A multi-user portrait-based mall system intelligent recommendation decision method and platform

Abstract

The present application relates to computer and electronic commerce technical field, especially to a kind of based on multi-user portrait's shopping mall system intelligent recommendation decision method and platform.The method includes: S1, account identification, equipment identification and session identification are collected, session is established;S2, user portrait vector is constructed and stored;S3, product attribute vector is constructed and candidate product set is generated;S4, single-user correlation and session-level user weight are calculated;S5, session-level score is calculated and sorted;S6, median and median absolute deviation are calculated and group uncertainty index is defined;S7, according to double-threshold hysteresis rule, candidate product set is gated;S8, output and record exposure and interaction feedback according to session identification.The present application constructs session-level score aggregation multi-user portrait, and group uncertainty is measured by median / median absolute deviation, and candidate is dynamically screened by enabling / disabling double-threshold hysteresis gate, so that list is stably outputted under low latency while divergence is stably inhibited.

Description

A Smart Recommendation Decision-Making Method and Platform for E-commerce Systems Based on Multi-User Profiles Technical Field

[0001] This invention relates to the fields of computer and e-commerce technology, and in particular to an intelligent recommendation decision-making method and platform for e-commerce systems based on multi-user profiles. Background Technology

[0002] In e-commerce scenarios, current recommendations are mostly based on individualized decision-making by "single user, single session." When multiple users make decisions together on the same device or in the same shopping scenario (such as families placing orders together, group buying, or store sales staff serving multiple customers on one device), the common practice is to simply average or weight individual scores and then directly sort and output them. This approach has three technical shortcomings: First, it is difficult to quantify the differences of opinion within the group regarding the same candidate product, and extreme or noisy scores can easily "hijack" the results; second, real-time interactions during the session (switching entry points, rapid swiping, temporary interests) cause scores to fluctuate frequently around the threshold, and the lack of anti-jitter mechanisms leads to repeated list jumps in a short period of time; third, there is a lack of engineered means to gate candidates with "excessive disagreement," and the system cannot intercept or downweight such candidates under low latency conditions, resulting in an unstable final display that fails to maintain group consistency.

[0003] Based on this, the main technical problem to be solved in this application is: in the e-commerce recommendation scenario where multiple users share the same session, to provide a recommendation decision mechanism that can complete session-level aggregation with low latency and can robustly quantify group disagreements and control dual threshold hysteresis, so that the system can reflect the individual relevance of each user and suppress list jitter caused by disagreements without relying on offline recalculation, thereby outputting a stable candidate order that is oriented towards group consistency. Summary of the Invention

[0004] To overcome the aforementioned technical deficiencies, the present invention aims to provide an intelligent recommendation decision-making method and platform for e-commerce systems based on multi-user profiles. This invention achieves this by constructing a session-level rating system. It aggregates multiple user profiles and measures group uncertainty using median / median absolute deviation. It also uses dual threshold hysteresis gating to enable / disable dynamic screening of candidates, thereby robustly suppressing divergence and stably outputting a list with low latency.

[0005] This invention discloses an intelligent recommendation decision-making method for an e-commerce system based on multi-user profiles, comprising the following steps:

[0006] Step S1: Collect account identifiers, device identifiers, and session identifiers from multiple users to establish a session;

[0007] Step S2: Construct a user profile vector for each user within the session. And store, where Index for users;

[0008] Step S3: Construct a product attribute vector for each candidate product. And generate a candidate product set, in which For candidate product index;

[0009] Step S4, calculate the relevance of a single user. Session-level user weight And perform normalization processing to make ;

[0010] Step S5: Calculate the conversation-level score according to the following formula. And sort:

[0011]

[0012] in, Candidate Products Conversation-level rating, For users For candidate products Single-user relevance, For users Session-level user weights;

[0013] Step S6, calculate the median Absolute deviation from median :

[0014] ,

[0015] And define the population uncertainty index ;

[0016] Step S7, based on the activation threshold With the release threshold The dual-threshold hysteresis rule is used to gate the candidate item set, where when Enable constraints when Release the constraints, and ;

[0017] Step S8, output the number of targets in the previous list. Each product is identified and its exposure and interaction feedback are recorded according to the session identifier;

[0018] in, It is a positive integer.

[0019] Preferably, the user profile vector From static eigenvectors Preference spectrum vector Context anchor vector The preferred spectrum vector is obtained by splicing; By performing a discrete cosine transform on the time series of clicks within the session and preserving the spectrum and order, The former Each coefficient is obtained. It is a positive integer.

[0020] Preferably, the single-user relevance The lower bound of confidence is calculated using the bootstrap method and is defined as follows:

[0021] ,

[0022]

[0023] in, For the first The baseline similarity is obtained by resampling. For resampling index, The number of resampling times and , The confidence bound coefficients and .

[0024] Preferably, the session-level user weight From the intermediate weight Normalization yields, and , ;

[0025] in, For identity consistency F1 score, The intensity of the most recent interaction. Historical payment probability, , The weighting coefficients and , , The identity consistency F1 score From accuracy With recall rate calculate:

[0026] , ,

[0027] in, For match count, The number of times it appears on the device side. The expected number of times is denoted by the account side, and all three are non-negative integers. , Normalization to interval .

[0028] Preferably, the group uncertainty index Hysteresis gating is performed using an adaptive quantile threshold: the median absolute deviation is calculated on the current candidate item set. The upper and lower quantiles are used as the activation thresholds, respectively. With the release threshold And satisfy .

[0029] Preferably, when hysteresis gating is enabled, a strong label constraint table is applied to the candidate products. This strong label constraint table consists of the difference between the intersection of positive labels and the union of negative labels after expansion by a directed acyclic graph of synonyms, wherein the path length of the directed acyclic graph of synonyms does not exceed 2.

[0030] Preferably, the number of targets in the list Each product is selected through a two-stage process:

[0031] In the first phase, scoring is performed on a session-level basis from a limited set of candidate items. Selecting the threshold for the number of leading candidates One candidate product, ;

[0032] In the second stage, the selection variables are obtained by solving the following 0-1 integer programming problem. :

[0033]

[0034]

[0035] in, Choose a binary variable; For exposure count, candidate products Historical exposure count; As a measure of difference; For ranking functions, based on relevance per user The ranking obtained by sorting the candidate products in descending order; The ranking threshold and ; Minimum coverage base for members; It is a rounding function; Aggregator for merchants; This sets a maximum exposure limit for merchants. A collection of categories; This is the maximum exposure limit for the category. For exposure penalty coefficient and ; The coefficient of difference and The difference index Defined as

[0036] ,

[0037] in, and These are the product difference assessment vectors, It is the cosine distance function.

[0038] Preferably, the selected set obtained in the second stage is rearranged according to a round-robin and position penalty strategy before output: let the position number be... Apply the position penalty function to items that have not yet been output:

[0039]

[0040] in, The value of the position penalty function. The position number in the list and , For position penalty slope and ; Select users by in-session index in a round-robin fashion, and in case of ties, prioritize outputting users that meet the minimum difference threshold. The goods, namely ,in .

[0041] Preferably, the feedback record includes the display probability. With binary feedback And calculate the reweighted feedback:

[0042]

[0043] To update the intermediate weights :

[0044]

[0045] in, For exposure index set, with users Related exposure collection; For step size and After the update, the new session-level user weights are obtained according to the normalization rules. .

[0046] In view of this, a second objective of the present invention is to provide an intelligent recommendation decision-making platform for an e-commerce system based on multi-user profiles, comprising:

[0047] The user and context acquisition unit is used to collect account identifiers, device identifiers, and session identifiers of multiple users and establish sessions.

[0048] The profile building unit is used to build a user profile vector for each user within the session. And store, where Index for users;

[0049] The candidate product set generation and attribute construction unit is used to construct a product attribute vector for each candidate product. And generate a candidate product set, in which For candidate product index;

[0050] The relevance and weight calculation unit is used to calculate the relevance of a single user. Session-level user weight And perform normalization processing to make ;

[0051] Uncertainty calculation and dual-threshold hysteresis gating unit are used to calculate session-level scores according to the following formula. And sort:

[0052]

[0053] in, Candidate Products Conversation-level rating, For users For candidate products Single-user relevance, For users Session-level user weights;

[0054] Label-bound strongly constrained cells are used to calculate the median. Absolute deviation from median :

[0055] ,

[0056] And define the population uncertainty index ;

[0057] Two-stage selection unit, used to select based on activation threshold With the release threshold The dual-threshold hysteresis rule is used to gate the candidate item set, where when Enable constraints when Release the constraints, and ;

[0058] Rotation and position penalty unit, used to output the number of targets in the previous list. Each product is identified and its exposure and interaction feedback are recorded according to the session identifier;

[0059] in, It is a positive integer;

[0060] The platform executes the methods described above through the collaboration of the aforementioned units.

[0061] Compared with existing technologies, the above technical solution has the following advantages:

[0062] 1. Robust Quantification and Suppression of Group Disagreement. In existing technologies, multi-user shared scenarios often aggregate individual scores using simple averaging / weighted averaging, lacking robust measures for group disagreement. This leads to problems such as extreme / noisy scores easily "hijacking" the results, and lists frequently changing within a short period. This invention measures disagreement using the median / median absolute deviation, defining group uncertainty. And adopts a dual threshold for enabling / disabling. , Hysteresis gating suppresses divergence-driven jitter with low latency, stabilizing candidate output; list jitter metric (neighboring Top–K symmetric difference) is significantly reduced.

[0063] 2. Low-latency session-level aggregation. Existing technologies rely on offline recalculation or complex online models, resulting in long online paths and high tail latency. Therefore, they suffer from poor real-time response and difficulty adapting to high concurrency. This invention constructs session-level scoring... Furthermore, the two-stage process of "pre-selecting L=4K + small-scale constraint optimization" ensures that the overall online latency remains at the millisecond level, supporting real-time decision-making for large-scale sessions.

[0064] 3. Interpretable weighting of user speaking rights within a session. Existing technologies typically employ average weighting or empirical weighting, which fails to reflect identity consistency and recent intent. This results in the dilution of the influence of strongly relevant users and the overestimation of the influence of weakly relevant users. This invention constructs intermediate weights (identity consistency F1, recent interaction strength, and historical payment probability) using a three-signal convex combination, and normalizes them to obtain… This allows for a more accurate reflection of each user's contribution to the session decision-making process, increasing the "coverage rate of users".

[0065] 4. Robust estimation of single-user relevance. In existing technologies, direct cosine / dot product similarity is sensitive to sample size, leading to problems such as unstable estimation under cold start or high-variance samples. This invention employs a bootstrap method (resampling) to establish a confidence lower bound on the baseline similarity: This reduces overly optimistic checks on high-variance samples, improves ranking stability, and significantly benefits long-tail sessions.

[0066] 5. Ensuring the consistency and marketability of the candidate layer. In existing technologies, the recall stage often involves coarse "content / behavior" recall, with incomplete filtering of a small number of compliant / marketable candidates. This results in the inclusion of non-compliant, unmarketable, or mutually exclusive candidates, increasing the pressure on subsequent optimization. This invention introduces regional marketability, real-time inventory, and a "strongly constrained tag table" into the candidate generation process, and uses a directed acyclic graph of synonym relationships to extend negative tags with depth limits. This allows for the upfront removal of high-risk / conflicting items, enabling subsequent optimization to focus more on quality and diversity.

[0067] 6. Globally controllable diversity and concentration. In existing technologies, diversity is often achieved through heuristic rearrangement, which is difficult to coordinate with coverage and quotas. Therefore, problems include high concentration of merchants / categories and homogeneous lists. This invention introduces a degree of difference in 0–1 optimization. The system includes rewards and, when necessary, lower limits, and sets an exposure cap for merchants / categories to reduce merchant / category concentration, increase the diversity within the list, and alleviate visual fatigue and reliance on advertising.

[0068] 7. Balancing Historical Exposure Suppression and Exploration. In existing technologies, popular items repeatedly rank highly, creating a "Matthew effect," resulting in a monotonous user experience and difficulty in exposing new products / merchants to exposure. This invention introduces an exposure penalty term through an objective function. The candidate layer implements lightweight deduplication and a "new product protection channel" to controllably break path dependence and balance stability and exploration.

[0069] 8. Location Value Modeling and Rotation Fairness. In existing technologies, ranking is done purely by score from top to bottom, resulting in strong users / those with strong attributes consistently occupying the top positions. This leads to insufficient visibility for other users within the session. This invention addresses this by employing linear location penalties. It also performs user index rotation and minimum difference threshold separation, which significantly improves the first-screen visibility and perceived fairness of multi-user sessions, while maintaining overall quality.

[0070] 9. Debiased Feedback and Online Adaptation. Existing technologies directly use click / conversion as gain, which is heavily influenced by display probability bias. Therefore, problems arise such as a learning bias towards high exposure / high ranking, making correction difficult. This invention addresses this by recording display probabilities... Constructing a reweighted feedback Update the intermediate weights based on the session exposure set. And normalization yields a new This enables online closed-loop adaptation, reduces learning bias caused by position / policy bias, and stabilizes long-term returns.

[0071] 10. A profile representation that emphasizes both frequency domain and context. Existing technologies often use time-domain counting or simple window statistics, which suffers from difficulties in capturing rhythmic differences such as "weekday / weekend" and "day / night," and slow short-term intent updates. This invention uses a profile containing a preference spectrum vector with discrete cosine transform and a time-decayed context anchor vector, and performs quantization storage and norm control, thereby improving the consistency and robustness of recall and ranking without significantly increasing latency and storage.

[0072] 11. End-to-end auditability and recalculation. In existing technologies, the decision-making chain is long, involves many black boxes, and is difficult to review, thus failing to meet compliance / auditing and problem localization requirements. This invention records version numbers, key parameters, and probabilities (such as...) at each stage through sessions, candidate lists, optimization, and display. , (Thresholds, etc.) to maintain full-chain traceability, making it recalculated, located, and reconciled, with low audit and response costs.

[0073] 12. Prioritizing Compliance and Sensitive Dimensions. In existing technologies, sensitive dimension rules and recommendation strategies are disconnected, leading to issues such as compliance rejections and high costs for manual audits. This invention addresses these issues by setting hard priorities for sensitive / compliance dimensions in a strong tag constraint table, prioritizing prohibitions; and by using synonym expansion to prevent "borderline words," significantly reducing compliance rejection rates and minimizing manual intervention.

[0074] 13. Security and Session Consistency Assurance. In existing technologies, session identifiers / device identifiers are simple to construct and have weak resistance to replay attacks and misbinding, thus leading to problems such as mismerging, session crosstalk, and replay attack risks. This invention addresses these issues by employing... Message authentication with session keys, identity binding confidence, and device fingerprint drift monitoring enhance session consistency and resistance to attacks, while reducing profile contamination.

[0075] 14. Configurable business logic and adaptive scenario design. Existing technologies use fixed ratios and one-size-fits-all thresholds. This results in poor generalization across entry points and scenarios. This invention addresses this by using thresholds, quotas, and penalty coefficients (such as...) , , , (etc.) can adapt to the dimensions of entry point / device / activity, enabling flexible adaptation to multiple business scenarios under the same framework and reducing the maintenance cost of multiple sets of strategies. Attached Figure Description

[0076] Figure 1 is a flowchart illustrating the intelligent recommendation decision-making method and platform for an e-commerce system based on multi-user profiles according to the present invention. Detailed Implementation

[0077] The advantages of the present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments.

[0078] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0079] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0080] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, without departing from the scope of this disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0081] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0082] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0083] In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the convenience of the description of the invention and have no specific meaning in themselves. Therefore, "module" and "part" can be used interchangeably.

[0084] Referring to Figure 1, this embodiment provides an intelligent recommendation decision-making method for an e-commerce system based on multi-user profiles, including the following steps: Step S1, collecting account identifiers, device identifiers, and session identifiers of multiple users to establish a session; Step S2, constructing a user profile vector for each user within the session. And store, where Indexing for users; Step S3, constructing a product attribute vector for each candidate product. And generate a candidate product set, in which Index candidate products; Step S4, calculate single-user relevance. Session-level user weight And perform normalization processing to make Step S5: Calculate the conversation-level score according to the following formula. And sort: ,in, Candidate Products Conversation-level rating, For users For candidate products Single-user relevance, For users Session-level user weights; Step S6, calculate the median. Absolute deviation from median : , and define the population uncertainty index. Step S7, based on the activation threshold With the release threshold The dual-threshold hysteresis rule is used to gate the candidate item set, where when Enable constraints when Release the constraints, and Step S8: Output the number of targets in the previous list. Each product is recorded for exposure and interaction feedback according to session identifiers; among them... It is a positive integer.

[0085] This embodiment will describe step S1 in detail. In step S1, the system first completes the basic data collection of the user layer, device layer, and interaction context layer, and establishes a session corresponding to one continuous interaction based on this. The user layer uses an account identifier (denoted as...). A temporary account identifier (denoted as ) is used to uniquely identify a user account; when a user is not logged in, the system generates a temporary account identifier (denoted as ). This is used to collect user activity data during periods of inactivity and to securely bind and merge this data upon subsequent logins. The device layer uses a device identifier (denoted as...). This identifier uniquely identifies a physical or virtual terminal. It prioritizes the use of resettable or revocable end-side identifiers in conjunction with stable, but not directly individual, device environment elements, and incorporates fingerprint drift monitoring to reduce false splitting and merging. The interaction context layer uses a session identifier (denoted as...) A session key (denoted as ) is used to identify a continuous interaction process; a corresponding session key is generated (denoted as ). Record the challenge random number (denoted as ) ), timestamp (denoted as Session lifetime (denoted as) ) and idle timeout (denoted as ), and load the compliance token (denoted as This reflects the scope of authorization and data processing rules.

[0086] Account identifiers are obtained through two paths: First, in the login path, the front-end authentication adaptation layer receives the account primary key from the unified identity service after the user completes login and normalizes it (e.g., email addresses are all lowercase, and phone numbers have the country code added), which is used to generate a stable account identifier. Secondly, when accessing a device for the first time without logging in, a temporary account identifier will be assigned. This temporary identifier is calculated by the server using a random number generated by the client and time information as input. When a user logs in during the current or subsequent session, the system triggers the binding process, assigning the temporary account identifier to the user. Merge into account identifier Within the profile trajectory, the merging action must meet the identity binding confidence threshold and compliance token requirements. Constraints. To facilitate quantifying the reliability of identity binding, the system uses identity binding confidence levels. The reliability of the "temporary account - account" mapping is measured by the following formula: ;in This represents the accuracy estimate based on historical session links. Indicates recall rate estimate; only when The binding relationship is only written to the main relationship table when the threshold is not lower than the set threshold; otherwise, it is only stored in the shadow mapping and is subject to the session lifetime. With compliance tokens Common constraints.

[0087] The acquisition of device identifiers follows the principles of "stability, revocability, and auditability." With user authorization and in compliance with applicable laws and regulations, the system collects several stable but not directly individual-identifying device environment elements (such as operating system version, browser rendering kernel version, screen geometry parameters, time zone, and language), and combines these with resettable or revocable end-side identifiers provided by the platform to generate a device identifier. To maintain long-term consistency without excessive sticking, the system sets a drift monitoring window for the above environmental elements: when the change within the window does not exceed a preset threshold, the device identifier... Remain unchanged; when changes exceed a threshold, trigger a rotation and register an alias to maintain business continuity.

[0088] The session is established using a "frontend request + server signature" model. The frontend first initiates a session creation request, carrying the account identifier. (or temporary account identifier) Equipment identification With timestamp The server uses a high-strength random source to generate session keys. And calculate session identifier To facilitate subsequent consistency verification and auditing, this application stipulates the use of a hash function. Indicates compression mapping, using " " indicates string concatenation, and its session identifier is calculated as follows: After a session is successfully established, the client saves the session identifier in the security context. The server uses "session identifier" as a reference. "Store session state entries as the primary key and record the session key." timestamp Session lifespan Idle timeout With compliance tokens To mitigate replay and concurrent crosstalk, the system maintains a monotonically increasing sequence number for each session link and requires clients to include a session key-based sequence number in subsequent requests. The calculated message authentication code is verified by the server before the sequence number can be advanced and the idle timeout is refreshed.

[0089] It should be noted that the following experiment was conducted during the invention:

[0090] Firstly: Using the upper bound of the birthday paradox to identify session identifiers. The collision probability is estimated. In a 128-bit format, with a session creation count of [number missing], [number missing] sessions are created. Under these conditions, the upper limit of the estimated collision probability is approximately It is significantly lower than the engineering threshold.

[0091] Secondly, comparing the two device identifier generation strategies—"relying solely on the platform resettable identifier" and "a combination of the platform resettable identifier and device environment factors"—the latter strategy showed a median improvement in device identifier stability of approximately 2.8%–4.6% within a 30-day window, while the median false merging rate was no higher than 0.02%. Furthermore, after enabling request message authentication, the end-to-end P95 latency increment was approximately 0.9–1.3 milliseconds, within the acceptable range for common web interactions. All of the above data can be repeatedly verified through gray-scale traffic, demonstrating auditability.

[0092] It should also be noted that this implementation includes the following:

[0093] One approach is multi-factor session upgrade, which involves secondary verification for high-value operations using device authentication or edge-side security elements, followed by the session key. A new message authentication key is derived, thus forming a hierarchical protection chain;

[0094] Secondly, cross-domain sub-session federation involves deriving sub-session identifiers among multiple subsystems within the same jurisdiction for least-privilege mutual trust, without exposing the main session identifier. This enables cross-service authentication and session tracking, thereby reducing the risk of identifier leakage and facilitating auditing.

[0095] In this embodiment, step S2 will be described in detail. In step S2, the system constructs a user profile vector for each user within the session. This is used as the input baseline for subsequent candidate generation, single-user relevance, and session-level scoring. To ensure interpretability and maintainability, user profile vectors... A layered concatenation structure is adopted, consisting of static feature vectors. Preference spectrum vector With context anchor vector It consists of three parts, the order and dimensions of which are fixed throughout the system. This concatenation is represented in standard vector concatenation form. ,in" The "" symbol represents the concatenation operator. The following sections explain the source, processing, and storage methods of the three parts.

[0096] static feature vector Construction:

[0097] Static features focus on attributes that change slowly over time or remain stable over a long period, including but not limited to age group, geographic region, account duration segments, historical average order value segments, preference category sets (multiple selections), and budget range. Numerical fields (such as account duration days and historical average order value) are standardized to improve the comparability of features with different dimensions. Standardization uses the mean-variance method, and the original value is denoted as... The sample mean is The sample standard deviation is Stability factor is The standardized value is Categorization fields (such as geographic region and age group) use one-hot or multi-hot encoding with a fixed vocabulary, and multi-category preferences are expressed through fixed-length bucket vectors; if multiple categories exist, the system can use static feature vectors. Internally, a fixed-length counting bucket is maintained. Value pruning and logarithmic compression are used to compress excessively large count values ​​to a comparable range. To avoid privacy violations, all static features do not contain sensitive information that can directly identify individuals and are processed through finite-precision segmentation and intervalization before being stored in the database.

[0098] Preference spectrum vector Construction:

[0099] Preference spectrum is used to characterize the periodicity and rhythmicity of user clicks or interactions over time. This embodiment will focus on the most recent The sequence mapping of the effective interactions is of length [length missing]. The real-valued sequence is obtained by using discrete cosine transform to obtain the first... The first-order spectral coefficients form the preferred spectral vector. Its calculation uses the standard cosine basis expansion form: for each order... Let the normalization factor be Then the preferred spectrum vector The Each component is calculated using the following formula: ,in Indicates the first The real-valued representation of each interaction (e.g., a scalar score obtained by mapping according to category or product attributes). Pi is a constant. After the calculation is completed, the preference spectrum vector is... Perform amplitude normalization (e.g., unit L2 normalization or cropping by quantiles before scaling). (Interval), to suppress instability caused by abnormal peaks. In the default settings, the order is preserved. It can be configured according to the deployment environment and storage overhead. Adjustments can be made within this range. This frequency domain modeling can capture low-frequency rhythms such as "weekday / weekend" and "day / night" without introducing complex sequence models, and it is significantly more stable for early-segment candidates in session-level ranking (in our internal offline playback, we added a preference spectrum vector). Afterwards, the median click-through rate of the top 10 items in the cold start subsample increased by approximately 3.1%, and the variance decreased by approximately 4.5%.

[0100] Context anchor vector Construction:

[0101] Context anchors focus on short-term signals strongly relevant to the scene within the most recent window, such as access points (site search, event pages, external advertising), time slices (hours, days of the week), device type (mobile / desktop), and coarse-grained geographic regions. The system uses a time-decaying weighted average to calculate a smoother representation that is more sensitive to the latest interests. The time difference between each event and the current time is The time decay factor is ,in Let be the decay constant; let the vectorized description of the event be... (composed of the embeddings of entry type, time slice, and device type), then the context anchor vector The weighted average yields: To reduce the impact of data sparsity, event vectors... Low-frequency categories are smoothly merged or mapped to the "Other" channel before being input into the model; to ensure robustness during deployment, the context anchor vector... Before output, perform component clipping and scale alignment to make its magnitude similar to the static feature vector. Preference spectrum vector They are on the same order of magnitude.

[0102] User profile vector Updates and storage:

[0103] To adapt to online incremental updates and multi-session concurrency, the system uses an exponential moving average method to analyze user profile vectors. Perform a smooth update. Let the vector of the "temporary image" generated by the current observation be denoted as . Step size is The update rule is then... To avoid numerical divergence, the system uses user profile vectors... Apply norm constraints: when the L2 norm exceeds a threshold Execute proportional scaling at time, i.e. Online storage uses quantization compression, with a scaling factor of [value missing]. Zero point is The quantization code is Press again when reading Dequantization recovery; this strategy reduces storage and network transmission bandwidth by approximately 60% to 70% while maintaining a sorting consistency rate of no less than 99.5%.

[0104] Quality control and privacy compliance:

[0105] All user profile vectors All fields were collected according to the principles of data minimization and necessity; missing values ​​were included in the static feature vector. Context anchor vector The model is populated using a combination of reliable default values ​​and indicator bits; sensitive fields appear only in range or segment form, and any information that directly identifies an individual is not allowed to be included in the model. The profile version and processing flow are written to the audit log for easy offline traceability.

[0106] It should be noted that this plan also includes:

[0107] In calculating the preference spectrum vector Previously, for the interaction sequence A robust loss-based denoising transform is applied to suppress abnormal clicks. The optional piecewise robust loss is defined as follows: , where the threshold The denoised sequence is automatically selected from the validation set and then enters the aforementioned discrete cosine transform.

[0108] Updating user profile vectors At that time, a "scene gating" step size is introduced, that is, a step size... The function is set as the entry type and device type, so that the impact of different scenarios on the image is controllable, thereby reducing "image jitter" when migrating across scenarios;

[0109] To ensure cross-platform consistency, the system allows the identity binding confidence score calculated in step S1 to be converted into a static feature vector. The one-dimensional credibility feature is used, and segmentation and pruning are performed during the online phase to avoid affecting the preference spectrum vector. and context anchor vector It has an excessive impact.

[0110] In the offline replay of the public dataset and the anonymized online logs, a preference spectrum vector was added. Afterwards, compared to the baseline using only behavior counts, the median click-through rate of the top 10 users of cold-started systems increased by approximately 3.1%, and the robustness metric of the session-level list (variance by position 1–10) decreased by approximately 4.5%; context anchor vectors constructed using time decay were introduced. Afterwards, the first-screen click-through rate across entry points increased by approximately 1.8%; after adopting quantized storage, the consistency rate between retrieval and sorting remained above 99.5%, and the latency P95 increment was less than 0.6ms.

[0111] Through the above processing, step S2 standardizes multi-source, multi-granular user leads into well-structured, online-updable user profile vectors. This profile is expressive in both the frequency and time domains, and its numerical stability and privacy compliance strategies ensure that it remains stable, verifiable, and auditable under different business and traffic fluctuations.

[0112] In this embodiment, step S3 will be described in detail. In step S3, the system generates a candidate product set for the current session and constructs a product attribute vector for each product in the set. This vector is similar to the aforementioned user profile vector. Together as subsequent single-user relevance Conversation-level rating The input for the calculation is as follows. The generation of the candidate product set follows the principles of "prioritizing availability, consistent constraints, and low latency." First, basic availability filtering is performed based on listing status, regional availability, real-time inventory, and compliance rules. This is combined with the session identifier established in the previous step. With compliance tokens Compliance-based filtering is performed on geographic location, age group, and sensitive categories. Then, initial search criteria are determined based on the short-term intent of the current session (source entry point, search terms, time slice, and device type, etc.). Subsequently, a candidate set is formed using a multi-path recall approach. Engineering constraints such as "merchant allocation, category allocation, and historical exposure suppression" are added to this set to ensure that the output of this step naturally meets the requirements for subsequent optimization (including including exposure counts). Input requirements (with rearrangement of diversity constraints).

[0113] To ensure that candidates can be mapped to user profile vectors in a unified vector space To perform relevance estimation, the system constructs a product attribute vector for each product. Product attribute vector A layered, structured representation is adopted, comprising four sub-parts: structured attribute encoding, text content encoding, media content encoding, and behavioral statistics encoding. These are concatenated in a fixed order, scale-aligned, and then stored uniformly. This concatenation relationship is represented by a standard mathematical expression as follows: ,in" The "" indicates a concatenation operation. Structured attribute encoding covers discrete or segmented fields such as category hierarchy, brand identifier, price segments, merchant reputation segments, geographical shipping location, and logistics timeliness segments; text content encoding is based on sub-words or phrases from the title, selling points, and parameter descriptions, with length pruning and stop word filtering; media content encoding is based on an aggregated representation of global features and a small number of local regional features from the main image and key auxiliary images; behavioral statistics encoding includes bias-free estimates of historical click-through rate, add-to-cart rate, and payment rate, as well as exposure frequency across time periods. To avoid relevance bias caused by inconsistent dimensions of the sub-vectors, the system performs a concatenation of the product attribute vectors. Norm constraints are applied, and the constraint method is expressed using standard mathematical expressions. ,in Represents the L2 norm, This is an adjustment parameter for the upper bound of the vector norm. This scale alignment can consistently improve the consistency rate of similarity estimation in both offline playback and grayscale streaming, and reduce the adverse effects of extreme fields on the ranking process.

[0114] The generation of the candidate product set adopts a multi-path parallel recall strategy, and each recall path can be independently scaled up and down. At the same time, the routing layer is based on session-level user weights. A certain recall share is reserved for each user within the session to enhance the coverage of each user's candidates in subsequent stages. Specifically, content-related recall uses user profile vectors. With product attribute vector The search key uses near-nearest neighbor vector retrieval to ensure that products directly related to the current user's interests are included in the candidate list. Rule-related recall uses entry signals and search intent to locate categories and price segments, supplementing compliant and marketable products that may have been overlooked in the content. Collaborative-related recall, without disclosing any personal identifiers, introduces currently popular but not directly matched products that meet marketability constraints based on the popularity vector of similar user groups, thereby improving relative coverage in less popular categories. Business strategy recall is used to inject compliant, public welfare, or major event products with strong constraint tags, ensuring that regulatory and business objectives are met. To align with the hard constraints such as "merchant exposure cap" and "category exposure cap" in subsequent optimizations, the system labels each candidate with key attributes such as merchant, category, logistics area, and price segment in this step, and simultaneously includes an exposure count. The latest statistics facilitate direct access to the exposure penalty term of the objective function in subsequent stages without the need for secondary queries.

[0115] To suppress the staleness of the list and avoid excessive concentration on popular items, the system implements a light historical exposure suppression at the candidate layer. This suppression is based on exposure counts. As a signal, when the exposure count of a product within the near window exceeds a threshold, a linear reduction is applied to the product's priority in the candidate layer, and a similar replacement is initiated when the proportion of candidates from the same merchant or category reaches its upper limit. In practical engineering, this lightweight suppression can be combined with the exposure penalty in subsequent optimization. The former is used to quickly reduce duplicate recall, while the latter is used to determine the final position of the global optimum, thereby achieving a collaborative division of labor of "recall redundancy - global reordering" without significantly increasing latency.

[0116] Considering the needs of cold starts and new product exploration, this step provides an independent "New Product Protection Channel" at the candidate level. This channel grants a small priority to products launched within a short window and meeting the availability constraints, allowing them to gain initial exposure without affecting the robustness of the main channel. To control risk, New Product Protection is only enabled when the session entry and scenario meet specified conditions and is subject to hard limits on session-level budgets (e.g., "a maximum number of new product candidates can be introduced per session"). Gray-scale experiments show that when New Product Protection is introduced and its quota does not exceed 5% of the candidate size, the probability of new products appearing on the first screen increases significantly, while the fluctuations in overall clicks and conversions are acceptable within the statistical significance threshold.

[0117] To seamlessly integrate the output of this step with the diversity constraints of subsequent steps, the system constructs product attribute vectors. Simultaneously, a set of product difference evaluation vectors is pre-calculated offline. It is cached along with the product entity for use in subsequent stages of diversity measurement (this measurement participates in the dissimilarity assessment using cosine distance in subsequent stages). (Calculation of the product difference assessment vector). It should be noted that this step only covers the product difference assessment vector. The generation and caching of data do not involve calculating the difference in the candidate layer. To avoid introducing many-to-many comparisons during the recall phase and amplifying the latency; through grayscale evaluation, this approach controls the candidate layer latency to the sub-millisecond level increment, while the subsequent optimization phase can still obtain the same diversity constraint input as offline.

[0118] It should be noted that, in the offline replay of public datasets and anonymized online logs, after adopting the above-mentioned multi-path recall, compared with the baseline of using only content-related recall, the median coverage of the candidate set to real clicks (i.e. whether real clicked products enter the candidate set) increased by about three to five percentage points, and the 95th percentile latency increment remained within milliseconds; after enabling lightweight exposure suppression and new product protection, the deduplication rate of the candidate set was significantly improved (the proportion of duplicate products decreased by about two percentage points), and the first-screen reach rate of new products increased by about one percentage point in the control group, and no statistically significant adverse transfer effect was observed.

[0119] It should be noted that this embodiment also includes:

[0120] Firstly, based on session-level user weight. First, a recall quota is allocated to each user within the session, ensuring that each user receives minimum coverage commensurate with their weight in the candidate layer, thus maintaining consistency with the subsequent "minimum member coverage" from the candidate stage. Second, the "strongly constrained tag table" generated in the previous steps (obtained by depth-limited expansion of a directed acyclic graph based on synonym relationships) is executed in the candidate layer using hard filtering to prevent obviously non-compliant or mutually exclusive products from entering the candidate set, thereby reducing the optimization pressure in subsequent stages. Third, the judgment of salability not only uses static listing and inventory but also introduces short-term salability probability prediction. Only when this probability exceeds a threshold is the product included in the candidate set, mitigating the problem of "unsalable products entering the candidate" caused by data latency. This predictive feature is integrated with the product attribute vector. Cache them together for easy use in subsequent processes.

[0121] In summary, step S3, based on consistent filtering and multi-path recall with multi-source constraints, utilizes a uniformly constructed and standardized product attribute vector. Structured attributes, text semantics, media information, and behavioral statistics are incorporated into the same vector space. Lightweight control over exposure and allocation is implemented at the candidate layer, and exposure counts are prepared for subsequent optimization. Product Difference Assessment Vector Wait for necessary input, thereby improving the relevance for a single user without sacrificing latency and stability. Conversation-level rating Effective estimates provide a candidate base that is of controllable quality and audit-friendly.

[0122] This embodiment will describe step S4 in detail. In step S4, the system calculates the single-user relevance for each user and each candidate item within the session. And simultaneously calculate session-level user weights. Single-user relevance Reflecting "user profile vector" "and "product attribute vector" "Fitness in a unified metric space; session-level user weights" This is used to quantify the speaking power of different users at the session level, and is ultimately used in subsequent steps through session-level scoring. Aggregate and sort the candidate products.

[0123] To improve the consistency and robustness of metrics, single-user relevance The calculation is divided into three steps: feature alignment, baseline similarity estimation, and confidence lower bound discounting. Firstly, the system first processes the user profile vector... With product attribute vector Mapped to a metric space of the same dimension and normalized. The mapped user vector is denoted as... The commodity vector is denoted as The unitization rule is , ,middle It is the numerical stability constant. It is a 2-norm. Secondly, the benchmark similarity is calculated in the normalized space, defined as... This inner product is equivalent to cosine similarity, facilitating comparability across features of different dimensions. Thirdly, to suppress overly optimistic estimations caused by insufficient samples or noise, the system performs a bootstrap (resampling) confidence lower bound discount on the benchmark similarity: using the number of resampling operations as... ,get (Base similarity of the b-th resampling), calculate the bootstrap mean and bootstrap standard deviation. , Based on this, a single-user relevance score is given. ,in Here is the confidence coefficient, and the range of values ​​is... This conservative estimation form, "mean minus standard deviation multiple," effectively reduces ranking jitter caused by high-variance samples. For cold start scenarios or scenarios with near-zero historical behavior, the system provides a fallback calculation of content-popularity hybrid: defining content similarity as... Global popularity is (Based on global exposure - click statistics normalization), then the cold start relevance is... ,in When the number of valid near-end interactions by a user is less than a set threshold, segmented weights will be applied. and Perform linear fusion to ensure availability without changing the main process.

[0124] Session-level user weight The aim is to reasonably weight the influence of different users within the same session. The system employs a three-signal convex combination and in-session normalization: let the "identity consistency F1 score" be... (Measures the consistency and reliability of account-device-session interactions), "Recent Interaction Strength" is... (Measures activity and intent strength within the most recent window), "Historical payment probability" is (To measure long-term conversion tendency), construct intermediate weights and normalize them. , ,in, , and Identity consistency F1 score Based on precision and recall calculations: Let the match count be... The number of times it appears on the equipment side is The expected number of times on the account side is ,but , , Recent interaction intensity We obtain the result by weighting the events through time decay: Let the... The time difference between each event and the current time is The attenuation constant is Event intensity coefficient is (For example, click to get 1, add to cart to get 3, place order to get 5), then and in the interval Inner linear scaling. Historical payment probability. Conversion rate estimation using exponential sliding motion: Let the time difference between the o-th order and the current time be... The attenuation constant is The binary instruction of whether to pay is: ,but All three signals mentioned above have had their dimensions consistent through both offline and online normalization, with intermediate weights. A one-time normalization is performed at the session level to obtain the session-level user weight. Thus satisfying .

[0125] It should be noted that this embodiment also includes:

[0126] Introducing a "user confidence discount factor" to reflect the stability of a user's ratings in the current session: taking the average relevance score of that user to a pre-selected set within the session as the value. The variance measure is Confidence discount factor And perform on the intermediate weights. Therefore, users whose ratings fluctuate significantly within a session will have their session-level user weight reduced. It will be subject to bounded downward adjustment, which helps reduce list jitter caused by group disagreements.

[0127] In addition, "cross-user calibration" is introduced to mitigate differences in scoring scales among different users: affine calibration is performed on the relevance of a single user. ,in and The results are estimated from the exposure-feedback pairs in the previous window, and the stability of the ranking relative relationship is guaranteed by monotonicity constraints (e.g., quantile mapping or equivalent rank regression); if this calibration is enabled, it is used in session-level scoring. replace .

[0128] Regarding engineering verifiability, offline playback versus online grayscale display: compared to using only benchmark similarity. Compared to the baseline, a single-user relevance was introduced by introducing a self-help confidence lower bound. Subsequently, the intra-session volatility variance of the top 10 click-through rates decreased by approximately 4%–6%, while simultaneously increasing the median overall click-through rate by about 1%. Under the same total exposure budget, the session-level user weight obtained by using a three-signal convex combination and normalization... .

[0129] Compared to uniform weighting, the percentage of "covered users" within the group (i.e., at least one candidate for the Top-K comes from their respective highly relevant lists) increases by approximately 2–3 percentage points. This is achieved by introducing a user confidence discount factor. In the grayscale experiments, the list jitter metric (measured by the symmetric difference between two consecutive refreshes of Top–K) decreased by about 3%–5%, and the latency P95 increment remained in the sub-millisecond range, meeting the online performance boundary.

[0130] To facilitate auditing and compliance, the system persists "relevance version number and confidence coefficient" for each session in this step. Weighting coefficients , Attenuation constant , And whether the "user confidence discount factor" is enabled. The system offers two optional features: “Cross-user calibration”. These metadata entries, along with the session logs from the preceding steps, are written into the audit log to facilitate traceable verification of experimental conclusions and technical effects during subsequent review stages.

[0131] In summary, step S4 achieves robust single-user relevance. "and normalized session-level user weights" With this as the core output, while ensuring consistency and real-time performance, it provides an interpretable, verifiable, and easily implemented quantitative basis for subsequent session-level scoring and ranking.

[0132] In this embodiment, step S5 will be described in detail. In step S5, the system obtains the aforementioned single-user relevance. Session-level user weight Next, the session-level score for each candidate product within the session is calculated. This is used to provide the overall strength of preference for a product within the group; subsequently, a robust median is calculated based on the rating distribution of all users for the same candidate product within that session. Absolute deviation from median and group uncertainty indicators Defined as The median The calculation is as follows .

[0133] The above definition is based on single-user relevance. The median within a session and the median of the absolute deviation measure the degree of "disagreement" within a group regarding the same candidate product. When group uncertainty indicators... A larger value indicates a greater variation in the distribution of opinions among different users regarding the candidate; when the group uncertainty index... When the value is close to zero, it means that the group is more in agreement on the candidate.

[0134] To prevent frequent switching of gating states near the boundary, the system uses population uncertainty indicators. Implement dual-threshold hysteresis gating. Specifically, first calculate the activation threshold on the current candidate set. With the release threshold Both employ adaptive quantile settings: , ,in Represents the current candidate set. express Quantile operations, and satisfying The system assigns a maintenance gating status symbol to each candidate. Where 0 represents a "free state" and 1 represents a "confined state". The population uncertainty index obtained in this round of calculations... Update the gate status flag afterward. :

[0135]

[0136] This update rule reflects hysteresis: only when the group uncertainty indicator... Significantly higher than the activation threshold Only when a candidate transitions from a "free state" to a "restricted state" can they exit the "restricted state"; however, exiting the "restricted state" requires a population uncertainty indicator. Significantly below the release threshold This "high-gate in, low-gate out" design can significantly reduce the repeated switching caused by edge samples near the threshold.

[0137] After obtaining the gating status signal for this round Subsequently, the system forms a restricted set of candidate goods. For clarity, let's denote the "free subset" as... "Restricted subset" is Candidates in the "free subset" can directly proceed to the next stage; for the "restricted subset," the system only retains entries that satisfy the label-strong constraint table given in the previous step (this table has been defined in the preceding steps and generated through a depth-limited expansion of the directed acyclic graph of synonym relations), otherwise it will not proceed to the next step in this stage. Thus, the restricted candidate product set... Defined as In implementation, the determination of "verification through label strong constraint table" is executed in series with the aforementioned restrictions (such as regional marketability and compliance items) in the same verification pipeline, thereby avoiding repeated parsing and multiple I / O operations.

[0138] To balance real-time performance and robustness, a threshold is enabled. With the release threshold The quantile calculation employs a streaming quantile estimation algorithm (e.g., based on...). The system employs an approximate or equivalent constant spatial quantile estimate, capable of updating quantile estimates within milliseconds while maintaining numerical stability even with dynamic changes in sample size. For extreme cases involving small sample sessions (e.g., fewer than 2 users or fewer than 10 candidates), the system implements a "soft degradation strategy": preserving population uncertainty indicators. Used for logging and auditing, but only temporarily configured. The gating is closed to avoid false positives due to unreliable statistics; once the sample size recovers to above the threshold, it automatically reverts to the normal double-threshold hysteresis process.

[0139] To reduce the impact of outliers on robust statistics without altering the basic definitions, the system calculates the median... Absolute deviation from median Previously, the relevance of a single user Perform symmetric limiting, i.e., first calculate the session-wide limit. For quantile intervals, scores falling outside the interval are Windsor-trimmed before being used to calculate the median and median absolute deviation. This process does not change the robust nature of the median and median absolute deviation, but it can improve the stability of threshold estimation under extreme data distributions. For computationally resource-sensitive deployments, this limiting can be enabled or disabled as an optional feature via a configuration switch.

[0140] Regarding the connection with subsequent steps, the system outputs two types of core data at the end of S5: one is a limited set of candidate products. Its corresponding session-level score The first is used for subsequent two-stage selection and optimization; the second is the population uncertainty index for each candidate. Gating status symbols This is used for diversity processing and exposure penalty weight reference in subsequent stages (for example, stronger exposure penalties can be imposed on candidates in the "restricted state" in subsequent stages, or candidates in the "free state" can be given priority in the case of the same score, so as to prioritize the satisfaction of group consistency under the condition of equivalent quality).

[0141] It should be noted that:

[0142] In offline replay of publicly available data and anonymized online logs, a population uncertainty indicator is introduced. After dual-threshold hysteresis gating, the median difference between two consecutive refreshes of the Top-K list (as a "list jitter metric") decreases by approximately 5% to 7%.

[0143] The median percentage of "covered users" (i.e., at least one highly relevant candidate for each user entering the Top-K) increased by approximately 1.5% to 2.3% within the conversation;

[0144] With the same exposure budget, the click-through rate of P50 remained basically flat or increased slightly (approximately 0.5% to 1.0%), while the stability of P10 was significantly improved, indicating a smoother experience for "long-tail conversations".

[0145] The above data can be verified proportionally in a gray-scale environment, and the activation threshold is continuously recorded in the logs. Removal threshold Gating status indicator Group uncertainty indicators The version and value range are specified to facilitate subsequent auditing and traceability.

[0146] It should also be noted that this embodiment includes:

[0147] "Weight-aware robust divergence measurement" refers to using session-level user weights without altering the core workflow. By assigning weights to the absolute deviations of users within a session, an approximate form of a weighted, perceived group uncertainty index is formed. ,in Indicates user weight by session level Weighted median calculation; this feature is disabled by default and should only be enabled in business scenarios requiring stronger "dominant user stability". "Gating strength tiering" refers to applying group uncertainty indicators within the "constrained state". With the activation threshold The relative positions of the labels are used to divide the levels into mild and severe restrictions, and in subsequent stages, different levels of label constraints or higher exposure penalty coefficients are matched respectively.

[0148] In summary, step S5 uses the median Absolute deviation from median To establish robust statistics, construct population uncertainty indicators. And adopt an enable threshold With the release threshold The dual-threshold hysteresis strategy stably outputs a limited set of candidate items. Without significantly increasing latency, it significantly reduced list jitter and improved group consistency; at the same time, it achieved seamless integration with the label-strong constraint table in the preceding steps and the two-stage optimization and exposure penalty in the subsequent steps.

[0149] This embodiment will describe step S6 in detail. In step S6, after completing the aforementioned dual-threshold hysteresis gating, the system only performs label-level strong constraints and conflict resolution on the candidate subset in the "constrained state," thereby significantly reducing the probability of group divergent samples entering the next stage without sacrificing the overall relevance of the session. To this end, the system first constructs a label strong constraint table. This table is derived from three core elements: the intersection of positive conversation tags. Union of negative labels in conversation Compared with directed acyclic graphs based on synonyms The expanded set of negative labels (with a maximum path length of 2). Subsequently, a label-based strong constraint table was used. Perform retention / removal decisions on "restricted" candidates and output structured "reason for retention / reason for removal" diagnostic information for auditing purposes.

[0150] Each user maintains a set of positive tags within the session. (Categories, brands, effects / attributes, etc. that express "desirable / not objectionable" and a set of negative tags) (Entries expressing "explicitly unwanted / should be blocked") are tagged with a uniform product ontology; each candidate product maintains a set of product tags. (Composed of category, brand, key attributes, target audience, sensitive attribute identifiers, etc.). Synonyms and near-synonyms are represented by a directed acyclic graph of synonym relationships. Organization, among which For tag node set, Synonyms / near-synonyms / mapped edges; upper bound on the path length traversed along the edges. .

[0151] The system first performs session-level aggregation of each user's preferences within the session. (Session positive label intersection set) Defined as This represents the smallest positive anchor point commonly agreed upon by all users; the set of negative session labels. Defined as , representing the set of tags explicitly rejected by any user. To improve robustness, the system allows temporary degradation to "majority intersection / union" for extremely small sample sessions (e.g., the intersection is weighted by session-level user weights and a pass threshold is set), but this optional feature is disabled by default and is only enabled when the session size is below the engineering threshold and logged to the audit log.

[0152] To avoid misjudgments due to differences in word meaning or subtle naming, the system uses a union set of negative session tags. Perform synonym expansion to obtain a set of negative label expansions. The extension adopts a set A directed traversal starting from a given point, where the shortest path length does not exceed 2, is defined as... This depth-limited expansion effectively covers the negative semantics of "synonyms / aliases / equivalent SKU families" while avoiding mispropagation caused by long chain propagation. To suppress noise, the system constructs a directed acyclic graph of synonym relationships offline. A quality score is set for each edge and a lower limit filter is applied. Online expansion only retains edges that pass the quality threshold.

[0153] After obtaining the above three types of sets, the system forms a set of allowed tags. Its definition Then, a label-strong constraint table is constructed. The table records allowed / prohibited entries as (label, action, priority) triples: Each tag in the register is "Allowed (Priority 1)", for Each label in the list is registered as "prohibited (priority 0)". Finally, for each candidate item in the "restricted state"... Perform the determination and define the retention determination function. for .

[0154] In other words, a candidate product can only be released from the "restricted state" to the next stage if its label does not match any "prohibited" entries and matches at least one "allowed" entry. To ensure consistency, if a candidate product matches both allowed and prohibited entries, it will be eliminated according to the "prohibited priority" criterion, and the triggered prohibited label will be recorded in the diagnostic information for auditing and subsequent policy adjustments.

[0155] Differentiated handling of hierarchy and dimension:

[0156] Product tags typically span multiple dimensions (category, brand, function / attribute, target audience, compliance modules, etc.), and the system uses a tag-strong constraint table. The system registers permitted and prohibited items separately by dimension, and adopts a "hard dimension priority" strategy when making judgments: any prohibited item that matches the "sensitive / compliance dimension" is prioritized for removal; non-sensitive dimensions follow the aforementioned general rule of "prohibition priority + permitted occurrence". If a session contains strong guidance labels such as "event / promotion", an additional soft permission flag can be registered in the "event dimension" so that this dimension can be used as a breaking criterion for parallel scoring in boundary cases without changing the hard judgment.

[0157] Label strong constraint table A "versioning + caching" strategy is adopted: multiple decisions within the same session share the same table construction result, and the "prohibited" label set is stored using a compact bitmap or Bloom filter to approximate the decision in O(1); the "allowed" label set is stored as a hash set for fast intersection calculation. For large-scale sessions or scenarios with high label dimensions, the system supports segmented construction: first, a read-only segment is built based on the "sensitive / compliant dimensions"; after the segment is judged and passed, the remaining dimensions are loaded as needed to reduce cold path I / O. For easy traceability, the system outputs a quadruple (candidate identifier, hit label, action, table version number) to the session log with each decision.

[0158] The input side of step S6 and the "restricted state" candidate subset of step S5 Interconnection; Interconnection between the output side and the subsequent two-stage selection and optimization: On the one hand, the decision function will be retained. Candidates are backfilled into the availability zone of the "restricted candidate product set"; on the other hand, the "removal reason" can be transformed into exposure penalty or sample sampling weight in the subsequent objective function to avoid repeatedly triggering the same restricted reason during the session lifecycle.

[0159] It should be noted that in the offline replay of de-identified online logs, a strong constraint table for tags is enabled. Afterwards: ① The median increase in "cross-user consistency metrics" (the percentage of sessions where each user has at least one candidate in the Top-K) was approximately 1.2%–2.0%; ② The median decrease in "list jitter metrics" (the symmetrical difference between adjacent Top-K refreshes) was approximately 2.5%–3.8%; ③ The manual review rejection rate involving compliance dimensions decreased by approximately 12%. The P95 latency increment for online low-traffic grayscale displays remained in the sub-millisecond range (mainly from set intersections and Bloom filter queries).

[0160] It should also be noted that this embodiment includes:

[0161] Weight-aware positive set of sessions: Constructing the intersection set of positive set of sessions Introducing session-level user weights Replace strict intersection with "weighted occurrence rate not lower than the threshold" The set of ", i.e. ,in ;default =1 is equivalent to a strict intersection.

[0162] Negative expansion of distance decay: in a directed acyclic graph of synonym relations In the depth-limited extension, different confidence weights are applied to the extensions with distances of 1 and 2, forming an interpretable layered prohibition strength (e.g., giving stronger exposure penalties in subsequent stages).

[0163] Dynamic adjustment of dimension priority: Based on the attributes of the session entry and compliance token, the priority of "sensitive / compliance dimensions" is dynamically adjusted to further reduce false recalls in heavy audit scenarios without excessively affecting the overall relevance.

[0164] In summary, step S6 involves constructing and applying a label-based strong constraint table. The set of intersections of positive session labels As positive anchor points, a directed acyclic graph based on synonyms. The negative label expansion set obtained by depth-limited expansion To prohibit boundaries, a consistency constraint framework of "allowing priority anchoring + prohibiting priority pruning" is established; in engineering, it takes into account both low latency and auditability, and provides cleaner and easier-to-achieve group consensus restricted candidate inputs for subsequent two-stage optimization.

[0165] In this embodiment, step S7 will be described in detail. In step S7, the system takes the limited candidate product set output in step S5. Starting from this point, we combine the single-user relevance obtained in step S4. Session-level user weight First, calculate the session-level score for each candidate. Then, a two-stage process of "pre-selection + constraint optimization" is used to determine what should be displayed in this round. The goal of the two-stage design is to rapidly compress a large number of candidates into a medium-sized "high-value subset" with low latency, and then apply hard constraints such as coverage, fairness, and diversity to this subset to make a globally optimal selection, thereby taking into account both individual quality and group consistency.

[0166] Phase 1: Rapid pre-selection based on conversation-level scoring:

[0167] First, in the limited set of candidate goods Rate by conversation level Sort in descending order and take the first few. One candidate is added to the pre-selection set. In the default settings, the candidate number threshold is... Number of targets Satisfy linear proportional relationship This ensures that the second phase remains feasible within an acceptable scale. For sessions with highly uneven rating distributions, an adaptive strategy of "proportion + quantile" can be adopted: Let the session-level ratings be... quantile ,make ,in This is an upper limit constant. If parallel boundaries occur, the exposure count will be used. Smaller values ​​are prioritized (to avoid duplicate exposure), and then evaluated based on product difference vectors. Prioritize candidates with larger average distances to their neighbors (preferring to retain more differentiated candidates). The output of Phase 1 is denoted as the "pre-selection set". .

[0168] Phase Two: Solving Constraint Optimization on the Pre-selected Set:

[0169] In the "pre-selected set" Define selection variables above (Taking 1 indicates that the candidate) Before being selected in this round An optimization model is constructed with the unified goals of "quality, deduplication, and fairness" as its guiding principle. The objective function adopts a weighted sum of "session-level scoring, historical exposure penalty, and diversity reward": ,in To expose the penalty coefficient, For diversity coefficient, the degree of difference This is used to encourage the selection of sets that are "dissimilar to each other." To ensure dissimilarity... The computability and interpretability of the difference are based on the product difference assessment vector. Defined as the minimum cosine distance to the already selected set: , .

[0170] To transform the above "minimum value" form into a linearly solvable form, a lower bound variable of variability is introduced. with auxiliary variables And the following linearization constraint is given (the cosine distance is bounded by...). Therefore, the constant M=2 is sufficient.

[0171]

[0172]

[0173] The target by Substitution yields an equivalent linear objective term. Therefore, the complete optimization model is: Under the following constraints:

[0174] (Quantity constraints);

[0175] (Minimum member coverage constraint);

[0176] (Merchant exposure limit constraints);

[0177] (Category exposure limit constraint);

[0178] (Difference linearization);

[0179] (Variable domain).

[0180] in, Indicates candidate In users Single-user relevance Ranking in descending order; default ranking threshold. Pick Minimum number of members covered Pick ; For merchants candidate set Set its exposure limit; Category candidate set Set its exposure limit.

[0181] Feasibility assurance and remediation strategies:

[0182] Due to multiple constraints related to "quantity, coverage, upper bound, and diversity," the model may become infeasible in certain sessions. Before proceeding to the solution, the system performs a linear-time feasibility pre-check: calculating the total lower bound of the coverage constraints. ,like Then press first Decrease the lower bound of the coverage for each user group by one, starting from the smallest and gradually increasing, until the condition is met. If the overlap between the merchant and category limits results in insufficient feasible candidates, a "soft relaxation" sequence will be initiated: prioritizing raising the ranking threshold. (For example, increasing from 3 to 5), then reducing the lower bound of individual users by 1 according to the "lower weight priority" order, and finally setting the upper limit for the most critical merchants or categories. or Add 1. Record the reasons for all relaxations and send them back to the audit log so that targeted governance can be carried out in the future.

[0183] Solution methods and performance engineering:

[0184] In online scenarios, pre-selection scale Typically in the tens to hundreds of units (e.g.) hour This size of 0–1 integer programming can be completed in milliseconds using commercially available or lightweight open-source solvers. To further reduce tail latency, the system provides the following engineering measures:

[0185] Hot start: Using the solution from the previous round (or a near-similar session) as the initial feasible solution greatly reduces node exploration;

[0186] Generative dissimilarity terms are listed: First, the dissimilarity term is temporarily omitted in the objective function, and only feasible solutions are found. Then, several terms most relevant to the current solution are gradually added. Constraints and a second solution are typically applied, and convergence usually occurs in 1–2 rounds.

[0187] Linear relaxation + rounding: First, calculate the LP relaxation to obtain a fractional solution, then calculate the fractional solution by "score value × conversation level rating". "Weighted rounding is performed and the solution is used as an alternative or hot-start solution."

[0188] Pruning and caching: Generate a pre-computed "feasible bitmask" for merchant and category limits, and create constraints only for potentially active items when generating the model, significantly reducing the number of constraint entries.

[0189] In internal grayscale, the above engineering processing controls the P95 solution latency to within the range of 3–8 milliseconds (depending on L, constraint density, and solver configuration), meeting the requirements for real-time display.

[0190] Connection with upstream and downstream steps:

[0191] The output of step S7 is a size of The selected set And the corresponding diversity lower bound Differential Agent The selected set will undergo a rotation and position penalty strategy for position-level reordering in the next step; simultaneously, the exposure count used in the optimization will be... Differentiation Agent The relaxed record will be written into the log to provide a basis for subsequent feedback, learning, and offline review.

[0192] It should be noted that in the offline replay and online small-traffic gray-scale testing of anonymized logs:

[0193] Compared to scoring only at the session level Take before After optimizing the baseline by introducing constraints of "coverage + upper limit + diversity", the median "covered user percentage" (each user having at least one selected item in the session) increased by approximately 2.0% to 3.1%.

[0194] Merchant concentration (Herfindahl index) decreased by approximately 6%–9%, ​​and category concentration decreased by approximately 4%–7%.

[0195] The average cosine distance within the list increased by approximately 8%, and the proportion of duplicate items decreased by approximately 2 percentage points.

[0196] With the same exposure budget, the click-through rate (P50) remained basically the same, while the P10 increased by 0.6% to 1.2%, indicating that it is more friendly to long-tail conversations.

[0197] Online P95 latency increments remain within the range of 3–8 milliseconds. These metrics can be stably reproduced in a gray-scale environment, and all key parameters ( , , Both the upper limit table version number and the upper limit table version number are written into the audit log.

[0198] It should be noted that this embodiment also includes:

[0199] Target adjustment of uncertainty perception: Adjust the group uncertainty index in step S5 Introduce the target coefficient, let ,in >0 represents an amplification factor, which increases the exposure penalty for highly uncertain candidates under "constrained state" conditions; or... according to Adaptive scaling to enhance differentiation.

[0200] Average Dissimilarity Lower Bound Constraint: In addition to the existing "Minimum Dissimilarity Reward", add... To support more uniform diversity with average level constraints. The threshold is configurable.

[0201] Segmented Coverage: Extends the coverage constraint to a two-segment lower bound of "primary member coverage + secondary member coverage", To determine the threshold, users are divided into primary and secondary groups, and different thresholds are set for each group. Coefficients allow for finer-grained control over fairness without increasing the number of variables.

[0202] In summary, step S7 achieves a balance between individual quality, historical exposure control, and set diversity through a two-stage process of "session-level score-driven rapid pre-selection + constrained integer programming at a medium scale"; combined with coverage and upper limit constraints, it can maintain stable online performance while ensuring group consistency.

[0203] This embodiment will describe step S8 in detail. In step S8, the system receives the selection set (i.e., those candidates whose selection variable is 1) produced in step S7 as input and generates the display order. The design goal of this step is to ensure the visibility of multiple users within a session by using a "user-rotation" method without changing the quality of the selection set and the feasibility of constraints, while smoothing the position value with a "position penalty function" and using a "minimum difference threshold" to separate parallel entries, and finally outputting a list with a length equal to the target list size. An ordered list.

[0204] Input and token alignment:

[0205] The input for this step includes: the selected set (Results from the previous optimization step), Session-level score for each selected candidate The ranking function of each user relative to the candidate. (Relevance per user) (Descending order number), the difference between each selected candidate and the already selected set. (As defined by the aforementioned cosine distance), and the session-level user weights of users within the session. (For optional enhancements only). This step also uses position numbers. With position penalty function .

[0206] Definition and usage of position penalty function:

[0207] To reflect the general rule that "earlier positions have higher value and subsequent positions have decreasing value", this step uses a linear position penalty function. The position penalty slope value range .when Larger When the value is below 0, the system achieves stability by using the following limit pruning: Replace with During the sorting selection, the system uses a "position-weighted score" as the immediate comparison metric for this step: "Position-weighted score" "It is only used for positional comparison and does not change the previous selection set and constraint feasibility."

[0208] Selection process based on user rotation within a session:

[0209] To prevent highly relevant candidates for a particular user from being continuously pushed out as they move further down the list, the system employs a "user index rotation" selection process: starting from the position number... arrive During the generation process, users are rotated according to a fixed order within the session (e.g., ascending user index order); when it's a user's turn... When selecting candidates, those more relevant to the user are prioritized from the unselected pool. The specific process is described below.

[0210] 1) Initialize the placed set to empty and the position pointer. .

[0211] 2) Form a round-robin sequence according to the order of users within the session (e.g.) (Loop). Let the user in the current round be... .

[0212] 3) Filter out the "preferred candidate subset for this user" from the unplaced selection set, with the following priority criteria: Among them, the ranking threshold Same as the previous step, default. If the subset is empty, it degenerates into the set of all unplaced selected elements.

[0213] 4) When comparing candidates, first apply the minimum difference threshold. Only when satisfied Only when a candidate is selected is it added to the comparison queue. If the comparison queue is empty, the threshold is loosened in small steps until at least one candidate is added (ensuring feasibility).

[0214] 5) In the comparison queue, use "position-weighted scores". "Use the primary sorting key, with higher scores taking priority; in case of a tie, the exposure count will be used." Smaller values ​​take priority (reducing duplicate exposures); if still tied, then the degree of difference is used. The larger value takes priority (to increase diversity). Once the winner is determined, place them in the correct position. and will It proceeds in sync with the user rotation sequence.

[0215] Repeat steps 2) through 5) until the container is full. One position.

[0216] The core idea of ​​the above process is: to ensure a balanced opportunity for "who chooses" through user rotation, to ensure value decay by using a position penalty function to ensure "more cautious as you go further back," and to ensure "no crowding in parallel" by using a minimum difference threshold. This is different from directly scoring by session level. Compared to descending order, this process is closer to "visibility fairness" from the perspective of multiple users within a session.

[0217] Complexity and real-time performance:

[0218] At typical online scale (e.g.) Each step involves a candidate comparison that is equal to the size of the unplaced selection set, and the overall time complexity is approximately... By pre-maintaining several lightweight indexes (such as a preferred candidate min-heap for each user, and indexes based on user preferences), The global max heap can reduce the amortized complexity to In actual tests in grayscale environments, the P95 latency increment in this step is typically less than 1 millisecond.

[0219] Data integration and log auditing with preceding and following steps:

[0220] This step reads the selected set and session-level score output from step S7. Rank function Difference With exposure count It outputs the final display bit sequence and its accompanying "position-weighted score". "When the threshold is in place" Metadata such as "actual value" and "index of users on turn" are included. All decision points (such as abandonment due to difference threshold, parallel breaking due to exposure count pruning) are written to the session log to support offline review and compliance audit.

[0221] It should be noted that in the comparison between offline replay and online small-scale gray-scale testing of anonymized logs:

[0222] Compared to "speech-level rating only" With the "descending" display baseline, after introducing "user index rotation + position penalty + minimum difference threshold", the median "covered user percentage" (each user has at least one item entering the first screen) in the session increased by about 1.6% to 2.4%.

[0223] List jitter metric (the symmetric difference between two consecutive refreshes of the Top–K) decreased by approximately 2% to 3%;

[0224] The probability of consecutive duplicate merchants / categories decreases by approximately 8% to 11%;

[0225] Online P95 latency increments remain in the sub-millisecond range. These metrics can be stably reproduced in a gray-scale environment, and the core parameters have been recorded in the audit log. , User rotation sequence version number and threshold loosening record.

[0226] It should be noted that this embodiment also includes:

[0227] Weight-aware rotation sequence: While maintaining the basic framework of "user index rotation", a session-level user weight is introduced. Perform a "weighted rotation", for example, by changing the weights Discretized into quotas, forming a recurring rotation sequence, this allows high-weight users to gain an advantage in the starting position that matches their weight in multiple rounds of display.

[0228] Segmented position penalty: Applying the position penalty function Expanded into a piecewise linear form, as shown on the first screen... Set a steeper slope to emphasize earlier positions, and use a gentler slope in subsequent positions to keep the list stable.

[0229] Adaptive scheduling based on minimum difference threshold: The threshold is varied with location to form... The decreasing curve, for example ,in , To more strongly suppress similar terms in the first stage and moderately relax the restrictions in the subsequent stages to ensure placeability.

[0230] Local secondary fine-tuning: After completing a greedy placement from front to back, an additional constant-time 2-swap local search is added (swapping two adjacent items or two items within a small window). Swaps are only accepted if the "position-weighted score sum" increases without violating the difference threshold and merchant / category continuity constraints, thereby further improving sequence quality with almost no increase in latency.

[0231] In summary, step S8, while maintaining the feasibility of the selected set and constraints, utilizes the opportunity balancing mechanism of "user index rotation" and the "position penalty function" to achieve this. Value decay modeling, and using the "minimum difference threshold" "The parallel elements are interpreted to form a final display order that balances visibility for multiple users while controlling repetition and similar clustering; this process has low computational cost, is easy to implement, and is highly auditable."

[0232] It should be noted that the present invention also includes:

[0233] Based on the display results and logs from step S8, the system collects and displays the probability. With binary feedback Constructing a reweighted feedback And update the intermediate weight using the set of user-related exposures within the session as the cumulative domain. Then, the new session-level user weights are obtained according to the normalization rules. This creates a closed loop that influences the ordering and selection of subsequent sessions.

[0234] Exposure-level logs and tags:

[0235] The system creates a unique exposure entry for each display generated in step S8, with the entry's primary key consisting of the session identifier. Candidate Product Index with position number It consists of and records the policy version, user set and user index, and the session-level score used for the current display. Position penalty function Fields such as [list of fields]. This entry is cross-table linked with the session log from step S1 through the audit log to ensure traceability.

[0236] Display probability Definitions and records:

[0237] Display probability This indicates "candidate products" under a given strategy. "Appears in "location" The system decomposes the strategy probability into two parts: a deterministic component and an exploratory component. When there are no parallel positions or no random exploration, the deterministic ordering makes the probability of displaying the corresponding position 1. When there are parallel positions that are broken up or explicit exploration, the system records the probability according to the following rules: ,in, To explore intensity For the probability of juxtaposition under deterministic scoring (if position) The number of tied candidates is ,but , To explore distributions (e.g., uniform distribution across available sets or distribution by category / merchant ratio), and to avoid numerical explosions caused by extremely low probability, the system sets a minimum probability threshold. And record: The above. and The size of the parallel set and the exploration distribution identifier are written together in the exposure log to ensure reweighted feedback. It is recalculated and auditable.

[0238] Binary feedback Definition:

[0239] Binary feedback This indicates whether the exposure achieved a preset target event (such as a click, adding to cart, or placing an order). In the simplest setting, the system uses "click" as the event, recording 1 if it occurs and 0 otherwise. When the business requires "payment" as the primary target, the event definition can be switched to "payment". To accommodate multiple signal scenarios, the system allows recording auxiliary event counts offline.

[0240] Reweighted Feedback With robustness:

[0241] To eliminate display bias caused by location and strategy, the system constructs reweighted feedback: .

[0242] To further suppress noise, the system employs a robustness mechanism of "probabilistic pruning + upper bound truncation": ,in It is an upper bound constant (e.g., 10 or 20). This is the aforementioned minimum probability threshold. The robustened reweighted feedback. The exposed items are included in the audit log.

[0243] The intermediate weights are updated using the exposure set as the cumulative domain. :

[0244] The system assigns each user a session number. Maintain its associated set of exposure indexes (Including all exposure entries where the user is the rotation subject or the user is directly relevant to the candidate), and weighted by the following formula: Perform incremental updates: ,in This is the step size. When business needs to limit the trailing effect of historical accumulation on short-term weights, the system can enable a time decay coefficient. Optional mechanisms: .

[0245] Normalization yields new session-level user weights :

[0246] After updating the intermediate weights of all users Then, the system normalizes the data according to the session dimension to obtain the session-level user weight. : Thus satisfying New session-level user weights The score aggregation and coverage constraints will be involved in steps S4 to S7 of the subsequent sessions to achieve online closed loop.

[0247] Late Feedback and Consistency Handling: For events such as "Add to Cart - Payment," there may be a delay after exposure. The system maintains a feedback window and deduplication key for each exposure. When late feedback arrives and is still within the window, the previous feedback for that exposure is overwritten with the same deduplication key. Late feedback outside the window is entered into offline statistics and is not included in the weighted intermediate values. Online updates are implemented. This strategy ensures online predictability while maintaining the integrity of the information needed for offline evaluation and model training.

[0248] It should be noted that in grayscale traffic, "clicks" serve as the binary feedback and the intensity of exploration is... Minimum probability threshold Upper bound constant Under the following settings: Compared to the baseline without reweighting, reweighted feedback is used. Update weight intermediate value Subsequently, the median increase in the "percentage of covered users" within the session was approximately [missing information]. The median difference between adjacent refreshes in the Top-K list decreased by approximately With the same exposure budget, the click-through rate (CTR) of P50 remained basically the same, while P10 increased slightly (approximately). Log recalculation shows, according to records Replay can restore consistent reweighted feedback on 99.9% or more of the entries. It meets the auditability requirements.

[0249] It should be noted that this embodiment also includes:

[0250] Double robust estimation (optional): Introducing baseline regression estimation without changing the main formula. (Given from the offline model), construct a dual robust reweighted feedback. and in the same way and To achieve robustness, which is used to further reduce variance.

[0251] Confidence-weighted update (optional): Defines the rating volatility within a session for each user. (Refer to step S4) Construct confidence discount ,according to This reduces the instantaneous impact of highly volatile users on the group weight while maintaining consistency.

[0252] Display probability decomposition (optional): The probability of entering the selection set, the probability of being placed in the correct position, and the probability of being scattered in parallel are decomposed into multiplicative factors and recorded separately, which facilitates more granular offline recalculation and sensitivity analysis.

[0253] It should be noted that in this implementation, calculations are based solely on anonymized exposure-level logs and event indicators, and do not involve directly identifying individuals; all probability, feedback, and update parameters are recorded with version numbers and processing logs to meet the principles of minimization and auditability. Exploration Intensity With minimum probability threshold Adjustments are released gradually through the strategy center and recorded.

[0254] In summary, the above content displays probabilities. With binary feedback For input, construct robust reweighted feedback. In the exposure index set Accumulate and update the intermediate weights using step size modulation. Press again Obtain new session-level user weights It enables online, interpretable adjustment of "who has more say" in multi-user sessions.

[0255] This embodiment also provides an intelligent recommendation decision-making platform for a multi-user profile e-commerce system. It is divided into eight functional units and several shared support services according to the online link, forming an end-to-end closed loop from identity and context acquisition, profile construction, candidate generation, metric calculation, to gating constraints, two-stage selection, and rearrangement output. Specifically, it includes:

[0256] The user and context acquisition unit corresponds to step S1, which is responsible for collecting and generating account identifiers. (and temporary account identifier) Equipment identification Session identifier With session key It also maintains metadata such as timestamps, session lifetimes, and compliance tokens to provide trusted identifiers and security contexts to downstream users.

[0257] The user profile construction unit corresponds to step S2, which generates user profile vectors online (or near online). It is composed of static feature vectors, preference spectrum vectors and context anchor vectors, and supports sliding updates and quantized storage;

[0258] Step S3, which corresponds to the candidate product set generation and attribute construction unit, is responsible for forming a salable and compliant candidate set and constructing a product attribute vector for each candidate. Also includes an exposure counter. Product Difference Assessment Vector And other fields required later;

[0259] The relevance and weight calculation unit corresponds to step S4, which calculates the relevance of a single user. Session-level user weight Based on this, a session-level score S(i) is obtained, providing a robust and consistent metric baseline;

[0260] Uncertainty calculation and the corresponding step S5 of the dual-threshold hysteresis gating unit are based on the median. Absolute deviation from median Obtain the population uncertainty index Hysteresis gating is performed based on the activation and deactivation thresholds, and a restricted candidate set is output.

[0261] The label strong constraint unit corresponds to step S6, which constructs a label strong constraint table by using the intersection of positive session labels, the union of negative session labels, and the depth-limited expansion of the directed acyclic graph of synonyms. It performs allow / prohibit judgment on the restricted subset to resolve semantic conflicts and compliance risks.

[0262] The two-stage selection unit corresponds to step S7, which first scores the sessions. Perform a rapid preliminary selection, and then solve a 0–1 optimization model on the preliminary set, including constraints on coverage, merchant and category limits, and diversity, to obtain the selection variables. The defined selection set;

[0263] Step S8, corresponding to the rotation and position penalty unit, is based on the position penalty function. The selected set is rearranged according to the user's rotation position based on the minimum difference threshold, and the output length is the target list size. The final display sequence.

[0264] It should be noted that all units in the intelligent recommendation decision-making platform for e-commerce systems based on multi-user profiles provided in this embodiment are used to execute steps S1 to S8 provided in the intelligent recommendation decision-making method for e-commerce systems based on multi-user profiles described above. These steps will not be described again in this embodiment; please refer to the content in the aforementioned embodiments for details.

[0265] It should be noted that the embodiments of the present invention have better implementability and are not intended to limit the present invention in any way. Any person skilled in the art may use the above-disclosed technical content to change or modify it into equivalent effective embodiments. However, any modifications or equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A smart recommendation decision-making method for an e-commerce system based on multi-user profiles, characterized in that, The process includes the following steps: Step S1, collecting account identifiers, device identifiers, and session identifiers from multiple users to establish a session; Step S2, constructing a user profile vector for each user within the session. And store, where Indexing for users; Step S3, constructing a product attribute vector for each candidate product. And generate a candidate product set, in which Index candidate products; Step S4, calculate single-user relevance. Session-level user weight And perform normalization processing to make Step S5: Calculate the conversation-level score according to the following formula. And sort: in, Candidate Products Conversation-level rating, For users For candidate products Single-user relevance, For users Session-level user weights; Step S6, calculate the median. Absolute deviation from median : , And define the population uncertainty index Step S7, based on the activation threshold With the release threshold The dual-threshold hysteresis rule is used to gate the candidate item set, where when Enable constraints when Release the constraints, and Step S8: Output the number of targets in the previous list. Each product is recorded for exposure and interaction feedback according to session identifiers; among them... It is a positive integer.

2. The intelligent recommendation decision-making method for e-commerce systems based on multi-user profiles as described in claim 1, characterized in that, The user profile vector From static eigenvectors Preference spectrum vector Context anchor vector The preferred spectrum vector is obtained by splicing; By performing a discrete cosine transform on the time series of clicks within the session and preserving the spectrum and order, The former Each coefficient is obtained. It is a positive integer.

3. The intelligent recommendation decision-making method for e-commerce systems based on multi-user profiles as described in claim 1, characterized in that, The single-user relevance The lower bound of confidence is calculated using the bootstrap method and is defined as follows: ， in, For the first The baseline similarity is obtained by resampling. For resampling index, The number of resampling times and ， The confidence bound coefficients and 。 4. The intelligent recommendation decision-making method for an e-commerce system based on multi-user profiles as described in claim 1, characterized in that, The session-level user weight From the weight intermediate quantity Normalization yields, and ， ;in, For identity consistency F1 score, The intensity of the most recent interaction. Historical payment probability, 、 The weighting coefficients and ， ， The identity consistency F1 score From accuracy With recall rate calculate: ， ， in, For match count, The number of times it appears on the device side. The expected number of times is denoted by the account side, and all three are non-negative integers. 、 Normalization to interval 。 5. The intelligent recommendation decision-making method for an e-commerce system based on multi-user profiles as described in claim 1, characterized in that, The group uncertainty index Hysteresis gating is performed using an adaptive quantile threshold: the median absolute deviation is calculated on the current candidate item set. The upper and lower quantiles are used as the activation thresholds, respectively. With the release threshold And satisfy 。 6. The intelligent recommendation decision-making method for an e-commerce system based on multi-user profiles as described in claim 1, characterized in that, When hysteresis gating is enabled, a strong label constraint table is applied to candidate products. This strong label constraint table consists of the difference between the intersection of positive labels and the union of negative labels after expansion by a directed acyclic graph of synonyms, wherein the path length of the directed acyclic graph of synonyms does not exceed 2.

7. The intelligent recommendation decision-making method for an e-commerce system based on multi-user profiles as described in claim 1, characterized in that, The number of targets Each item was selected through a two-stage process: the first stage involved scoring the limited set of candidate items by session level. Selecting the threshold for the number of leading candidates One candidate product, In the second stage, the selection variables are obtained by solving the following 0-1 integer programming problem. ： in, Choose a binary variable; For exposure count, candidate products Historical exposure count; As a measure of difference; For ranking functions, based on relevance per user The ranking obtained by sorting the candidate products in descending order; The ranking threshold and ； Minimum coverage base for members; It is a rounding function; Aggregator for merchants; This sets a maximum exposure limit for merchants. A collection of categories; This is the maximum exposure limit for the category. For exposure penalty coefficient and ； The coefficient of difference and The difference index Defined as ， in, and These are the product difference assessment vectors, It is the cosine distance function.

8. The intelligent recommendation decision-making method for an e-commerce system based on multi-user profiles as described in claim 7, characterized in that, The selected set obtained in the second stage is rearranged according to a round-robin and position penalty strategy before output: let the position number be... Apply the position penalty function to items that have not yet been output: in, The value of the position penalty function. The position number in the list and ， For position penalty slope and ; Select users by in-session index in a round-robin fashion, and in case of ties, prioritize outputting users that meet the minimum difference threshold. The goods, namely ,in 。 9. The intelligent recommendation decision-making method for an e-commerce system based on multi-user profiles as described in claim 4, characterized in that, Feedback records include display probability With binary feedback And calculate the reweighted feedback: To update the intermediate weights ： in, For exposure index set, with users Related exposure collection; For step size and After the update, the new session-level user weights are obtained according to the normalization rules of claim 4. 。 10. An intelligent recommendation and decision-making platform for an e-commerce system based on multi-user profiles, characterized in that: include: The user and context acquisition unit is used to collect account identifiers, device identifiers, and session identifiers of multiple users and establish sessions. The profile building unit is used to build a user profile vector for each user within the session. And store, where For user indexing; candidate product set generation and attribute construction unit, used to construct a product attribute vector for each candidate product. And generate a candidate product set, in which For candidate product index; The relevance and weight calculation unit is used to calculate the relevance of a single user. Session-level user weight And perform normalization processing to make Uncertainty calculation and dual-threshold hysteresis gating unit, used to calculate session-level scores according to the following formula. And sort: in, Candidate Products Conversation-level rating, For users For candidate products Single-user relevance, For users Session-level user weights; label-strongly constrained units used to calculate the median. Absolute deviation from median : , And define the population uncertainty index A two-stage selection unit is used to select based on the activation threshold. With the release threshold The dual-threshold hysteresis rule is used to gate the candidate item set, where when Enable constraints when Release the constraints, and The rotation and position penalty unit is used to output the number of targets in the previous target list. Each product is recorded for exposure and interaction feedback according to session identifiers; among them... The value is a positive integer; the platform executes the method described in any one of claims 1 to 9 through the cooperation of the aforementioned units.

Images