E-commerce debiasing recommendation method and system based on funnel heat and sensitivity stratification

Abstract

The application discloses an e-commerce debiasing recommendation method and system based on funnel heat and sensitivity layering, and belongs to the technical field of commodity recommendation. In a preset statistical window, multiple types of user behaviors are weighted and time-decay aggregated to generate commodity funnel heat; according to the heat bucket to which the user historical interactive commodity belongs, the heat preference grade of the user is determined; dynamic disturbance is introduced in combination with the interactive time interval and the behavior type to train a hypergraph recommendation model. In the online stage, the user preference grade and the candidate commodity funnel heat are read, the original prediction logarithmic value is calculated through the model, is monotonically normalized into the online basic ranking score through the Sigmoid function, and differential rearrangement is performed; when the user historical interaction is insufficient, the cold start bypass is automatically switched to. After the online behavior backflow, the related commodity heat and the user preference grade are incrementally updated. The scheme effectively alleviates the interference of heat bias on the recommendation result while guaranteeing the online response efficiency.

Description

Technical Field

[0001] This invention relates to the field of product recommendation technology, specifically to an e-commerce recommendation method and system based on funnel heat and sensitivity stratification. Background Technology

[0002] As the number of users and products on e-commerce platforms grows, recommendation systems typically generate user-oriented recommendation lists from candidate products through stages such as recall, coarse ranking, fine ranking, and re-ranking. Graph convolution and hypergraph convolution methods can learn representations using high-order interaction relationships between users and products, but their training data may still be influenced by historical exposure mechanisms, product popularity, and user behavior tendencies.

[0003] Existing graph recommendation methods typically perform neighbor aggregation on the user-product graph or user interaction hypergraph; some debiasing methods reduce the dominance of popular products through popularity features, causal decomposition, reweighting, or regularization. Different methods handle interaction types, behavior time decay, differences in user popularity preferences, and online update mechanisms in different ways.

[0004] Existing technologies for online e-commerce deployment have at least the following issues that require further resolution: 1. Multi-layer graph aggregation may reduce the distinguishability of node representations; for behaviors at different times and different transformation stages within the same hyperedge, if fixed structural weights are used, it is difficult to reflect their timeliness and confidence level. 2. Historical interactions are influenced by both platform exposure strategies and product popularity. If the same debiasing intensity is applied to all users, it may not be possible to take into account the different user preferences. Furthermore, inconsistencies between offline training features and online reordering features may lead to differences in prediction accuracy. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this invention provides an e-commerce recommendation method and system based on funnel heat and sensitivity stratification.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A method for bias-free recommendation in e-commerce based on funnel heat and sensitivity stratification includes the following steps: S1. Collect various types of user behavior data in e-commerce scenarios, aggregate the data within a preset statistical window according to behavior conversion weight and time decay rules to obtain product funnel heat, and discretize the product funnel heat into heat buckets according to preset bucketing rules. S2. For users whose historical interaction product set is not empty, calculate the proportion of top popular products based on the popularity bucket corresponding to their historical interaction products, and map the user to the corresponding popularity preference level according to the proportion; for users whose historical interaction product set is empty, use the preset default level. S3. Construct a user hyperedge based on the user's historical interaction with products, generate dynamic perturbations based on the interaction time interval and behavior type between products within the hyperedge, add the dynamic perturbations to the structure aggregation process of hypergraph convolution, train the recommendation model using hypergraph convolution with dynamic perturbations and a unified prediction function, and send the trained model parameters to the online feature library. S4. In response to an online recommendation request, read the requesting user's popularity preference level, the funnel heat of candidate products, and the online parameters of the recommendation model. Calculate the original model prediction logarithm of the candidate products using the online parameters. Monotonically normalize the original model prediction logarithm using the Sigmoid function to obtain an online normalized basic ranking score with a value range of [0, 1]. Perform differential correction on the online normalized basic ranking score according to the popularity preference level and the funnel heat of candidate products. When the number of historically interacted products of the requesting user is lower than a preset cold start threshold, switch to cold start bypass to generate recommendation results. S5. Collect real-time user behavior data generated in response to the recommendation results, feed the real-time user behavior data back to the behavior data path, and perform incremental updates on the product funnel heat and user heat preference levels related to the real-time user behavior data for subsequent online recommendation requests.

[0007] Furthermore, the specific method for aggregating the product funnel heat according to behavioral conversion weights and time decay rules in S1 is as follows: S11. Define the set of e-commerce user behavior types. , for set Each type of behavior in Preset the corresponding funnel weight With time decay constant Among them, the behavior types are sorted by conversion priority, and the behaviors at the end of the funnel are... Values ​​greater than the behavior at the top of the funnel Funnel end behavior Values ​​greater than the behavior at the top of the funnel ; S12. For any product At any moment Product funnel heat Calculated using the following formula:

[0008] In the formula, For goods exist Behavioral types occurring at any time Number of times, It is a natural exponential function.

[0009] Furthermore, the specific method for calculating the user's head heat ratio and mapping the user to the corresponding sensitivity level in S2 is as follows: S21, Preset head heat bucket threshold Total number of heat barrels Number the heat barrels in ascending order of heat, with the numbers greater than... The heat tank is the head heat tank; S22. For users whose historical interaction item set is not empty. Calculate the percentage of products belonging to the top trending categories in the historical product interaction set. , Calculated using the following formula:

[0010] In the formula, For users A collection of historical interactive products. This refers to the heat bucket number corresponding to product i. A counting operator for the number of elements in a set; S23, Preset conservative to neutral threshold From neutral to conformity threshold , will users Mapped to sensitivity level The mapping rule is: when hour For conservative purposes, when hour For neutral settings, when hour To conform to the crowd; S24. For users whose historical interactive product set is empty. The default sensitivity level is directly assigned, and the default level is the neutral level.

[0011] Furthermore, the specific method for generating dynamic perturbations and injecting them into the hypergraph convolution process in S3 is as follows: S31, For users Historical interactive product collection Any two items and Calculate the interaction time between the two. And determine the dominant behavior type of their interaction. ; Among them, the behavior priority is arranged from high to low in the order of placing an order, adding to cart, favorite, clicking, and exposure; when a product involves multiple behaviors in the statistics window, the behavior with the highest weight is determined as the dominant behavior, and when the weights are the same, the behavior that occurs later is selected as the dominant behavior. S32, using a multi-scale time decay function With behavioral confidence function Jointly modulate self-attention related terms to generate dynamic perturbations of the product pair. , Calculated using the following formula:

[0012] In the formula: The number of scales for multi-scale time decay. For the first Weights for each time scale For the first The decay constant corresponding to each time scale It is a natural exponential function. For product interaction time interval; S33, Dynamically disturbed elements Adding the symmetric normalized adjacency submatrix of the hypergraph convolution to obtain the perturbed adjacency matrix, and then performing the perturbed hypergraph convolution using the following formula, yields the first... Layer embedding :

[0013] In the formula, It is a symmetric normalized adjacency matrix. It is a non-linear activation function. For the first Layer embedding The intermediate embedding matrix obtained after perturbation propagation and nonlinear transformation The attention value projection matrix, For multi-head attention aggregation operators; Furthermore, the unified prediction function and model training method in S3 include: S34. Preset thermal coupling coefficients for different sensitivity levels. Among them, the heat coupling coefficient of the crowd Positive value, neutral range The thermal coupling coefficient is zero or within a preset neutral range, representing the conservative setting. The value is within the preset negative correction interval; S35, the same prediction function is used in the forward training phase and the online main model path to calculate the user's prediction. For goods The original prediction logarithmic value of the model The original logarithmic prediction value of the model is calculated using the following formula:

[0014] In the formula, For users The representation vector, For goods The representation vector, For vector dot product operators, For users Popularity embedding, For goods Popularity embedding; S36. The model is trained using a hierarchical gradient equalization loss function, which is calculated using the following formula:

[0015]

[0016]

[0017]

[0018] In the formula, For users The training product sample set, For sample weights, For users For goods The true label, It is the logical sigmoid function. These are intermediate predicted values ​​with individual bias. For users A collection of future interactive products For goods The heat correlation regularization coefficient, For users A collection of historical interactive products. The L2 norm square operator for vectors, The basic regularization coefficient, For users Individual bias embedding, This is the thermal regularization gain coefficient. The normalization transformation function for heat is... For goods exist The temperature of the funnel at any given moment. It is a commodity i The product embedding vector after bias correction, For users Sensitivity level The corresponding weight coefficients of the already interacted samples, For users Sensitivity level The corresponding future interaction sample weight coefficients, It is a commodity At any moment The temperature of the funnel; Furthermore, step S4 specifically includes the following steps: S41. Preset the rearrangement reward / penalty coefficients for different sensitivity levels. Among them, the crowd Positive value, neutral range Values ​​that are zero or fall within a preset neutral range, conservative settings. The value is within the preset negative correction interval; S42, For online request candidate product set The original predicted logarithmic values ​​obtained in step S35 are monotonically normalized using the Sigmoid function to obtain an online normalized basic ranking score with a value range of [0, 1]. Based on the online normalized basic ranking score, each candidate product... i Calculate rearrangement score The calculation formula is as follows:

[0019] In the formula, Candidate Products Online normalized base sorting score, The normalization transformation function for heat is... For goods exist The temperature of the funnel at any given moment; S43. Sort the candidate products in descending order according to the rearrangement sorting, and select the top... The recommended items are output as the rearranged recommendation results. This is the preset recommended output length; S44, Preset cold start threshold Requesting users Length of historical interactive product collection : or When the value is empty, a cold start bypass is triggered. For users A collection of historical interactive products; S45, Based on the requesting user The default sensitivity level is used to retrieve candidate products from the corresponding popularity bucket and select the top ones. Each item is output as a cold start recommendation result.

[0020] Furthermore, the specific method for feeding back the real-time user behavior data generated by online recommendations in S5 is as follows: S51. Collect the exposure, click, favorite, add to cart and order behavior logs generated by the e-commerce terminal for the recommendation results, and record each behavior log as a five-tuple containing user identifier, product identifier, behavior type, timestamp and context identifier; S52. Write the behavior log to the message queue, perform funnel heat increment aggregation for products with new behaviors, and recalculate the heat preference level for users with new behaviors; write the updated results to the online feature library for subsequent online recommendation requests to call.

[0021] A bias-free e-commerce recommendation system based on funnel heat and sensitivity stratification includes a memory, a processor, and a program stored in the memory and executed by the processor. When the processor executes the program, it implements bias-free e-commerce recommendations through the following functional modules: The e-commerce behavior flow collection module is used to collect various types of user behavior data and write behavior records composed of user identifier, product identifier, behavior type, timestamp and context identifier into the message queue. The funnel heat calculation module is used to aggregate the product funnel heat according to the behavior conversion weight and time decay rules within a preset statistical window, generate heat buckets according to the preset bucketing rules, and write the product funnel heat and heat buckets into the online feature library. The user popularity preference layering module is used to read the user's historical interaction with products and their popularity buckets, calculate the proportion of the user's top popular products, map non-empty historical users to corresponding popularity preference levels, and provide preset default levels for users with empty historical data. The offline model training module is used to construct user hyperedges based on users' historical interactions with products, generate dynamic perturbations according to the interaction time interval and behavior type between products, train the recommendation model using hypergraph convolution with dynamic perturbations and a unified prediction function, and write the model parameters into the online feature library. The online re-ranking module is used to respond to online recommendation requests, read user popularity preference levels, candidate product funnel popularity, and model parameters, calculate the original model prediction logarithm of candidate products using the model parameters, monotonically normalize the original model prediction logarithm of the model using the Sigmoid function to obtain an online normalized basic ranking score, and perform differential correction on the online normalized basic ranking score according to the user popularity preference level and candidate product funnel popularity to obtain the re-ranking score and output the recommendation result. The cold start bypass module is used to bypass the requesting user when the number of historically interacted products is less than 5: when the historical interactions are empty, a neutral level is used; when the number of historical interactions is 1 to 4, the level is determined by the proportion of head popularity after Laplace smoothing; the high popularity preference level, neutral level, and low popularity preference level retrieve products from the 8th to 10th bucket, the 4th to 8th bucket, and the 1st to 5th bucket, respectively. After being filtered for availability, category compliance, SPU deduplication, and in-bucket quality sorting, the cold start recommendation results are output. An online feature library is used to store product funnel heat, heat bucket, user heat preference level and model parameters according to version identifier, and provides read and write interfaces to the funnel heat calculation module, user heat preference layering module, offline model training module, online reordering module and cold start bypass module; The real-time feedback update module is used to consume the new behavior records written by the e-commerce behavior flow collection module, incrementally update the funnel popularity of related products and the popularity preference levels of related users, and write the update results into the online feature library.

[0022] The present invention has the following beneficial effects: 1. Dynamic perturbation adjusts the structural aggregation weights in the hypergraph convolution based on the interaction time and behavior type of the product pair, so that interactions with different timeliness and different behavioral stages can be distinguished during the representation learning process.

[0023] 2. Funnel heat index integrates conversion levels and time decay information from multiple behaviors. Compared with heat index that only uses a single interaction count, it can reduce the dominance of high exposure and low conversion behaviors on product heat index estimation.

[0024] 3. User popularity preference stratification enables the online re-ranking coefficient to be configured differently according to the proportion of top popular products in the user's historical interactions, thereby providing user-level adjustment capabilities between hit rate, recommendation diversity and long-tail product exposure.

[0025] 4. The same definition of the original model prediction logarithm value is used in both the training and online phases. The online normalized base ranking score is obtained by monotonically normalizing the original model prediction logarithm value using the Sigmoid function. Then, in the online re-ranking phase, the online normalized base ranking score is corrected for heat to obtain the re-ranking score, thereby reducing the inconsistency between the offline training caliber and the online service caliber.

[0026] 5. Product funnel popularity, user tier, model parameters, and online re-ranking strategies are versioned and accessed through an online feature library, enabling popularity and tier to be updated without retraining the main model.

[0027] 6. Cold start bypass bypasses the main model path that depends on the user's superedge when there is insufficient user historical interaction. It generates candidate products through the default level and candidate popularity range, reducing the instability of prediction caused by the superedge being too short.

[0028] 7. After online behavior feedback, only the popularity of related products and the user tiers are incrementally updated, so that subsequent recommendation requests can use the updated popularity features and user tiers.

[0029] 8. The above-mentioned technical effects are generated by the synergistic relationship between funnel heat calculation, user-level tiers, the logarithmic values ​​of the model's original predictions and their normalized mapping relationship, and online re-ranking, and are not based on unverified specific improvement ratios as necessary evidence. During verification, ablation studies were conducted using the same dataset, the same candidate set, and the same evaluation cutoff point to compare Recall@K, NDCG@K, in-list diversity, and the proportion of long-tail exposures. When online conditions are available, a randomized A / B test was then used to evaluate click-through rate and conversion rate. Attached Figure Description

[0030] Figure 1 This is a system structure block diagram of an embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram of the method flow of an embodiment of the present invention.

[0032] Figure 3 This diagram illustrates the core calculation and online branching of an embodiment of the present invention. In the diagram, K_rec represents the recommended output length, K_bucket represents the number of hot buckets, and K_head represents the threshold of the head bucket; a larger hot bucket number indicates a more popular bucket. Detailed Implementation

[0033] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0034] The application scenario of this invention is specifically targeted at the Top-K recommendation link of e-commerce platforms. That is, in typical product list positions such as "Homepage You May Like / Category Channel Page / Search Results Page / Related Recommendations on Details Page / Shopping Cart Combinations / Post-Order Recommendations" in e-commerce apps or websites, the top-K recommendations are returned based on the predicted matching degree between the user and the candidate products. The invention relates to the engineering process of displaying a product on the front end. The modified part of this invention consists of three segments: fine-sorting, reordering, and cold-start bypass. These segments are further modified by four preceding modules: "behavior flow collection → funnel heat engine → sensitivity layering service → offline main model," which provide training data, heat features, and user tags. The prerequisites for this invention are as follows: Among them, the recommended output length and the offline evaluation truncation point use different symbols; the output of equation (4) is the original predicted log value of the model, and the online normalized basic ranking score of equation (5) is fixedly obtained by monotonically normalizing the original predicted log value of the model using the Sigmoid function; equation (5) performs heat preference correction on the online normalized basic ranking score and outputs the rearranged ranking score.

[0035] Data access requirements: E-commerce platforms must be able to record at least three types of behaviors, such as exposure, clicks, favorites, adding to cart, and placing orders, and write the behavior logs into the behavior data path as a five-tuple containing user identifier, product identifier, behavior type, timestamp, and context identifier; Product identification criteria: Candidate products must have stable product identification and be able to obtain category, inventory status, regional availability, and other attributes required for online filtering; Link connection conditions: The recommendation link can provide candidate products, the recommendation model can calculate the original prediction log value of the model according to formula (4), and allow the online re-ranking module to read the user level and product funnel heat before outputting the recommendation list; Computational requirements: The offline side provides model training resources, while the online side provides message queues, stream processing components, and low-latency feature storage; specific product implementations can use components with equivalent functionality.

[0036] This embodiment deploys multi-behavior funnel heat, user heat preference stratification, hypergraph model with dynamic perturbation, online reordering, and cold start bypass in the e-commerce recommendation chain. Each stage forms a closed loop through clear data input, calculation results, and versioned feature access.

[0037] like Figure 1 The diagram shown is a system structure block diagram of the present invention. It illustrates the connections between the e-commerce behavior collection module, the funnel heat calculation module, the user sensitivity layering module, the offline bias-reduction main model module, the online sensitivity reordering module, the cold start bypass module, and the online feature library. Module 1: E-commerce behavior flow collection module, which is responsible for unifying behaviors such as exposure, click, favorites, add to cart and place of order into five-tuple records and writing them into the message queue, as the data source for funnel heat and user tier updates; Module 2: Funnel Heat Calculation Module. This module records consumer behavior and maintains the product funnel heat according to behavior weights and time decay rules within a preset statistical window. It also generates heat buckets according to preset bucketing rules. Its output is written to Module 6 and read by Modules 3, 4, and 5. Module 3: User Popularity Preference Layering Module. This module reads the user's historical interaction with products and their corresponding popularity buckets, calculates the percentage of top popular products, and generates user tiers. For users with no historical data, a preset default tier is written to their profile. Module 4: Offline model training module, reads interaction data, product funnel heat and user tier, performs user hyperedge construction, dynamic perturbation calculation, hypergraph convolution with dynamic perturbation and model training, and writes the model parameters with version identifier into Module 6; Module 5: Online Re-ranking Module. Responding to recommendation requests, it reads user tiers and product funnel heat, calculates the original model prediction logarithm of candidate products using model parameters, monotonically normalizes the original model prediction logarithm using the Sigmoid function to obtain an online normalized base ranking score, performs differential correction on the online normalized base ranking score, obtains the re-ranking score, and outputs the final recommendation list; it forms a mutually exclusive online branch with Module 7. Module Six: Online Feature Library. This library caches product funnel heat, heat bucket, user tiers, and model parameters according to version identifiers, and provides a consistent read interface to both the offline training module and the online service module. Module 7: Cold Start Bypass Module. When the number of historical products interacted with by a user is less than 5, the module enters the bypass mode. Users with no historical products are assigned a neutral level, while users with 1 to 4 historical products interacted with are assigned a level using Laplace smoothing. Products within a preset popularity bucket are retrieved according to the level, and a recommendation list is output after filtering by availability, category compliance, SPU deduplication, and quality sorting within the bucket.

[0038] After receiving online behavior, Module 1 performs incremental popularity updates on relevant products, and Module 3 recalculates the tiers for relevant users; Module 4 updates model parameters according to the offline cycle; Modules 5 and 7 read the same version of features during the online request phase. This forms a closed-loop processing mechanism of behavior feedback, feature update, online reading, and recommendation output.

[0039] This embodiment describes the execution process in four stages: offline training, near-line feature refresh, online recommendation request, and behavior reflow. The P series represents the offline training step, the N series represents the near-line refresh step, the R series represents the online request step, and the F series represents the behavior reflow step. The step numbers correspond to... Figure 2 The following text is consistent with this.

[0040] like Figure 2The diagram shows the overall flowchart of the method according to an embodiment of the present invention. This diagram illustrates the steps of the method in four stages: offline training, near-line refresh, online request, and second-level behavior backflow. A unique notation system is used to describe the core formulas in subsequent steps. The evaluation cutoff point is uniformly denoted as... It is recommended that the output length be uniformly denoted as . Candidate sets are uniformly denoted as User sensitivity levels are uniformly denoted as: The online normalized base sort score and rearrangement score of equation (5) are denoted as follows: and As shown in Table 1. The original predicted logarithmic values ​​of the model in Equation (4) are monotonically normalized by the Sigmoid function to obtain the online normalized basic ranking score, the symbols and meanings of which are also shown in Table 1.

[0041] In this embodiment, the user tiers refer to the popularity preference tiers obtained based on the proportion of top-selling products; "conformity," "neutrality," and "conservatism" are category labels and do not represent a judgment on the user's psychological attributes. To avoid ambiguity, in engineering implementation, they can be named high popularity preference tier, neutral tier, and low popularity preference tier, respectively.

[0042] Table 1 Definition of Main Symbols

[0043] Equation (1) is for funnel thermal polymerization (used in subsequent step N1):

[0044] In the formula, the behavior count term represents the number of times a product performs a specific behavior at a certain historical moment, and the behavior weight and time decay constant describe the importance and timeliness of the funnel behavior, respectively.

[0045] In this invention, the statistical window is 7 to 90 days; the weights of each behavior are all located in (0, 1] and the weight of the order behavior is normalized to 1, and the weights of exposure, click, favorite, add to cart, and order behavior increase sequentially; the time decay constants are 1 to 30 days, and the time decay constant of the behavior at the end of the funnel is not less than the time decay constant of the behavior at the top of the funnel. In this embodiment, formula (1) is calculated within the statistical window of the most recent 30 days; the initial value of the funnel heat of new products with no behavior records in the window is 0, and they are temporarily assigned to the first heat bucket. The product freshness item in the cold start sort provides a non-zero exploration opportunity for new products; after the behavior is generated, the heat is updated according to formula (1) and reassigned to the bucket when the bucket boundary is refreshed next time. In this embodiment, the weights of exposure, click, favorite, add to cart, and order behavior are 0.05, 0.20, 0.45, 0.70, and 1.00 respectively, and the corresponding time decay constants are 1 day, 3 days, 7 days, 14 days, and 30 days respectively. This makes the behavior at the later conversion level have higher weights and slower decay speeds.

[0046] The total number of heat buckets is an integer from 5 to 20. Heat buckets are divided according to the equal frequency quantiles of the funnel heat of all candidate products, and numbered in ascending order of heat. Let the m-th quantile be q_m. The product heat bucket number is determined by adding 1 to the number of quantiles whose funnel heat is greater than q_m. Therefore, products with the same quantile are placed into lower-numbered buckets. The threshold for the head heat bucket is 80% of the total number of heat buckets, rounded down. Buckets with a bucket number greater than this threshold are head heat buckets. In this embodiment, the total number of heat buckets is 10. Each time the bucket boundaries are recalculated, 10%, 20% to 90% quantiles are calculated based on the funnel heat of the entire candidate product set, and the bucket number is determined by adding 1 to the number of quantiles strictly less than the product heat. Products with the same quantile are placed into lower-numbered buckets, allowing for empty buckets due to a large number of identical heat values. The threshold for the head heat bucket is 8, therefore buckets 9 and 10 are head heat buckets.

[0047] The difference in this embodiment is that product popularity considers the funnel levels and time decay of exposure, clicks, favorites, add-to-cart, and order placement behaviors simultaneously, rather than relying solely on a single interaction number to form a popularity metric. This difference is directly applied to subsequent bucketing, user tiers, and online re-ranking.

[0048] Equations (2), (2a), and (2b) demonstrate dynamic perturbation and perturbed hypergraph convolution (used in subsequent step P2):

[0049] Equation (2) satisfies the non-negativity, symmetry and normalization constraints: the diagonal elements of the perturbation matrix are 0, and the two-way perturbation is the arithmetic mean; after adding with the structural adjacency matrix, the negative values ​​are truncated to 0, the original structural self-loops are retained and unit self-loops are added, and then node degree symmetric normalization is performed.

[0050]

[0051]

[0052] In the formula, the confidence of the dominant behavior type of the product pair, the multi-scale time decay function, and the attention query matrix, attention key matrix and attention value matrix are all defined according to the symbol table; Equation (2) generates dynamic perturbations for each product pair. The diagonal elements of the dynamic perturbation matrix are 0, and the off-diagonal elements are symmetric according to the arithmetic mean of the bidirectional calculation results of the product pair. After adding the symmetric perturbation to the structural adjacency matrix, the negative values ​​are truncated to 0, the original structural self-loops are retained and a unit self-loop matrix is ​​added, and then symmetric normalization is performed according to the node degree; the normalized matrix is ​​used as the perturbation structure matrix in Equation (2b).

[0053] The difference in this embodiment is that the dynamic perturbation is generated pairwise according to the interaction time interval and dominant behavior type of each product pair, and symmetry and normalization are performed before each hypergraph convolution layer, so that the structural weights change with the interaction timeliness and behavior confidence. The behavior confidence function is the ratio of the corresponding behavior weight to the maximum behavior weight, and the output range is (0, 1). The multi-scale time decay uses three time scales, with scale weights of 0.6, 0.3 and 0.1 respectively, and decay constants of 1 day, 7 days and 30 days respectively; all scale weights are non-negative and their sum is 1.

[0054] Equations (3) and (3a) indicate the user sensitivity levels (used in subsequent steps N3-N4):

[0055] The denominator of equation (3) is calculated only when the historical interaction set is not empty; empty historical users directly enter the default gear and cold start bypass to avoid the denominator being zero.

[0056]

[0057] In the formula, the starting threshold of the head heat bucket, the conservative to neutral threshold, and the neutral to conformity threshold are all defined according to the symbol table; formulas (3) and (3a) are calculated only when the user's historical interaction product set is not empty. If the user's history is empty, formula (3) / (3a) is not substituted, and the cold start bypass is directly entered and the default or prior level is adopted; the default level for users with empty history is the neutral level. The user's historical interaction window is 7 to 90 days, and the historical interaction set only includes clicked, favorited, added to cart, or ordered products, and is deduplicated according to the product identifier; the exposure behavior is only used for product funnel heat calculation and does not constitute user active interaction alone. The conservative to neutral threshold is 0.35, the neutral to conformity threshold is 0.65, and 0 < 0.35 < 0.65 < 1 is satisfied.

[0058] The difference in this embodiment is that the bias reduction intensity is determined by the proportion of top-ranked products in each user's historical interactions, rather than using the same coefficient for all users; the same level simultaneously constrains the heat coupling of equation (4), the online rearrangement of equation (5), and the sample weight of equation (6).

[0059] Equation (4) is the formula for calculating the original logarithmic prediction value of the model, which is used for the training forward step P3; the online main model path adopts the same definition of the original logarithmic prediction value of the model, and the original logarithmic prediction value of the model is monotonically normalized by the Sigmoid function to obtain the online normalized basic ranking score. The online re-ranking step then performs heat correction on the online normalized basic ranking score according to Equation (5):

[0060] Equation (4) defines the original predicted logarithmic value of the model; in the online stage, Equation (4) is used as the output, and the original predicted logarithmic value of the model is monotonically normalized by the Sigmoid function to obtain the online normalized basic ranking score of Equation (5); the ranking scores of existing candidate generation models are not included in Equation (5).

[0061] In the formula, the heat coupling coefficient is looked up according to the user level; the high heat preference level, the neutral level, and the low heat preference level are 0.20, 0, and -0.20, respectively. The user representation, product representation, and heat embedding are all L2 normalized so that each inner product term is in the range [-1, 1]. The rearrangement reward and punishment coefficient is not included in equation (4) and is only used in equation (5). After the user representation, product representation, and heat embedding are all L2 normalized, the heat coupling coefficient of the high heat preference level is 0.05 to 0.30, the neutral level is -0.02 to 0.02, and the low heat preference level is -0.30 to -0.05; a set of specific values ​​are 0.20, 0, and -0.20, respectively.

[0062] The difference in this embodiment is that: both the training forward pass and the online main path use equation (4) to generate the original predicted logarithmic values ​​of the model, and obtain the online normalized basic ranking score by monotonically normalizing it using the same Sigmoid function; the online stage does not switch to another set of unlabeled fine ranking scores.

[0063] Equation (5) Online differentiated reward and punishment rearrangement (used in step R3):

[0064] Equation (5) is only executed for the main route candidate; the cold start bypass does not calculate Equation (5), but instead retrieves products according to the heat bucket range corresponding to the default level, and sorts them using the in-bucket quality score composed of smooth order conversion rate and product freshness.

[0065] In the formula, the online normalized base ranking score of formula (5) is fixed by the original predicted logarithmic value of the model in formula (4) and monotonically normalized by the Sigmoid function, with a value range of [0, 1]. The heat transformation function adopts log normalization: first, the heat of the product funnel is truncated to the 99th percentile of the heat of all candidate products in the current refresh window, and then the ratio of log(1 + truncated heat) to log(1 + the quantile value) is calculated; when the quantile value is 0, the heat transformation function is 0. H_i^f(t) in formula (6b) and P_i(t) in formula (5) are both taken as the heat of the product funnel at the same time and the same version, and the same heat transformation function is used. The function is monotonic and the output range is [0, 1]. The re-ranking reward and penalty coefficients are 0.10, 0 and -0.10 respectively; the re-ranking score calculated by formula (5) is truncated to [0, 1]. The online main path is fixed with the online normalized basic ranking score as the input of equation (5); the existing candidate generation model is only used for recall or candidate generation, and its ranking score is not used as the online normalized basic ranking score, so as to ensure that the training and online inference are consistent.

[0066] The difference in this embodiment is that the online re-ranking module is decoupled from the offline model parameters, and the product popularity, user level and re-ranking coefficient can be updated independently; after the update, the version identifier still ensures that the same version is read in the same request.

[0067] Equations (6), (6a), and (6b) represent the hierarchical gradient equalization loss (used in step P4):

[0068] The future interaction set in Equation (6) comes only from the 7-day label window after the sample anchor point; the 30 days before the sample anchor point is the feature window, and the 7 days after the label window that do not overlap are the validation window. The validation window and online future behavior are not included in the training loss. Taking the sample anchor point as the boundary, the 30 days before the anchor point are used as the feature window, the 7 days after the anchor point are used as the label window, and the 7 days after the label window that do not overlap are used as the validation window; only samples whose label windows have completely ended are used for training. Each positive sample is a negative sample drawn from products of the same product category that have not interacted in the feature window and label window at a ratio of 1:4. The weights of the interacted samples in the high-popularity preference category, the neutral category, and the low-popularity preference category are 1.0, 1.0, and 0.5, respectively, and the weights of the future interaction samples are 0.5, 1.0, and 1.5, respectively.

[0069] In the formula, the user true label vector, the user training product sample set, the training sample weight, the user future interaction product set, the logical Sigmoid function, and the basic regularization coefficient are all defined according to the symbol table; Equation (6) sums up the user and training product samples item by item, and the cross-entropy term includes both positive label terms and negative label terms.

[0070] The difference in this embodiment is that the weights of already interacted samples, the weights of future interacted samples, and the regularization coefficients of product popularity are determined according to user tiers and product funnel popularity, so that the training objective is consistent with user-level popularity preferences.

[0071] The connection between formulas and processes is as follows Figure 3 As shown, Equation (1) is calculated and written into the online feature library in step N1; Equations (2), (2a), and (2b) are calculated in each hypergraph convolution layer in step P2; Equations (3) and (3a) are periodically calculated for non-empty historical users in steps N3 and N4 and written into the online feature library; Equation (4) calculates the original predicted log value of the model in step P3, and the online main path calculates the original predicted log value of the model using the same function in step R1 and monotonically normalizes it using the Sigmoid function to obtain the online normalized basic ranking score; Equation (5) is input in step R3 with the online normalized basic ranking score, and the online reordering machine calculates the reordering score; Equations (6), (6a), and (6b) are used to update the embedding in reverse in step P4. When the value is empty, the online phase switches to step R4 cold start bypass; this branch does not calculate the gear mapping of equation (3) / (3a), but adopts the default gear fixed as neutral gear, and equation (5) does not participate in this branch, so the logical closed loop is not contradictory.

[0072] Process boundary descriptions: Use preset initial values ​​when product popularity is missing; enter bypass when user history is empty or the number of history records is lower than the cold start threshold; continue using the previous complete version when online feature version is incomplete.

[0073] like Figure 2 The diagram shown illustrates the process flow of the present invention, which specifically includes the following stages: Phase 1: Offline training phase (executed by Module 4 according to a preset training cycle) This stage updates the recommendation model based on the behavior logs within the training window, outputting user embeddings, product embeddings, individual bias embeddings, and popularity embeddings with version identifiers, and writing them into the online feature library. This stage includes the following five steps: Step P1 (Constructing User Hyperedge): Module 4 reads the e-commerce interaction data from Module 6 from the previous day, and for each active user, constructs a hyperedge composed of their historical interactive products, sorted by time; the length of this hyperedge is denoted as... This step achieves "aggregating e-commerce behavior sequences by user".

[0074] Step P2 (Generate dynamic perturbations and perform perturbed hypergraph convolution): Module 4 calls equation (2) for each pair of goods within each hyperedge to generate dynamic perturbations based on the interaction time interval and dominant behavior type of the goods pair; after adding the dynamic perturbation elements to the symmetric normalized adjacency submatrix within the hyperedge, it performs convolution aggregation with the goods embedding to produce the hypergraph convolution output of this layer. This step amplifies the impact of recent high-confidence behaviors and suppresses the impact of long-term or low-confidence behaviors.

[0075] Step P3 (Calculation of the original model prediction logarithm): Module 4 calculates the original model prediction logarithm for the user and product pairs in the training samples according to Equation (4), where the heat coupling coefficient is looked up in a table according to the user level; the online main model path adopts the same definition of the original model prediction logarithm, and the original model prediction logarithm is monotonically normalized by the Sigmoid function and used as the online normalized base sorting score of Equation (5).

[0076] Step P4 (Loss Calculation and Parameter Update): Module 4 calculates the loss according to equations (6), (6a), and (6b). The feature window is taken from the 30 days before the sample anchor point, the label window is taken from the 7 days after the anchor point, and the validation window is taken from the subsequent 7 days that do not overlap; training is only conducted using samples whose label windows have been fully completed. The ratio of positive to negative samples is 1:4. The weights of the interacting samples in the high-popularity preference category, neutral category, and low-popularity preference category are 1.0, 1.0, and 0.5, respectively, and the weights of the future interacting samples are 0.5, 1.0, and 1.5, respectively. The basic regularization coefficient is 10^-4, and the popularity regularization gain coefficient is 1.0.

[0077] Step P5 (Version Distribution): Module 4 writes the trained model parameters into the online feature library according to the version identifier; online requests only switch to the new version when the parameters and dependent features of the same version are fully available, otherwise continue to use the previous complete version.

[0078] Phase Two: Near-Line Feature Refresh (Executed by Module Two and Module Three according to their respective update cycles) This phase updates the product funnel popularity, popularity bucket, and user tier without retraining the main model. Product popularity is updated incrementally based on new behaviors; popularity bucket boundaries and user tiers are recalculated according to a preset period. This phase includes the following four steps: Step N1 (funnel heat aggregation): The e-commerce behavior flow of module two consumes module one, and maintains the funnel heat status for each product according to formula (1); the behavior weight satisfies the partial order that the weight at the end of the funnel is higher than the weight at the top of the funnel, and the behavior decay constant satisfies the partial order that the behavior at the end of the funnel is not less than the behavior at the top of the funnel; the specific values ​​in this embodiment are taken according to Table 1 and the aforementioned parameter configuration.

[0079] Step N2 (Discretization of Popularity Buckets): Module 2 discretizes the popularity of the product funnel by performing quantile analysis on the set number of buckets to obtain the popularity bucket number, and writes it into Module 6; the popularity buckets are numbered in ascending order of popularity, and the larger the number, the more popular the product.

[0080] In this embodiment, the product funnel heat status is updated incrementally every minute; the equal-frequency bucket boundaries of all candidate products are recalculated every 15 minutes. After the bucket boundaries change, only users whose bucket numbers in the historical set have changed are recalculated in terms of tier, and the bucket boundaries and user tiers are switched atomically using version numbers.

[0081] Step N3 (User Head Popularity Ratio): Module three-step formula (3) calculates the proportion of products located in the head popularity bucket in the historical interaction set of each e-commerce user.

[0082] Step N4 (User Tier Mapping): Module 3 maps the proportion of top-ranked popular products to user tiers according to formula (3a) and dual thresholds; only users whose historical product popularity buckets have changed or who have generated new behaviors are recalculated, and the updated tiers are written into the online feature library.

[0083] Phase 3: Online Recommendation Request (Executed by Modules 5, 6, and 7) This phase executes whenever an e-commerce recommendation request is triggered, and its goal is to differentiate the ranking candidates and output the Top-ranked candidates within tens of milliseconds. There are five steps in total: In this stage, when a recommendation request is triggered, the online normalized base ranking score of candidate products is adjusted based on popularity at each tier to obtain a re-ranking score; if historical interactions are insufficient, a cold start bypass is switched on, and a recommendation list is output. This stage includes the following five steps: Step R1 (Preparation of candidate and online normalized basic ranking scores): The recommendation model calculates the original predicted logarithmic value of the candidate product according to equation (4), and monotonically normalizes it to an online normalized basic ranking score in the range of [0, 1] using the Sigmoid function; Module 5 reads the user level and product funnel heat consistent with the model parameter version from the online feature library. Existing candidate generation models are only used for candidate generation, and their scores are not included in equation (5).

[0084] Step R2 (Branch Decision): Module 5 reads the length of the user's super-edge. With cold start threshold ;like If the threshold is reached, the system will proceed to the main route for rerouting; otherwise, it will proceed to the cold start bypass route.

[0085] The online normalized base sorting score is fixed and obtained by monotonically normalizing the original predicted logarithmic value of the model in equation (4) using the Sigmoid function; the online model parameters, user level and product funnel heat must be read from compatible versions.

[0086] Step R3 (Main Path Differentiation Reordering): When the number of historical interactive products of a user reaches the cold start threshold, Module 5 corrects the online normalized basic ranking score of each candidate product according to Equation (5) to obtain the reordering ranking score; the reordering reward and punishment coefficients of high popularity preference category, neutral category and low popularity preference category are 0.10, 0 and -0.10 respectively, and the reordering ranking score is truncated to [0,1] and sorted in descending order.

[0087] Step R4 (Cold Start Bypass): The cold start threshold is taken from 5 deduplicated actively interacted products. When historical interactions are empty, a neutral tier is used; when the number of historical interactions is 1 to 4, let the number of top-performing products be N_head, and use Laplace smoothing ratio (N_head+1) / (L_u+2), determining the tier according to two thresholds of 0.35 and 0.65. Candidates for the high-popularity preference tier are selected from buckets 8 to 10, for the neutral tier from buckets 4 to 8, and for the low-popularity preference tier from buckets 1 to 5. Within each bucket, products are sorted in descending order by the sum of 0.7 times the smoothed order conversion rate and 0.3 times the product freshness; the smoothed order conversion rate is calculated by dividing the number of orders placed in the last 7 days by the number of clicks placed by 20, and the product freshness is the exponential decay value of the product's listing duration over a 7-day time constant. After completing inventory, regional availability, category compliance, and SPU deduplication, the top K_rec products are output. When the number of products a user has interacted with in the past is less than 5, module seven is bypassed. Users with no historical data are assigned a neutral tier, while users with 1 to 4 historical interaction products are assigned a tier based on the proportion of Laplace smoothing. Candidates for the three tiers are selected from the 8th to 10th buckets, the 4th to 8th buckets, and the 1st to 5th buckets, respectively. After completing inventory, regional availability, category compliance, and SPU deduplication, the top K_rec products are selected in descending order based on the sum of 0.7 times the smoothed order conversion rate and 0.3 times the product freshness.

[0088] Step R5 (Top- Output): Module 5 or Module 7 will be rearranged in descending order of scores, with the first result taken. The selected items are returned to the e-commerce frontend as part of the final recommendation list.

[0089] Phase 4: Second-level behavior reflow (from Module 1 back to Modules 2 and 3) This stage writes the new behaviors generated by the recommendation results back into the behavioral data pipeline and triggers partial updates of relevant products and users. This stage includes the following two steps: Step F1 (Behavior Write-back): After the e-commerce front-end generates exposure, click, favorite, add to cart or place an order based on the recommendation results, it writes the corresponding five-tuple behavior record into the message queue of Module 1; the write-back latency depends on the end-side reporting and message queue configuration.

[0090] Step F2 (Incremental Update): Module 2 updates the funnel heat of related products every minute for new consumption behaviors; every 15 minutes, it recalculates the full heat bucket boundaries and only recalculates the tiers for affected users. Model parameters are updated daily. The online feature library maintains version numbers for heat, bucket boundaries, user tiers, and model parameters respectively, and only one complete and compatible set of versions is read in a single request.

[0091] The roles of each stage and model mentioned above are shown in Table 2 below: Table 2. Role of Each Stage and Model

[0092] Specific application scenarios are as follows: In the product recommendation section of the e-commerce platform, the original recall and coarse ranking stages remain unchanged; the recommendation model outputs the original predicted logarithmic value of the candidate products according to equation (4), and obtains the online normalized basic ranking score by monotonically normalizing it using the Sigmoid function; Module 5 performs a heat preference-aware re-ranking on the online normalized basic ranking score before outputting the list, and Module 7 provides cold start results when there is insufficient user historical interaction. The product category, user scale, and log scale are determined according to the actual deployment environment.

[0093] Based on chronological order, it can be divided into the following four stages: Phase 1: Module 4 triggers offline training daily at 02:00. It constructs a feature window for the 30 days before the anchor point, a label window for the 7 days after the anchor point, and a subsequent 7-day non-overlapping validation window using the sample anchor point whose most recent validation window has been fully completed. After successful validation, the model parameters are written to Module 6. Atomic switching is performed using a dual-version approach, and the previous version is retained for fallback in case of failure.

[0094] Phase 2: Module 2 adds new consumption behavior and updates the popularity of related products every minute; the full popularity bucket boundary is recalculated every 15 minutes, and Module 3 only recalculates the tier for affected users; the bucket boundary and user tier are atomically released using the same batch version number.

[0095] The third stage (millisecond-level response for each recommendation request): After obtaining the fine-ranked candidates, Module 5 reads the tier and funnel heat from Module 6, and rearranges them according to formula (5) or When switching to module seven cold start bypass, output the Top-K list.

[0096] The third stage: After each recommendation request arrives, Module 5 reads the online normalized basic ranking score of the candidate products of the same version, the user level and the funnel heat of the products; when the main road conditions are met, the re-ranking score is calculated according to formula (5); when the conditions are not met, it switches to the cold start bypass of Module 7 and outputs the recommendation list. Phase 4: New behaviors generated by the recommendation results are written to the message queue of Module 1, triggering partial updates of the popularity of related products and the user tiers.

[0097] After completing a control experiment with the same dataset, the same candidate set, and the same evaluation conditions, this embodiment can be evaluated from the following aspects: Recommendation diversity: Compare the diversity metrics within the list for different user tiers and analyze the impact of popularity correction on users with low popularity preferences; Long-tail exposure: The number of orders placed in the last 30 days is sorted by product in ascending order. Products that are in the bottom 80% of the total number of products are defined as long-tail products. Within a 7-day evaluation window, the proportion of exposure times of long-tail products to the total number of recommended exposure times is calculated and compared with the baseline under the same candidate set and the same recommendation length.

[0098] Offline vs. Online Consistency: Compare the correlation between changes in offline ranking metrics and changes in online click or conversion metrics; Strategy update efficiency: Record the effective latency when only popularity, tier, or rearrangement coefficients are updated, and distinguish it from the time required to retrain the main model; Cold start performance: Comparing the click, conversion, and coverage metrics predicted by the cold start bypass and the main model directly in scenarios with insufficient user historical interaction; Real-time updates and comprehensive evaluation: Record the actual latency of edge behavior write-back, 1-minute product popularity increment, 15-minute bucket boundary recalculation, affected user tier updates, and daily model updates. Offline experiments use the same data segmentation and candidate set, reporting Recall@K, NDCG@K, list diversity, and long-tail exposure ratio; if conducting online experiments, use user-level random traffic splitting, fixed traffic ratio, and an observation window of no less than one complete business cycle, reporting click-through rate, conversion rate, and confidence interval, without pre-setting an increase ratio when there is a lack of actual test data.

[0099] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0100] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0101] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0102] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0103] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for bias-free recommendation in e-commerce based on funnel heat and sensitivity stratification, characterized in that, Includes the following steps: S1. Collect various types of user behavior data in e-commerce scenarios, aggregate the data within a preset statistical window according to behavior conversion weight and time decay rules to obtain product funnel heat, and discretize the product funnel heat into heat buckets according to preset bucketing rules. S2. For users whose historical interaction product set is not empty, calculate the proportion of top-ranked popular products based on the popularity bucket corresponding to their historical interaction products, and map the user to the corresponding popularity preference level according to the proportion; for users whose historical interaction product set is empty, a preset default level is used. The specific method for generating dynamic perturbations and injecting them into the hypergraph convolution process is as follows: S31, For users Historical interactive product collection Any two items and Calculate the interaction time between the two. And determine the dominant behavior type of their interaction. ; Among them, the priority of behavior is arranged from high to low in the order of placing an order, adding to cart, adding to favorites, clicking, and exposure; when a product involves multiple behaviors in the statistics window, the behavior with the highest weight is determined as the dominant behavior, and when the weights are the same, the behavior that occurs later is selected as the dominant behavior. S32, using a multi-scale time decay function With behavioral confidence function Jointly modulate self-attention related terms to generate dynamic perturbations of the product pair. , Calculated using the following formula: In the formula: The number of scales for multi-scale time decay. For the first Weights for each time scale For the first The decay constant corresponding to each time scale It is a natural exponential function. For product interaction time interval; S33, Dynamically disturbed elements Adding the symmetric normalized adjacency submatrix of the hypergraph convolution to obtain the perturbed adjacency matrix, and then performing the perturbed hypergraph convolution using the following formula, yields the first... Layer embedding : In the formula, It is a symmetric normalized adjacency matrix. It is a non-linear activation function. For the first Layer embedding The intermediate embedding matrix obtained after perturbation propagation and nonlinear transformation The attention value projection matrix, For multi-head attention aggregation operators; S3. Construct a user hyperedge based on the user's historical interaction with products, generate dynamic perturbations based on the interaction time interval and behavior type between products within the hyperedge, add the dynamic perturbations to the structure aggregation process of hypergraph convolution, train the recommendation model using hypergraph convolution with dynamic perturbations and a unified prediction function, and send the trained model parameters to the online feature library. S4. In response to an online recommendation request, read the requesting user's popularity preference level, the funnel heat of candidate products, and the online parameters of the recommendation model. Calculate the original model prediction logarithm of the candidate products using the online parameters. Monotonically normalize the original model prediction logarithm using the Sigmoid function to obtain an online normalized basic ranking score with a value range of [0, 1]. Perform differential correction on the online normalized basic ranking score according to the popularity preference level and the funnel heat of candidate products. When the number of historically interacted products of the requesting user is lower than a preset cold start threshold, switch to cold start bypass to generate recommendation results. S5. Collect real-time user behavior data generated in response to the recommendation results, feed the real-time user behavior data back to the behavior data path, and perform incremental updates on the product funnel heat and user heat preference levels related to the real-time user behavior data for subsequent online recommendation requests.

2. The e-commerce recommendation method based on funnel heat and sensitivity stratification according to claim 1, characterized in that, The specific method for aggregating product funnel popularity according to behavioral conversion weights and time decay rules in S1 is as follows: S11. Define the set of e-commerce user behavior types. , for set Each type of behavior in Preset the corresponding funnel weight With time decay constant Among them, the behavior types are sorted by conversion priority, and the behaviors at the end of the funnel are... Values ​​greater than the behavior at the top of the funnel Funnel end behavior Values ​​greater than the behavior at the top of the funnel ; S12. For any product At any moment Product funnel heat Calculated using the following formula: In the formula, For goods exist Behavioral types occurring at any time Number of times, It is a natural exponential function.

3. The e-commerce recommendation method based on funnel heat and sensitivity stratification according to claim 1, characterized in that, The specific method for calculating the user's head heat ratio and mapping the user to the corresponding sensitivity level in S2 is as follows: S21, Preset head heat bucket threshold Total number of heat barrels Number the heat barrels in ascending order of heat, with the numbers greater than... The heat tank is the head heat tank; S22. For users whose historical interaction item set is not empty. Calculate the percentage of products belonging to the top trending categories in the historical product interaction set. , Calculated using the following formula: In the formula, For users A collection of historical interactive products. This refers to the heat bucket number corresponding to product i. A counting operator for the number of elements in a set; S23, Preset conservative to neutral threshold From neutral to conformity threshold , will users Mapped to sensitivity level The mapping rule is: when hour For conservative purposes, when hour For neutral settings, when hour To conform to the crowd; S24. For users whose historical interactive product set is empty. The default sensitivity level is directly assigned, and the default sensitivity level is set to neutral.

4. The e-commerce recommendation method based on funnel heat and sensitivity stratification according to claim 1, characterized in that, The unified prediction function and model training method in S3 include: S34. Preset thermal coupling coefficients for different sensitivity levels. Among them, the heat coupling coefficient of the crowd Positive value, neutral range The thermal coupling coefficient is zero or within a preset neutral range, representing the conservative setting. The value is within the preset negative correction interval; S35, the same prediction function is used in the forward training phase and the online main model path to calculate the user's prediction. For goods The original prediction logarithmic value of the model The original logarithmic prediction value of the model is calculated using the following formula: In the formula, For users The representation vector, For goods The representation vector, For vector dot product operators, For users Popularity embedding, For goods Popularity embedding; S36. The model is trained using a hierarchical gradient equalization loss function, which is calculated using the following formula: In the formula, For users The training product sample set, For sample weights, For users For goods The true label, It is the logical sigmoid function. These are intermediate predicted values ​​with individual bias. For users A collection of future interactive products For goods The heat correlation regularization coefficient, For users A collection of historical interactive products. The L2 norm square operator for vectors, The basic regularization coefficient, For users Individual bias embedding, This is the thermal regularization gain coefficient. The normalization transformation function for heat is... For goods exist The temperature of the funnel at any given moment. It is a commodity i The product embedding vector after bias correction, For users Sensitivity level The corresponding weight coefficients of the already interacted samples, For users Sensitivity level The corresponding future interaction sample weight coefficients, It is a commodity At any moment The temperature of the funnel.

5. The e-commerce recommendation method based on funnel heat and sensitivity stratification according to claim 4, characterized in that, S4 specifically includes the following steps: S41. Preset the rearrangement reward / penalty coefficients for different sensitivity levels. Among them, the crowd Positive value, neutral range Values ​​that are zero or fall within a preset neutral range, conservative settings. The value is within the preset negative correction interval; S42, For online request candidate product set The original predicted logarithmic values ​​of the model obtained in step S35 are monotonically normalized using the Sigmoid function to obtain the online normalized basic sorting score with a value range of [0, 1]. Based on the online normalized basic ranking score, for each candidate product... i Calculate rearrangement score The calculation formula is as follows: In the formula, Candidate Products Online normalized base sorting score, The normalization transformation function for heat is... For goods exist The temperature of the funnel at any given moment; S43. Sort the candidate products in descending order according to the rearrangement sorting, and select the top... The recommended items are output as the rearranged recommendation results. This is the preset recommended output length; S44, Preset cold start threshold Requesting users Length of historical interactive product collection : or When the value is empty, a cold start bypass is triggered. For users A collection of historical interactive products; S45, Based on the requesting user The default sensitivity level is used to retrieve candidate products from the corresponding popularity bucket and select the top ones. Each item is output as a cold start recommendation result.

6. The e-commerce recommendation method based on funnel heat and sensitivity stratification according to claim 1, characterized in that, The specific method for feeding back the real-time user behavior data generated by online recommendations in S5 is as follows: S51. Collect the exposure, click, favorite, add to cart and order behavior logs generated by the e-commerce terminal for the recommendation results, and record each behavior log as a five-tuple containing user identifier, product identifier, behavior type, timestamp and context identifier; S52. Write the behavior log to the message queue, perform funnel heat increment aggregation for products with new behaviors, and recalculate the heat preference level for users with new behaviors; write the updated results to the online feature library for subsequent online recommendation requests to call.

7. An e-commerce bias-free recommendation system based on the method of any one of claims 1-6, characterized in that, It includes a memory, a processor, and a program stored in the memory and executed by the processor; when the processor executes the program, it implements e-commerce bias-reduction recommendations through the following functional modules: The e-commerce behavior flow collection module is used to collect various types of user behavior data and write behavior records composed of user identifier, product identifier, behavior type, timestamp and context identifier into the message queue. The funnel heat calculation module is used to aggregate the product funnel heat according to the behavior conversion weight and time decay rules within a preset statistical window, generate heat buckets according to the preset bucketing rules, and write the product funnel heat and heat buckets into the online feature library. The user popularity preference layering module is used to read the user's historical interaction with products and their popularity buckets, calculate the proportion of the user's top popular products, map non-empty historical users to corresponding popularity preference levels, and provide preset default levels for users with empty historical data. The offline model training module is used to construct user hyperedges based on users' historical interactions with products, generate dynamic perturbations according to the interaction time interval and behavior type between products, train the recommendation model using hypergraph convolution with dynamic perturbations and a unified prediction function, and write the model parameters into the online feature library. The online re-ranking module is used to respond to online recommendation requests, read user popularity preference levels, candidate product funnel popularity, and model parameters, calculate the original model prediction logarithm of candidate products using the model parameters, monotonically normalize the original model prediction logarithm of the model using the Sigmoid function to obtain an online normalized basic ranking score, and perform differential correction on the online normalized basic ranking score according to the user popularity preference level and candidate product funnel popularity to obtain the re-ranking score and output the recommendation result. The cold start bypass module is used to bypass the requesting user when the number of historically interacted products is less than 5: when the historical interactions are empty, a neutral level is used; when the number of historical interactions is 1 to 4, the level is determined by the proportion of head popularity after Laplace smoothing; the high popularity preference level, neutral level, and low popularity preference level retrieve products from the 8th to 10th bucket, the 4th to 8th bucket, and the 1st to 5th bucket, respectively. After being filtered for availability, category compliance, SPU deduplication, and in-bucket quality sorting, the cold start recommendation results are output. An online feature library is used to store product funnel heat, heat bucket, user heat preference level and model parameters according to version identifier, and provides read and write interfaces to the funnel heat calculation module, user heat preference layering module, offline model training module, online reordering module and cold start bypass module; The real-time feedback update module is used to consume the new behavior records written by the e-commerce behavior flow collection module, incrementally update the funnel popularity of related products and the popularity preference levels of related users, and write the update results into the online feature library.

Images