Product recommendation system and method based on cross-border e-commerce platform

Abstract

The invention relates to the technical field of cross-border e-commerce intelligent recommendation, in particular to a product recommendation system and method based on a cross-border e-commerce platform, and the method comprises the steps: obtaining a historical behavior track of a target user, and processing the historical behavior track to form a pre-selected commodity pool, a dynamic environment variable factor is obtained by integrating the real-time exchange rate fluctuation value of the destination country, customs clearance timeliness data and international logistics trunk transportation cost weighting, commodity information, the user activeness level and the factor are input into a multi-modal reasoning calculation engine, joint probability distribution is constructed, and an instant recommendation score is calculated; and then starting a state inversion mechanism by taking the score as an input, reversely deducing an unexplicit potential demand state of the user, and fusing the score and the demand state to generate an enhanced recommendation list containing explicit recommendation reasons and implicit demand mining. The method is adaptive to a cross-border scene dynamic environment, reverse mining of potential demands of the user is realized, and the recommendation list composition dimension is optimized.

Description

Technical Field

[0001] This invention relates to the field of intelligent recommendation technology in cross-border e-commerce, and in particular to a product recommendation system and method based on a cross-border e-commerce platform. Background Technology

[0002] Currently, product recommendation technologies on cross-border e-commerce platforms largely rely on processing users' historical behavior patterns to create a pre-selected set of products. This is combined with user activity levels from basic user profiles to calculate the matching between products and users. This type of technology calculates recommendation scores by building a user-product association model, and is a common approach in the industry. Some technologies incorporate explicit user behavior data to optimize recommendation results, but all rely on inherent characteristics of both the user and product sides as the core calculation basis, failing to develop differentiated calculation logic specifically for the unique scenarios of cross-border transactions.

[0003] Existing cross-border e-commerce product recommendation technologies fail to incorporate cross-border-specific parameters such as real-time exchange rate fluctuations in the target user's destination country, local customs clearance time, and international logistics trunk line transportation costs into their weighted calculation systems. This disconnects the recommendation score calculation process from the real-time external environment of cross-border transactions, resulting in calculation results that cannot adapt to the dynamic changes in cross-border scenarios. Such technologies can only derive recommendation scores forward from existing data, failing to analyze user needs using generated recommendation scores. Consequently, recommendation results only reflect explicit user behavioral preferences and cannot uncover unexpressed latent needs. This invention aims to address the problem of lacking weighted integration of cross-border-specific dynamic environmental variables into recommendation calculations, while also resolving the inability to infer unexpressed latent user needs using recommendation scores as input. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a product recommendation system and method based on a cross-border e-commerce platform.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a product recommendation method based on a cross-border e-commerce platform, comprising:

[0006] Obtain the historical behavior trajectory of the target user, process the historical behavior trajectory, and form a pre-selected product pool;

[0007] A dynamic environmental variable factor is introduced, which is derived by comprehensively weighting the real-time exchange rate fluctuation value of the target user's destination country, the customs clearance time data of the local customs, and the trunk transportation cost of international logistics.

[0008] The product information in the pre-selected product pool, the user activity level in the basic user profile, and the environmental variable factors are jointly input into the multimodal inference computing engine. The multimodal inference computing engine calculates the real-time recommendation score of each candidate product after considering environmental factors by constructing a joint probability distribution of user intent and product features.

[0009] After calculating the instant recommendation score, the state inversion mechanism is activated. The state inversion mechanism takes the currently calculated recommendation score as input and reverses the unexpressed potential demand state of the target user under the current recommendation score.

[0010] The calculated instant recommendation score is fused with the potential demand state obtained through the state inversion mechanism to generate an enhanced recommendation list that includes explicit recommendation reasons and implicit demand mining.

[0011] As a further aspect of the present invention, the historical behavior trajectory of the target user is obtained, and the historical behavior trajectory is processed to form a pre-selected product pool, including:

[0012] The historical behavior trajectory consists of the duration of browsing product detail pages, the set of product icons added to the shopping cart, and the logistics status of paid orders;

[0013] The historical behavior trajectory is anonymized and standardized to generate a basic user profile containing user consumption preference vectors and user activity levels.

[0014] From the global product database of the cross-border e-commerce platform, retrieve several candidate products with the highest similarity to the consumption preference vector in the basic user profile to form a pre-selected product pool;

[0015] The historical behavior trajectories are anonymized and standardized to generate a basic user profile containing user consumption preference vectors and user activity levels, specifically including:

[0016] The set of product identifiers in the historical behavior trajectory is deduplicated, and the frequency of occurrence under each product category is counted. The frequency is then transformed into a high-dimensional sparse category preference vector.

[0017] The dwell time on the product details page in the historical behavior trajectory is segmented. Products with dwell time exceeding a set threshold are regarded as high interest points, and their corresponding price range and brand characteristics are extracted and added to the high-dimensional sparse category preference vector to form a dense user consumption preference vector.

[0018] The logistics status of paid orders in the historical behavior trajectory is analyzed to calculate the average cycle from order placement to receipt. Based on the deviation of the average cycle from the platform average level, the user activity level is determined. The user activity level is used to distinguish between impulsive consumption and planned consumption.

[0019] As a further aspect of the present invention, the introduction of dynamic environmental variable factors is derived by a weighted average of the real-time exchange rate fluctuations in the target user's destination country, local customs clearance timeliness data, and international trunk transportation costs, including:

[0020] By pulling the real-time exchange rate of the target user's country currency relative to the platform's settlement currency through an external data interface, the exchange rate volatility over the past seven days is calculated and mapped to a risk coefficient.

[0021] Connect to the data interface of cross-border logistics service providers to obtain the current average customs clearance time in the target user's region, and convert the average customs clearance time into a time delay coefficient.

[0022] Based on the weight and volume of goods in the pre-selected goods pool and the current international route fuel surcharge, the estimated trunk transportation cost is calculated and converted into a price fluctuation coefficient.

[0023] The risk coefficient, the time delay coefficient, and the price fluctuation coefficient are each assigned a weight based on business strategy. The risk coefficient, the time delay coefficient, and the price fluctuation coefficient are then weighted and summed to obtain the environmental variable factor. The environmental variable factor is used to adjust the basic price attractiveness of the product.

[0024] As a further aspect of the present invention, the multimodal inference computing engine calculates the instant recommendation score for each candidate product after considering environmental factors by constructing a joint probability distribution of user intent and product features, including:

[0025] The user consumption preference vector in the basic user profile is concatenated with the user activity level to form a feature representation on the user side;

[0026] For each product in the pre-selected product pool, extract its price, rating stars, listing time, and place of origin information to form a feature representation on the product side;

[0027] The user-side feature representation, the product-side feature representation, and the environmental variable factors are all input into a conditional random field model. The conditional random field model learns the probability that a user will choose a specific product under given environmental conditions.

[0028] The forward algorithm of the conditional random field model is used to calculate the selection probability of all products in the pre-selected product pool. The top few values ​​with the highest selection probabilities are used as the instant recommendation scores and stored in a temporary score matrix.

[0029] As a further aspect of the present invention, the state inversion mechanism takes the currently calculated recommendation score as input and reversely infers the unexpressed potential demand state of the target user under the current recommendation score, including:

[0030] From the temporary score matrix, extract products with recommendation scores higher than the average but which have not yet been clicked by the target user, and define these products as the inversion trigger set;

[0031] For each product in the inversion trigger set, analyze the degree of difference between it and the user consumption preference vector in the basic user profile, and identify the abnormal features that cause the product to be recommended but not clicked;

[0032] All identified abnormal features are summarized, and combined with the cultural characteristics and holiday information of the target user's country of origin, a potential demand state that explains the abnormal features is deduced. The potential demand state includes price-sensitive, quality-seeking, or specific scenario-based demand.

[0033] As a further aspect of the present invention, the step of fusing the calculated instant recommendation score with the potential demand state obtained through the state inversion mechanism includes:

[0034] Each potential demand state obtained through the state inversion mechanism is assigned a state weight, the value of which depends on the number and degree of anomalous features explained by the state.

[0035] Iterate through each product in the temporary score matrix and check whether its product features match the potential demand state. If they match, multiply the product's instant recommendation score by the state weight of the matching potential demand state to enhance the score.

[0036] For products that do not match any potential demand status, their original instant recommendation scores remain unchanged;

[0037] The scores of all products, whether enhanced or unchanged, are normalized, and then sorted from highest to lowest according to the normalized scores to generate the initial draft of the enhanced recommendation list.

[0038] As a further aspect of the present invention, the method further includes a compliance filtering step for the enhanced recommendation list:

[0039] From the enhanced recommendation list, the country of origin information of all products is extracted and compared with the trade compliance access list of the target user's country, excluding products whose country of origin is not on the trade compliance access list;

[0040] The prices of products in the enhanced recommendation list are recalculated by adding the original price of the products to the estimated cross-border logistics costs and tariffs to obtain the total landed cost, and filtering out products whose total landed cost exceeds twice the historical average order value of the target user.

[0041] The enhanced recommendation list, after compliance and price filtering, is then matched one last time with the user activity level in the basic user profile. For users with low activity levels, the length of the recommendation list is further reduced, retaining only the top three ranked products.

[0042] As a further aspect of the present invention, the forward algorithm of the conditional random field model is used to calculate the selection probability of all items in the pre-selected item pool, including:

[0043] The user consumption preference vector is extracted from the feature representation on the user side. The user consumption preference vector is generated by encoding the product category preference, price range preference and brand preference that frequently appear in the user's historical behavior trajectory.

[0044] Extract product feature vectors from the feature representation of the product side. The product feature vectors are generated by encoding the product's price, average rating star, shelf duration, country of origin, and logistics method.

[0045] The exchange rate risk coefficient, time delay coefficient, and price fluctuation coefficient are extracted from the environmental variable factors, and then combined into an environmental feature vector.

[0046] The user consumption preference vector, the user activity level in the user-side feature representation, the product feature vector, and the environmental feature vector are concatenated to form an input observation sequence.

[0047] The input observation sequence is input into a trained conditional random field model, which internally defines the transition weights and state feature functions between the product selection state and the aforementioned features;

[0048] The conditional random field model uses its forward algorithm to recursively calculate the probability value of a user making a final selection of each candidate product in the pre-selected product pool under the given current observation sequence. The probability value is the selection probability of each candidate product, and the selection probabilities of all candidate products constitute a selection probability distribution list.

[0049] As a further aspect of the present invention, for each item in the inversion trigger set, the analysis of its difference from the user consumption preference vector in the basic user profile, and the identification of abnormal features that cause the item to be recommended but not clicked, include:

[0050] Obtain a normalized user consumption preference vector from the basic user profile. The user consumption preference vector has specific preference values ​​in multiple product feature dimensions.

[0051] Extract the corresponding product feature vector from the target product in the inversion trigger set, and map the product feature vector to multiple product feature dimensions that are the same as the user consumption preference vector to obtain the mapped target product feature vector;

[0052] Calculate the absolute difference between the target product feature vector and the user consumption preference vector in each dimension, and obtain the difference value in multiple dimensions.

[0053] Select the items with the largest differences among multiple dimensions and mark the corresponding product feature dimensions as candidate abnormal features;

[0054] The specific values ​​of the target product feature vector on the labeled product feature dimension are analyzed. If the specific values ​​exceed the set threshold of the user's historical preference range, the product feature dimension is determined to be an abnormal feature that causes the product to be recommended but not clicked.

[0055] All the identified abnormal features of the target product, along with the high recommendation score calculated in the multimodal reasoning engine, are recorded to form the inversion analysis result entries of the target product.

[0056] As a further aspect of the present invention, the present invention also includes a product recommendation system based on a cross-border e-commerce platform. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the product recommendation method based on a cross-border e-commerce platform as described above.

[0057] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0058] The dynamic environmental variable factor is derived by comprehensively weighting the real-time exchange rate fluctuations of the target user's destination country, the customs clearance timeliness data of the local customs, and the trunk transportation costs of international logistics. The product information of the pre-selected product pool, the user activity level of the basic user profile, and this dynamic environmental variable factor are input together into the multimodal inference calculation engine. The real-time recommendation score is calculated by constructing a joint probability distribution of user intent and product features. The calculation dimension of the recommendation score is aligned with the specific attributes of cross-border e-commerce transactions. The real-time recommendation score corresponds to the real-time external constraints in the cross-border transaction scenario. The calculation result of the recommendation score can reflect the real-time environmental changes in the cross-border transaction process. The calculation logic is adapted to the actual operation scenario of cross-border e-commerce.

[0059] Using the calculated instant recommendation score as input data for the state inversion mechanism, the mechanism reversely infers the potential needs of the target user that are not explicitly expressed. After fusing the instant recommendation score with the potential needs, an enhanced recommendation list is generated that includes explicit recommendation reasons and implicit needs mining. The logic for generating recommendation results breaks through the single path of conventional forward calculation. The recommendation list can carry the needs information that users have not actively presented. The content of the recommendation list covers both the explicit behavioral association and implicit needs mining dimensions. The presentation format of the recommendation results breaks out of the single content framework of conventional recommendation lists. Attached Figure Description

[0060] Figure 1 This is a flowchart of a product recommendation method based on a cross-border e-commerce platform according to the present invention;

[0061] Figure 2 A flowchart for the formation of the pre-selected product pool and the generation of basic user profiles;

[0062] Figure 3 A flowchart for real-time recommendation score calculation in a multimodal reasoning computation engine;

[0063] Figure 4 Frequency graph of abnormal features for different potential demand states;

[0064] Figure 5 A bar chart showing the environmental variables of cross-border e-commerce. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0066] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0067] See Figure 1 The system acquires the target user's historical behavioral trajectory and processes this information to form an initial pre-selected product pool. During this process, the system dynamically incorporates an environmental variable factor calculated by weighting the real-time exchange rate fluctuations of the target user's destination country, local customs clearance timeliness data, and international trunk transportation costs. The product information in the pre-selected product pool, the user activity level within the basic user profile analyzed from historical behavior, and the aforementioned environmental variable factor are all input into a multimodal inference engine. This engine calculates the instant recommendation score for each candidate product in the pre-selected product pool under the influence of comprehensive environmental factors by constructing a joint probability distribution model between user intent and product features. After calculation, the system initiates a state inversion mechanism. This mechanism uses the currently calculated instant recommendation score as input to inversely infer the potential unexpressed needs of the target user under this score distribution. The instant recommendation score output by the multimodal inference engine and the potential need state inferred by the state inversion mechanism are then fused to generate an enhanced recommendation list that simultaneously contains explicit recommendation reasons and implicit need insights.

[0068] In one embodiment of the present invention, the historical behavior trajectory of the target user is obtained, wherein the historical behavior trajectory specifically consists of the dwell time on the product details page, the set of product identifiers added to the shopping cart, and the logistics status data of paid orders. See also... Figure 2The historical behavioral data is anonymized and standardized to generate a structured basic user profile, which includes a user consumption preference vector and user activity level. When processing the product identifier set, deduplication is performed, and the frequency of occurrence under each product category is counted, transforming the statistical frequency into a high-dimensional sparse category preference vector. Simultaneously, browsing dwell time is analyzed, and products with dwell time exceeding a preset threshold are marked as high-interest points. The price range and brand characteristics corresponding to these high-interest point products are extracted and integrated into the aforementioned sparse category preference vector, thus forming a denser and more comprehensive user consumption preference vector. On the other hand, by analyzing the logistics status of paid orders, the average cycle from order placement to final receipt is calculated. Based on the deviation of this average cycle from the overall average level of the cross-border e-commerce platform, a user activity level is determined to distinguish between impulsive and planned consumption. After completing the basic user profile construction, several candidate products with the highest similarity to the user consumption preference vector in the profile are retrieved from the global product database of the cross-border e-commerce platform. These candidate products constitute the pre-selected product pool required for subsequent steps.

[0069] In practical implementation, the target user's historical behavior trajectory acquired by the system contains three types of structured data: the duration of browsing product detail pages, the set of product identifiers added to the shopping cart, and the logistics status of paid orders. Desensitizing and standardizing this raw data is a prerequisite for generating a basic user profile. The processing of the product identifier set includes deduplication and frequency statistics. The system iterates through each product identifier in the set, extracts its product category, and accumulates the frequency of each category, forming a frequency statistics dictionary indexed by category. This process transforms the raw behavioral data into quantifiable preference signals. In some embodiments, the method of converting product category frequencies into a high-dimensional sparse category preference vector involves a mapping and assignment process. The system predefines a high-dimensional vector containing all product categories, with each dimension corresponding to a product category. The calculated frequency of each category is normalized and assigned to the corresponding dimension vector. Most category dimensions in the high-dimensional sparse category preference vector that are unrelated to user behavior have values ​​of zero.

[0070] In specific implementation, the processing of browsing dwell time aims to identify high points of interest. The system sets a fixed threshold for the dwell time on the product details page. When the dwell time exceeds the fixed threshold, the corresponding product is marked as a high-interest product. The system extracts the price range and brand identifier from the high-interest products. Optionally, the logic for supplementing the high-dimensional sparse category preference vector is to add the price range and brand characteristics of the high-interest products as new feature dimensions. For the price range feature, the system maps the product price to discrete ranges such as "low price", "medium price", or "high price"; for the brand characteristic, the brand identifier is one-hot encoded. These newly generated feature vectors are concatenated with the original category preference vector to form a dense user consumption preference vector containing multi-dimensional information such as category, price range, and brand. In some embodiments, the analysis of the logistics status of paid orders is used to calculate the user activity level. The system parses the order placement timestamp and the final receipt timestamp in the order data, calculates the difference between the two to obtain the logistics cycle of a single order, and calculates the average cycle by averaging the logistics cycles of all completed orders. The formula for calculating the average cycle is expressed as:

[0071]

[0072] Where: symbol Indicates the average period, symbol Indicates the total number of completed orders, symbol This represents the timestamp of the i-th order upon receipt, symbol [symbol missing]. This represents the order timestamp of the i-th order. It's understandable that determining user activity levels requires comparing the average cycle with the platform's average level. The platform's average level refers to the historical average of the logistics cycle for all user orders on the cross-border e-commerce platform. The deviation between the average cycle and the platform's average level is calculated, and based on a preset deviation threshold, user activity levels are categorized into high activity levels (representing impulsive consumption) or low activity levels (representing planned consumption). In practice, the step of forming a pre-selected product pool involves similarity retrieval based on the generated user consumption preference vector. The system accesses the cross-border e-commerce platform's global product database. Each product in the database has its own feature vector representation, which resides in the same feature space as the user consumption preference vector. The system calculates the cosine similarity between the user consumption preference vector and the feature vector of each product in the global product database, selects several candidate products with the highest similarity scores, and outputs the set of these candidate products as the pre-selected product pool for subsequent processes.

[0073] In one embodiment of the present invention, the process of introducing dynamic environmental variable factors involves the integration and calculation of multiple external data sources. By calling an external financial data interface, the real-time exchange rate of the target user's destination country's currency relative to the platform's settlement currency is retrieved, and based on this, the exchange rate volatility over the past seven days is calculated, mapping this volatility to a quantified risk coefficient. Simultaneously, a data interface provided by a cross-border logistics service provider is accessed to obtain the current average customs clearance time data for the target user's region, converting this time data into a timeliness delay coefficient. Furthermore, based on the weight and volume parameters of specific goods in the pre-selected commodity pool, combined with the current fuel surcharge standards for major international routes, the estimated trunk transportation cost of the commodity is calculated, and this cost is converted into a price fluctuation coefficient. Weights are assigned to the risk coefficient, timeliness delay coefficient, and price fluctuation coefficient according to preset weights based on the business strategy. By performing a weighted summation of these three coefficients, the environmental variable factors used to adjust the basic price attractiveness of the commodity are finally obtained.

[0074] In its implementation, a product recommendation method based on a cross-border e-commerce platform involves the calculation and introduction of dynamic environmental variable factors. These factors are derived by weighting the real-time exchange rate fluctuations of the target user's destination country, local customs clearance timeliness data, and international logistics trunk transportation costs. The core of this method lies in integrating multi-source external data and quantifying it into a unified correction coefficient. In practice, real-time exchange rate fluctuations are obtained to calculate the risk coefficient. The system periodically retrieves real-time exchange rate quotes between the target user's destination country's currency and the cross-border e-commerce platform's settlement currency through a pre-defined external financial data application interface, recording and maintaining a historical sequence of exchange rates for the past seven calendar days. Calculating the exchange rate volatility over the past seven days involves analyzing this historical sequence. Volatility reflects the degree of exchange rate instability in the short term, and its calculation formula is expressed as:

[0075]

[0076] Where: symbol This represents the calculated exchange rate volatility over the past seven days, with the symbol [symbol missing]. This represents the exchange rate value on day d in the historical sequence, with the symbol... This represents the arithmetic mean of the exchange rate over the past seven days. It's understandable that mapping volatility to a risk coefficient requires a pre-defined mapping function. This function could be a piecewise linear function that maps continuous volatility... The numerical input is mapped to a continuous risk coefficient between 0 and 1. The higher the volatility, the greater the risk coefficient. In some embodiments, to obtain customs clearance timeliness data to calculate the timeliness delay coefficient, the system connects to one or more application programming interfaces (APIs) provided by cross-border logistics service providers, requests the specific region or port code of the destination country where the target user is located from the API, and the logistics service provider's API returns the average customs clearance time in that region within the current time window. The average customs clearance time is a value in hours. Optionally, converting the average customs clearance time into a timeliness delay coefficient involves a standardization process. The system presets a standard customs clearance time benchmark value, and the timeliness delay coefficient... The calculation method is based on the actual average customs clearance time. Divide by the standard customs clearance time benchmark. ,Right now Delay coefficient A value greater than 1 indicates a customs clearance delay, while a value less than 1 indicates that customs clearance efficiency is higher than the benchmark. In practice, the system calculates international trunk line transportation costs to derive a price fluctuation coefficient. Based on the physical attributes of each specific commodity in the pre-selected commodity pool, including the commodity's weight and outer packaging volume, and combined with the commodity's destination country and region code, the system queries the currently valid international route freight rate table. The freight rate table includes basic freight rates and fuel surcharge rates for different weight and volume segments. The system calculates the applicable basic freight rate based on the commodity's weight and volume, and then adds the fuel surcharge to obtain the estimated trunk line transportation cost for that commodity. In some embodiments, converting the estimated trunk transportation cost into a price fluctuation factor requires a calculation process related to the intrinsic value of the commodity. The system obtains the commodity's base selling price from its characteristics. Price fluctuation coefficient The calculation method is to estimate the trunk transportation cost. Divide by the base price of the product ,Right now This ratio reflects the relative proportion of logistics costs in the total cost of goods. In practice, a comprehensive weighted average is used to derive the final environmental variable factors, with the system representing a risk coefficient. Delay coefficient With price fluctuation coefficient Weight values ​​based on business strategies were configured separately, and these weight values ​​are denoted as follows: , and And satisfy Conditions. Environmental variable factors. The formula is derived from the weighted summation formula: Calculated environmental variable factors It is a dimensionless numerical value that is subsequently called by the multimodal inference engine to adjust the basic price attractiveness of the product to the target user from the perspectives of cost and timeliness risk.

[0077] In one embodiment of the present invention, the core of the multimodal inference computing engine for calculating real-time recommendation scores lies in constructing a conditional probability model. See also... Figure 3 The system concatenates and merges the user consumption preference vector from the basic user profile with the user activity level to form a unified user-side feature representation. For each product in the pre-selected product pool, its price, rating, listing time, and place of origin are extracted and encoded as a product-side feature representation. The user-side feature representation, product-side feature representation, and environmental variable factors are combined as an input observation sequence for a conditional random field (CRF) model. This CRF model has been pre-trained and internally defines the transition weights and state feature functions between the state of product selection and user features, product features, and environmental features. During computation, the user consumption preference vector generated from the encoding of the user's historical behavior is extracted from the user-side feature representation; the product feature vector generated from the encoding of product attributes is extracted from the product-side feature representation; and the real-time exchange rate risk coefficient, time delay coefficient, and price fluctuation coefficient are extracted from the environmental variable factors and merged into an environmental feature vector. The model uses its forward algorithm to recursively calculate the probability value of a user's final selection behavior for each candidate product in the pre-selected product pool under the given current input observation sequence. This probability value is the selection probability of each product, and the selection probabilities of all products constitute a selection probability distribution list. The engine extracts the top few values ​​with the highest probabilities from the list and uses them as the instant recommendation scores for the corresponding products, storing them in a temporary score matrix.

[0078] In practical implementation, the preparation of feature representation involves the concatenation of multi-source data. From the basic user profile, the generated user consumption preference vector and user activity level are extracted. The user consumption preference vector is a high-dimensional numerical vector generated by encoding frequently occurring product category preferences, price range preferences, and brand preferences from the user's historical behavior trajectory. The user activity level is a discrete label used to distinguish consumption types. The user consumption preference vector and user activity level are concatenated to form a comprehensive user-side feature representation. In some embodiments, the product-side feature representation originates from the attribute extraction of products in a pre-selected product pool. For each candidate product in the pre-selected product pool, the system extracts the product's price value, average rating star rating, product's shelf life in days, product's country of origin code, and the product's supported logistics method code. These attribute information are standardized and normalized, and encoded into a unified product feature vector. In practical implementation, the construction of the environmental feature vector is based on dynamic environmental variable factors. The system extracts the real-time exchange rate risk coefficient, the real-time time delay coefficient, and the real-time price fluctuation coefficient from the calculation results contained in the environmental variable factors. The three scalar values ​​of the exchange rate risk coefficient, the time delay coefficient, and the price fluctuation coefficient are sequentially concatenated and merged to form a three-dimensional environmental feature vector. It can be understood that forming the input observation sequence requires integrating all the above feature representations. The comprehensive user-side feature representation, product feature vector, and environmental feature vector are concatenated along the feature dimension to form a complete input observation sequence for inference in the conditional random field model. Each candidate product corresponds to a specific input observation sequence composed of the user-side feature representation, the product's own product feature vector, and the environmental feature vector.

[0079] In some embodiments, the inference of the Conditional Random Field (CRF) model relies on its internally defined probabilistic graphical structure and parameters. The CRF model has completed its training phase and internally defines the transition weight matrix and set of state feature functions between the implicit state of a user selecting a product and the features of the input observation sequence. In a specific implementation, a forward algorithm is used to calculate the selection probability. The input observation sequence, constructed for a specific candidate product, is input into the trained CRF model. The model uses its forward algorithm to perform recursive calculations. The forward algorithm progressively calculates the probability of the model being in each possible state given the current input observation sequence. Its recursive formula involves model parameters and observation features, and is expressed as follows:

[0080]

[0081] Where: symbol Indicates the position in the sequence In state Forward probability, symbol The set of all possible states, symbol Indicates the possible states of the previous position, symbol Indicates the position in the sequence Observational characteristics, symbols It is a state transition function that combines state transition features and observation features. The recursive calculation ultimately yields the probability value at the end of the sequence of the user's specific state of making a selection for the current candidate product; this probability value is the selection probability of that candidate product. Optionally, calculating the selection probability of all products in the pre-selected product pool is a batch processing procedure. The system iterates through the pre-selected product pool, repeatedly executing the above process of constructing the input observation sequence and calculating the forward algorithm for each candidate product, obtaining the selection probability value corresponding to each candidate product. The selection probability values ​​of all candidate products constitute an ordered selection probability distribution list. In specific implementation, generating the instant recommendation score and temporary score matrix is ​​the output of the calculation process. The system reads all probability values ​​from the selection probability distribution list, directly using each probability value as the instant recommendation score of the corresponding candidate product, or linearly scaling the probability value before using it as the instant recommendation score. The correspondence between all candidate product identifiers and their calculated instant recommendation scores is stored in a temporary score matrix created in memory. This temporary score matrix is ​​used by the subsequent state inversion mechanism.

[0082] In one embodiment of the present invention, when the state inversion mechanism is activated, the system uses the temporary score matrix calculated by the above scheme as input. First, products with recommendation scores higher than the average value but which the target user did not click on after being generated are extracted from the score matrix, and the set of these products is defined as the inversion trigger set. For each product in the inversion trigger set, a difference analysis is performed to identify abnormal features. A normalized user consumption preference vector is obtained from the basic user profile, which has specific preference values ​​in multiple product feature dimensions. At the same time, the product feature vector corresponding to the target product is extracted and mapped to the same feature dimensions as the user consumption preference vector to obtain a comparable target product feature vector. The absolute difference between the value of the target product feature vector in each dimension and the preference value of the user consumption preference vector in the corresponding dimension is calculated to obtain a series of difference values. Several items with the largest difference values ​​are selected, and their corresponding product feature dimensions are marked as candidate abnormal features. The specific values ​​of the target product in these candidate abnormal feature dimensions are further analyzed. If the value exceeds the threshold set according to the user's historical preference range, the dimension is determined to be an abnormal feature that caused the product to be recommended but not clicked. All identified anomalous features of the target product are associated with its high recommendation score and recorded to form inversion analysis result entries. Finally, the anomalous features of all products in the inversion trigger set are summarized, and combined with the cultural characteristics of the target user's destination country and current holiday information, a potential demand state that can reasonably explain these anomalous features is deduced. This state may be categorized as price-sensitive, quality-seeking, or demand for specific scenarios.

[0083] In practical implementation, defining the inversion trigger set relies on the analysis of the temporary score matrix. The system reads the temporary score matrix storing the instant recommendation scores, calculates the arithmetic mean of the instant recommendation scores of all products in the temporary score matrix, traverses the temporary score matrix, and filters out products whose instant recommendation scores are higher than the arithmetic mean and whose target users have not clicked within a set time window after the score was generated. The set of identifiers of the filtered products is defined as the inversion trigger set. In some embodiments, analyzing the difference to identify abnormal features requires obtaining and comparing vector data. A normalized user consumption preference vector is obtained from the basic user profile. The user consumption preference vector is a point in a multi-dimensional product feature space, where each dimension corresponds to a specific product feature and has a specific preference value. For each target product in the inversion trigger set, the system extracts its corresponding product feature vector. The product feature vector shares the same feature dimension definition as the user consumption preference vector. The values ​​of the product feature vector are mapped to multiple product feature dimensions that are exactly the same as those of the user consumption preference vector to obtain the mapped target product feature vector. An example of a data table 1 used to illustrate the difference calculation process is as follows:

[0084] Table 1: Calculation of Vector Difference in the Inversion Analysis Process

[0085] Product feature dimensions Preference values ​​in the user consumption preference vector Eigenvalues ​​in the feature vector of the target product absolute difference Dimension A: Price Range (Coded Value) 0.85 0.20 0.65 Dimension B: Brand Preference (Coded Value) 0.10 0.90 0.80 Dimension C: Category Preference (Coded Value) 0.60 0.55 0.05 Dimension D: Origin Preference (Coded Value) 0.30 0.85 0.55

[0086] In practice, calculating the difference value and labeling candidate abnormal features is a quantitative comparison process. The system calculates the absolute difference between the feature value of the target product feature vector in each feature dimension and the preference value of the user consumption preference vector in the corresponding dimension. The formula for calculating the absolute difference is as follows:

[0087]

[0088] Where: symbol This represents the difference value on the k-th product feature dimension, with the symbol... This represents the preference value in the k-th dimension of the user's consumption preference vector, with the symbol... This represents the feature value of the target product feature vector in the k-th dimension. The system iterates through all dimensions to calculate a set of difference values. From this set of difference values, the top N values ​​with the largest values ​​are selected, and the product feature dimensions corresponding to each of the top N difference values ​​are marked as candidate abnormal features. The value N is preset by the system. It can be understood that judging abnormal features requires comparison with the user's historical preference range. The system presets a preference range threshold for each product feature dimension. This threshold is derived from statistical analysis of the user's historical behavior data. For each product feature dimension marked as a candidate abnormal feature, the system checks whether the specific feature value of the target product feature vector in that dimension exceeds the preset threshold range of the user's consumption preference vector in that dimension. If the specific feature value exceeds the threshold range, the product feature dimension is ultimately judged as an abnormal feature that causes the product to be recommended but not clicked. Optionally, the inversion analysis result entry is the output of the difference analysis. The system creates an inversion analysis result entry for the target product. The entry records the identifier of the target product, its high real-time recommendation score calculated in the multimodal inference computing engine, and the identifiers and specific feature values ​​of all judged abnormal feature dimensions.

[0089] In some embodiments, summarizing anomalous features and inferring potential demand states combines external knowledge. The system traverses all inversion analysis result entries of all products in the inversion trigger set, statistically summarizes the anomalous features recorded in all entries, and identifies the most frequent combinations of anomalous features. In specific implementation, state inference based on external information is the final step of the inversion mechanism. The system accesses an external database to obtain cultural feature data and current holiday information of the target user's destination country. The cultural feature data includes local consumption habits and taboos, and the holiday information includes promotional cycles and gift demand cycles. The system matches and analyzes the summarized high-frequency anomalous feature combinations with the obtained cultural features and holiday information to infer a potential demand state that can explain the existence of most anomalous features. For example, price-sensitive anomalous features are concentrated in the dimensions of price and logistics costs, quality-seeking anomalous features are concentrated in the dimensions of brand and origin, and specific scenario demand anomalous features are concentrated in the dimensions of product attributes related to specific holidays or cultural customs.

[0090] See Figure 4 This is a frequency chart of abnormal features for different potential demand states, used to show the frequency of abnormal features in different product feature dimensions for the three types of potential demand states. Price-sensitive users' abnormal features are highly concentrated in price range (85) and logistics timeliness (75), indicating that these users are extremely sensitive to the final cost of the goods, which is the core reason for the recommendation not being clicked. Quality-seeking users' abnormal features are concentrated in brand preference (90), place of origin preference (80), and rating (85), indicating that these users pay more attention to the brand value, place of origin trust, and reputation of the goods. Specific scenario demand users' abnormal features are concentrated in category preference (85), and also have a high frequency in brand, rating, and logistics timeliness, indicating that the needs of these users are strongly correlated with specific scenarios and have the highest requirements for category matching.

[0091] In one embodiment of the present invention, during the data fusion stage, a state weight is assigned to each potential demand state inferred through the state inversion mechanism. The value of the weight depends on the number and degree of difference of the abnormal features that the state can explain. Next, each product in the temporary score matrix storing instant recommendation scores is traversed, and it is checked whether the product's features match any inferred potential demand state. If a match is found, the instant recommendation score corresponding to the product is multiplied by the state weight of the matched potential demand state to enhance the product's recommendation score. For products whose features do not match any potential demand state, their instant recommendation scores remain unchanged. After completing the score enhancement or preservation processing for all products, the scores of all products are normalized, and sorted from high to low according to the normalized scores to generate a draft of the enhanced recommendation list. Subsequently, a compliance filtering step is performed, extracting the country of origin information of all products from the draft list and comparing it with the trade compliance access list in effect in the target user's country, excluding products whose country of origin is not on the trade compliance access list. The prices of the products in the list are recalculated by adding the original price, estimated cross-border logistics costs, and customs duties to obtain the total landed cost of the products. Products with a total landed cost exceeding twice the historical average order value of the target users are filtered out. The list, after compliance and price filtering, is then matched with the user activity level in the basic user profile. For users with low activity levels, the length of the recommendation list is further compressed, retaining only the top three products after sorting, forming the final enhanced recommendation list.

[0092] In practice, data fusion combines real-time recommendation scores with potential demand states, while compliance filtering filters and adjusts the generated list. The basis for assigning state weights to potential demand states is the anomalous characteristics they explain. The system reads one or more potential demand states obtained through a state inversion mechanism. For each potential demand state, it analyzes the number of anomalous characteristics it can explain and the degree of difference between these anomalous characteristics. The formula for calculating the state weight is expressed as:

[0093]

[0094] Where: symbol The symbol represents the state weight assigned to the j-th potential demand state. The symbol represents the number of anomalous features that can be explained by the j-th potential demand state. The symbol represents the average degree of difference among all anomalous features explained by the j-th potential demand state. and These are preset coefficients used to adjust the proportion of influence between quantity and degree. In some embodiments, traversing the score matrix and performing score enhancement is an iterative comparison process. The system reads a temporary score matrix storing instant recommendation scores. For each product item in the temporary score matrix, the system extracts the product feature vector and performs a matching check with the feature pattern corresponding to each potential demand state. The matching check determines whether the value of the product feature vector in the key dimension falls within the numerical range defined by the potential demand state or conforms to the feature combination defined by the state. If the product feature vector matches a certain potential demand state, the system reads the original instant recommendation score of the product from the temporary score matrix and multiplies the instant recommendation score by the state weight corresponding to the potential demand state to obtain the enhanced new score. The new score updates the score value corresponding to the product item in the temporary score matrix. Optionally, for products that do not match any potential demand state, their instant recommendation scores stored in the temporary score matrix remain unchanged, and the system does not modify the scores of these products.

[0095] In practice, normalization and generating the initial draft list are the final steps in data fusion. After completing score checks and possible enhancement operations for all products, the system reads the latest scores of all products from the temporary score matrix, normalizes all the latest scores to map them to a unified range, and sorts all products from highest to lowest according to the normalized scores, generating an ordered initial draft of the enhanced recommendation list. In some embodiments, the first step of compliance filtering is origin information comparison. The system extracts the country code of origin for each product from the initial draft of the enhanced recommendation list and accesses a database storing a trade compliance access list for the target user's country. This trade compliance access list contains a list of country and region codes that can enjoy normal tariff treatment and are not subject to trade restrictions. The system compares the country code of the product with the trade compliance access list and removes product entries whose country code is not in the trade compliance access list from the initial draft of the enhanced recommendation list. It is understandable that the secondary calculation of commodity prices involves total cost calculation. For each commodity remaining after filtering by country of origin, the system obtains the commodity's base selling price, obtains the estimated trunk transportation cost of the commodity from the calculation process of environmental variable factors, and calculates the estimated tariff amount according to the commodity category and the tariff rules of the destination country. The base selling price, the estimated trunk transportation cost and the estimated tariff amount are added together to obtain the estimated total landed cost of the commodity.

[0096] In practice, filtering based on total landed cost and historical average order value is crucial for price compliance. The system obtains the target user's historical average order value from the user's basic profile. The historical average order value is the average of all the user's past order amounts. The estimated total landed cost for each product is calculated, and the total landed cost is compared with twice the target user's historical average order value. Products with a total landed cost exceeding twice the target user's historical average order value are filtered out from the product list. Optionally, compressing the list length based on user activity level is the final personalized adjustment step. The system reads the user activity level from the basic user profile. The user activity level is used to distinguish between impulsive and planned consumption. For users marked as low-activity, the system truncates the enhanced recommendation list remaining after all the aforementioned filtering steps, retaining only the top three product items as the final output enhanced recommendation list. For users whose activity level is not low-activity, the complete filtered list is output.

[0097] See Figure 5 This is a bar chart of environmental variables in cross-border e-commerce, focusing on three core indicators: exchange rate risk coefficient, customs clearance delay coefficient, and transportation cost coefficient for five destination countries: the United States, the United Kingdom, Germany, Japan, and Australia. It visually reflects the differences in the operating environment of cross-border e-commerce in different countries. The transportation cost coefficient is the core pain point in the five countries, with the UK (0.25) having the highest transportation cost and Japan (0.18) having the lowest. This reflects that the logistics trunk line costs for Oceania and Europe are significantly higher than those for East Asia and North America. The exchange rate risk coefficient is highest in Germany (0.18) and lowest in Japan (0.10). This reflects that exchange rate fluctuations in the Eurozone have the greatest impact on cross-border settlement, while the Japanese yen exchange rate is relatively stable. The customs clearance delay coefficient is highest in the UK (0.12) and lowest in Japan (0.06). This indicates that the UK customs clearance process takes the longest, and the logistics timeliness is the most uncertain.

[0098] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A product recommendation method based on a cross-border e-commerce platform, characterized in that, include: Obtain the historical behavior trajectory of the target user, process the historical behavior trajectory, and form a pre-selected product pool; A dynamic environmental variable factor is introduced, which is derived by comprehensively weighting the real-time exchange rate fluctuation value of the target user's destination country, the customs clearance time data of the local customs, and the trunk transportation cost of international logistics. The product information in the pre-selected product pool, the user activity level in the basic user profile, and the environmental variable factors are jointly input into the multimodal inference computing engine. The multimodal inference computing engine calculates the real-time recommendation score of each candidate product after considering environmental factors by constructing a joint probability distribution of user intent and product features. After calculating the instant recommendation score, the state inversion mechanism is activated. The state inversion mechanism takes the currently calculated recommendation score as input and reverses the unexpressed potential demand state of the target user under the current recommendation score. The calculated instant recommendation score is fused with the potential demand state obtained through the state inversion mechanism to generate an enhanced recommendation list that includes explicit recommendation reasons and implicit demand mining.

2. The product recommendation method based on a cross-border e-commerce platform as described in claim 1, characterized in that, Obtain the target user's historical behavior trajectory, process the historical behavior trajectory to form a pre-selected product pool, including: The historical behavior trajectory consists of the duration of browsing product detail pages, the set of product icons added to the shopping cart, and the logistics status of paid orders; The historical behavior trajectory is anonymized and standardized to generate a basic user profile containing user consumption preference vectors and user activity levels. From the global product database of the cross-border e-commerce platform, retrieve several candidate products with the highest similarity to the consumption preference vector in the basic user profile to form a pre-selected product pool; The historical behavior trajectories are anonymized and standardized to generate a basic user profile containing user consumption preference vectors and user activity levels, specifically including: The set of product identifiers in the historical behavior trajectory is deduplicated, and the frequency of occurrence under each product category is counted. The frequency is then transformed into a high-dimensional sparse category preference vector. The dwell time on the product details page in the historical behavior trajectory is segmented. Products with dwell time exceeding a set threshold are regarded as high interest points, and their corresponding price range and brand characteristics are extracted and added to the high-dimensional sparse category preference vector to form a dense user consumption preference vector. The logistics status of paid orders in the historical behavior trajectory is analyzed to calculate the average cycle from order placement to receipt. Based on the deviation of the average cycle from the platform average level, the user activity level is determined. The user activity level is used to distinguish between impulsive consumption and planned consumption.

3. The product recommendation method based on a cross-border e-commerce platform as described in claim 1, characterized in that, The introduction of dynamic environmental variable factors is derived by a weighted average of real-time exchange rate fluctuations in the target user's destination country, local customs clearance timeliness data, and international trunk transportation costs, including: By pulling the real-time exchange rate of the target user's country currency relative to the platform's settlement currency through an external data interface, the exchange rate volatility over the past seven days is calculated and mapped to a risk coefficient. Connect to the data interface of cross-border logistics service providers to obtain the current average customs clearance time in the target user's region, and convert the average customs clearance time into a time delay coefficient. Based on the weight and volume of goods in the pre-selected goods pool and the current international route fuel surcharge, the estimated trunk transportation cost is calculated and converted into a price fluctuation coefficient. The risk coefficient, the time delay coefficient, and the price fluctuation coefficient are each assigned a weight based on business strategy. The risk coefficient, the time delay coefficient, and the price fluctuation coefficient are then weighted and summed to obtain the environmental variable factor. The environmental variable factor is used to adjust the basic price attractiveness of the product.

4. The product recommendation method based on a cross-border e-commerce platform as described in claim 1, characterized in that, The multimodal inference computing engine calculates the instant recommendation score for each candidate product after considering environmental factors by constructing a joint probability distribution of user intent and product features, including: The user consumption preference vector in the basic user profile is concatenated with the user activity level to form a feature representation on the user side; For each product in the pre-selected product pool, extract its price, rating stars, listing time, and place of origin information to form a feature representation on the product side; The user-side feature representation, the product-side feature representation, and the environmental variable factors are all input into a conditional random field model. The conditional random field model learns the probability that a user will choose a specific product under given environmental conditions. The forward algorithm of the conditional random field model is used to calculate the selection probability of all products in the pre-selected product pool. The top few values ​​with the highest selection probabilities are used as the instant recommendation scores and stored in a temporary score matrix.

5. The product recommendation method based on a cross-border e-commerce platform as described in claim 4, characterized in that, The aforementioned state inversion mechanism takes the currently calculated recommendation score as input and reverse-engineers the unexpressed latent demand state of the target user under the current recommendation score, including: From the temporary score matrix, extract products with recommendation scores higher than the average but which have not yet been clicked by the target user, and define these products as the inversion trigger set; For each product in the inversion trigger set, analyze the degree of difference between it and the user consumption preference vector in the basic user profile, and identify the abnormal features that cause the product to be recommended but not clicked; All identified abnormal features are summarized, and combined with the cultural characteristics and holiday information of the target user's country of origin, a potential demand state that explains the abnormal features is deduced. The potential demand state includes price-sensitive, quality-seeking, or specific scenario-based demand.

6. The product recommendation method based on a cross-border e-commerce platform as described in claim 5, characterized in that, The process of fusing the calculated instant recommendation score with the potential demand state obtained through the state inversion mechanism includes: Each potential demand state obtained through the state inversion mechanism is assigned a state weight, the value of which depends on the number and degree of anomalous features explained by the state. Iterate through each product in the temporary score matrix and check whether its product features match the potential demand state. If they match, multiply the product's instant recommendation score by the state weight of the matching potential demand state to enhance the score. For products that do not match any potential demand status, their original instant recommendation scores remain unchanged; The scores of all products, whether enhanced or unchanged, are normalized, and then sorted from highest to lowest according to the normalized scores to generate the initial draft of the enhanced recommendation list.

7. The product recommendation method based on a cross-border e-commerce platform as described in claim 1, characterized in that, The method also includes a compliance filtering step for the enhanced recommendation list: From the enhanced recommendation list, the country of origin information of all products is extracted and compared with the trade compliance access list of the target user's country, excluding products whose country of origin is not on the trade compliance access list; The prices of products in the enhanced recommendation list are recalculated by adding the original price of the products to the estimated cross-border logistics costs and tariffs to obtain the total landed cost, and filtering out products whose total landed cost exceeds twice the historical average order value of the target user. The enhanced recommendation list, after compliance and price filtering, is then matched one last time with the user activity level in the basic user profile. For users with low activity levels, the length of the recommendation list is further reduced, retaining only the top three ranked products.

8. The product recommendation method based on a cross-border e-commerce platform as described in claim 4, characterized in that, The forward algorithm of the conditional random field model is used to calculate the selection probability of all items in the pre-selected item pool, including: The user consumption preference vector is extracted from the feature representation on the user side. The user consumption preference vector is generated by encoding the product category preference, price range preference and brand preference that frequently appear in the user's historical behavior trajectory. Extract product feature vectors from the feature representation of the product side. The product feature vectors are generated by encoding the product's price, average rating star, shelf duration, country of origin, and logistics method. The exchange rate risk coefficient, time delay coefficient, and price fluctuation coefficient are extracted from the environmental variable factors, and then combined into an environmental feature vector. The user consumption preference vector, the user activity level in the user-side feature representation, the product feature vector, and the environmental feature vector are concatenated to form an input observation sequence. The input observation sequence is input into a trained conditional random field model, which internally defines the transition weights and state feature functions between the product selection state and the aforementioned features; The conditional random field model uses its forward algorithm to recursively calculate the probability value of a user making a final selection of each candidate product in the pre-selected product pool under the given current observation sequence. The probability value is the selection probability of each candidate product, and the selection probabilities of all candidate products constitute a selection probability distribution list.

9. The product recommendation method based on a cross-border e-commerce platform as described in claim 5, characterized in that, For each item in the inversion trigger set, the difference between it and the user consumption preference vector in the basic user profile is analyzed to identify abnormal features that cause the item to be recommended but not clicked, including: Obtain a normalized user consumption preference vector from the basic user profile. The user consumption preference vector has specific preference values ​​in multiple product feature dimensions. Extract the corresponding product feature vector from the target product in the inversion trigger set, and map the product feature vector to multiple product feature dimensions that are the same as the user consumption preference vector to obtain the mapped target product feature vector; Calculate the absolute difference between the target product feature vector and the user consumption preference vector in each dimension, and obtain the difference value in multiple dimensions. Select the items with the largest differences among multiple dimensions and mark the corresponding product feature dimensions as candidate abnormal features; The specific values ​​of the target product feature vector on the labeled product feature dimension are analyzed. If the specific values ​​exceed the set threshold of the user's historical preference range, the product feature dimension is determined to be an abnormal feature that causes the product to be recommended but not clicked. All the identified abnormal features of the target product, along with the high recommendation score calculated in the multimodal reasoning engine, are recorded to form the inversion analysis result entries of the target product.

10. A product recommendation system based on a cross-border e-commerce platform, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the product recommendation method based on a cross-border e-commerce platform as described in any one of claims 1 to 9.

Images