effects: This precision marketing method and system based on multi-channel user profiles replaces plaintext ID splicing with de-identified dynamic association tags, cutting off the direct flow of personal identity information at the source and meeting user privacy protection requirements. Simultaneously, the dynamic update and inheritance mechanism of tags ensures accurate cross-channel identity matching without data leaving the domain, overcoming the industry challenge of balancing privacy protection and data integration. Addressing the cold start challenge in low-fragmentation phases, a dynamic fuzzy inference based on tiered settings is designed to adapt matching accuracy under different data densities. More importantly, a verification observation period is introduced as a firewall, preventing erroneous fragments from contaminating the core profile through behavioral consistency comparison and even virtual marketing probes. This closed-loop logic of isolation, merging, and retraining completely avoids the cascading error risk of a single mis-spelling in existing technologies. Abandoning traditional static profiles, it combines incremental retraining with full retraining to enable the profile to have endless self-iterative capabilities, capturing real-time shifts in user interests. At the marketing end, setting up dual verification gates for tag matching and cross-channel compliance not only ensures semantic alignment between content and user needs, but also prevents multi-channel bombardment through frequency and authorization control, significantly improving user experience and conversion efficiency, which is different from the blind push of existing technologies.

[0017] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0018] Figure 1 This is a flowchart of the precision marketing method based on multi-channel user profiling according to the present invention.

[0019] Figure 2 This is a schematic diagram illustrating the logic of the precision marketing method based on multi-channel user profiling in this invention.

[0020] Figure 3 This is a flowchart illustrating the logic of de-identification and dynamic association tag update in the precision marketing method based on multi-channel user profiles of the present invention.

[0021] Figure 4 This is a schematic diagram illustrating the dynamic fuzzy inference gear adjustment in the precision marketing method based on multi-channel user profiles of this invention.

[0022] Figure 5 This is a schematic diagram illustrating the management of the observation period in the precision marketing method based on multi-channel user profiling of this invention.

[0023] Figure 6 This is a schematic diagram illustrating the dual verification of precision marketing in the precision marketing method based on multi-channel user profiling of the present invention.

[0024] Figure 7 This is a module diagram of the precision marketing system based on multi-channel user profiling of the present invention. Detailed Implementation

[0025] This application embodiment, through a precision marketing method and system based on multi-channel user profiles, achieves a complete closed-loop mapping from fragmented data collection to precision marketing by using de-identified dynamic associated tags as a unified anchor point, through phased adaptive matching, access control during the verification observation period, continuous iteration of version profiles, and double-verification precise output.

[0026] The overall concept of this application's embodiments is as follows: User behavior data generated across various channels is defined as profile fragments. Each fragment is bound to an irreversibly encrypted, de-identified, dynamically associated tag as a unified anchor point for cross-channel identity alignment. All fragments of the same user are associated through the same tag. The matching process consists of two phases: when the number of fragments is small, three-layer fuzzy inference is used for matching; when the number of fragments is sufficient, the latest version of the profile is used as a reference for matching. Any matched fragments are not directly merged but must enter a verification and observation period. Only after passing dual checks for behavioral consistency and attribute consistency during the verification and observation period can they be merged into the profile. Each merge triggers incremental retraining to generate the next version of the profile, continuously iterating without end. Finally, based on the latest version of the profile, precise marketing is performed with dual verification of profile tag matching degree and cross-channel compliance.

[0027] Example 1, please refer to Figure 1 , Figure 2 This embodiment provides a precision marketing method based on multi-channel user profiles, which includes the following steps.

[0028] Step S1.1: Collect fragments of user behavior from various channels and bind de-identified dynamic association tags to each fragment.

[0029] Channels include, but are not limited to: proprietary brand mobile applications, WeChat mini-programs, offline store POS systems, third-party e-commerce platform flagship stores, official websites, official social media accounts, and SMS interaction platforms. Every recordable behavioral trace a user generates on any of these channels constitutes a fragment.

[0030] The specific information carried by the fragment includes: which channel the fragment came from, when the user performed the action, what type of action it was, what the content involved in the action was, and the device and network environment information at the time the action occurred. The type of action is, for example, browsing a product details page, adding a product to the shopping cart, submitting an order, completing payment, adding a product to favorites, writing a review, or clicking on an advertisement. The content involved in the action includes which product was viewed and which category of advertisement was clicked. The network environment information includes the device model, operating system version, and the location of the network IP address.

[0031] The method for generating de-identified dynamic association tags is as follows: When the system first collects a user's behavior fragment from any channel, it triggers the generation of a de-identified, dynamically associated tag for that user. The generation process is as follows: The system first extracts information from the fragment that uniquely identifies the user. For example, in mobile applications, this is usually the phone number bound during registration; in mini-programs, it's usually the user's unique WeChat Open Platform identifier; and in e-commerce platforms, it's usually the user's platform account ID. However, the system does not directly use this raw identification information as a tag. Instead, it irreversibly encrypts this raw information. The encryption process is as follows: First, a random string is generated. This random string is called a salt value in cryptography. This salt value is independent and randomly generated for each user and is stored separately in a secure location by the system, managed separately from the tag. Then, the system concatenates the user's raw identification information with this salt value to form a longer string, and then performs a secure hash calculation on this concatenated string. A secure hash calculation, simply put, is a mathematical transformation that can compress input data of arbitrary length into a fixed-length output result. Moreover, this transformation is unidirectional; the output can be calculated from the input, but it is almost impossible to deduce the input from the output. The fixed-length result obtained after hash calculation is the user's de-identified dynamic association tag.

[0032] The hash algorithm uses either the national cryptographic algorithm SM3 or the SHA-256 algorithm, selected and fixed during system initialization. The salt value is generated by the system's built-in secure random number generator, which reads at least 128 bits of random entropy from the system's entropy pool each time it is generated, ensuring the unpredictability of the salt value. After generation, the salt value is concatenated with the user identifier, appended to the user identifier string to form the complete input string. The salt value itself is stored in a separate secure storage area, physically isolated from the tag data, and is only accessed by the system when tag verification or updates are needed; no external query interface returns the original salt value. When a user appears for the first time almost simultaneously on multiple channels, the system triggers tag generation calculations based on the user's identifier information on each channel. During the calculation process, concurrent requests from the same user are handled with locking: the user's mobile phone number or device fingerprint is used as the locking key to ensure that only one tag generation task is executed for the same user at any given time, preventing multiple different tags from being generated for the same user due to concurrency contention. If the locking wait times out, the system processes requests in the order they arrive, and later requests directly reuse the tag results generated earlier.

[0033] For example, suppose user "Zhang San" registers in a mobile application using his phone number "12345678987". The system generates a random salt value "s7f9k2m" for Zhang San, then concatenates "12345678987" and "s7f9k2m" into "12345678987s7f9k2m". A secure hash is then performed on this concatenated string, resulting in a hash value similar to "a3f8c9e2d1b4...". This hash value is Zhang San's de-identified dynamic association tag. This tag itself is just a string of seemingly meaningless characters; no one can deduce Zhang San's phone number from this string.

[0034] After a tag is generated, the system binds this tag to the fragment currently generated by Zhang San, and simultaneously records an association: the tag "a3f8c9e2d1b4..." corresponds to user Zhang San. When Zhang San subsequently performs actions within the mini-program, the system uses the same salt value and the unique identifier from the WeChat Open Platform to perform a hash calculation, resulting in the same tag value. Therefore, the system knows that the fragment in the mini-program and the fragment in the mobile application belong to the same user. This is the meaning of all fragments of the same user being associated with the same tag.

[0035] It's important to note that the salt value can be changed periodically. The system has a preset update cycle, for example, changing the salt value every 30 days. When the update cycle arrives, the system recalculates the tags for all active users using the new salt value. After the update, the system needs to establish an inheritance relationship between the old and new tags. That is, the new tag must be able to identify which old tag it corresponds to, and all associated fragments under the old tag must be transferred to the new tag. This is done to prevent tags from being used consistently for a long time, indirectly exposing user identity information through large-scale data association analysis. However, after the tags are updated, the old tags are not immediately deleted but marked as expired so that the original associations can still be found when historical data needs to be traced.

[0036] Step S1.2: Set a preset fragment quantity threshold and perform fragment matching in two stages.

[0037] The system has a pre-defined fragmentation threshold, which serves as the dividing line between the first and second stages. The specific value of this threshold can be flexibly set according to the business scenario; for example, it can be set to 30 fragments. The threshold is set based on the following: before the system goes live, it is simulated in a test environment, and the rate of change of core profile features of virtual users at different fragmentation levels is statistically analyzed. For example, the top 5 tags with the highest weights and their weight values ​​are analyzed. When the number of fragments reaches 30, the rate of change of core profile features drops below 5%, indicating that the profile has stabilized; therefore, 30 fragments are chosen as the threshold.

[0038] When a user's fragment count has not yet reached the fragment count threshold, the system is in the first stage.

[0039] The difficulty in the first stage lies in the scarcity of available fragments, lacking sufficient information to build a reliable user profile as a reference. In this context, this embodiment employs fuzzy reasoning to address the fragment matching problem. Fuzzy reasoning avoids absolute yes / no judgments, instead providing a confidence score between 0 and 1, indicating the degree of certainty that the fragment belongs to the user.

[0040] In the first phase, the system performs three levels of reasoning for each fragment in sequence.

[0041] The first layer of reasoning is single-fragment feature evaluation. The system examines the information carried by the fragment itself: its source channel, the time of occurrence, the type of behavior, and the content involved. Then, it compares this information with the tags associated with the fragment. For example, if the fragment comes from a channel frequently used by the user, the match rate for the channel dimension is high; if the behavior occurs during the user's typically active time period, the match rate for the time dimension is high; if the content of the behavior aligns with the user's historical preferences, the match rate for the content dimension is high. The match rates across all dimensions are combined to generate a score, which is the first score for the single-fragment feature.

[0042] The second layer of reasoning is the evaluation of inter-fragment correlation. The system not only considers the characteristics of a single fragment but also examines whether there are inherent connections between this fragment and other fragments already grouped under the same label. For example, if the timing of this fragment's actions forms a reasonable timeline with the actions of other fragments—for instance, a user first browsed a certain type of product on a certain day, added the same product to their shopping cart a few days later on another channel, and then completed the payment two days later on a third channel—the more coherent this behavioral chain is and the more it conforms to a normal consumer decision-making path, the higher the correlation between these fragments. Another example is if the geographical location of this fragment coincides with the geographical locations of other fragments—for instance, if the user consistently makes actions in the same city—then the spatial trajectory coordination is high. The system integrates the degree of coordination across these dimensions and assigns a score; this is the second score for inter-fragment correlation.

[0043] The third layer of reasoning is a comprehensive scoring. The first score for individual fragment features and the second score for inter-fragment correlation are weighted and fused according to certain weights to obtain a final third score for matching confidence. This weight is not fixed; it changes dynamically as the number of collected fragments increases. When the number of fragments is small, the first score has a higher weight because there are not enough other fragments for reference, and the judgment mainly relies on the features of the fragments themselves. As the number of fragments gradually increases, the weight of the second score gradually increases because more fragments can corroborate each other, and the correlation between fragments becomes increasingly reliable.

[0044] The dynamic adjustment mechanism for fuzzy inference levels works as follows: The system presets at least three inference levels—high, medium, and low. Each level corresponds to a different set of inference parameters. Simply put, high level means that the inference is very sensitive; even a slight difference in fragment features will lead to a significant change in the score. Low level means that the inference is relatively sluggish; a relatively large difference in fragment features is required for the score to change. Medium level is in between.

[0045] The method for determining which sensitivity level to use is as follows: First, consider the number of fragments currently collected and the range in which the number of fragments falls. This corresponds to a baseline sensitivity level. For example, if the number of fragments is between 0 and 10, the baseline sensitivity level is set to high; between 11 and 20, it's set to medium; and between 21 and above, it's set to low. The rationale behind this design is that fewer fragments mean less information, requiring amplification of subtle differences between fragments to aid differentiation, hence the use of a high-sensitivity level. More fragments mean more information, necessitating a focus on stability and avoiding misjudgments caused by oversensitivity, hence the use of a low-sensitivity level. Second, consider the dispersion of the inference results from the current batch of fragments. Dispersion, simply put, is the degree of difference between the scores of each fragment. If the scores of each fragment are very close and indistinguishable, it indicates insufficient differentiation at the current sensitivity level, so the level is adjusted one step towards higher sensitivity. If the scores of each fragment are already significantly different, it means the current sensitivity level provides sufficient differentiation, so the level is adjusted one step towards lower sensitivity.

[0046] In the first stage, after obtaining the third score of the matching confidence of a fragment through three layers of reasoning, this fragment will not be immediately merged into the user's profile, but will enter a verification and observation period, where it will not be merged for the time being and will be observed for a period of time.

[0047] When a user's fragment count reaches or exceeds a preset threshold, the system enters the second stage.

[0048] After entering the second phase, the system has accumulated enough fragments to build an initial profile. This initial profile includes all the system's knowledge of the user so far: what types of products they prefer, when they are usually active, which channels they mainly use, their approximate spending power, and so on.

[0049] Starting from the second phase, the matching method differs from the first phase whenever the system collects a new fragment. In the first phase, without a user profile as a reference, fuzzy reasoning relied solely on the fragment's own characteristics and the relationships between fragments. In the second phase, with a user profile as a reference, new fragments can be directly compared to the latest version of the user profile to see if the fragment matches the user characteristics in the profile. Even after a match, the new fragment is not immediately merged but enters a verification and observation period. Only after successful verification can it be incorporated into the current user profile.

[0050] Step S1.3: Verify the specific operating mechanism of the observation period.

[0051] In both the first and second phases, the matched fragments will not be used immediately, but will undergo a verification and observation period. The purpose is to prevent incorrect fragments from contaminating the user profile.

[0052] The duration of the observation period is not fixed, but is dynamically calculated based on three factors.

[0053] The first factor is the credibility of the source channel. Data quality varies across different channels. For example, a channel requiring user login and authentication before any action is likely to have higher data credibility; conversely, a channel allowing anonymous operation has relatively lower data credibility. The system assigns a preset credibility score to each channel, ranging from 0 to 1. Channels with lower credibility require longer observation periods for their data fragments.

[0054] The second factor is the completeness of attributes. The more information a fragment carries, the more confident the judgment will be. If a fragment only carries a channel identifier and a timestamp, without any other information, it is difficult to determine who it belongs to and requires a longer observation period. If a fragment carries a lot of information such as a channel identifier, timestamp, behavior type, product information, device information, and geographical location, it is relatively easy to judge and the observation time can be shorter.

[0055] The third factor is the matching confidence score. If the matching score of this fragment is very high, close to 1, it means that the system is very confident that this fragment belongs to this user, and the observation period can be shorter; if the score is just barely above the passing mark, just reaching the threshold for entering the verification observation period, it means that the system is not very confident, and then a longer time is needed for verification.

[0056] By combining these three factors, we start with a standard observation period, such as 24 hours. Then, we adjust the duration based on channel credibility, attribute completeness, and third-party score to calculate a suitable observation period for this segment. The standard observation period is set based on the following criteria: covering the typical user behavior cycle across various channels within a complete natural day, ensuring sufficient time window to collect user behavior feedback data during the verification observation period.

[0057] The verification process is successful if the consistency between the fragment and the current user profile is continuously monitored during the verification observation period. Specifically, comparisons are made across at least two preset dimensions. For example, regarding the behavior time dimension, if the user profile shows that the user is typically active between 8 PM and 11 PM, but the fragment's behavior occurred at 3 AM, the consistency index for behavior time is very low. Similarly, regarding content preferences, if the user profile shows that the user prefers electronic products, but the fragment involves baby and maternity products, the consistency index for content preferences is also very low. Verification is only successful when the consistency index across all monitored dimensions consistently reaches or exceeds its corresponding threshold, and the fragment can then be incorporated into the user profile. If any dimension fails to meet the standard, verification fails, and the fragment is discarded.

[0058] Virtual Marketing Verification Mechanism: During the verification observation period, a virtual marketing message can be proactively pushed to assist in verification. Virtual marketing content is content that appears to be a normal marketing push but will not actually generate any real transactions. Based on the user's historical behavior data corresponding to this message, the system selects a content direction that the user has historically shown a clear preference for and generates a virtual marketing message to push to the user. If the user's response to this content, such as clicking or viewing details, is consistent with the user's historical behavior pattern, it indicates that the message indeed belongs to this user, and the verification credibility is higher; if the user's response is significantly inconsistent with historical patterns, it indicates that the message may have been mismatched, and the verification fails.

[0059] Step S1.4: Iterative updates and version management of the profile.

[0060] Whenever a fragment passes the validation observation period and is officially incorporated into the current profile, the system will trigger a retraining of the profile. Retraining means putting the newly incorporated fragment together with all the previously collected fragments and recalculating the weights and feature vectors of each label in the profile.

[0061] Here, labels refer to various markers describing user characteristics, such as electronics enthusiast, high-spending user, nighttime active user, new user, etc. Each label has a weight, indicating its importance in describing the user. The feature vector is a list of values ​​with a fixed format that organizes these labels and their weights, facilitating subsequent matching calculations by the computer.

[0062] After each retraining, a new version of the profile is generated, and a version number and a timestamp are assigned to this version. When the system needs to match marketing content based on the profile later, it will always call the latest version of the profile.

[0063] There are two update methods: Incremental updates are used under normal circumstances, which only make partial adjustments to the existing profile based on the newly incorporated fragment information, without recalculating all historical data. This is fast and consumes fewer resources. However, if the core features of the profile remain almost unchanged after multiple consecutive merges, it indicates that the profile has stabilized. At this point, the system will trigger a full retraining, which reconstructs the profile from scratch using all historical fragments, completely eliminating the small errors that may have accumulated during the incremental update process.

[0064] Step S1.5: Accurate matching and dual-verification output of marketing content.

[0065] When the system receives a marketing trigger request, such as when a user opens a mobile application, enters a mini-program, or triggers a scheduled marketing campaign, it calls the latest version of the user's profile and filters out content that matches that profile from the marketing content library.

[0066] The matching process works as follows: each piece of marketing content carries one or more target tags. For example, an advertisement for a digital product might carry the target tags of "electronic product enthusiasts" and "high-spending power." The system compares the target tags of the marketing content with the tags in the user profile one by one, calculating the matching degree between the two levels.

[0067] The first level is semantic similarity, which is how similar the target label and the profile label are in meaning. For example, although digital products and electronic products are not exactly the same in words, their meanings are very similar, so the semantic similarity is high.

[0068] The second aspect is weighted matching, which refers to the weight of the target tag in the user profile. If the tag "electronics enthusiast" has a weight of 0.8 in the profile, it means that this tag is very important in describing the user, and the matching degree of marketing content with this target tag is high; if the weight is only 0.2, it means that this tag is not so important, and the matching degree is relatively low.

[0069] The semantic similarity and weighted matching scores are combined to obtain a matching score between the marketing content and the current user profile. Only when this score reaches the preset precision targeting threshold is the content considered to have passed the first verification.

[0070] After passing the first verification, a second verification is required: cross-channel compliance verification. This verification checks two aspects: first, how many marketing pushes the user has received on the current channel recently, and whether it exceeds the channel's allowed frequency limit; second, whether the user has authorized receiving this type of marketing content on this channel. Only after both verifications are passed will the marketing content be actually pushed to the user's corresponding channel.

[0071] Example 2, please refer to Figure 3 This embodiment further explains the update mechanism of de-identified dynamic association tags.

[0072] How tag updates are triggered: Two preset methods for triggering tag updates are provided, which can be used individually or in combination.

[0073] The first method is timed triggering. The system presets a fixed time period, such as automatically triggering a tag update every 30 days. The advantage of this method is that the rules are simple and predictable, making system operation and maintenance easier. However, the disadvantage is the lack of flexibility. If a user generates a large amount of new behavioral data within 30 days, the tags should be updated more frequently to reflect the latest relationships; if a user has almost no new behavior within 30 days, updating the tags is meaningless.

[0074] The second method is event-triggered. The system monitors the growth rate of each user's fragment count. When a user's new fragment count reaches a certain threshold within a preset period—for example, if a user adds more than 50 fragments in 7 days—the system triggers an early update of the user's tag. This method can dynamically adjust the update frequency based on the user's actual activity level, updating tags more frequently when user activity is high and reducing unnecessary updates when user activity is low.

[0075] In practical applications, two methods can be used in combination: use a fixed period of 30 days, monitor the user's fragment growth rate, and if the number of new fragments reaches the trigger threshold ahead of schedule within the fixed period, execute the update ahead of schedule.

[0076] The specific steps for tag updates are as follows: S2.1, retrieve all fragments associated with the user's current tag. These fragments are scattered across various channels, but are all linked together through the current tag.

[0077] S2.2 Extract the user's original identification information from these fragments. It is important to note that extracting the original identification information requires obtaining the user's legitimate authorization; tag updates will not be performed on unauthorized users.

[0078] S2.3, Generate a new random salt value.

[0079] S2.4, Re-encrypt the original identification information with the new salt value to generate a new tag value.

[0080] S2.5, Establish the inheritance mapping relationship between the new and old tags—that is, record a message inside the system: the new tag “b7e2f9a3c6d1…” is updated from the old tag “a3f8c9e2d1b4…”.

[0081] S2.6, associate the new tag with all the fragments associated with the original tag. Now all fragments are associated with the new tag.

[0082] S2.7 Mark the old label as expired. However, the old label will not be deleted; its associated records will be retained so that the original association can still be found when historical data needs to be traced.

[0083] The basis for setting the tag update cycle and trigger threshold is: The values ​​for the tag update cycle and trigger threshold can be set based on the following considerations.

[0084] Setting the tag update cycle requires a balance between privacy protection and system overhead. A cycle that's too short, such as every 3 days, while offering higher privacy protection, consumes significant system resources due to frequent recalculation and mapping updates. Conversely, a cycle that's too long, such as every six months, while having lower system overhead, increases the risk of indirectly exposing user identities through data analysis due to prolonged use of the same tag. Considering all factors, 30 days is a more reasonable cycle.

[0085] The threshold setting for event triggering needs to take into account typical user behavior patterns. If an average user generates 5 to 10 fragments per day, then adding 50 fragments within 7 days would be considered abnormally active, warranting an early tag update. This threshold can be dynamically adjusted based on actual operational data: if triggering is found to be too frequent or too infrequent, the threshold can be appropriately increased or decreased.

[0086] Example 3, please refer to Figure 4 This embodiment provides a detailed explanation of the dynamic adjustment mechanism for fuzzy inference gears.

[0087] The physical meaning of inference gears: In essence, inference gears control the sensitivity of the membership function during fuzzy inference. Simply put, a membership function is a converter: it transforms the input—the matching degree between fragment features and labels—into a membership output between 0 and 1.

[0088] Different gear positions correspond to different membership function shapes.

[0089] The membership function shape corresponding to high inference levels is narrow-band high sensitivity. This means that the output is only high when the input value is very close to the ideal value; even a slight deviation from the input value will cause the output to drop rapidly. This black-and-white characteristic is very useful when there are few fragments, because with limited information, it's necessary to amplify the subtle differences between fragments to help distinguish different candidate users.

[0090] The membership function shape corresponding to the medium inference level is wide-band and moderately sensitive. When the input value varies over a relatively wide range, the output maintains a relatively high level; only after deviating to a certain extent will the output gradually decrease. This characteristic is suitable for situations with a moderate number of fragments, retaining a certain level of discriminative ability without being overly sensitive and causing misjudgments.

[0091] The membership function shape corresponding to the low inference level is wide bandwidth and low sensitivity. Even if the input value deviates significantly from the ideal value, the output will not drop too quickly. This sluggish characteristic is suitable for situations where the number of fragments is already sufficient: because there are enough other fragments to corroborate each other, even if the features of a single fragment deviate slightly, it will not affect the overall judgment, so there is no need to be overly sensitive to subtle differences.

[0092] In this embodiment, the membership function adopts a Gaussian function. At the high inference level, the standard deviation of the Gaussian function is set to a small value, making the function curve narrow and steep. Even a slight deviation of the input value near the center point causes a rapid decrease in membership, suitable for scenarios with very few fragments and requiring high discriminative power. At the medium inference level, the standard deviation is set to a moderate value, resulting in a moderately wide function curve. When the input value changes within a certain range, the membership changes smoothly, suitable for scenarios with a moderate number of fragments. At the low inference level, the standard deviation is set to a large value, resulting in a wide and gentle function curve. Even if the input value deviates significantly from the center point, the membership remains at a certain level, suitable for scenarios with a sufficient number of fragments and a focus on stability. The specific values ​​of the standard deviation for each level are preset according to the business scenario during system initialization. For example, the standard deviation for the high level is set to 0.1, for the medium level to 0.3, and for the low level to 0.6, with units representing the units of the matching input value. The above parameters can be optimized and adjusted based on historical matching accuracy during system operation: the system periodically calculates the matching accuracy for each level. If the accuracy for a certain level is significantly lower than expected, the standard deviation for that level will be fine-tuned, with each adjustment not exceeding 10% of the current value.

[0093] The baseline inference level is determined as follows: The system maintains an internal table mapping fragment quantity ranges to baseline levels. This table is set based on historical data analysis and business experience. Initially, an initial mapping can be established based on experience. For example, a high level is used when the number of fragments is between 0 and 10, a medium level between 11 and 30, and a low level for 31 or more. As the system runs longer and accumulates sufficient actual matching data, these level settings can be optimized. Specifically, the matching accuracy of each level within different fragment quantity ranges in historical data is analyzed. If the accuracy of a certain level within a certain range is significantly higher than other levels, that level is set as the baseline level for that range. Through this data-driven approach, the level settings can be continuously optimized.

[0094] The calculation method for the dispersion of inference results: Dispersion measures the degree of difference between the matching scores of fragments in the current batch. Calculating dispersion does not require complex mathematical formulas; an intuitive approach is used: First, look at the third score of all fragments in the current batch, find the highest and lowest scores, and calculate the difference between them. If the highest and lowest scores differ significantly, it indicates a significant difference in scores among the fragments, resulting in high dispersion; if the highest and lowest scores are very close, it indicates that the scores of the fragments are similar, resulting in low dispersion. Besides the range method, another approach is to calculate the average of the third scores of all fragments in the current batch, and then observe how much each fragment's score deviates from this average. The greater the deviation, the higher the dispersion; the smaller the deviation, the lower the dispersion.

[0095] This embodiment defaults to using the range method as the standard calculation method for dispersion. Specifically, the range is calculated by subtracting the minimum value from the maximum third score of all fragments in the current batch, and this range is used as the dispersion. When the number of fragments in a batch is less than 3, the range method may not effectively reflect the dispersion. In this case, the standard deviation method is used: the standard deviation of the third score of each fragment relative to the average of the batch is calculated. The dispersion calculated by both methods is an absolute value and needs to be normalized before comparison between different batches. The normalization method is to divide the calculated range or standard deviation by the theoretical maximum range of scores in the batch, i.e., 1 minus 0 equals 1, to obtain the normalized dispersion, with a value range of 0 to 1. The initial values ​​for the high and low dispersion thresholds are set to 0.4 and 0.2, respectively. These settings are based on offline analysis of simulated data in the test environment before the system went live: the analysis results show that when the dispersion is higher than 0.4, the current gear level generally lacks sufficient differentiation, while when the dispersion is lower than 0.2, the current gear level provides sufficient differentiation. After the system is running, it recalculates the dispersion distribution corresponding to the best gear in the historical data every 100 matching records, and updates the two thresholds with the latest distribution characteristics.

[0096] The specific rules for gear adjustment are as follows: The system presets two dispersion thresholds: a high threshold and a low threshold. After calculating the dispersion of the current batch, the system compares it with these two thresholds: if the dispersion is higher than the high threshold, it means the current gear's discrimination is insufficient, and it needs to be adjusted one level towards higher sensitivity. If the dispersion is lower than the low threshold, it means the current gear is sufficiently discriminative, and it can be adjusted one level towards lower sensitivity to pursue stability. If the dispersion is between the two thresholds, it means the current gear is just right and no adjustment is needed. The initial values ​​of these two thresholds can be set based on experience, for example, the high threshold can be set to 0.4 and the low threshold to 0.2. As the system accumulates more data, these thresholds can be optimized and adjusted based on the actual matching results.

[0097] Example 4: This example provides a detailed explanation of the specific meanings and acquisition methods of the three scores.

[0098] How to obtain the first score for a single fragment feature: The first score for a single fragment feature measures how well the fragment's own features match its associated label. The system extracts information from the fragment in the following dimensions: The first dimension is channel identification. Each channel has a pre-defined level of credibility and typical user characteristics. For example, users of mobile application channels are usually more active core users, while users of mini-program channels may include a large number of casual users. If this fragment comes from a channel that the user has frequently used in the past, this dimension will score higher.

[0099] The second dimension is behavior type. Different behavior types contribute differently to identifying a user. Payment behavior is more telling than browsing behavior because it involves funds and is more likely to be performed by the user themselves. The system assigns different weights to different behavior types; the higher the weight, the more telling the behavior type is.

[0100] The third dimension is the timestamp. The system calculates a user's active periods based on the distribution of behavior times in historical fragments: for example, if it finds that more than 80% of a user's behavior occurs between 8 pm and 11 pm. If the behavior time of a new fragment falls within the active period, the time dimension score is high; if it falls outside the active period, the score is low.

[0101] The fourth dimension is content attributes. The system extracts content preferences from user history fragments: for example, which categories of products the user has browsed and purchased in the past. If the content of new fragments belongs to these preferred categories, the content dimension score will be high.

[0102] The scores of the four dimensions are summed up according to their respective importance weights to obtain the first score of the single fragment feature.

[0103] The method for obtaining the second score based on the correlation between fragments: The second score for fragment correlation measures how well a fragment is coordinated with other already aggregated fragments. The system evaluates this from the following dimensions: The first dimension is time series synergy. User behavior usually follows a sequence, such as browsing, then saving, then adding to cart, and finally paying—this is a natural decision-making process. If the timing of new fragmented actions can be inserted into the time series of existing fragments to form a reasonable storyline, the time series synergy is high.

[0104] The second dimension is the synergy of behavioral sequences. Different types of actions have natural pairings. For example, browsing a product and then adding it to your favorites, or adding a product to your cart and finally completing the payment, are all logical behavioral chains. If the behavior type of a new fragment can form such a logical chain with existing fragments, the synergy of the behavioral sequence is high.

[0105] The third dimension is spatial trajectory coordination. Users typically have relatively fixed activity ranges. If the geographical location of a new fragment coincides with the geographical location of existing fragments, such as in the same city or even the same business district, the spatial trajectory coordination is high.

[0106] By combining the synergy of the three dimensions according to their respective weights, we obtain the second score for the correlation between fragments.

[0107] The method for obtaining the third score of the match confidence: The third score of the matching confidence is the result of combining the first and second scores. The combination method is a weighted sum: the third score equals the first score multiplied by its weight plus the second score multiplied by its weight, and the sum of the two weights equals 1.

[0108] The key point is that these two weights are not fixed. When the number of fragments is still small, the first score has a larger weight, for example, the first score accounts for 0.7 and the second score accounts for 0.3. This is because there are not enough other fragments to use as a reference at this time, and the judgment can only rely mainly on the characteristics of the fragments themselves.

[0109] As the number of fragments gradually increases, the weight of the second score gradually increases. When the number of fragments reaches a certain scale, the weight ratio may become 50 / 50, or even the second score may have a larger share. This is because at this point, there are enough fragments to corroborate each other, and the correlation between the fragments becomes increasingly reliable.

[0110] The weight changes are smooth and will not jump abruptly. The system will gradually adjust the ratio of the two weights according to the current number of fragments, following a preset curve.

[0111] Example 5, please refer to Figure 5 This embodiment provides a detailed description of the management mechanism for the verification observation period.

[0112] Steps for determining the duration of the observation period: The duration of the verification observation period is dynamically determined according to the following steps: S5.1, obtain a standard observation period. This standard period is a preset baseline value of the system and can be flexibly configured according to business scenarios. In the normal configuration, the standard observation period is 24 hours.

[0113] S5.2, Obtain the credibility of the source channel for this fragment. Each channel is assigned a credibility score during system initialization, with a value ranging from 0 to 1. The credibility score is set based on the data quality of the channel: channels requiring user login authentication have high credibility, while anonymous channels have low credibility; channels with standardized data collection have high credibility, while channels with unstable data quality have low credibility.

[0114] S5.3 Adjust duration based on channel credibility. If channel credibility equals the standard value (e.g., the standard credibility is set to 0.8), no adjustment is made. If channel credibility is lower than the standard value, the standard duration is extended by 10% for every 0.1 decrease. If channel credibility is higher than the standard value, the standard duration is shortened by 10% for every 0.1 increase.

[0115] S5.4, Obtain the attribute completeness of the fragment. The attribute completeness is calculated by dividing the actual number of attribute fields carried by the fragment by the total number of attribute fields that a complete fragment should contain. For example, if the system considers a complete fragment to contain 7 fields: channel identifier, user identifier, behavior type, behavior time, content identifier, device information, and geographic location, but the current fragment only contains 3 fields: channel identifier, behavior type, and behavior time, then the attribute completeness is 3 divided by 7, approximately equal to 0.43.

[0116] S5.5 Adjust the duration based on attribute completeness. The lower the attribute completeness, the less information the fragment carries, making judgment more difficult, and the observation period should be longer. The adjustment logic is similar to that of channel credibility adjustment: for every certain percentage the completeness falls below the standard value, the duration is extended by the corresponding proportion.

[0117] S5.6, Obtain the third score of the matching confidence for this fragment. If the third score is very high, such as above 0.9, it indicates that the system is very confident in this matching result, and the observation period can be significantly shortened. If the third score just reaches the threshold for entering the validation observation period, such as around 0.6, it indicates that the system is not very confident, and the observation period should be significantly extended.

[0118] S5.7, the duration obtained after the above adjustments shall be taken as the final verification observation period duration for the fragment.

[0119] Verify the conditions for merging with the current image: During the verification and observation period, the system will continuously monitor the consistency between the fragment and the current profile in at least two preset dimensions.

[0120] Optional monitoring dimensions include, but are not limited to: Behavioral time dimension: Whether the fragmented behavior time falls within the user's active time period in the profile.

[0121] Behavior type dimension: Does the behavior type of the fragment match the user's behavior pattern in the profile?

[0122] Content preference dimension: Whether the content involved in the fragment falls within the user's preference category in the profile.

[0123] Device consistency dimension: Whether the device information of the fragment is consistent with the user's commonly used devices in the profile.

[0124] For each monitored dimension, the system calculates a consistency metric: a value between 0 and 1, representing the degree of match between the fragment and the profile in that dimension. The system presets a threshold for each dimension; only when the consistency metric for that dimension consistently reaches or exceeds this threshold is the dimension considered validated. Only after all monitored dimensions have been validated can the fragment be incorporated into the profile.

[0125] The specific implementation method of the virtual marketing verification mechanism: During the verification and observation period, the system can perform virtual marketing verification to assist in the judgment.

[0126] The specific approach is as follows: The system first extracts the user's preference characteristics from the user's historical behavior data corresponding to the fragment: such as which product categories they prefer, what price range they typically spend in, and what types of marketing campaigns they respond to most frequently. Then, based on these preference characteristics, the system generates a virtual marketing message. The term "virtual" means that although this content looks like a normal marketing push, it will not actually generate any real transactions: for example, it might push a discount information for a virtual product, or a promotional message for a product that has already been discontinued.

[0127] The generation rules for virtual marketing content are as follows: The system first extracts preference features from the historical behavioral data of the user corresponding to the fragment. Specifically, the extraction method involves counting the frequency of each product category in the user's historical behavior and taking the top three most frequent categories as the user's preferred categories. Then, the system matches templates corresponding to the above preferred categories from a pre-set virtual marketing content template library. Each category in the template library has several pre-set virtual marketing copy templates, such as "Your favorite XX category is currently on sale, click to view details." The generated virtual marketing content carries a "virtual tag" field at the data level. This field is automatically added by the system when the content is entered into the database, and its value is "TRUE," while the value of this field for real marketing content is "FALSE." When pushing content, the system uses this field to identify virtual content, ensuring that virtual content does not generate actual order conversions and is not included in real marketing conversion statistics. When a user responds to virtual marketing content, the system collects the response results, such as whether they clicked, viewed details, or had any subsequent interactions. The system then compares the response results with the user's historical average response rate to similar real marketing content. If the deviation exceeds a preset deviation range, the verification is deemed unsuccessful. In this embodiment, the preset deviation range is set to fluctuate by 30% above or below the average.

[0128] After the system pushes this virtual marketing content, it monitors whether the user responds: for example, whether they click to view details, whether they stay for a period of time, or whether they engage in further interaction. If the user's response pattern is consistent with their historical behavior pattern, it indicates that the fragment does indeed belong to that user, increasing the credibility of the verification. If the user does not respond at all, or the response pattern is significantly inconsistent with their historical pattern, it indicates that the fragment may have been mismatched, and the verification fails.

[0129] Feedback collection mechanism: During the verification and observation period, in addition to proactively pushing virtual marketing content for verification, the system will also passively monitor the user's natural behavioral feedback on various channels.

[0130] The system continuously monitors the user's behavioral events across all integrated channels: browsing, clicking, searching, adding to cart, and making payments. Simultaneously, based on the user's profile in the current version, the system builds a predictive behavior model. This model can predict the user's most likely behavior under specific conditions based on their historical behavioral patterns.

[0131] Once the system monitors a user's actual behavior, it compares it with the predictions provided by the predictive behavior model. If the actual behavior matches the prediction—for example, the model predicts the user will browse electronic devices at night, and the user actually does so—it indicates the user profile is accurate, and the remaining observation period can be shortened. If the actual behavior does not match the prediction—for example, the model predicts the user prefers electronic devices, but the user actually browses baby and maternity products—it suggests the user profile may be flawed, and the remaining observation period needs to be extended to gather more evidence for confirmation.

[0132] When subsequent behavioral feedback matches the predicted behavior model, the system shortens the remaining duration of the validation observation period according to preset shortening rules. The specific rules are: each time the remaining duration is shortened by 20%, with a maximum shortening of 6 hours per instance. For example, if the remaining duration is 24 hours, a 20% shortening (4.8 hours) does not exceed the 6-hour limit, leaving 19.2 hours. If the remaining duration is 40 hours, a 20% shortening (8 hours) exceeds the 6-hour limit, but only 6 hours are actually shortened, leaving 34 hours. When subsequent behavioral feedback does not match the predicted behavior model, the system extends the remaining duration of the validation observation period according to preset extension rules. The specific extension rules are: each time the remaining duration is extended by 30%, with a maximum extension of 12 hours per instance. For example, if the current remaining time is 24 hours, extending it by 30% (7.2 hours) does not exceed the 12-hour limit, and the remaining time becomes 31.2 hours. If the current remaining time is 50 hours, extending it by 30% (15 hours) exceeds the 12-hour limit, but only 12 hours are actually extended, and the remaining time becomes 62 hours. The above-mentioned reduction ratio of 20%, extension ratio of 30%, maximum reduction of 6 hours per instance, and maximum extension of 12 hours per instance are all system preset values. These settings are based on statistical analysis of simulated verification data in the test environment before the system went live. The analysis shows that the above parameters can achieve a balance between verification efficiency and verification sufficiency. Each preset value can be adjusted by the system administrator according to the actual operation situation. The adjustment method is to modify the corresponding parameters in the system configuration interface, and the modification takes effect immediately. The total duration of the verification observation period has a lower limit and an upper limit. The lower limit is 2 hours to ensure that the system has at least a 2-hour observation window to collect user behavior feedback under any circumstances. The upper limit is 3 times the initial verification observation period to prevent indefinite delays in verification due to repeated extensions.

[0133] The predictive behavior model is built based on the current version profile. Its core logic is to extract user behavior preference patterns from the profile's tag weights and feature vectors. Specifically, the system extracts the highest-weighted behavior-related tags and their weight values ​​from the current version profile, organizing these tags and weights into a behavior preference vector. The dimension of this vector matches the number of behavior-type tags in the profile tag system, and each dimension's value represents the current weight of the corresponding tag. The model's input is the user identifier and the current scenario identifier, and the output is the most likely behavior types the user will exhibit in the current scenario, along with their probabilities. The specific method for determining whether subsequent behavior feedback is consistent with the predictive behavior model is as follows: After the system detects the user's actual subsequent behavior, it extracts the behavior type and searches for the predicted probability of that behavior type in the predictive behavior model's output. If this probability is greater than a preset consistency threshold, it is considered consistent; if the probability is less than or equal to the threshold, it is considered inconsistent. The consistency threshold is preset to 0.5, meaning that only when the model's predicted probability for a certain behavior type exceeds 50% is the behavior considered within the model's expectations.

[0134] Example 6: This example provides a detailed description of the version management and retraining mechanism in the iterative update of the portrait.

[0135] Version profile generation and identification: Whenever a fragment passes the verification observation period and is officially incorporated into the current profile, the system will trigger a retraining to generate a new version of the profile.

[0136] Each version of the profile has two identifiers: a version number, which is an incrementing number: the first version is V1, the second is V2, and so on; and a timestamp, which records the specific time when the version was generated.

[0137] The version number allows the system to clearly identify which version is the latest, and the latest version of the profile is used when a marketing trigger request is received. The timestamp allows operations personnel to trace the history of profile changes: for example, to know what a user's profile looked like at a certain point in time, the corresponding version can be found through the timestamp.

[0138] The specific execution method of incremental update: Under normal circumstances, the system uses an incremental update method. The characteristic of this method is that only the parts that need to be modified are changed, without recalculating all the data.

[0139] Specifically, when a new fragment is incorporated, the system analyzes what new information it brings. For example, if the fragment indicates that a user has shown interest in a certain product category, but this category doesn't yet have a corresponding tag in the user profile, the system will create a new tag. Or, if the fragment further confirms a user's preference for a particular product category, the system will increase the weight of that tag.

[0140] The advantages of incremental updates are speed and low resource consumption, because each update only needs to process newly added fragment information. However, the disadvantage is that small errors may accumulate over time: each update only makes local adjustments and never recalculates from scratch, so the weights of each tag may deviate from their true values ​​over time.

[0141] Triggering conditions and execution method for full retraining: To eliminate the errors that may accumulate during incremental updates, the system will trigger full retraining under certain conditions.

[0142] The trigger condition is that after multiple consecutive merges, the changes in the core features of the profile are all below a preset change threshold. Here, core features refer to the highest-weighted labels and their weight values ​​in the profile. If these core features show almost no change after multiple merges, it indicates that the profile has stabilized, and the accumulated error from incremental updates may have become quite large, requiring a complete reconstruction.

[0143] The full retraining process involves the system extracting all historical fragments of the user's data, from the first fragment to the latest, and recalculating the weights of each label and regenerating the feature vectors from scratch to create a completely new user profile. This new version then replaces the previous version and becomes the latest version.

[0144] After full retraining, the accumulated errors from incremental updates are completely eliminated, and the profile returns to an accurate state. The system then continues to perform subsequent updates using an incremental approach until the conditions for triggering full retraining are met again.

[0145] Example 7, please refer to Figure 6 This embodiment provides a detailed explanation of the profile tag matching degree verification in marketing content matching.

[0146] Target Tag Extraction: When each piece of marketing content is added to the database, the operations team will label it with one or more target tags. These tags describe the type of users the marketing content is suitable for being pushed to. For example, an advertisement for the latest smartphone might target electronics enthusiasts and high-spending consumers; a promotional message for maternity and baby products might target families with infants and young children and those interested in childcare.

[0147] When the system needs to match marketing content to a user, it will extract all the target tags carried by the candidate marketing content.

[0148] How semantic similarity is calculated: Semantic similarity measures how similar the target tags of marketing content are in meaning to the tags in user profiles.

[0149] The system maintains an internal tag semantic library. In this library, each tag is represented as a semantic vector: which can be understood as a string of numbers that reflects the tag's position in the semantic space. Tags with similar meanings also have semantic vectors that are located close to each other in the semantic space.

[0150] When calculating the semantic similarity between two labels, the system finds the semantic vectors corresponding to each label and then calculates the distance between them. The closer the distance between the two vectors, the closer the meanings of the two labels are, and the higher the semantic similarity; the farther apart they are, the lower the semantic similarity.

[0151] For example, the tags "digital products" and "electronic products" have a close semantic vector distance, indicating high semantic similarity. Conversely, the tags "digital products" and "baby products" have a large semantic vector distance, indicating low semantic similarity.

[0152] The semantic vector of the tag is generated as follows: The system preloads a word embedding model trained on Chinese corpus. This model can map any Chinese word or phrase into a fixed-dimensional real vector, with either 128 or 256 dimensions. When a semantic vector for a tag needs to be generated, the system inputs the text of the tag into the word embedding model, and the model outputs the corresponding vector value. If the tag is a multi-word phrase, the average value of each word vector in its corresponding dimension is taken as the vector of the phrase. When calculating the semantic similarity between two tags, cosine distance is used as the distance metric: that is, the cosine value of the angle between two vectors is calculated. This value ranges from -1 to 1, with the closer to 1 indicating that the two vectors are more consistent in direction and more semantically similar. To facilitate subsequent weighted fusion, the system normalizes the cosine distance value: when the cosine value is less than 0, it is taken as 0; when it is greater than 1, it is taken as 1; and the intermediate value remains unchanged, compressing the value range to the interval between 0 and 1, which is used as the final semantic similarity.

[0153] How to calculate weighted matching degree: Weighted matching measures how important the target tags of marketing content are in the user profile.

[0154] After the system receives a target tag for the marketing content, it searches the user profile for identical or similar tags. If found, it checks the tag's weight within the user profile. A higher weight indicates the tag is more important to describing the user, resulting in a higher weight match. If no matching tag is found in the user profile, the weight match is 0.

[0155] The overall matching score and the setting of the precise targeting threshold: The system combines semantic similarity and weighted matching score to obtain a matching score between the marketing content and the current user profile. The combination method is a weighted sum: semantic similarity multiplied by its weight plus weighted matching score multiplied by its weight.

[0156] The higher the match score, the more suitable the marketing content is to be pushed to that user.

[0157] The precise targeting threshold is a preset score threshold. Only marketing content with a matching score that reaches or exceeds this threshold can pass the profile tag matching verification. This threshold is set based on historical data—the system analyzes the marketing conversion rates corresponding to different matching score ranges in history and selects the score with the most significant increase in conversion rate as the threshold. For example, if the analysis finds that marketing content with a matching score below 0.5 has a conversion rate of only 2%, while the conversion rate of a score above 0.5 increases to 8%, then 0.5 can be used as the precise targeting threshold.

[0158] Example 8, please refer to Figure 6 This embodiment provides a detailed explanation of the compliance verification for cross-channel distribution.

[0159] Acquisition and Management of Marketing Authorization Status: When a user uses a channel for the first time, the system will display options for receiving marketing content through the privacy policy or marketing preference settings page. Users can choose to authorize receiving, refuse to receive, or only receive specific types of marketing content.

[0160] A user's authorization choices are recorded in the system, forming a marketing authorization status record for that user on that channel. Each authorization record includes the following information: user identifier, channel identifier, content type, authorization status, authorization time, and validity period.

[0161] When the system performs cross-channel compliance verification, it queries the user's authorization status on the target advertising channel. The authorization verification is only considered successful if the authorization status is "authorized" and the current time is within the validity period.

[0162] Obtaining and managing frequency limits: Each channel has a frequency cap—the maximum number of times marketing content can be pushed to a single user within a specific time period.

[0163] The frequency cap is set based on the characteristics of the channel and user experience considerations. Mobile app push notifications can be disruptive if too frequent, so the frequency cap is usually low, such as no more than 3 times per day. SMS messages are more disruptive, so the frequency cap should be even lower, such as no more than 2 times per week. Emails are relatively less disruptive, so the frequency cap can be higher, such as no more than 5 times per week.

[0164] The frequency limit can be configured by the system administrator according to business needs, or it can be dynamically adjusted based on user feedback: if a user frequently gives negative feedback on push notifications from a certain channel, the system can automatically reduce the user's frequency limit on that channel.

[0165] Complete execution process of compliance verification: Once the candidate marketing content passes the profile tag matching verification, it enters the cross-channel campaign compliance verification stage. The verification is performed according to the following steps: S8.1, obtain the current user identifier and the target delivery channels for candidate marketing content.

[0166] S8.2, query how many marketing pushes the user has received on this channel within the current time period, such as the last 24 hours or the last 7 days.

[0167] S8.3, retrieves the frequency limit for this user on this channel.

[0168] S8.4, determine if the number of received messages is less than the frequency limit. If the limit has been reached, it means that the user has received enough marketing content in the current period, and no more messages can be pushed to them; the verification fails.

[0169] S8.5, query the user's authorization status for this type of marketing content on this channel.

[0170] S8.6 Determine if the authorization status is "authorized" and the current time is within the validity period. If not, the verification fails.

[0171] S8.7 After both verifications are passed, the candidate marketing content passes the cross-channel delivery compliance verification and can be pushed to the user's corresponding channel.

[0172] Example 9, please refer to Figure 7 This embodiment provides a precision marketing system based on multi-channel user profiles.

[0173] The system includes the following modules: The multi-channel fragment collection module is used to collect fragments of user behavior from various channels and bind a de-identified dynamic association tag to each fragment as a unified anchor point for cross-channel identity alignment. All fragments of the same user are associated with the same tag.

[0174] The fragment matching module performs fragment matching in two stages based on a preset fragment quantity threshold. The first stage occurs when the fragment quantity threshold is not reached. This stage dynamically adjusts the fuzzy inference level based on the current number of collected fragments and the dispersion of the inference results. Fragments are processed based on the adjusted level, outputting a first score for single fragment features, a second score for inter-fragment correlation, and a third score for matching confidence. Matched fragments enter a validation observation period and are not merged. The second stage begins when the fragment quantity threshold is reached. An initial profile is generated, and subsequent new fragments are matched against the latest version of the profile. After matching, a validation observation period begins. Once validation is successful, the fragments are merged with the current profile. Each merge triggers retraining and updating of the profile to the next version, continuing iteratively without end.

[0175] The precision marketing matching module is used to call the latest version of the profile when a marketing trigger request is received. It performs profile tag matching degree verification and cross-channel delivery compliance verification on the candidate marketing content in sequence. If both verifications pass, the precise push to the corresponding channel will be executed.

[0176] The data flow between the modules is as follows: The fragments and tags collected by the multi-channel fragment collection module are output to the fragment matching module; after matching, verification, merging, and retraining, the fragment matching module outputs the latest version of the profile to the precision marketing matching module; the precision marketing matching module matches marketing content based on the latest version of the profile, and outputs the final push result after double verification.

[0177] As the central hub of the system, the control unit is responsible for coordinating the working sequence of each module and ensuring the orderly flow of data between modules.

[0178] In summary, this application has at least the following effects: By dynamically associating tags with de-identified identifiers, unified anchoring of cross-channel identities is achieved. Tags are generated using irreversible encryption and updated regularly, fundamentally ensuring user privacy and security. This also addresses the shortcomings of traditional marketing matching processes, which rely on strong features like phone numbers and fail when those features are missing. After matching, verification is performed before merging. A firewall is in place before fragmented profiles are incorporated into the main profile: only fragments that pass both behavioral and attribute consistency checks during the verification observation period are allowed to be merged, fundamentally preventing the cumulative error problem of a single matching error continuously polluting the profile. This mechanism is a completely new design never before seen in existing technologies. The phased matching strategy employs dynamic fuzzy reasoning in the first stage when data is sparse, and uses the latest version of the profile as a reference in the second stage when data is abundant. This achieves adaptive matching throughout the entire lifecycle from cold start to mature operation, significantly improving the accuracy of cross-channel user identification.

[0179] Those skilled in the art will understand that embodiments of the present invention can be provided as methods or systems. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0180] This invention is described with reference to flowchart illustrations and structural diagrams of methods and systems according to embodiments of the invention. It should be understood that the combination of each process and module in the flowchart and structural diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 One or more processes and structures Figure 1 A device for a function specified in one or more modules.

[0181] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and structures Figure 1 The function specified in one or more modules.

[0182] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and structures Figure 1 The steps of a specified function in one or more modules.

[0183] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0184] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A precision marketing method based on multi-channel user profiles, characterized in that: Includes the following steps: Collect fragments of user behavior from various channels, bind each fragment with a de-identified dynamic association tag, and use it as a unified anchor point for cross-channel identity alignment. All fragments of the same user are associated with the same tag. Fragment matching is performed in two stages when the fragment quantity threshold is preset: the first stage is when the fragment quantity threshold is not reached. The fuzzy inference level is dynamically adjusted according to the current fragment quantity and the dispersion of the inference results. Fragments are processed based on the adjusted level, and the first score of single fragment feature and the second score of inter-fragment correlation are output. The third score of matching confidence is obtained by fusion. Matched fragments enter the verification and observation period and are not merged. After the fragment quantity threshold is reached, the second stage is entered, and an initial portrait is generated. Subsequent new fragments are matched with the latest version of the portrait as a reference. After matching, they enter the verification and observation period. Once the verification is passed, they are merged with the current portrait. After each merge, the portrait is retrained and updated to the next version of the portrait. The iteration continues without end. When a marketing trigger request is received, the latest version of the profile is invoked, and profile tag matching degree verification and cross-channel delivery compliance verification are performed on the candidate marketing content in sequence. If both verifications pass, the corresponding channel's precise push is executed.

2. The precision marketing method based on multi-channel user profiles according to claim 1, characterized in that, The de-identified dynamic association tag is generated based on the user's irreversible encrypted identifier in each channel. There is no reversible mapping relationship between the de-identified dynamic association tag and the user's identity information. It is dynamically updated according to a preset period as the user's cross-channel behavior data accumulates. The updated de-identified dynamic association tag inherits all the fragments associated with the original de-identified dynamic association tag.

3. The precision marketing method based on multi-channel user profiles according to claim 1, characterized in that, The dynamic adjustment of the fuzzy inference level includes: At least three inference levels are preset, each corresponding to a different set of fuzzy inference parameters. The inference level is the sensitivity level of the membership function during fuzzy inference. A high inference level corresponds to a narrowband high-sensitivity membership function, a medium inference level corresponds to a wideband medium-sensitivity membership function, and a low inference level corresponds to a wideband low-sensitivity membership function. The baseline inference level is determined based on the current range of collected fragments. Then, the baseline inference level is adjusted based on the dispersion of the inference results. The higher the dispersion, the higher the inference level is shifted, and the lower the dispersion, the lower the inference level is shifted.

4. The precision marketing method based on multi-channel user profiles according to claim 1, characterized in that, The first score of the single fragment feature is calculated based on the membership degree of the channel identifier, behavior type, timestamp, and content attributes carried by the fragment with the de-identified dynamic association tag; The second score for the correlation between fragments is calculated based on the degree of coordination between the fragment and other aggregated fragments in terms of time series, behavioral series, and spatial trajectory. The third score of matching confidence is obtained by dynamically weighting and fusing the first score and the second score, wherein the weight of the second score gradually increases as the amount of currently collected fragments increases.

5. The precision marketing method based on multi-channel user profiles according to claim 1, characterized in that, The duration of the verification observation period is dynamically determined based on the credibility of the source channel, the completeness of the attributes, and the third score of the matching confidence of the fragment. The condition for the verification to be merged with the current profile is that during the verification observation period, the behavior consistency index and attribute consistency index of the fragment and the current profile in at least two preset dimensions continuously reach or exceed the corresponding dimension thresholds. During the verification observation period, a virtual marketing verification mechanism is also set up. Virtual marketing content is determined and pushed based on the historical behavior of the user corresponding to the fragment. The user's response to the virtual marketing content is collected and compared with the user's historical behavior pattern. When the comparison result exceeds the preset deviation range, the fragment is determined to fail the verification. The historical behavior pattern is the behavior pattern in the historical behavior data collected before the user corresponding to the fragment enters the verification observation period.

6. The precision marketing method based on multi-channel user profiles according to claim 1, characterized in that, The process of triggering retraining and updating the portrait to the next version after each merge includes: using the merged fragment set as training samples, performing incremental updates on the label weights and feature vectors of the current version portrait to generate the next version portrait. Each version portrait carries a version number and a timestamp. When a marketing trigger request is received, the portrait corresponding to the latest version number is called. When the core feature change of the portrait is lower than a preset change threshold after a preset number of consecutive merges, full retraining is triggered to replace the incremental update.

7. The precision marketing method based on multi-channel user profiles according to claim 1, characterized in that, The profile tag matching degree verification specifically includes: extracting the target tags carried by the candidate marketing content, matching each target tag with the tags in the latest version of the profile, calculating the semantic similarity and weight matching degree between each target tag and the tags in the latest version of the profile, and comprehensively obtaining the matching degree score between the candidate marketing content and the current user profile. When the matching degree score reaches the preset accurate delivery threshold, the candidate marketing content passes the profile tag matching degree verification.

8. The precision marketing method based on multi-channel user profiles according to claim 1, characterized in that, The cross-channel delivery compliance verification includes: obtaining the user's marketing authorization status and frequency limit on each channel; verifying, based on the target delivery channel corresponding to the candidate marketing content, whether the number of marketing messages the user has received on that channel within the current time period has not exceeded the frequency limit of that channel, and whether the user has valid authorization to receive that type of marketing content on that channel. Valid authorization means that the user has authorization for that type of marketing content on that channel and that the authorization is valid. When the number of marketing messages the user has received on that channel within the current time period has not exceeded the frequency limit of that channel, and the user has valid authorization to receive that type of marketing content on that channel, the candidate marketing content passes the cross-channel delivery compliance verification.

9. The precision marketing method based on multi-channel user profiles according to claim 1, characterized in that, The verification observation period also includes a feedback collection mechanism: when a fragment enters the verification observation period, the subsequent behavior feedback of the user corresponding to the fragment on at least one channel is monitored. When the subsequent behavior feedback is consistent with the predicted behavior model of the current profile, the remaining duration of the verification observation period is shortened according to a preset shortening rule. When the subsequent behavior feedback is inconsistent with the predicted behavior model of the current profile, the remaining duration of the verification observation period is extended according to a preset extension rule.

10. A precision marketing system based on multi-channel user profiles, employing the precision marketing method based on multi-channel user profiles as described in any one of claims 1-9, characterized in that, include: The multi-channel fragment collection module is used to collect fragments of user behavior from various channels, bind each fragment to a de-identified dynamic association tag, and use it as a unified anchor point for cross-channel identity alignment. All fragments of the same user are associated with the same tag. The fragment matching module performs fragment matching in two stages with a preset fragment quantity threshold: the first stage is when the fragment quantity threshold is not reached. The fuzzy inference level is dynamically adjusted according to the current number of collected fragments and the dispersion of the inference results. Fragments are processed based on the adjusted level, and the first score of single fragment features and the second score of inter-fragment correlation are output. The third score of matching confidence is obtained by fusion. Matched fragments enter the verification and observation period and are not merged. After the fragment quantity threshold is reached, the second stage is entered, and an initial portrait is generated. Subsequent new fragments are matched with the latest version of the portrait as a reference. After matching, they enter the verification and observation period. Once the verification is passed, they are merged with the current portrait. Each merge triggers the portrait to be retrained and updated to the next version of the portrait. The iteration continues without end. The precision marketing matching module is used to call the latest version of the profile when a marketing trigger request is received. It then performs profile tag matching degree verification and cross-channel delivery compliance verification on the candidate marketing content in sequence. If both verifications pass, the precise push to the corresponding channel is executed.

Images