An e-commerce traffic distribution method and system based on user behavior sequence

Abstract

The application provides an e-commerce traffic distribution method and system based on user behavior sequence, relating to the technical field of electronic commerce. The method comprises: forming a standardized user cognitive state database; generating a standard candidate traffic set; generating a traffic distribution log; obtaining a behavior sequence disorder degree, correcting the behavior sequence disorder degree in combination with cognitive buffer intervention conditions, calculating actual causal purity based on the corrected behavior sequence disorder degree, and finally generating a feedback sample; and outputting a new LTV causal effect parameter and a category conflict penalty coefficient table. The technical problem of the prior art that the traffic distribution strategy deviates from the user's true long-term preference due to the pollution of feedback signals caused by neglecting user cognitive residues is solved.

Description

Technical Field

[0001] This application relates to the field of e-commerce technology, and in particular to an e-commerce traffic allocation method and system based on user behavior sequences. Background Technology

[0002] Traffic allocation is a core component of e-commerce recommendation systems, directly impacting the platform's conversion rate and long-term user value. With the increasing maturity of multi-level recommendation system architectures, accurately allocating limited high-quality traffic to the most suitable users during the fine-tuning phase has become crucial for improving platform business efficiency. Current mainstream traffic allocation methods are generally based on modeling user behavior sequences, capturing users' real-time interests and preferences to allocate traffic to products or advertisements with the highest estimated conversion rates, thereby maximizing short-term business benefits.

[0003] However, existing traffic allocation methods have the following technical limitations: Firstly, they all assume that user behavior sequences are continuous and complete signals of genuine intent, neglecting the semantic breaks in these sequences caused by external interference during the shopping process. When a user returns after an interruption, their prior cognitive residue continues to interfere with their current decision, making their feedback to subsequent traffic not purely reflect the quality of the current recommendations, but rather a reaction to cognitive switching. Secondly, existing methods treat all user feedback signals regarding traffic allocation as reliable training samples, directly using them to update the long-term value model. They fail to recognize that when users are in a state of high cognitive residue, negative feedback to traffic from distant sources is cognitive noise rather than a genuine preference expression. Directly using such contaminated samples for training causal inference models can lead to the model incorrectly misjudging cognitive conflict as a natural aversion to product categories, resulting in causal collapse driven by cognitive noise and causing the system to deviate from users' true interests in the long term. Thirdly, existing methods lack mechanisms for evaluating and filtering the reliability of feedback signals, and fail to establish a complete closed loop from cognitive conflict identification to sample purity assessment to causal model feedback. This results in the inability to guarantee the quality of training data for the long-term value model, limiting the long-term optimization effect of traffic allocation strategies.

[0004] In summary, existing technologies suffer from limitations due to neglecting the contamination of feedback signals by user cognitive residues. This results in training samples for long-term value causal inference models being mixed with a large amount of cognitive noise, causing traffic allocation strategies to deviate from users' true long-term preferences. Therefore, a method is urgently needed to address these issues. Summary of the Invention

[0005] This disclosure provides an e-commerce traffic allocation method and system based on user behavior sequences to solve the technical problem in the prior art where the training samples of the long-term value causal inference model are mixed with a large amount of cognitive noise due to the neglect of the pollution of feedback signals by user cognitive residues, which in turn causes the traffic allocation strategy to deviate from the user's true long-term preferences.

[0006] According to a first aspect of this disclosure, an e-commerce traffic allocation method based on user behavior sequences is provided, comprising: Receive raw real-time behavior logs from external data sources, perform streaming filtering on the raw real-time behavior logs to obtain the current session's valid behavior event sequence, and extract cognitive residual vectors, time decay coefficients, behavior inertia coefficients, and residual category sets based on the current session's valid behavior event sequence to form a standardized user cognitive state database. The candidate traffic set is obtained based on the candidate traffic pool embedding library and the coarse-ranked candidate set output by the coarse-ranking stage of the recommendation engine. The standard candidate traffic set is generated by combining the cognitive residual vector, the time decay coefficient and the residual category set. The final traffic to be distributed is defined based on the standard candidate traffic set and behavioral inertia coefficient. At the same time, cognitive buffering intervention is performed according to the conflict level of the final traffic to be distributed. The final traffic to be distributed after cognitive buffering intervention is then distributed to the user client, and a traffic allocation log is generated. The feedback behavior collection window is set according to the time decay coefficient. The feedback behavior sequence and three types of feedback indicators are obtained according to the feedback behavior collection window. The three types of feedback indicators are fused and calculated to obtain the disorder of the behavior sequence. The disorder of the behavior sequence is corrected in combination with the cognitive buffer intervention. The actual causal purity is calculated based on the corrected disorder of the behavior sequence, and finally the feedback sample is generated. The system obtains user historical behavior confidence profiles from external data sources, combines them with feedback samples, and outputs new LTV causal effect parameters and category conflict penalty coefficient tables, which are then fed back into the comprehensive allocation score and comprehensive cognitive distance calculation formulas.

[0007] According to a second aspect of this disclosure, an e-commerce traffic allocation system based on user behavior sequences is provided, comprising: The cognitive extraction module is used to receive raw real-time behavior logs provided by external data sources, perform streaming filtering on the raw real-time behavior logs to obtain the current session's valid behavior event sequence, and extract cognitive residual vectors, time decay coefficients, behavior inertia coefficients, and residual category sets based on the current session's valid behavior event sequence to form a standardized user cognitive state database. The conflict detection module is used to obtain a candidate traffic set based on the candidate traffic pool Embedding library and the coarse-ranked candidate set output by the coarse-ranking stage of the recommendation engine, and to generate a standard candidate traffic set by combining the cognitive residual vector, time decay coefficient and residual category set. The allocation buffer module is used to define the final traffic to be sent based on the standard candidate traffic set and the behavior inertia coefficient. At the same time, it performs cognitive buffer intervention according to the conflict level of the final traffic to be sent, and sends the final traffic to be sent after cognitive buffer intervention to the user client, generating a traffic allocation log. The feedback posterior module is used to set the feedback behavior acquisition window according to the time decay coefficient, acquire the current feedback behavior sequence and three types of feedback indicators according to the feedback behavior acquisition window, fuse the three types of feedback indicators to obtain the behavior sequence disorder, correct the behavior sequence disorder based on the cognitive buffer intervention, calculate the actual causal purity based on the corrected behavior sequence disorder, and finally generate feedback samples. The dual-path feedback module is used to obtain user historical behavior confidence profiles from external data sources, combine them with feedback samples, output new LTV causal effect parameters and category conflict penalty coefficient tables, and feed them back into the comprehensive allocation score and comprehensive cognitive distance calculation formula.

[0008] One or more technical solutions provided in this disclosure have at least the following technical effects or advantages: receiving raw real-time behavior logs from external data sources, performing streaming filtering on the raw real-time behavior logs to obtain a sequence of valid behavior events in the current session, extracting cognitive residual vectors, time decay coefficients, behavior inertia coefficients, and residual category sets based on the sequence of valid behavior events in the current session to form a standardized user cognitive state database; obtaining a candidate traffic set based on the candidate traffic pool embedding library and the coarse-ranked candidate set output by the recommendation engine's coarse-ranking stage, and generating a standard candidate traffic set by combining the cognitive residual vectors, time decay coefficients, and residual category sets; defining the final traffic to be distributed based on the standard candidate traffic set and the behavior inertia coefficient, and simultaneously... Cognitive buffering intervention is implemented for the conflict level of traffic to be distributed. The final traffic after cognitive buffering intervention is then distributed to the user client, generating a traffic allocation log. A feedback behavior collection window is set based on a time decay coefficient. The feedback behavior sequence and three types of feedback indicators are obtained based on this window. These three types of feedback indicators are then fused and calculated to obtain the disorder of the behavior sequence. The disorder of the behavior sequence is corrected based on the cognitive buffering intervention. The actual causal purity is calculated based on the corrected disorder of the behavior sequence, ultimately generating a feedback sample. User historical behavior confidence profiles are obtained from external data sources. Combined with the feedback sample, new LTV causal effect parameters and category conflict penalty coefficient tables are output and fed back into the comprehensive allocation score and comprehensive cognitive distance calculation formulas. This solves the technical problem in existing technologies where neglecting user cognitive residues contaminates the feedback signal, leading to a large amount of cognitive noise in the training samples of the long-term value causal inference model, thus causing the traffic allocation strategy to deviate from the user's true long-term preferences. This achieves the technical effect of improving the prediction accuracy of the long-term value causal inference model and the long-term benefits of traffic allocation.

[0009] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0011] Figure 1 A flowchart illustrating an e-commerce traffic allocation method based on user behavior sequences, provided as an embodiment of this application; Figure 2 This is a schematic diagram of the structure of an e-commerce traffic allocation system based on user behavior sequences, provided in an embodiment of this application.

[0012] Figure labeling: Cognitive extraction module 11, conflict detection module 12, allocation buffer module 13, feedback posterior module 14, dual-path feedback module 15. Detailed Implementation

[0013] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0014] Example 1: This disclosure provides an e-commerce traffic allocation method based on user behavior sequences, which is referred to below. Figure 1 The methods include: S1: Receive raw real-time behavior logs from external data sources, perform streaming filtering on the raw real-time behavior logs to obtain the current session's valid behavior event sequence, and extract cognitive residual vectors, time decay coefficients, behavior inertia coefficients, and residual category sets based on the current session's valid behavior event sequence to form a standardized user cognitive state database. Furthermore, step S1 also includes: Receive raw real-time behavior logs from external data source Ext1, and perform streaming filtering on the raw real-time behavior logs according to event type to form a sequence of valid behavior events for the current session; Based on the current session's sequence of valid behavioral events, an involvement score is calculated for each valid behavioral event. The specific calculation formula is as follows: ; in, This represents the involvement score for the i-th valid behavioral event. This indicates the weight of the corresponding event type. This indicates the duration of time the user stayed on the corresponding page. This represents the normalization coefficient for the corresponding page scroll depth; All valid behavioral events within the current session are sorted in descending order of involvement score. Only the top K high involvement events with the highest involvement scores are retained, while other valid behavioral events with lower involvement scores are considered as brief browsing events and are discarded. Use an offline pre-trained semantic embedding model to obtain the semantic vectors corresponding to each highly involved event; The cognitive residual vector is obtained by weighted aggregation of all semantic vectors. The specific calculation formula is as follows: ; Where R represents the cognitive residual vector of the current user session, and V i Eng represents the semantic vector corresponding to the i-th high-involvement event. i This represents the corresponding event involvement score, where K represents the number of high involvement events participating in the aggregation, and i represents the index value of the high involvement event. During the weighted aggregation process, the category identifiers corresponding to all high-involvement events participating in the aggregation are recorded simultaneously to form a residual category set. set ; Extract the occurrence time of the most recent high-involvement event from the above K high-involvement events, and read the current system time. Calculate the time interval between the two using the following formula: ; Where ΔT represents the time difference since the most recent high-involvement behavior, t last Indicates the time of the most recent high-involvement event, t now Indicates the current system time; The time decay coefficient of current cognitive residue is calculated using an exponential decay model. The specific calculation formula is as follows: ; Where α represents the time decay coefficient of cognitive residue, ΔT represents the time difference from the most recent high-involvement behavior, and H represents the half-life parameter of the corresponding product category; Extract the 10 most recent valid behavior events from the current session's valid behavior event sequence, extract the corresponding category identification sequence, and use an offline pre-trained semantic embedding model to map the various categories of identification to a unified semantic space to form a category semantic identifier sequence. Calculate the moving variance of the category semantic identifier sequence, and map the moving variance to the 0-1 interval through the Sigmoid function to obtain the behavior inertia coefficient; The cognitive residual vector, time decay coefficient, behavioral inertia coefficient, and residual category set are uniformly encapsulated to form a standardized user cognitive state database.

[0015] Specifically, this step aims to perform streaming filtering and structuring processing on the raw real-time behavior logs generated during the user's current session, transforming discrete and chaotic behavioral events into structured data that can characterize the user's current cognitive state, and providing a unified data input for subsequent candidate traffic cognitive conflict detection.

[0016] First, the system receives raw real-time behavior logs from the external data source Ext1. These logs are generated and pushed in real-time by the e-commerce platform's existing event tracking system. Specifically, they consist of continuous user behavior events within the current session. Each event includes the event type, event time, product identifier, category identifier, page dwell time, page scroll depth, and corresponding session identifier. Since the raw real-time behavior logs typically contain system background heartbeats, page preload requests, cache refreshes, and other system behavior events that don't reflect the user's true shopping intent, they are first filtered by streaming based on event type. This streaming filtering uses a whitelist-based mechanism. A pre-defined whitelist of valid behavior event types (browse, click, favorite, add to cart, return to page, exit page) is used. For each incoming event, its event type is extracted and matched against the whitelist. Successful matches are retained, while unsuccessful matches are discarded. Ultimately, only valid behavior events reflecting the user's active interactions are retained, forming a sequence of valid behavior events for the current session. The streaming filtering operation is completed at the data access layer of the real-time computing engine.

[0017] After generating the sequence of valid behavioral events for the current session, an engagement score is calculated for each valid behavioral event to measure the user's actual level of attention to the corresponding product or category identifier. The engagement score comprehensively considers three factors: the importance of the event type, page dwell time, and page scroll depth. The specific calculation formula is as follows: ; in, This represents the involvement score for the i-th valid behavioral event. This indicates the weight of the corresponding event type. This indicates the duration of time the user stayed on the corresponding page. This represents the normalization coefficient for the corresponding page scroll depth.

[0018] To further explain, the principle for setting the weights of event types is that the more explicit the behavioral intent and the closer the event type is to the final purchase decision, the higher its weight. This invention involves six valid behavioral event types: browsing, clicking, favorites, adding to cart, returning to the previous page, and exiting the previous page. Specifically, favorites indicate that the user actively saves the product for later consideration; adding to the cart indicates that the user has a clear purchase intention, both of which are high-intent behaviors and are assigned a weight of 2.0; clicking indicates that the user is interested in the product and actively views the details, which is a medium-intent behavior and is assigned a weight of 1.2; browsing indicates that the user passively browses, which is a low-intent behavior and is assigned a weight of 0.8; returning to the previous page and exiting the previous page indicate that the user interrupts the current browsing, which is a negative behavior and is assigned a weight of 0.3. In this embodiment, the above six event types and their corresponding weight values ​​are used as the preferred solution. Those skilled in the art can also recalibrate the above weights using objective weighting methods such as the analytic hierarchy process or entropy weighting method, depending on the specific business scenario, as long as the basic ranking relationship of favorites and adding to cart being greater than click, click being greater than browsing, and returning and exiting having the lowest weight is satisfied.

[0019] To further clarify, page dwell time is calculated from the event occurrence time in the original real-time behavior log, including both the entry and exit times. In this calculation method, the page dwell time is compressed using a logarithmic function. The rationale for using logarithmic compression is that while user attention to a product increases with dwell time, this increase exhibits diminishing marginal returns; for example, the increase in attention from 5 seconds to 30 seconds is far greater than the increase from 300 seconds to 600 seconds. The logarithmic function effectively characterizes this diminishing marginal return while preventing the abnormal amplification of engagement scores caused by extremely long dwell times.

[0020] To further explain, the page scroll depth normalization coefficient is normalized based on the ratio of the actual scroll distance the user views to the total page length, and its value ranges from 0 to 1. Specifically, this coefficient is obtained in real time through the page scroll monitoring interface provided by the front-end event tracking SDK. When the user scrolls the page, the front-end SDK continuously records the ratio of the current scroll bar position to the total scrollable height of the page, and reports the final scroll depth ratio along with the behavior event to the log collection system when the user leaves the page. When the user views the entire page, the page scroll depth normalization coefficient is close to 1, and when only viewing the top of the page, the page scroll depth normalization coefficient is close to 0.

[0021] After calculating the involvement scores for all valid behavioral events, all valid behavioral events in the current session are sorted in descending order of their involvement scores. Only the top K high-involvement events are retained as the primary source of the user's current cognitive state. K is preferably set to 5 for the following reasons: Firstly, users have limited short-term memory capacity; psychological research shows that human working memory can typically process approximately 4-7 units of information simultaneously. Selecting 5 high-involvement events falls within this range, effectively covering the primary source of the user's current cognitive state. Secondly, if K is too large, it will introduce many low-involvement noise events, diluting the accuracy of the core intent representation; if K is too small, it may miss some dimensions of the user's shopping intent. Considering these two factors, K of 5 is the optimal value. If the total number of valid behavioral events in the current session is less than K, all are retained and used in subsequent calculations based on the actual number. Other valid behavioral events with lower involvement scores are considered brief browsing events and discarded to reduce the impact of noise behavior on the modeling of the user's cognitive state.

[0022] After obtaining the high-involvement events, the offline pre-trained semantic embedding model provided by the external data source Ext4 is further used to perform semantic mapping on the product identifiers or category identifiers involved in all high-involvement events. The offline pre-trained semantic embedding model can employ a Graph Embedding model, a BERT e-commerce semantic model based on the Transformer structure, or other product semantic representation models. These models are all existing technologies and commonly used semantic representation methods in the e-commerce recommendation field. Their outputs are all low-dimensional dense vectors with fixed dimensions, which can be directly used for subsequent similarity calculations. For the i-th high-involvement event, the corresponding semantic vector V can be obtained. i .

[0023] Subsequently, using the involvement scores corresponding to each high-involvement event as weighting coefficients, all semantic vectors are weighted and aggregated to obtain the cognitive residual vector of the current user session. The specific calculation formula is as follows: ; Where R represents the cognitive residual vector of the current user session, and V i Eng represents the semantic vector corresponding to the i-th high-involvement event. i This represents the corresponding event involvement score, where K represents the number of high involvement events participating in the aggregation, and i represents the index value of the high involvement event.

[0024] In the aforementioned weighted aggregation process, events with higher involvement contribute more to the final cognitive residual vector with their corresponding semantic vectors. Therefore, the resulting cognitive residual vector can represent the semantic center of the user's current, lingering main shopping intent. Simultaneously, during the weighted aggregation process, the category identifiers corresponding to all high-involvement events participating in the aggregation are recorded, forming a residual category set. set The residual category set is saved as incidental output data to retain information about the category composition that constitutes the current cognitive residual.

[0025] After obtaining the cognitive residual vector, the time decay coefficient α of the current cognitive residual is further calculated. Specifically, the occurrence time of the most recent high-involvement event is extracted from the above K high-involvement events, and the current system time is read. The time interval is calculated from the two, and the specific calculation formula is as follows: ; Where ΔT represents the time difference since the most recent high-involvement behavior, t last Indicates the time of the most recent high-involvement event, t now Indicates the current system time.

[0026] Next, the time decay coefficient α of the current cognitive residue is calculated using an exponential decay model. The specific calculation formula is as follows: ; Where α represents the time decay coefficient of cognitive residue, ΔT represents the time difference from the most recent high-involvement behavior, and H represents the half-life parameter of the corresponding product category. The time decay coefficient α ranges from 0 to 1. When a high-involvement event has just occurred, α is close to 1, indicating that the user's current cognitive residue still has a high influence; as time increases, α gradually decays to around 0, indicating that the influence of previous behaviors on the current decision gradually weakens.

[0027] It should be noted that, considering the significant differences in user decision-making cycles across different product categories, the system pre-configures a corresponding half-life parameter H based on the product category to ensure that cognitive residue decay more closely aligns with actual shopping patterns. The half-life parameter H is set based on the statistical patterns of average product decision-making cycles. Offline data analysis from e-commerce platforms shows significant differences in the average time users experience from initial browsing to final purchase across different product categories. Fast-moving consumer goods (FMCG) have short decision-making cycles, typically completed within minutes to hours; apparel and footwear have medium decision-making cycles, typically completed within hours to 1-2 days; and high-priced durable goods such as home appliances and 3C digital products have long decision-making cycles, often requiring users to compare and consider for days or even weeks. Based on the above statistical patterns, this invention sets the half-life parameter H to 1 / 4 of the average decision-making cycle for the corresponding category. The rationale is that in the exponential decay model, after one half-life, the residual value decays to 50% of the initial value, and after two half-lives, it decays to 25%. Using 1 / 4 of the average decision-making cycle as the half-life ensures that cognitive residue decays to approximately 6.25% within a complete average decision-making cycle, essentially dissipating. In this embodiment, the specific settings are as follows: for fast-moving consumer goods such as snacks and beverages, the average decision-making cycle is approximately 10 minutes, so H is set to 2.5 minutes; for clothing, shoes, and bags, the average decision-making cycle is approximately 1 day, so H is set to 6 hours; for large appliances and 3C digital products, the average decision-making cycle is approximately 3-5 days, so H is set to 18 hours. Those skilled in the art can statistically analyze the average decision-making cycle for each product category based on the offline data analysis results of specific e-commerce platforms and set the corresponding H value according to the above 1 / 4 rule for the average decision-making cycle, without needing to list the specific values ​​for each category individually.

[0028] To further characterize whether users have a habitual tendency to continuously browse similar products, this step continues to calculate the behavioral inertia coefficient I using the current session's effective behavioral event sequence. Unlike the aforementioned cognitive residue modeling, this step no longer only retains high-involvement events, but directly extracts the 10 most recent effective behavioral events from the current session's effective behavioral event sequence, extracts the corresponding category identification sequences, and uses the aforementioned offline pre-trained semantic embedding model to map various category identifications to a unified semantic space, forming corresponding category semantic label sequences.

[0029] The moving variance of the category semantic identifier sequence is then calculated to reflect the dispersion of user browsing category changes. Specifically, the mean vector of the category semantic identifier sequence is first calculated, and then the mean of the squared Euclidean distances between each vector and the mean vector is calculated to obtain the moving variance σ². This moving variance reflects the dispersion of user browsing categories in the semantic space. When the moving variance is small, it indicates that the user browses similar categories continuously, and their browsing behavior has strong inertia; when the moving variance is large, it indicates that the user frequently switches between different categories, and their behavioral inertia is weak.

[0030] The system maps the aforementioned moving variance to the 0-1 interval using the Sigmoid function to obtain the behavioral inertia coefficient, the specific calculation formula of which is as follows: ; Where I represents the behavioral inertia coefficient, σ represents the moving variance of the category semantic identifier sequence, σ0 represents the preset variance benchmark value, determined based on the statistical median of the moving variance of all users in historical data, and k represents the slope parameter for adjusting the steepness of the curve, preferably set to 1. The physical meaning of this mapping method is that when the moving variance is less than the historical benchmark value, I tends towards 1; when the moving variance is greater than the historical benchmark value, I tends towards 0. Specifically, the closer I is to 1, the higher the user's behavioral inertia, and the more inclined they are to continue browsing the current category; the closer I is to 0, the more divergent the user's browsing path, and the more receptive they are to recommendations from different categories.

[0031] After completing the above calculations, this step will combine the cognitive residual vector R, the time decay coefficient α, the behavioral inertia coefficient I, and the residual category set. set A unified encapsulation is performed to form a standardized user cognitive state database. The cognitive residue vector R describes the semantic state of the user's current dominant shopping intention; the time decay coefficient α describes the real-time impact of cognitive residue; the behavioral inertia coefficient I describes the stability of the user's browsing behavior; and the residual category set is... set Used to record the main categories that form the current cognitive residue.

[0032] S2: Based on the candidate traffic pool Embedding library and the coarse-ranked candidate set output by the recommendation engine in the coarse-ranking stage, a candidate traffic set is obtained. Combined with the cognitive residual vector, time decay coefficient and residual category set, a standard candidate traffic set is generated. Furthermore, step S2 also includes: It receives the candidate traffic pool Embedding library provided by the external data source Ext2 and the coarse-ranked candidate set output by the coarse-ranking stage of the recommendation engine, integrates the candidate traffic pool Embedding library and the coarse-ranked candidate set to obtain the candidate traffic set. For each candidate traffic in the candidate traffic set, based on the product identifier, the corresponding semantic embedding vector is queried in the candidate traffic pool embedding library. If there is no corresponding cache in the candidate traffic pool embedding library, the semantic representation of the candidate traffic is calculated in real time using the offline pre-trained semantic embedding model to obtain the semantic embedding vector corresponding to the candidate traffic.

[0033] Obtain cognitive residual vectors from a standardized user cognitive state database, calculate the cosine distance between the semantic embedding vector and the cognitive residual vector, and obtain the basic semantic distance. Based on the category to which the candidate traffic belongs and the residual category set. set The hierarchical relationship between categories determines the category hierarchy penalty items; Based on the basic semantic distance and category-level penalty, the comprehensive cognitive distance corresponding to candidate traffic is calculated using the following formula: ; Among them, D j D represents the comprehensive cognitive distance of the j-th candidate traffic. cos P represents the basic semantic distance. cat β represents the category-level penalty item, and β represents the category penalty adjustment hyperparameter. Real-time detection of the user's current attention state and calculation of the user's real-time perturbation parameters; For each candidate traffic, the corresponding estimated causal confidence weight is calculated by comprehensively utilizing the time decay coefficient, comprehensive cognitive distance, and real-time perturbation parameters. The specific calculation formula is as follows: ; in, Let α represent the estimated causal confidence weight of the j-th candidate flow, and let D represent the time decay coefficient output from step S1. j This represents the overall cognitive distance of the corresponding candidate traffic. λ1 represents the real-time perturbation parameter, λ2 represents the real-time perturbation suppression weight, and Sigmoid(·) represents the Sigmoid activation function. Candidate traffic is classified into conflict levels based on comprehensive cognitive distance, and the conflict levels include green, yellow and red levels; The product identifier, comprehensive cognitive distance, estimated causal confidence weight, and conflict level corresponding to each candidate traffic are uniformly encapsulated to generate a standard candidate traffic set.

[0034] Specifically, this step aims to analyze the candidate traffic set item by item based on the user's current cognitive state obtained in step S1, to pre-assess the degree of conflict between each candidate traffic and the user's current cognitive state before formally allocating traffic, and to further predict the credibility of the user's feedback behavior in truly reflecting their preferences after the candidate traffic is allocated.

[0035] First, the system receives the candidate traffic pool embedding library from the external data source Ext2 and the coarse-ranked candidate set output from the coarse-ranking stage of the recommendation engine. Both the candidate traffic pool embedding library and the coarse-ranked candidate set are standard outputs from existing e-commerce recommendation systems; this invention directly reads and uses them without additional computation. Integrating the candidate traffic pool embedding library and the coarse-ranked candidate set yields a candidate traffic set. This candidate traffic set contains M candidate traffic items, with the candidate traffic expression being Item. j , where j represents the candidate traffic index, j=1,2,…,M. Since the cognitive residual vector R in step S1 is constructed in the semantic space corresponding to the offline pre-trained semantic embedding model, in order to ensure that the subsequent semantic distance calculation has a unified spatial metric, it is necessary to first complete the real-time semantic mapping of the candidate traffic.

[0036] Specifically, for each candidate traffic in the candidate traffic set, the system queries its corresponding semantic embedding vector in the candidate traffic pool embedding library based on the product identifier. When the corresponding cache is not found in the candidate traffic pool embedding library, the semantic representation of the candidate traffic is calculated in real time using the same offline pre-trained semantic embedding model as in step S1, thereby obtaining the semantic embedding vector corresponding to the candidate traffic. Since the semantic embedding model used is consistent with the model used to generate the cognitive residual vector in step S1, the obtained semantic embedding vector and the cognitive residual vector are located in the same semantic space, and subsequent distance calculations can be performed directly.

[0037] After obtaining the semantic embedding vectors corresponding to all candidate traffic, the cognitive distance between each candidate traffic and the current user's cognitive state is calculated sequentially. First, using the cognitive residual vector output in step S1 as the semantic center of the user's current dominant shopping intention, the cosine distance between the semantic embedding vector and the cognitive residual vector is calculated to obtain the basic semantic distance D. cos Cosine distance is used to reflect the similarity between candidate traffic and the user's current cognitive semantics. The smaller the cosine distance, the closer the candidate traffic is to the current shopping intent; the larger the cosine distance, the greater the semantic deviation between the two.

[0038] Based solely on basic semantic distance, the category relationships in e-commerce scenarios cannot be fully reflected. Therefore, a category hierarchy penalty term P is further introduced on top of the basic semantic distance. cat Category-level penalty items are based on the category to which the candidate traffic belongs and the residual category set (Category) output from step S1. set The hierarchical relationship between them is determined, and the specific calculation process is as follows: First, the categories to which the candidate traffic belongs are compared with the residual category set. setThe system performs hierarchical matching on each residual category. When candidate traffic belongs to a direct subcategory of a residual category or has a clear upstream / downstream relationship, for example, if the residual category is "mobile phone" and the candidate traffic belongs to accessory categories such as "phone case" or "charger," then the candidate traffic is considered to have a strong correlation with current understanding, and a category-level penalty item P is assigned. cat Negative values ​​are used to shorten the overall cognitive distance. When candidate traffic and residual categories belong to the same category, the category-level penalty term P is applied. cat Set to zero. When the candidate traffic and the residual category belong to different first-level categories, for example, the residual category belongs to "home appliances" and the candidate traffic belongs to "beauty", it is considered that there is a significant cognitive conflict between the two, and a category-level penalty item P is assigned. cat Positive values ​​are used to increase the overall cognitive distance. The penalty values ​​corresponding to the above category hierarchy are derived from the category conflict penalty coefficient table formed by the periodic update in step S5, and can be preset initial values ​​when the system is first deployed. The logic for setting the preset initial values ​​is as follows: for the same category, P cat =0; For direct subcategories or upstream / downstream related categories, P cat =-0.15; For different first-level categories, P cat =+0.30. The above initial value design was based on offline historical data analysis. The average semantic cosine distance between different first-level categories is approximately 0.35 to 0.45. P was set... cat Setting P = +0.30 increases the overall cognitive distance across first-level categories by approximately 0.3, effectively amplifying the semantic distance to near its maximum value and thus suppressing the allocation priority of cross-category traffic. Meanwhile, the semantic distance between upstream and downstream related categories is typically closer, with a cosine distance of approximately 0.1 to 0.2. cat Setting the value to -0.15 further reduces the overall distance, strengthening the allocation priority of related categories. Those skilled in the art can adjust the initial value accordingly based on the category hierarchy depth and semantic distribution characteristics of specific business scenarios. Subsequently, as the category conflict knowledge base is continuously updated in step S5, the corresponding penalty value is updated synchronously to achieve continuous optimization of the cognitive distance measurement.

[0039] After obtaining the basic semantic distance and category-level penalty, the comprehensive cognitive distance corresponding to the candidate traffic is calculated using the following formula: ; Among them, D j D represents the comprehensive cognitive distance of the j-th candidate traffic. cos P represents the basic semantic distance. catThe term β represents the category-level penalty, and β represents the category penalty adjustment hyperparameter, used to control the influence of category-level factors on the overall cognitive distance. Its preferred value range is 0.5–2.0; in this embodiment, it is set to 1.0. This setting ensures that the category-level penalty and cosine distance are on the same order of magnitude, and are weighted and fused in a 1:1 ratio, preserving the basic similarity information of the semantic vectors while providing sufficient adjustment power for the category-level relationship. If β is too large, for example, greater than 2.0, the category-level factor will dominate the cognitive distance, weakening the fine-grained distinguishing ability of the semantic vectors. If β is too small, for example, less than 0.5, the category-level penalty is almost ineffective and cannot effectively correct cognitive conflicts in cross-category scenarios. Those skilled in the art can determine the optimal value of β based on historical recommendation data using offline optimization methods such as grid search.

[0040] Through the above calculations, the comprehensive cognitive distance not only reflects the distance relationship in the semantic space, but also takes into account the degree of business association between e-commerce categories, making the cognitive conflict assessment more in line with the actual recommendation scenario.

[0041] After completing the comprehensive cognitive distance calculation, to further improve the real-time performance of causal purity prediction, the system performs real-time detection of the user's current attention state before officially executing traffic allocation, thereby calculating the user's real-time perturbation parameter ρ. Specifically, in the last 2-second time window before the traffic allocation decision is triggered, the system continuously monitors the user's micro-interactions on the page, including slight page swipes, mouse hovering, changes in page dwell time, and short pauses during page refresh.

[0042] Let nmicro be the total number of micro-interactions detected by the system within the 2-second time window, and let d1, d2, …, d be the durations of each micro-interaction. nmicro The interaction amplitudes of each micro-interaction are a1, a2, ..., a nmicro First, calculate the weighted interaction strength using the following formula: ; Where S represents the weighted interaction intensity, m represents the index of the micro-interaction, nmicro represents the total number of micro-interactions, and d m a represents the duration of the m-th micro-interaction. m Let S represent the interaction amplitude of the m-th micro-interaction, and 2000 be the total length of the time window. Through the above calculations, the value of S ranges from 0 to 1. When no micro-interaction is detected within the window, S = 0; when high-amplitude interactions continuously occur within the window, S approaches 1.

[0043] Then, the weighted interaction strength is mapped to real-time perturbation parameters using the Sigmoid function. The specific calculation formula is as follows: ; Where ρ represents the real-time perturbation parameter, Sigmoid(∙) represents the activation function, and S represents the weighted interaction strength. The physical meaning of this mapping method is that when the weighted interaction strength S is less than 0.3, ρ tends to 0, and the user's attention is stable at this time; when S is greater than 0.3, ρ rises rapidly to close to 1 as S increases, and the user's attention is significantly perturbed at this time.

[0044] The real-time perturbation parameter ρ is obtained, with a value ranging from 0 to 1. The closer ρ is to 1, the more easily the user's current attention is affected by the real-time content of the page, indicating that there is a large fluctuation in the current cognitive state; the closer ρ is to 0, the more stable the user's current cognitive state is, and the original cognitive residue can continue to influence subsequent behavioral decisions.

[0045] Subsequently, for each candidate traffic, the credibility of its future feedback behavior is predicted by comprehensively utilizing the time decay coefficient output in step S1, the currently calculated comprehensive cognitive distance, and the real-time perturbation parameters, and the corresponding estimated causal confidence weight is calculated. The specific calculation formula is as follows: ; in, Let α represent the estimated causal confidence weight of the j-th candidate flow, and let D represent the time decay coefficient output from step S1. j ρ represents the comprehensive cognitive distance of the corresponding candidate traffic, λ1 represents the cognitive residual influence weight, λ2 represents the real-time perturbation suppression weight, and Sigmoid(·) represents the Sigmoid activation function, which is used to normalize the calculation results to the range of 0 to 1. and The preferred values ​​are respectively =2.0、 =1.5. The basis for choosing the above value is that when α=1, D j When =0 and ρ=0, =2.0, Sigmoid(2.0)≈0.88, meaning the estimated causal confidence weight is at most about 0.88, and will not reach 1.0, to retain a certain conservative margin. When α=0 or D j When it is very large, this term approaches 0; at the same time, if ρ=1, then... =1.5, Sigmoid(-1.5)≈0.18, that is, the lowest confidence weight is about 0.18. > The design ensures that the positive contribution of cognitive residuals is slightly greater than the negative contribution of perturbations, reflecting the physical law that residuals can be buffered and diluted, but perturbations will immediately interfere with feedback. Those skilled in the art can perform cross-validation based on actual business data to further refine the design. and Optimize.

[0046] The above calculation logic embodies the formation mechanism of cognitive purity: when the user's cognitive residual is strong, that is, α is large, and the candidate traffic is close to the current cognitive distance, that is, D j is small, it is predicted that the feedback behavior generated by the candidate traffic mainly comes from the user's real shopping intention, and at this time, the estimated causal confidence weight is relatively high; on the contrary, when the cognitive distance is relatively far or the user's attention is greatly affected by the disturbance of the real-time page, that is, ρ is relatively high, the user's feedback is more susceptible to accidental factors, and the feedback behavior cannot truly reflect long-term interests. At this time it is reduced correspondingly. The calculated will be used as an important constraint basis for the traffic allocation score in step S3.

[0047] After completing the calculation of the comprehensive cognitive distance and the estimated causal confidence weight corresponding to all candidate traffic, the system divides the conflict level of the candidate traffic according to the comprehensive cognitive distance. Specifically, a first distance threshold T1 and a second distance threshold T2 are preset, where T1<T2. The preferred values of T1 and T2 are T1=0.4 and T2=0.7 respectively. The basis for selecting the above values is that through offline statistical analysis of the historical recommendation data of the e-commerce platform, when the semantic cosine distance between the candidate traffic and the user's historical behavior sequence is less than 0.4, the correlation between the candidate traffic and the user's interest is significantly higher than the random level, and the credibility of user feedback is relatively high at this time, which is classified as low conflict; when the cosine distance is greater than 0.7, the correlation between the candidate traffic and the user's interest is close to the random level, and the user feedback is extremely susceptible to cognitive conflict interference at this time, which is classified as high conflict; the state between the two is moderate conflict. Those skilled in the art can adjust T1 and T2 accordingly according to the distribution characteristics of semantic vectors of a specific platform, such as the average inter-class distance of commodity vectors of different categories. When the candidate traffic satisfies D j <T1, it is determined that the candidate traffic belongs to a low-conflict state, and Level is marked j as the green level; when T1≤D is satisfied j <T2, it is determined to be a moderate conflict state, and Level is marked j as the yellow level; when D is satisfied j ≥T2, it is determined to be a high-conflict state, and Level is marked j as the red level.

[0048] Finally, the commodity identifier, comprehensive cognitive distance, estimated causal confidence weight and conflict level corresponding to each candidate traffic are uniformly encapsulated to generate a standard candidate traffic set.

[0049] S3: Define the final traffic to be sent based on the standard candidate traffic set and behavioral inertia coefficient. At the same time, perform cognitive buffer intervention according to the conflict level of the final traffic to be sent. Send the final traffic to be sent after cognitive buffer intervention to the user client and generate traffic allocation log. Furthermore, step S3 also includes: Read the standard candidate traffic set, estimate the pCVR conversion rate, and the new LTV causal effect parameters; For each candidate traffic in the standard candidate traffic set, a comprehensive allocation score is calculated. The specific calculation method is as follows: ; Among them, Score j pCVR represents the comprehensive allocation score of the j-th candidate traffic. j This represents the estimated conversion rate (LTV) output by the main model of the recommendation engine. j This represents the long-term value coefficient, which is equivalent to the new LTV causal effect parameter obtained from the feedback in step S5. The estimated causal confidence weights calculated in step S2 are represented by δpenalty, which represents the purity loss penalty base. The category of the current candidate traffic is compared with the category of the user's most recent behavior. When the category span between the candidate traffic and the most recent behavior exceeds two levels, a behavior inertia coefficient is introduced to correct the overall allocation score. If the span between the category of the candidate traffic and the most recent behavior does not exceed two levels, the overall allocation score remains unchanged. The candidate traffic flows are sorted according to their comprehensive allocation scores after inertia constraint correction, and the highest-scoring candidate traffic flows are selected as the final traffic flows to be distributed. Item Simultaneously, the conflict level corresponding to this traffic is extracted from the standard candidate traffic set. j The final recommendation results are then validated again. When the conflict level corresponding to the final traffic to be distributed is green or yellow, the traffic distribution phase begins directly. When the conflict level corresponding to the final traffic to be distributed is red, a cognitive buffer intervention is further implemented, inserting a buffer duration of Δt before the final traffic to be distributed is officially displayed. buf The cognitive buffering phase, buffering duration Δt buf Determined based on the half-life parameter H; Based on the cognitive buffering intervention process, a buffering execution record is generated. At the same time, the final traffic to be distributed, the official time of this traffic distribution, and the comprehensive allocation score are integrated to generate this traffic allocation log.

[0050] Specifically, this step aims to optimize the candidate traffic based on cognitive conflict analysis and causal purity prediction, taking into account commercial benefits, long-term value, and feedback data quality, to determine the final traffic content to be distributed to users.

[0051] First, the system reads the standard candidate traffic set output in step S2. Simultaneously, the system calls the estimated conversion rate pCVR output by the main model of the recommendation engine. The estimated conversion rate pCVR is the standard output of the main model of the recommendation engine in the e-commerce recommendation system, and this invention directly reads and uses it. The system also reads the new LTV causal effect parameters obtained through continuous training and updating in step S5, which are used to obtain the long-term value coefficient of the category to which the candidate traffic for the current user belongs. It should be noted that in the initial system deployment phase, step S5 has not yet started running, and the new LTV causal effect parameters do not yet exist. At this time, the initial LTV model parameters, pre-trained offline, are used as a substitute. The initial LTV model can use historical platform data, with the user's access frequency and total spending over the next 7 days as the prediction targets, to train a baseline regression model, such as XGBoost or a deep neural network. The output LTV estimate is the initial LTV model parameter. As step S5 continues to run, the initial LTV model parameters are gradually replaced by the new LTV causal effect parameters.

[0052] After obtaining the above data, a comprehensive allocation score is calculated for each candidate traffic in the standard candidate traffic set. The comprehensive allocation score considers short-term conversion benefits, long-term value benefits, and feedback data purity constraints. The specific calculation method is as follows: ; Among them, Score j pCVR represents the comprehensive allocation score of the j-th candidate traffic. j This represents the estimated conversion rate (LTV) output by the main model of the recommendation engine. j This represents the long-term value coefficient, which is equivalent to the new LTV causal effect parameter obtained from the feedback in step S5. δpenalty represents the estimated causal confidence weight calculated in step S2, and δpenalty represents the purity loss penalty base, which is used to measure the adverse effect of low purity feedback on model training. Its preferred value is 0.3. γ1, γ2 and γ3 represent the short-term conversion benefit weight, long-term value benefit weight and feedback data purity constraint weight, respectively. Their parameter values ​​are determined based on offline optimization of historical recommendation effects.

[0053] In the above calculation formula, the first term reflects the immediate short-term conversion benefits brought by candidate traffic, the second term reflects the contribution of candidate traffic to the long-term value growth of users, and the third term reflects the constraint of feedback data quality. When the estimated causal confidence weight is low, it means that even if the candidate traffic receives user feedback, it is difficult to truly reflect user interests. Therefore, the corresponding purity loss penalty term increases, causing the overall score to decrease, thereby reducing the probability of high cognitive conflict traffic being prioritized and achieving a unified optimization between business goals and training sample quality.

[0054] To further explain, the value of δpenalty is based on the fact that δpenalty needs to be combined with... and Being of the same order of magnitude (both are typically values ​​between 0 and 1), the purity penalty term... The value of δpenalty ranges from 0 to 0.3, and can generate an adjustment range of approximately 10% to 30% in the overall score, which is sufficient to affect the ranking results without completely suppressing the contribution of business indicators. Those skilled in the art can adjust δpenalty within the range of 0.1 to 0.5 according to the intensity of business requirements for data quality.

[0055] To further explain, , , The preferred values ​​are respectively =0.4、 =0.4、 =0.2, all three conditions are met. + + =1. The basis for selecting the above value is that short-term conversion revenue and long-term value revenue are equally important, each accounting for 40%, while the purity constraint of feedback data, as an auxiliary optimization objective, accounts for 20%. This proportion design ensures that the system, while pursuing commercial benefits, maintains sufficient attention to the quality of training samples, avoiding sacrificing long-term model performance for excessive pursuit of short-term conversion. Those skilled in the art can adjust the above weights accordingly based on specific business objectives, such as placing more emphasis on short-term GMV or on long-term user retention.

[0056] After calculating the comprehensive allocation score, the behavioral inertia coefficient output from step S1 is further introduced to perform a second correction on the comprehensive score. The behavioral inertia coefficient reflects the stability of the user's current browsing behavior. When I is close to 1, it indicates that the user is continuously browsing around the same category, and their shopping intention is relatively stable. When I is close to 0, it indicates that the user's browsing behavior is more divergent, and they can accept larger content switching. Therefore, the system reads the behavioral inertia coefficient I output from step S1 and compares the category of the current candidate traffic with the category of the user's most recent behavior. When the category span between the candidate traffic and the category of the most recent behavior exceeds two levels, and the behavioral inertia coefficient I is at a high level, it is considered that the current recommended content jumps too much, which is likely to disrupt the user's continuous cognitive state. Therefore, an inertia smoothing penalty is further added to the comprehensive allocation score of the corresponding candidate traffic. The correction method is as follows: ; in, This represents the overall allocation score after inertia constraint correction. The score represents the overall allocation score, I represents the behavioral inertia coefficient output in step S1, and η represents the category span penalty base, the value of which is preset according to different business scenarios, with a preferred value of 0.15. The basis for selecting this value is that η needs to be consistent with the overall allocation score. The magnitudes are matched so that the maximum value of the inertia penalty term I×η is 0.15, which can produce a downward adjustment of no more than 15% in the overall score, enough to affect the ranking but not completely exclude cross-category recommendations. Those skilled in the art can adjust η in the range of 0.05 to 0.30 according to the business's sensitivity to the continuity of recommendations.

[0057] If the gap between the category of the candidate traffic and the category of the most recent behavior does not exceed two levels, the overall allocation score remains unchanged. This inertial smoothing constraint allows the recommendation system to dynamically control the switching speed of recommended content based on the user's current browsing stability, avoiding cognitive interruptions caused by excessively large recommendation gaps.

[0058] After correcting the overall allocation scores of all candidate traffic, the system sorts them according to the overall allocation scores after inertia constraint correction, and selects the candidate traffic with the highest score as the final traffic to be distributed. Item Simultaneously, the conflict level corresponding to this traffic is extracted from the standard candidate traffic set. j The final recommendation results are then validated again.

[0059] When the conflict level corresponding to the final traffic to be distributed is green or yellow, it means that there is no obvious conflict between the current recommended content and the user's cognitive state. The system does not need to perform any additional processing and can directly enter the traffic distribution stage.

[0060] When the conflict level corresponding to the final traffic to be distributed is red, it indicates that although the candidate traffic has high commercial value, there is a strong conflict between it and the user's current cognitive state. If this content is displayed directly, it is easily influenced by prior cognitive residue, causing subsequent feedback behavior to fail to truly reflect the user's interests. Therefore, this invention further implements cognitive buffering intervention.

[0061] Specifically, before formally displaying the final traffic to be distributed, the system first inserts a buffer for a duration of Δt. buf The cognitive buffer phase. This phase uses low-information-density, neutral interactive content, such as brand atmosphere display pages, general coupon redemption pages, and gradual loading animations for recommended content. Their common characteristic is that they do not actively reinforce any specific product awareness; they are only used to extend the decision-making time interval before the user officially encounters the target traffic. Buffer duration Δt buf The half-life parameter H corresponding to the product category in step S1 is determined, and the specific calculation formula is Δt. buf = min(H / 4, 3s), where H is the half-life parameter for the product category set in step S1. This formula is designed based on the fact that after H / 4 hours, cognitive residual decays to approximately (1 / 2)^(1 / 4) ≈ 84% of its initial value. This indicates that the residual has not yet significantly decayed but has already entered the decay phase. Inserting target traffic at this point utilizes some of the residual decay benefit without waiting too long and affecting the user experience. The maximum of 3 seconds is to avoid user churn due to excessive buffering time. Since the cognitive retention time varies for different product categories, the buffering duration is adjusted according to the half-life parameter of the corresponding product category. This ensures that the buffering phase covers the rapid decay range of cognitive residual, allowing prior cognition to decay naturally during the buffering period and reducing the impact of cognitive conflict on subsequent feedback behavior.

[0062] After cognitive buffering intervention is completed, the system generates a corresponding buffer execution record. Log The buffer execution log includes the buffer action identifier, buffer duration, and buffer execution status, used to fully record the cognitive buffer intervention process. If the current final traffic to be distributed corresponds to a conflict level of green or yellow, the buffer execution log will record an empty buffer state to maintain consistency in subsequent data structures.

[0063] Finally, the system sends the final determined traffic volume to be distributed and the corresponding display content to the user's client for display. When cognitive buffering intervention exists, the buffered content is displayed first, and after the buffering ends, the system automatically switches to the traffic page corresponding to the final traffic volume to be distributed; when there is no cognitive buffering intervention, the traffic display is completed directly.

[0064] While completing the traffic distribution, the system generates an Allocation log for this traffic allocation. Log The traffic allocation log serves as the data source for the feedback collection phase in step S4. It internally records information about the entire traffic allocation process, including: the final traffic to be allocated (Selected). Item The official time for this traffic distribution is T. alloc Buffer execution record Log And the overall allocation score.

[0065] S4: Set the feedback behavior acquisition window according to the time decay coefficient, obtain the current feedback behavior sequence and three types of feedback indicators according to the feedback behavior acquisition window, integrate and calculate the three types of feedback indicators to obtain the behavior sequence disorder, correct the behavior sequence disorder based on the cognitive buffer intervention, calculate the actual causal purity based on the corrected behavior sequence disorder, and finally generate feedback samples. Furthermore, step S4 also includes: The feedback behavior collection window is dynamically adjusted based on the degree of cognitive residual impact. The specific calculation method is as follows: ; in, This indicates the feedback behavior collection window. This represents the system's baseline acquisition window, and α represents the time decay coefficient. Read the traffic allocation log and record the official traffic distribution time T. alloc As the starting time for feedback behavior collection, based on the starting time and the feedback behavior collection window, all feedback behaviors generated by the user within the corresponding time range are captured to form the current feedback behavior sequence. At the same time, based on the current feedback behavior sequence, three types of feedback indicators are collected, including explicit conversion tags, bounce behavior indicators, and category jump distance. The disorder of the behavior sequence is calculated based on explicit conversion behavior, exit behavior, and category jump span. The specific calculation formula is as follows: ; Where Turb represents the disorder of the behavioral sequence, C bounce This indicates a bounce-out behavior flag. C represents the result of normalized category jump distance. conv The explicit conversion label is represented by ω1, ω2, and ω3, which represent the weight coefficients corresponding to the bounce behavior, category jump span, and explicit conversion behavior, respectively. The buffer execution record attached to the traffic allocation log is read to verify the buffer execution effect. When the buffer execution record is empty, the disorder of the behavior sequence after buffer correction is directly set to the original disorder of the behavior sequence. When there is a valid buffer action recorded in the buffer execution record, the time of the user's first interaction after the buffer ends is further read. And compare it with the buffer duration. If the conditions are met... The disorder of the behavioral sequence is corrected, and its calculation method is as follows: ; Where Turb' represents the disorder of the behavior sequence after buffer correction, Turb represents the disorder of the behavior sequence, and 0.7 is the correction coefficient; If the above conditions are not met, the original disorder of the behavioral sequence remains unchanged; The actual causal purity of this feedback behavior is further calculated using the following formula: ; in, This indicates the actual causal purity of the feedback behavior, while Turb' indicates the disorder of the behavior sequence after buffer correction. Based on the traffic allocation records and feedback analysis results, a feedback sample is generated.

[0066] Specifically, this step aims to collect and analyze feedback behavior after traffic allocation is completed in step S3. Instead of directly judging the recommendation effect based on whether users click or convert, it further analyzes the degree to which user feedback behavior is interfered with by prior cognitive residues, thereby assessing whether the feedback sample can truly reflect the user's current preferences.

[0067] First, the system reads the time decay coefficient α output in step S1 and determines the feedback behavior acquisition window based on the current level of cognitive residual influence. Since the duration of current cognitive residual varies among different users, using a fixed feedback behavior acquisition window may result in the acquisition of behaviors still influenced by cognitive residual, thus reducing the quality of subsequent training samples. Therefore, this invention dynamically adjusts the feedback behavior acquisition window based on the level of cognitive residual influence. The specific calculation method is as follows: ; in, This indicates the feedback behavior collection window. The system baseline acquisition window is represented by a default value of 5 seconds, and α represents the time decay coefficient of cognitive residuals calculated in step S1.

[0068] The value is based on empirical statistics from user behavior analysis in the e-commerce industry, which show that after browsing a product details page, if a user becomes interested, they will typically take the next step of interaction within 3 to 8 seconds, such as adding to the shopping cart, returning, or continuing to browse. A 5-second window serves as a baseline to cover the timeframe for most effective feedback behaviors, while avoiding the introduction of irrelevant noise due to an excessively long window. Those skilled in the art can adjust the window within the 3-8 second range based on the average decision-making speed of users on a specific platform. .

[0069] The calculations above show that when the user's current cognitive residual is strong (α is close to 1), the feedback behavior acquisition window is extended accordingly, enabling the system to collect changes in user behavior as cognitive residual gradually decays. When cognitive residual is weak (α is close to 0), a acquisition window length close to the baseline is sufficient to complete feedback acquisition, thus balancing feedback completeness and system real-time performance. The calculated feedback behavior acquisition window will serve as the unified time range for subsequent user feedback behavior data acquisition.

[0070] After the feedback behavior collection window is determined, the system reads the traffic allocation log output in step S3 and obtains the final traffic to be distributed (Selected) corresponding to this traffic allocation. Item The official time for this traffic distribution is T. alloc and Buffer execution record Log Among them, T alloc As the starting time for feedback behavior collection, Selected Item Used to associate the target traffic corresponding to this feedback, Buffer Log Used for subsequent buffering effect verification.

[0071] Subsequently, the system receives external data source Ext Interact The provided real-time interaction logs after allocation have the same source as the original real-time behavior logs in step S1, requiring no additional collection or deployment. Based on the start time and window of feedback behavior collection, all user feedback behaviors generated within the corresponding time range are captured to form the current feedback behavior sequence. Simultaneously, based on this feedback behavior sequence, the following three types of feedback metrics are collected.

[0072] The first type of feedback metric is the explicit conversion tag, which represents explicit conversion behavior. The system monitors whether the user has selected the final traffic to be delivered. Item Perform actions indicating a clear purchase intention, such as adding to cart or buying immediately, and generate explicit conversion tags C accordingly. conv When a user completes any of the above conversion actions, assign the value C. conv =1, otherwise assign the value C conv=0. This metric indicates whether this traffic resulted in a direct commercial conversion.

[0073] The second type of feedback metric is the bounce behavior identifier, which indicates bounce behavior. The system continuously monitors whether the user actively returns to the previous page, closes the page, or exits the current browsing process within a short period after the data is displayed. When the user performs the above actions within a specified time, a bounce behavior identifier C is generated. bounce C bounce Assign a value to a binary variable, and assign the value C when the user performs a back, close, or exit action. bounce =1, otherwise assign the value C bounce =0. This metric reflects whether users quickly abandon the currently recommended content and is an important basis for subsequent evaluation of cognitive conflict.

[0074] The third type of feedback metric is the category jump distance, which represents the category jump span. The system continues to monitor the user's next product browsing behavior after completing the current page view, extracts the primary category of the next viewed product, and compares it with the final selected traffic to be distributed. Item By comparing the two items within their respective first-level categories, the distance between their first-level categories is calculated, resulting in the category jump distance C. catdist To maintain a consistent data structure, C will be used when the user does not perform another browsing action. catdist The value is set to the maximum category distance allowed by the system, indicating that the user has not formed a continuous browsing behavior.

[0075] After collecting the above three types of feedback indicators, the system integrates and calculates them to quantify the degree of cognitive interference affecting the user's current feedback behavior. Specifically, the disorder of the behavior sequence is calculated based on explicit conversion behavior, bounce behavior, and category jump span, using the following formula: ; Where Turb represents the disorder of the behavioral sequence, C bounce This indicates a bounce-out behavior flag. C represents the result of normalized category jump distance. conv This represents the explicit conversion label, where ω1, ω2, and ω3 represent the weight coefficients corresponding to bounce behavior, category jump range, and explicit conversion behavior, respectively. , , The preferred values ​​are respectively =0.35、 =0.35、 =0.30, all three conditions are met. + + =1. The basis for selecting the above values ​​is that bounce rate and category jump span are the most direct external manifestations of cognitive conflict, each assigned a weight of 35%. Although the lack of conversion behavior also reflects insufficient user interest, it may be affected by non-cognitive factors such as price and logistics, and therefore is assigned a slightly lower weight of 30%. Those skilled in the art can adjust the above weights accordingly based on the correlation analysis results between various indicators and cognitive conflict in specific business scenarios.

[0076] The above formula comprehensively considers multiple dimensions of feedback behavior. When a user quickly exits the page, browses a wide range of categories, and fails to complete a conversion, the Turbidity (Turbo) of the behavioral sequence increases significantly, indicating that user feedback behavior is more influenced by cognitive conflict or random factors. Conversely, when a user continues to browse the currently recommended content and completes a conversion, the Turbidity remains at a low level, indicating that the current feedback behavior has high credibility.

[0077] To further determine whether the cognitive buffering intervention measures are truly effective, this implementation method continues to read the buffering execution records attached to the traffic allocation log in step S3 to verify the buffering execution effect. Specifically, it first determines whether there are valid buffering records in the buffering execution records. When the buffering execution records are empty, it indicates that cognitive buffering was not performed in this traffic allocation, and this step directly sets the disorder of the behavior sequence after buffering correction to the original disorder of the behavior sequence, i.e., Turb'=Turb. When there are valid buffering actions recorded in the buffering execution records, the first interaction time after the user's buffering ends is further read. And compare it with the buffer duration. If the conditions are met... This indicates that the user still needs a certain amount of time after the buffering period ends before starting new interactive behaviors, demonstrating that cognitive buffering intervention effectively extends the user's decision-making interval and further attenuates the cognitive residue formed in step S1. Therefore, the disorder of the behavioral sequence is corrected, and its calculation method is as follows: ; Here, Turb' represents the disorder of the behavioral sequence after buffering correction, and 0.7 is the correction coefficient. This value is based on cognitive psychology research showing that when the decision interval is effectively extended, the interference of prior cognitive residues on subsequent decisions can be reduced by approximately 20% to 40%. Taking the midpoint of 30% as the reduction, the disorder after correction is reduced to 70% of the original, reflecting both the effectiveness of the buffering and preserving reasonable fluctuations due to individual differences. Those skilled in the art can adjust this correction coefficient within the range of 0.5 to 0.9 based on actual A / B test results.

[0078] If the above conditions are not met, it is considered that the buffering intervention has not produced a significant effect, and the original behavioral sequence disorder remains unchanged, i.e., Turb'=Turb.

[0079] After obtaining the corrected behavioral sequence disorder degree Turb', the actual causal purity of this feedback behavior is further calculated using the following formula: ; in, This indicates the actual causal purity of the feedback behavior, while Turb' indicates the disorder of the behavior sequence after buffer correction.

[0080] Through the above calculations, It always remains between 0 and 1. The closer the sample size is to 1, the more likely it is that the current feedback behavior primarily stems from the user's genuine shopping intent, and is less affected by cognitive conflict and random perturbations, making it suitable for use as a high-quality training sample; when... The closer the value is to 0, the more the feedback behavior is affected by cognitive residue or page perturbation, and the feedback cannot truly reflect user preferences. Therefore, its importance in model training should be reduced.

[0081] Finally, based on the traffic allocation records and feedback analysis results, the system generates feedback samples. The feedback samples include Allocation... ID Item, C conv C bounce C catdist W real Among them, Allocation ID The traffic allocation log generated in step S3 is the Allocation log. Log Item represents the final traffic to be sent, which is equivalent to the final traffic to be sent (Selected) obtained in step S3. Item Used to associate corresponding traffic content, C conv Indicates an explicit conversion label, C bounce Indicates the exit behavior flag, C catdist W represents the category jump distance. real This indicates the actual causal purity calculated in this step.

[0082] S5: Obtain user historical behavior confidence profiles from external data sources, combine them with feedback samples, and output new LTV causal effect parameters and category conflict penalty coefficient tables, which are then fed back into the comprehensive allocation score and comprehensive cognitive distance calculation formula.

[0083] Furthermore, step S5 also includes: Receive user historical behavior confidence profiles provided by the external data source Ext3; The purity threshold θ corresponding to the current user is dynamically calculated based on the confidence level of the user's historical behavior. Read the actual causal purity recorded in each feedback sample and compare it with the current purity threshold; When satisfied If the feedback sample is determined to be a high-purity sample, it will be completely stored in the online causal training pool. causal ; When satisfied When the current feedback sample is determined to be a low-purity sample, it is stored in the noise buffer pool. noise Simultaneously, for each low-purity sample, its corresponding comprehensive cognitive distance and conflict category pair are saved. residual Category item}; The long-term value causal inference model is updated using high-purity samples from the online causal training pool to obtain new LTV causal effect parameters. These parameters are then fed back into the long-term value coefficient in the comprehensive allocation score calculation formula. Traverse the noise cache secondary pool noise All conflict category pairs are stored in the database. The frequency of each category combination is counted. Based on the comprehensive cognitive distance stored in the corresponding samples, the average cognitive distance corresponding to the category combination is calculated. The cognitive conflict penalty coefficient is then calculated. The specific calculation formula is as follows: ;of in, Freq represents the cognitive conflict penalty coefficient corresponding to the category combination. AB This indicates the number of times the category combination appears. This indicates the average cognitive distance of this category combination; A new category conflict penalty coefficient table is formed based on the cognitive conflict penalty coefficient, which is then fed back into the category-level penalty item in the comprehensive cognitive distance calculation process.

[0084] Specifically, this step aims to stratify the feedback samples based on the actual causal purity calculated in step S4, thoroughly separating high-purity samples that truly reflect user preferences from low-purity samples that are affected by cognitive conflict. High-purity samples are used to continuously update the long-term value causal inference model, improving the long-term revenue prediction capability of subsequent traffic allocation. Low-purity samples do not participate in model training but are used to uncover patterns of cognitive conflict and form cognitive conflict knowledge, continuously optimizing the cognitive distance calculation process in step S2, thereby constructing a data closed-loop learning mechanism for the entire e-commerce traffic allocation system.

[0085] First, the system receives a user historical behavior confidence profile from an external data source, Ext3. This profile is a standard data asset in e-commerce platform user profiling systems, recording statistical information on the consistency between user behavior and actual conversion results during historical recommendations, and can be used to measure the stability of user behavior. This profile is updated and maintained daily by the platform's data warehouse team through an offline ETL process and stored in data storage systems such as Hive or ClickHouse. This invention directly reads the target user's historical behavior data from this profile, without requiring additional data collection.

[0086] Because the behavioral stability of different users varies significantly, using a fixed purity threshold uniformly can easily lead to a large number of low-quality samples from some users with stable behavior being included in the training model, or some users with highly volatile behavior failing to form effective training samples for a long time. Therefore, this invention first dynamically calculates the purity threshold θ corresponding to the current user based on the confidence level of the user's historical behavior.

[0087] Specifically, the system calculates the consistency variance between the current user's historical behavior and their actual conversion rate. The calculation method is as follows: Extract the user's past nh recommendation exposure records from the user's historical behavior confidence profile. Each recommendation exposure record contains the estimated conversion probability p of that exposure from the long-term value causal inference model. i And the actual conversion result y i Based on these two data points, the prediction error e is calculated. i = y i - p i Then calculate the variance σ of the prediction error for nh times. user ². The variance σ user ² reflects the degree of fluctuation in user behavior relative to the model's expectations. When the consistency variance is small, it indicates that the user's historical behavior has high stability, and its historical feedback can accurately reflect true interests. In this case, the purity threshold θ is set to a higher level. The specific value of θ is determined by a piecewise function. When σ user When ² < 0.05, θ = 0.75; when 0.05 ≤ σ user When ² < 0.12, θ = 0.60; when σ userWhen ² ≥ 0.12, θ = 0.45. The basis for the above segmented values ​​is that the statistical distribution of the variance of user behavior prediction errors shows that the user group with variance less than 0.05 (approximately 30% of the total users) exhibits highly stable behavior, and a high threshold of 0.75 ensures extremely high sample quality entering the training pool; the user group with variance between 0.05 and 0.12 (approximately 50% of the total users) exhibits moderately stable behavior, and a medium threshold of 0.60 is used; the user group with variance greater than 0.12 (approximately 20% of the total users) exhibits more volatile behavior, and a low threshold of 0.45 is used to ensure sufficient samples are still available for training. The calculated purity threshold θ is output as the current user sample quality judgment standard to the next processing stage.

[0088] After obtaining the purity threshold θ, the system begins to perform dual-path splitting of the feedback samples. Specifically, the actual causal purity recorded in each feedback sample output from step S4 is read and compared with the current purity threshold.

[0089] When satisfied At that time, the feedback sample was determined to be a high-purity sample. The system then completely stored the sample in the online causal training pool. causal Since these samples have undergone actual purity verification in step S4, their feedback behavior mainly stems from users' genuine interests, thus serving as the data foundation for training long-term value causal inference models.

[0090] When satisfied If the current feedback sample is deemed to be of low purity, it is considered to be a low-purity sample. Such samples are highly susceptible to cognitive conflict, page disturbances, or random behavior. Directly involving them in long-term value model training could easily introduce erroneous causal relationships. Therefore, this invention prohibits such samples from entering the main online causal training pool and instead stores them in a noise cache secondary pool. noise .

[0091] To facilitate subsequent analysis of the causes of cognitive conflict, the system writes to the noise cache secondary pool. noise At the same time, for each low-purity sample, its corresponding comprehensive cognitive distance and conflict category pair are associated and saved. residual Category item}. The comprehensive cognitive distance originates from the standard candidate traffic set in step S2; the Category in the conflict category pair. residual The residual category set derived from step S1 set Record all categories that form the current cognitive residue. item The final traffic to be sent from step S3 (Selected) ItemThe system records the category information when distributing traffic in step S3 and uses it in conjunction with traffic distribution in step S5. This information is not used in the model training of the current step, but rather serves as an important basis for subsequent analysis of cognitive conflict patterns.

[0092] After completing the dual-path sample splitting, the system begins updating the long-term value causal inference model using high-purity samples from the online causal training pool. The long-term value causal inference model is one of the existing technological foundations upon which this invention relies. Its function is to estimate, based on a user's historical behavior and recommendation exposure records, "how a user's visit frequency and total spending will change over a future period if a certain category of goods is recommended to the user." This model can be implemented using existing causal inference methods such as the Doublely Robust Estimator or Causal Forest. Its training objective is to estimate the true causal impact of different recommendation traffic on the user's future long-term value.

[0093] Since the samples entering the online causal training pool have already passed the purity screening in step S4, effectively eliminating false negative feedback caused by cognitive conflict, the long-term value relationships learned by the model can more realistically reflect the long-term impact of the recommended content itself, without misinterpreting cognitive conflict as a decrease in product value. After completing this round of incremental training, the system updates the long-term value model parameters, including the weight coefficients of the propensity score model and the outcome regression model. The system extracts the long-term value coefficients corresponding to each product category from the updated model. Specifically, for each category c, the model outputs the increase in the user's expected total spending over the next 7 days, ΔLTV, after that category is recommended to the current user. c The increase ΔLTV c This is the LTV causal effect parameter for this category. The ΔLTV for all categories... c In summary, the new LTV causal effect parameter NewLTV = {ΔLTV} is obtained. c | c∈All Categories}. The new LTV causal effect parameter NewLTV will serve as the source of the long-term value coefficient in the comprehensive allocation score calculation formula of step S3, continuously feeding back into step S3 through the feedback loop, so that subsequent traffic allocation strategies can continuously improve their long-term revenue prediction capabilities.

[0094] Meanwhile, the system continues to utilize the noise buffer secondary pool. noise The analysis of low-purity samples reveals patterns of cognitive conflict. Specifically, the system offline traverses the noisy cache secondary pool. noise All conflicting category pairs are stored in the database, and the frequency of each category combination is counted (Freq). AB And calculate the average cognitive distance corresponding to the category combination based on the comprehensive cognitive distance saved in the corresponding samples. When the frequency of a certain category combination is significantly higher than that of random statistics, it indicates that this category combination is more likely to induce cognitive conflict in users. For this category combination, the system calculates the corresponding cognitive conflict penalty coefficient, and the calculation formula is as follows: ;of in, Freq represents the cognitive conflict penalty coefficient corresponding to the category combination. AB This indicates the number of times the category combination appears. This indicates the average cognitive distance of this category combination.

[0095] As shown by the formula above, when a category combination frequently generates cognitive conflict and has a large average cognitive distance, its corresponding penalty coefficient will increase significantly. Conversely, for category combinations with occasional conflict or close proximity, the corresponding penalty coefficient remains at a low level. After completing all statistics, the system generates a new table of category conflict penalty coefficients.

[0096] To enable cognitive distance calculation to continuously adapt to changes in user behavior, this invention further utilizes the aforementioned conflict knowledge to periodically calibrate the cognitive distance measurement function in step S2. Specifically, the system reads the latest generated category conflict penalty coefficient table according to a preset update time period, and writes the cognitive conflict penalty coefficients of each category combination stored therein into the corresponding category-level penalty item P during the cognitive distance calculation process in step S2. cat When step S2 subsequently calculates the comprehensive cognitive distance between candidate traffic and user cognitive residue, if it detects that the corresponding category combination between the candidate category and the cognitive residue category already exists in the category conflict penalty coefficient table, the corresponding combination will be automatically read. As a category-level penalty item P cat Calculation. As the system continuously accumulates new feedback samples, the category conflict penalty coefficient table is continuously updated, thereby continuously optimizing the cognitive distance calculation capability.

[0097] This step ultimately outputs two continuously updated knowledge outcomes. One is the new LTV causal effect parameter, NewLTV, which originates from Pool. causal The long-term value model update results obtained from training with medium-to-high purity samples are used to feed back into the long-term value coefficient in the comprehensive allocation score calculation formula of step S3, thereby continuously optimizing the long-term business benefits of subsequent traffic allocation. Secondly, there is the category conflict penalty coefficient table, which is derived from Pool... noise The category conflict knowledge obtained from the statistical analysis of samples with low to medium purity is used to feed back into the category hierarchy penalty item P in the comprehensive cognitive distance calculation process of step S2. cat This enables the system to continuously revise the comprehensive cognitive distance measurement method based on historical cognitive conflict experience, thereby reducing the generation of low-purity feedback samples in the future from the source.

[0098] Example 2: Based on the same inventive concept as the e-commerce traffic allocation method based on user behavior sequences in the previous examples, this application also provides an e-commerce traffic allocation system based on user behavior sequences. Please refer to the appendix. Figure 2 The system includes: The cognitive extraction module 11 is used to receive raw real-time behavior logs provided by external data sources, perform streaming filtering on the raw real-time behavior logs to obtain the current session's valid behavior event sequence, and extract cognitive residual vectors, time decay coefficients, behavior inertia coefficients, and residual category sets based on the current session's valid behavior event sequence to form a standardized user cognitive state database. The conflict detection module 12 is used to obtain a candidate traffic set based on the candidate traffic pool Embedding library and the coarse-ranked candidate set output by the coarse-ranking stage of the recommendation engine, and to generate a standard candidate traffic set by combining the cognitive residual vector, time decay coefficient and residual category set. The allocation buffer module 13 is used to define the final traffic to be sent based on the standard candidate traffic set and the behavior inertia coefficient, and at the same time, to perform cognitive buffer intervention according to the conflict level of the final traffic to be sent, and to send the final traffic to be sent after the cognitive buffer intervention to the user client, and generate a traffic allocation log. The feedback posterior module 14 is used to set the feedback behavior acquisition window according to the time decay coefficient, acquire the current feedback behavior sequence and three types of feedback indicators according to the feedback behavior acquisition window, perform fusion calculation on the three types of feedback indicators to obtain the behavior sequence disorder, correct the behavior sequence disorder based on the cognitive buffer intervention, calculate the actual causal purity based on the corrected behavior sequence disorder, and finally generate a feedback sample. The dual-path feedback module 15 is used to obtain the user's historical behavior confidence profile from the external data source, combine it with the feedback samples, output the new LTV causal effect parameters and category conflict penalty coefficient table, and feed them back into the comprehensive allocation score and comprehensive cognitive distance calculation formula.

[0099] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0100] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for e-commerce traffic allocation based on user behavior sequences, characterized in that, The method includes: Receive raw real-time behavior logs from external data sources, perform streaming filtering on the raw real-time behavior logs to obtain the current session's valid behavior event sequence, and extract cognitive residual vectors, time decay coefficients, behavior inertia coefficients, and residual category sets based on the current session's valid behavior event sequence to form a standardized user cognitive state database. The candidate traffic set is obtained based on the candidate traffic pool embedding library and the coarse-ranked candidate set output by the coarse-ranking stage of the recommendation engine. The standard candidate traffic set is generated by combining the cognitive residual vector, the time decay coefficient and the residual category set. The final traffic to be distributed is defined based on the standard candidate traffic set and behavioral inertia coefficient. At the same time, cognitive buffering intervention is performed according to the conflict level of the final traffic to be distributed. The final traffic to be distributed after cognitive buffering intervention is then distributed to the user client, and a traffic allocation log is generated. The feedback behavior collection window is set according to the time decay coefficient. The feedback behavior sequence and three types of feedback indicators are obtained according to the feedback behavior collection window. The three types of feedback indicators are fused and calculated to obtain the disorder of the behavior sequence. The disorder of the behavior sequence is corrected in combination with the cognitive buffer intervention. The actual causal purity is calculated based on the corrected disorder of the behavior sequence, and finally the feedback sample is generated. The system obtains user historical behavior confidence profiles from external data sources, combines them with feedback samples, and outputs new LTV causal effect parameters and category conflict penalty coefficient tables, which are then fed back into the comprehensive allocation score and comprehensive cognitive distance calculation formulas.

2. The e-commerce traffic allocation method based on user behavior sequences as described in claim 1, characterized in that, Establish a standardized user cognitive state database, including: Receive raw real-time behavior logs from external data source Ext1, and perform streaming filtering on the raw real-time behavior logs according to event type to form a sequence of valid behavior events for the current session; Based on the current session's sequence of valid behavioral events, an involvement score is calculated for each valid behavioral event. The specific calculation formula is as follows: ； in, This represents the involvement score for the i-th valid behavioral event. This indicates the weight of the corresponding event type. This indicates the duration of time the user stayed on the page. This represents the normalization coefficient for the corresponding page scroll depth; All valid behavioral events within the current session are sorted in descending order of involvement score. Only the top K high involvement events with the highest involvement scores are retained, while other valid behavioral events with lower involvement scores are considered as brief browsing events and are discarded. Use an offline pre-trained semantic embedding model to obtain the semantic vectors corresponding to each highly involved event; The cognitive residual vector is obtained by weighted aggregation of all semantic vectors. The specific calculation formula is as follows: ； Where R represents the cognitive residual vector of the current user session, and V i Eng represents the semantic vector corresponding to the i-th high-involvement event. i This represents the corresponding event involvement score, where K represents the number of high involvement events participating in the aggregation, and i represents the index value of the high involvement event. During the weighted aggregation process, the category identifiers corresponding to all high-involvement events participating in the aggregation are recorded simultaneously to form a residual category set. set ; Extract the occurrence time of the most recent high-involvement event from the above K high-involvement events, and read the current system time. Calculate the time interval between the two using the following formula: ； Where ΔT represents the time difference since the most recent high-involvement behavior, t last Indicates the time of the most recent high-involvement event, t now Indicates the current system time; The time decay coefficient of current cognitive residue is calculated using an exponential decay model. The specific calculation formula is as follows: ； Where α represents the time decay coefficient of cognitive residue, ΔT represents the time difference from the most recent high-involvement behavior, and H represents the half-life parameter of the corresponding product category; Extract the 10 most recent valid behavior events from the current session's valid behavior event sequence, extract the corresponding category identification sequence, and use an offline pre-trained semantic embedding model to map the various categories of identification to a unified semantic space to form a category semantic identifier sequence. Calculate the moving variance of the category semantic identifier sequence, and map the moving variance to the 0-1 interval through the Sigmoid function to obtain the behavior inertia coefficient; The cognitive residual vector, time decay coefficient, behavioral inertia coefficient, and residual category set are uniformly encapsulated to form a standardized user cognitive state database.

3. The e-commerce traffic allocation method based on user behavior sequences as described in claim 1, characterized in that, Generate a standard candidate traffic set, including: It receives the candidate traffic pool Embedding library provided by the external data source Ext2 and the coarse-ranked candidate set output by the coarse-ranking stage of the recommendation engine, integrates the candidate traffic pool Embedding library and the coarse-ranked candidate set to obtain the candidate traffic set. For each candidate traffic in the candidate traffic set, based on the product identifier, the corresponding semantic embedding vector is queried in the candidate traffic pool embedding library. If there is no corresponding cache in the candidate traffic pool embedding library, the semantic representation of the candidate traffic is calculated in real time using the offline pre-trained semantic embedding model to obtain the semantic embedding vector corresponding to the candidate traffic. Obtain cognitive residual vectors from a standardized user cognitive state database, calculate the cosine distance between the semantic embedding vector and the cognitive residual vector, and obtain the basic semantic distance. Based on the category to which the candidate traffic belongs and the residual category set. set The hierarchical relationship between categories determines the category hierarchy penalty items; Based on the basic semantic distance and category-level penalty, the comprehensive cognitive distance corresponding to candidate traffic is calculated using the following formula: ； Among them, D j D represents the comprehensive cognitive distance of the j-th candidate traffic. cos P represents the basic semantic distance. cat β represents the category-level penalty item, and β represents the category penalty adjustment hyperparameter. Real-time detection of the user's current attention state and calculation of the user's real-time perturbation parameters; For each candidate traffic, the corresponding estimated causal confidence weight is calculated by comprehensively utilizing the time decay coefficient, comprehensive cognitive distance, and real-time perturbation parameters. The specific calculation formula is as follows: ； in, Let α represent the estimated causal confidence weight of the j-th candidate flow, and let D represent the time decay coefficient output from step S1. j This represents the overall cognitive distance of the corresponding candidate traffic. λ1 represents the real-time perturbation parameter, λ2 represents the real-time perturbation suppression weight, and Sigmoid(·) represents the Sigmoid activation function. Candidate traffic is classified into conflict levels based on comprehensive cognitive distance, and the conflict levels include green, yellow and red levels; The product identifier, comprehensive cognitive distance, estimated causal confidence weight, and conflict level corresponding to each candidate traffic are uniformly encapsulated to generate a standard candidate traffic set.

4. The e-commerce traffic allocation method based on user behavior sequences as described in claim 1, characterized in that, Generate traffic allocation logs, including: Read the standard candidate traffic set, estimate the pCVR conversion rate, and the new LTV causal effect parameters; For each candidate traffic in the standard candidate traffic set, a comprehensive allocation score is calculated. The specific calculation method is as follows: ； Among them, Score j pCVR represents the comprehensive allocation score of the j-th candidate traffic. j This represents the estimated conversion rate (LTV) output by the main model of the recommendation engine. j This represents the long-term value coefficient, which is equivalent to the new LTV causal effect parameter obtained from the feedback in step S5. The estimated causal confidence weights calculated in step S2 are represented by δpenalty, which represents the purity loss penalty base. The category of the current candidate traffic is compared with the category of the user's most recent behavior. When the category span between the candidate traffic and the most recent behavior exceeds two levels, a behavior inertia coefficient is introduced to correct the comprehensive allocation score. If the span between the category of the candidate traffic and the most recent behavior does not exceed two levels, the comprehensive allocation score remains unchanged. The candidate traffic flows are sorted according to their comprehensive allocation scores after inertia constraint correction, and the highest-scoring candidate traffic flows are selected as the final traffic flows to be distributed. Item Simultaneously, the conflict level corresponding to this traffic is extracted from the standard candidate traffic set. j The final recommendation results are then verified again. When the conflict level corresponding to the final traffic to be distributed is green or yellow, the traffic distribution phase begins directly. When the conflict level corresponding to the final traffic to be distributed is red, a cognitive buffer intervention is further implemented, inserting a buffer duration of Δt before the final traffic to be distributed is officially displayed. buf The cognitive buffering phase, buffering duration Δt buf Determined based on the half-life parameter H; Based on the cognitive buffering intervention process, a buffering execution record is generated. At the same time, the final traffic to be distributed, the official time of this traffic distribution, and the comprehensive allocation score are integrated to generate this traffic allocation log.

5. The e-commerce traffic allocation method based on user behavior sequences as described in claim 1, characterized in that, Generate feedback samples, including: The feedback behavior collection window is dynamically adjusted based on the degree of cognitive residual impact. The specific calculation method is as follows: ； in, This indicates the feedback behavior collection window. This represents the system's baseline acquisition window, and α represents the time decay coefficient. Read the traffic allocation log and record the official traffic distribution time T. alloc As the starting time for feedback behavior collection, based on the starting time and the feedback behavior collection window, all feedback behaviors generated by the user within the corresponding time range are captured to form the current feedback behavior sequence. At the same time, based on the current feedback behavior sequence, three types of feedback indicators are collected, including explicit conversion tags, bounce behavior indicators, and category jump distance. The disorder of the behavior sequence is calculated based on explicit conversion behavior, exit behavior, and category jump span. The specific calculation formula is as follows: ； Where Turb represents the disorder of the behavioral sequence, C bounce This indicates a bounce-out behavior flag. C represents the result of normalized category jump distance. conv The explicit conversion label is represented by ω1, ω2, and ω3, which represent the weight coefficients corresponding to the bounce behavior, category jump span, and explicit conversion behavior, respectively. The buffer execution record attached to the traffic allocation log is read to verify the buffer execution effect. When the buffer execution record is empty, the disorder of the behavior sequence after buffer correction is directly set to the original disorder of the behavior sequence. When there is a valid buffer action recorded in the buffer execution record, the time of the user's first interaction after the buffer ends is further read. And compare it with the buffer duration. If the conditions are met... The disorder of the behavioral sequence is corrected, and its calculation method is as follows: ； Where Turb' represents the disorder of the behavior sequence after buffer correction, Turb represents the disorder of the behavior sequence, and 0.7 is the correction coefficient; If the above conditions are not met, the original disorder of the behavioral sequence remains unchanged; The actual causal purity of this feedback behavior is further calculated using the following formula: ； in, This indicates the actual causal purity of the feedback behavior, while Turb' indicates the disorder of the behavior sequence after buffer correction. Based on the traffic allocation records and feedback analysis results, a feedback sample is generated.

6. The e-commerce traffic allocation method based on user behavior sequences as described in claim 1, characterized in that, Output the new LTV causal effect parameters and category conflict penalty coefficient table, including: Receive user historical behavior confidence profiles provided by the external data source Ext3; The purity threshold θ corresponding to the current user is dynamically calculated based on the confidence level of the user's historical behavior. Read the actual causal purity recorded in each feedback sample and compare it with the current purity threshold; When satisfied If the feedback sample is determined to be a high-purity sample, it will be completely stored in the online causal training pool. causal ; When satisfied When the current feedback sample is determined to be a low-purity sample, it is stored in the noise buffer pool. noise Simultaneously, for each low-purity sample, its corresponding comprehensive cognitive distance and conflict category pair are saved. residual Category item }; The long-term value causal inference model is updated using high-purity samples from the online causal training pool to obtain new LTV causal effect parameters. These parameters are then fed back into the long-term value coefficient in the comprehensive allocation score calculation formula. Traverse the noise cache secondary pool noise All conflict category pairs are stored in the database. The frequency of each category combination is counted. Based on the comprehensive cognitive distance stored in the corresponding samples, the average cognitive distance corresponding to the category combination is calculated. The cognitive conflict penalty coefficient is then calculated. The specific calculation formula is as follows: ;of in, Freq represents the cognitive conflict penalty coefficient corresponding to the category combination. AB This indicates the number of times the category combination appears. This indicates the average cognitive distance of this category combination; A new category conflict penalty coefficient table is formed based on the cognitive conflict penalty coefficient, which is then fed back into the category-level penalty item in the comprehensive cognitive distance calculation process.

7. An e-commerce traffic allocation system based on user behavior sequences, characterized in that, The system is used to implement the e-commerce traffic allocation method based on user behavior sequences as described in any one of claims 1 to 6, and the system comprises: The cognitive extraction module is used to receive raw real-time behavior logs provided by external data sources, perform streaming filtering on the raw real-time behavior logs to obtain the current session's valid behavior event sequence, and extract cognitive residual vectors, time decay coefficients, behavior inertia coefficients, and residual category sets based on the current session's valid behavior event sequence to form a standardized user cognitive state database. The conflict detection module is used to obtain a candidate traffic set based on the candidate traffic pool Embedding library and the coarse-ranked candidate set output by the coarse-ranking stage of the recommendation engine, and to generate a standard candidate traffic set by combining the cognitive residual vector, time decay coefficient and residual category set. The allocation buffer module is used to define the final traffic to be sent based on the standard candidate traffic set and the behavior inertia coefficient. At the same time, it performs cognitive buffer intervention according to the conflict level of the final traffic to be sent, and sends the final traffic to be sent after cognitive buffer intervention to the user client, generating a traffic allocation log. The feedback posterior module is used to set the feedback behavior acquisition window according to the time decay coefficient, acquire the current feedback behavior sequence and three types of feedback indicators according to the feedback behavior acquisition window, fuse the three types of feedback indicators to obtain the behavior sequence disorder, correct the behavior sequence disorder based on the cognitive buffer intervention, calculate the actual causal purity based on the corrected behavior sequence disorder, and finally generate feedback samples. The dual-path feedback module is used to obtain user historical behavior confidence profiles from external data sources, combine them with feedback samples, output new LTV causal effect parameters and category conflict penalty coefficient tables, and feed them back into the comprehensive allocation score and comprehensive cognitive distance calculation formula.

Images