Commodity tasting system for e-commerce

Abstract

A commodity tasting system for e-commerce, comprising: a rating data acquisition module, configured to receive ratings from multiple users for multiple commodities across multiple dimensions, providing raw rating data for correction processing; a user bias correction module, configured to model and correct individual bias in user ratings; the rating data acquisition module is connected to the user bias correction module, outputting corrected rating data; a feature extraction module, performing dimensionality reduction on the rating data based on principal component analysis; the user bias correction module is connected to the feature extraction module, outputting a reduced-dimensional feature representation; and a feature mapping module, configured to perform nonlinear mapping on a dimensionality reduction result to enhance feature expression. By adopting an output module design driven by group profiles, combined with weighted fusion of ranked rating results and group preference tags, recommendations are made more aligned with user group characteristics.

Description

Product vetting system for e-commerce Technical Field

[0001] This invention relates to the field of e-commerce system technology, specifically to a product appraisal system for e-commerce. Background Technology

[0002] E-commerce is a trade activity conducted between buyers and sellers without face-to-face interaction, based on a browser / server. This browser / server model expands the scope of transactions, allowing people to access information on goods from all over the world simply by logging into their browser / server accounts. With the rapid development of e-commerce, users often struggle to make effective judgments when faced with a large amount of product information, leading to a continuously increasing demand for personalized recommendation and decision-making support systems. As a mechanism to assist users in understanding and choosing, product evaluation systems are widely used on various online platforms to support product display, review generation, and consumer decision-making. These systems typically analyze product attributes, user feedback, and behavioral data to achieve product ranking and recommendation output based on ratings or tags, thereby improving user acceptance of recommended content and click-through conversion rates.

[0003] In existing technologies, product evaluation models generally suffer from insufficient characterization of group features. Most recommendation strategies rely on individual ratings to predict product ranking, lacking interest modeling at the user group level and ignoring the weighted guiding role of group tags from an evaluation perspective. Furthermore, traditional output strategies struggle to optimize recommendations from the perspective of collective behavior, resulting in highly homogenized recommendation results and failing to demonstrate personalized adjustment capabilities for different consumer groups. This problem is particularly evident in e-commerce platforms' promotional recommendations and vertical category recommendations, impacting the system's performance in group-focused recommendation tasks.

[0004] Traditional scoring models are mostly based on linear regression or deep learning methods for prediction. When dealing with high-dimensional sparse feature data, they are prone to problems such as high model complexity and uncontrollable parameters, especially in scenarios with limited sample size and many noisy features. Existing solutions often fail to provide interpretable control over regression parameters, making the model sensitive to changes in input features and exhibiting unstable generalization ability. Furthermore, parameter training often relies on a single error minimization mechanism, lacking prior constraints and dynamic sparsity mechanisms. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a product evaluation system for e-commerce, which solves the problem of the lack of group-oriented scenario optimization in recommended content.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a product appraisal system for e-commerce, comprising:

[0007] The rating data acquisition module is used to receive ratings from multiple users for multiple products across multiple dimensions, and to provide raw rating data for correction processing.

[0008] The user bias correction module is used to model and correct individual biases in user ratings; the rating data acquisition module is connected to the user bias correction module and outputs the corrected rating data.

[0009] The feature extraction module performs dimensionality reduction on the scoring data based on principal component analysis; the user bias correction module is connected to the feature extraction module and outputs the dimensionality-reduced feature representation.

[0010] The feature mapping module is used to perform non-linear mapping on the dimensionality reduction results to enhance feature representation; the feature extraction module is connected to the feature mapping module and provides the mapped features for score prediction modeling.

[0011] The scoring modeling module constructs a scoring prediction model based on sparse Bayesian regression; the feature mapping module is connected to the scoring modeling module and provides feature input to optimize model parameters.

[0012] The model training module is used to optimize model parameters through maximum a posteriori estimation; the scoring modeling module is connected to the model training module and outputs the model's predicted scoring results.

[0013] The output module is used to sort products based on the predicted rating results and make personalized recommendations in combination with group profile information. The model training module is connected to the output module, and the output module is directly connected to the system interface or data storage unit to display or save the final recommendation results. The system constructs a product appreciation perspective based on the group profile to realize group preference recommendation output under multi-dimensional integration.

[0014] Preferably, the rating data acquisition module constructs a three-dimensional rating tensor, wherein the first dimension is the product identifier, the second dimension is the rating dimension, and the third dimension is the user ID. The rating tensor is used to support subsequent dimensionality reduction and modeling.

[0015] Preferably, the user bias correction module centralizes the ratings of each user, removing individual rating biases, which are obtained by modeling the mean or statistical distribution of the user's historical ratings.

[0016] Preferably, the feature extraction module uses principal component analysis to reduce the dimensionality of the two-dimensional rating matrix formed after the rating tensor is expanded, as shown in the formula:

[0017] Z = RU q ;

[0018] in:

[0019] R represents a two-dimensional rating matrix obtained by expanding the rating tensor, with dimensions n×d, where n is the number of samples and d is the rating dimension;

[0020] U q The eigenvector matrix formed by the directions of the first q principal components has a dimension of d×q and is obtained by the covariance decomposition of the scoring matrix.

[0021] q represents the number of principal components selected, used to retain the main change information in the scoring data;

[0022] Z represents the dimensionality-reduced rating representation matrix with dimensions n×q, which is used in the subsequent feature mapping and rating modeling process.

[0023] Preferably, the feature mapping module uses a kernel function to perform a nonlinear transformation on the dimensionality-reduced score vector. The kernel function includes a polynomial kernel function or a radial basis function to improve the model's ability to fit nonlinear structures.

[0024] Preferably, the scoring modeling module represents the product score as a weighted combination of features, wherein the weights of each feature dimension follow a zero-mean Gaussian distribution, and the sparsity constraint of the scoring model is achieved by controlling its variance parameter.

[0025] Preferably, the model training module uses the expectation-maximization algorithm for iterative optimization, wherein the posterior distribution of the model parameters is estimated in the E-step, and the feature precision parameter and noise variance parameter are updated in the M-step, until convergence. The expectation-maximization algorithm formula is as follows:

[0026] E-step (expected step):

[0027] in:

[0028] w represents the feature weight vector in the scoring model, and represents the coefficient of each feature in the score prediction;

[0029] q(w) represents the posterior distribution of the feature weight vector, which is calculated in the E-step;

[0030] Represents a normal distribution;

[0031] μ represents the mean vector of the posterior distribution of the feature weights, and μ represents the optimal estimate of each feature weight in the current iteration.

[0032] Σ represents the covariance matrix of the posterior distribution of feature weights, and represents the uncertainty relationship between features;

[0033] M-step (maximization step):

[0034] in:

[0035] α j The precision parameter represents the j-th feature, which measures how well the feature is retained in the model;

[0036] Σ jj The diagonal element corresponding to the j-th feature in the covariance matrix represents the uncertainty of the estimated value of that feature;

[0037] μ j This represents the current estimate of the j-th feature in the posterior mean vector;

[0038] σ 2 The noise variance parameter represents the overall fluctuation level of the scoring error;

[0039] y represents the rating vector, which represents the actual rating value of a user or group for a product in the training samples;

[0040] X T X represents the product of the feature matrix and its transpose, used to measure the joint variance of the feature space;

[0041] Xμ represents the model's predicted score, i.e., the score result under the current feature estimation;

[0042] ∥y-Xμ∥ 2 This represents the sum of squares of the score prediction error, and this represents the magnitude of the residuals from the model fit.

[0043] Tr(·) represents the trace operator, used to calculate the sum of the diagonal elements of a matrix;

[0044] n represents the number of training samples, used to normalize the overall error.

[0045] Preferably, the output module sorts the products according to the predicted ratings and adjusts the rating sorting by combining the user group profile to generate a group product review list with preference adjustment capabilities.

[0046] Preferably, the group profile includes the user's preference distribution, rating volatility, similar group clustering labels, and historical behavior trajectory under each rating dimension, which are used to guide the personalized reconstruction of rating output.

[0047] Preferably, the generation method of the group evaluation perspective includes modeling the rating contribution of multiple similar profile groups, and adjusting the final evaluation ranking result of the products with the aggregation result as the weight.

[0048] This invention provides a product evaluation system for e-commerce. It has the following beneficial effects:

[0049] 1. This invention employs a group profile-driven output module design, combining ranked scoring results with weighted fusion of group preference tags to make recommendations more closely aligned with user group characteristics. In existing technologies, most recommendation systems rank users based solely on single-point predicted scores, failing to consider differences in group-level preferences. This invention addresses the problem of recommended content lacking optimization for group-specific scenarios.

[0050] 2. This invention employs a scoring modeling mechanism based on sparse Bayesian regression. By introducing feature compression and kernel mapping, it constructs a regression prediction structure, achieving the technical effect of maintaining modeling stability and sparsity even under high-dimensional feature conditions. Compared to existing technologies that rely on traditional linear regression or deep model schemes without prior control, this invention overcomes the technical shortcomings of being prone to overfitting in small sample scenarios and being unable to interpret parameter behavior.

[0051] 3. This invention simultaneously optimizes regression parameters and prior control vectors using the maximum a posteriori estimation method, constructing a dynamic adjustment mechanism during the training phase, thereby improving the accuracy and controllability of parameter learning. Compared with the traditional scoring model that simply uses the least squares method to fit parameters, this effectively solves the problems of unstable training convergence and lack of global adjustment capability.

[0052] 4. This invention constructs an end-to-end closed-loop structure, forming a traceable, feedback-enabled, and self-updating model link from feature input to result output. In particular, it establishes a residual-driven update interface between the model training module and the output module, which overcomes the limitation of the model's inability to adapt to the drift of user interests compared to the traditional static inference structure. Attached Figure Description

[0053] Figure 1 is a system framework diagram of the present invention. Detailed Implementation

[0054] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Please refer to Figure 1. An embodiment of the present invention provides a product appraisal system for e-commerce, comprising:

[0056] The rating data acquisition module is used to receive ratings from multiple users for multiple products across multiple dimensions, and to provide raw rating data for correction processing.

[0057] The user bias correction module is used to model and correct individual biases in user ratings; the rating data acquisition module is connected to the user bias correction module and outputs the corrected rating data.

[0058] The feature extraction module performs dimensionality reduction on the scoring data based on principal component analysis; the user bias correction module is connected to the feature extraction module and outputs the dimensionality-reduced feature representation.

[0059] The feature mapping module is used to perform non-linear mapping on the dimensionality reduction results to enhance feature representation; the feature extraction module is connected to the feature mapping module and provides the mapped features for score prediction modeling.

[0060] The scoring modeling module constructs a scoring prediction model based on sparse Bayesian regression; the feature mapping module is connected to the scoring modeling module and provides feature input to optimize model parameters.

[0061] The model training module is used to optimize model parameters through maximum a posteriori estimation; the scoring modeling module is connected to the model training module and outputs the model's predicted scoring results.

[0062] The output module is used to sort products based on the predicted rating results and make personalized recommendations in combination with group profile information. The model training module is connected to the output module, and the output module is directly connected to the system interface or data storage unit to display or save the final recommendation results. The system constructs a product appreciation perspective based on the group profile to realize group preference recommendation output under multi-dimensional integration.

[0063] In practical applications, the rating data collection module first receives rating data from multiple users on multiple products across multiple dimensions. This module supports multi-terminal access, receiving real-time ratings from users on platforms such as websites and apps, as well as integrating historical evaluation records. Rating dimensions can include product quality, cost-effectiveness, appearance design, after-sales service, etc., and can be flexibly adjusted according to the scenario.

[0064] User ratings are often subjective. To address this, the system introduces a user bias correction module. This module models user rating preferences and learns their rating habits. For example, some users habitually give high scores, while others tend to give conservative scores. The system corrects these biases through Bayesian inference and standardization, making the input for subsequent model training more objective and reliable.

[0065] Next, the corrected scoring data is dimensionality-reduced using a feature extraction module. This module employs Principal Component Analysis (PCA) to compress data dimensionality and reduce redundancy while preserving key scoring information. The dimensionality-reduced features are easier for modeling algorithms to process, effectively avoiding the curse of dimensionality and improving model computational efficiency.

[0066] To enhance the expressive power of the data, the system includes a feature mapping module. This module, based on kernel methods or deep neural network structures, performs nonlinear mapping on the PCA-reduced data. Its purpose is to more fully express the potential preferences of users for products in the new feature space, compensating for the information structure lost during linear dimensionality reduction.

[0067] The following rating modeling module is one of the core components of the system. It uses a sparse Bayesian regression algorithm to build a rating prediction model. This algorithm can automatically select the features that are most valuable for rating prediction, while suppressing the interference of redundant or irrelevant variables. The model is more concise and has stronger generalization ability. In addition, by introducing an adjustable prior distribution, the model still has stable output capability in scenarios with insufficient data or new products.

[0068] The model training module optimizes the parameters of the regression model using Maximum A posteriori estimation (MAP). During this process, the system calculates the mean and covariance of the posterior distribution, iteratively updates the feature weights, and, combined with the EM (Expectation-Maximization) algorithm, further improves the accuracy of parameter estimation. The training results converge after multiple iterations, and the model performance continues to improve.

[0069] After training, the prediction results are processed and displayed by the output module. This module intelligently sorts products based on the predicted scores and generates a personalized recommendation list by combining user profiles. User profiles may include factors such as age, gender, geographic location, and purchase history, thus achieving recommendations that better match user preferences. The final results can be presented through a front-end interface or pushed to the user's device.

[0070] Module 1: Scoring Data Acquisition Module

[0071] In the overall system operation flow, the rating data acquisition module serves as the starting point for data input, and its design directly impacts the processing quality and modeling effectiveness of subsequent modules. Generally, to ensure the rating data has broad representativeness and sufficient dimensional coverage, this module needs stable data access capabilities and close data transmission linkage with the user bias correction module to ensure that the collected rating data can be promptly incorporated into the modeling process. The rating data acquisition module connects to the e-commerce platform's user interaction layer, forming the main channel for product review information. Simultaneously, through interface protocol specifications, it enables structured data transmission with subsequent modules, improving the overall data consistency and response efficiency of the system.

[0072] In this embodiment, the rating data acquisition module receives ratings from multiple users for multiple products across multiple dimensions. These dimensions may include, but are not limited to, attributes such as product quality, cost-effectiveness, packaging, delivery speed, user experience, and after-sales service. Ratings for different dimensions can use a uniform integer system, such as a five-point or ten-level system, or floating-point ratings depending on platform settings. The system standardizes the rating data structure, including user identifier, product identifier, rating dimension, rating value, and timestamp fields, and encapsulates them into a unified format for transmission to the user deviation correction module.

[0073] In one possible implementation, the module interfaces with the e-commerce system's database backend to automatically retrieve historical rating records submitted by users. Simultaneously, it can be embedded in the front-end user interface to achieve real-time rating capture. At the data synchronization layer, the acquisition module employs an asynchronous communication mechanism to avoid impacting system performance due to front-end interaction delays.

[0074] Alternatively, the rating data acquisition module can be supplemented with a data preprocessing unit to identify anomalous ratings, such as extreme values ​​significantly exceeding the mean or records where the rating is inconsistent with the behavior (e.g., high rating but return). These anomalous ratings can be flagged or removed to enhance data quality in subsequent modeling stages. In some embodiments, the module also supports a unified processing mechanism for anonymous and logged-in ratings, ensuring the integrity of the data structure even in scenarios where user identity is untraceable.

[0075] Specifically, the rating data is sent to the user bias correction module immediately after being collected. The rating vector can be represented as:

[0076] r i =[r i1 ,r i2 ,…,r iM ];

[0077] Where, r i This represents the rating vector of user i for the set of items, where M is the number of items, and r is the number of items. ij The rating value of user i for product j on a specific dimension will be directly passed as input to the user bias correction module.

[0078] During the connection with the user bias correction module, the rating data acquisition module also needs to include rating time information and rating context information (such as terminal type, geographical location, and rating channel) as additional reference features for the correction module when modeling bias. The advantage of this is that when user rating behavior is affected by the scenario or device, the system can more flexibly identify potential biases, thereby improving the fitting ability of the individual bias model.

[0079] In some embodiments, this module can also combine user behavior logs (such as browsing, clicks, and add-to-cart data) with rating information for joint collection. Although behavioral data is not part of standard rating content, it has a high correlation with ratings and can be used as an external factor introduced in subsequent modeling processes to enrich the data feature space.

[0080] As a technological extension, the scoring data acquisition module is configured with a data caching mechanism at the acquisition end. This mechanism retains unuploaded data in the event of network anomalies or service interruptions, and automatically re-uploads it once the system recovers, ensuring data continuity and integrity. Furthermore, the module includes a scoring validity verification mechanism, including scoring value range checks, duplicate scoring detection, and scoring dimension matching, to prevent data noise introduced due to user errors or system malfunctions.

[0081] In summary, the scoring data acquisition module is not only the entry point for raw data, but also plays multiple roles in the entire system, including stabilizing input, filtering noise, format conversion, and structural correction. Through direct data link with the user bias correction module, this module can ensure that the scoring data has high accuracy and consistency before entering the modeling process, thus laying a solid foundation for feature extraction and predictive modeling.

[0082] Module 2: User Deviation Correction Module

[0083] In the system, user rating data exhibits significant subjective fluctuations. Without proper handling, this can severely impact the quality of subsequent modeling and prediction accuracy. Therefore, after the rating data acquisition module completes the collection of raw rating data, the user bias correction module must immediately address individual subjective biases in the ratings. This module occupies a crucial position in the system's data flow, as its output is directly used for feature extraction and subsequent dimensionality reduction analysis; thus, it needs to possess strong robustness and universality. Generally, different users exhibit certain patterns in their product ratings, such as some users giving excessively high or low ratings, having narrow rating ranges, or having ratings concentrated within a specific interval. These non-systematic biases need to be corrected through modeling.

[0084] In this embodiment, the user bias correction module is used to model and correct individual biases in user ratings. Its core is to construct a rating residual model, perform individualized analysis of each user's rating behavior, and correct the subjective offset in their rating vector accordingly.

[0085] Specifically, let the original rating matrix be... Where U represents the total number of users, I represents the total number of products, and element r ui This represents the rating value of user u for product i. When modeling this rating value, a user bias term b is introduced. u , representing the average rating tendency of user u, can be estimated by the following formula:

[0086] in:

[0087] b u This indicates the rating bias of user u, representing a systematic tendency for that user to give a higher or lower rating overall.

[0088] This represents the set of items that user u has rated;

[0089] Represents a set The number of elements in the data, i.e., the number of products rated by user u;

[0090] r ui This represents the rating value of user u for product i;

[0091] μ i This represents the average rating of product i among all users.

[0092] In one possible implementation, the system also introduces user rating variance. As an auxiliary feature, it is used to measure the dispersion of user ratings, so as to distinguish and normalize under extreme user behavior (such as extreme raters). This variance can be calculated and defined as follows:

[0093] in:

[0094] This represents the variance of user u's ratings, measuring the range of fluctuation in their ratings.

[0095] This represents the average rating value for user u.

[0096] In some embodiments, to enhance robustness, a product deviation term may be introduced to jointly model the scoring behavior:

[0097] r ui ≈μ+b u +b i +ε ui ;

[0098] in:

[0099] r ui This represents the rating value of user u for product i;

[0100] μ represents the global rating mean, which is the overall average rating of all users for all products;

[0101] b i The deviation term represents product i, reflecting the adjustment value of the product based on the average score;

[0102] b u This indicates the rating bias of user u, representing a systematic tendency for that user to give a higher or lower rating overall.

[0103] ε ui This represents the error term, which typically follows a zero-mean, Gaussian distribution and is used to model the uninterpretable part of the scoring.

[0104] The score after correction Relative objective ratings that remove subjective bias:

[0105] in:

[0106] This represents the corrected rating of user u for product i;

[0107] r ui This represents the original score;

[0108] b u This represents the user rating bias term, as defined above. This corrected rating will be used as standard input and sent to the subsequent feature extraction module for dimensionality reduction processing, ensuring that the data used in the subsequent modeling stage has a unified reference benchmark.

[0109] As an alternative, to improve computational efficiency, deviation modeling can employ matrix factorization and batch updates of deviation terms, which is particularly suitable for e-commerce platform environments with a large user base. In the iterative deviation estimation, a penalty term is used to control the deviation update range to avoid overfitting.

[0110] In some implementations, the user bias correction module also supports cross-dimensional rating fusion. If a product has multiple rating dimensions (such as price, functionality, aesthetics, etc.), user biases under each dimension can be modeled separately, corrected, and then aggregated to calculate the comprehensive rating. For example, the consistency of the rating structure can be improved by fusing dimensional results through weighted average or principal component weighting.

[0111] Extending this further, the user bias correction module can also provide prior parameters to support the generation of subsequent group profiles. For example, users with high rating preferences can be classified as "appreciative" users, while users with low scores and large fluctuations can be identified as "picky" users. This serves as an important input basis for forming user group profiles during the final recommendation process.

[0112] Through the above methods, this module not only corrects the subjectivity of the scoring, but also provides objective and structured data input for subsequent feature modeling and prediction tasks, ensuring that the expressiveness and interpretability of the scoring data are preserved throughout the entire process.

[0113] Module 3: Feature Extraction Module

[0114] After correcting for user bias in the rating data, the system obtains a relatively objective corrected rating matrix. While removing subjective bias, this matrix still suffers from high dimensionality and sparse structure, hindering modeling. To address this, the system further introduces a feature extraction module to reduce the dimensionality and vectorize the rating matrix, embedding user behavior and product attributes into a compact feature space. This provides standardized modeling input for the subsequent rating prediction module. This module plays a role in compressing the input space and enhancing representational capabilities within the system's information flow, exhibiting a certain degree of structural learning functionality.

[0115] In this embodiment, the feature extraction module is used to model the behavioral patterns of users and products in the calibration rating matrix and extract their quantifiable representation in a low-dimensional space.

[0116] In this embodiment, the feature extraction module is used to model the behavioral patterns of users and products in the correction rating matrix and extract their quantifiable representation in a low-dimensional space. This module receives the rating matrix output by the user bias correction module.

[0117] As an option, a rating matrix It can be approximately decomposed into the following form:

[0118] in:

[0119] P is the user feature matrix, representing the behavioral vector representation of each user in the K-dimensional latent feature space;

[0120] Q T This is the product feature matrix, which corresponds to the attribute vector representation of each product in the same space;

[0121] K is the defined dimension of the low-dimensional space.

[0122] This feature extraction module constructs a residual squared loss function for the rating matrix, introduces a regularization term to limit model complexity, and minimizes the set K of user-item pairs with ratings. Specifically, it takes the following form:

[0123] The meanings of each parameter and its symbol are as follows:

[0124] The total loss function value represents the sum of the prediction error and the regularization penalty term for the entire scoring matrix.

[0125] u,i represents the set of all user-product pairs indexes with existing rating records, where u represents the user index and i represents the product index;

[0126] This represents the corrected rating value of user u for product i, i.e., the rating input after user bias processing;

[0127] q i Let i represent the latent feature vector of product i, which is a K-dimensional column vector;

[0128] This represents the inner product of user u and item i in the latent space, used to approximate the predicted rating.

[0129] λ represents the regularization coefficient, usually a non-negative real number, used to control the user feature vector p. u and product feature vector q i The norm size is used to prevent the model from overfitting;

[0130] ∥p u ∥、∥q i ∥ represent the eigenvectors p respectively u q i The Euclidean norm of a vector is defined as the square root of the sum of the squares of all its elements.

[0131] In one possible implementation, optimization strategies such as stochastic gradient descent (SGD) or alternating least squares (ALS) are used to iteratively solve the objective function to obtain the latent feature vector. The optimization process supports parallel processing and batch training to adapt to large-scale data scenarios.

[0132] In practice, each user's feature vector represents a compressed representation of their historical rating behavior, while each product's feature vector represents its comprehensive attribute expression across multiple dimensions. The dot product of the two vectors reflects the user's preference for the product, and the subsequent rating prediction module is based on this structure for prediction modeling.

[0133] In some embodiments, to further improve the representational and contextual modeling capabilities of features, auxiliary attributes such as user rating time, rating frequency, and rating dimension can be feature-encoded and incorporated as additional inputs into the latent space vector.

[0134] As a technological extension, the feature extraction module can construct rating matrices for different rating dimensions, such as constructing rating sub-matrices for products in dimensions such as "quality", "price" and "packaging", and perform feature learning on each sub-matrices. Finally, the multi-dimensional feature vectors are merged by splicing to form a unified user behavior profile. This structure can more accurately express the differences in user preferences under different dimensions.

[0135] Furthermore, the system can also introduce a clustering mechanism into the extracted feature space, such as using the K-means algorithm to cluster user feature vectors for subsequent recommendation system or rating prediction group modeling. Based on this, different user groups can use different prediction models or rating rules to improve the system's flexibility.

[0136] In summary, the feature extraction module plays a crucial role in information compression and structure extraction within the entire rating modeling system. By vectorizing the rating patterns of users and products, it transforms the original rating matrix into structured modeling input, providing a low-dimensional, embedded input foundation for the rating prediction module.

[0137] Module 4: Rating Prediction Module

[0138] After extracting features from user rating data, the system obtains feature vector representations of users and products in a unified low-dimensional space. This feature representation is trained based on actual user rating data and includes the structural relationship between user historical preferences and product attributes. To further complete the modeling task, the system needs to use this feature representation to estimate the value of unrated products, thus realizing a closed-loop logic between rating prediction and subsequent recommendation functions. The rating prediction module is the structural component responsible for this task, calculating the estimated rating values ​​based on the learned user and product feature vectors, forming the system's final output rating matrix.

[0139] In this embodiment, the rating prediction module is used to estimate the rating value of the unrated items by the user based on the feature representation results obtained in module 3.

[0140] Generally, rating prediction is based on a similarity metric function between user and product feature vectors. This function can be expressed as the inner product of feature vectors, i.e.:

[0141] in:

[0142] This represents the predicted rating result for user u for product i;

[0143] This represents the estimated similarity between the user and the product in the latent feature space. The higher the value, the greater the user's interest in the product.

[0144] In its implementation, this prediction function is used to construct the missing items in the entire rating matrix, thereby completing the sparse rating matrix. Based on this, the system can perform an estimated ranking of all unrated items for generating a recommendation candidate set.

[0145] As an alternative, to improve the stability of the rating prediction, bias terms for users and products can be introduced into the above rating function. The bias terms are used to compensate for rating bias trends that may be missed by the feature inner product model, such as some users prefer to give high scores, or some products have overall higher scores due to popularity bias.

[0146] Specifically, in this bias modeling approach, the system defines the global score mean μ and the user bias term b. u With the product bias item b i In the prediction process, the aforementioned bias term is jointly modeled with the inner product of the eigenvectors. The final predicted score can be expressed as:

[0147] In the above expression:

[0148] μ is the system's global score mean, which can be obtained by averaging all historical scores.

[0149] b u This is a user bias parameter that reflects the degree to which user ratings deviate from the system mean.

[0150] b i This is the product bias parameter, used to represent the adjustment of product ratings in the global context;

[0151] All bias parameters can be obtained during the feature learning phase by minimizing the loss function along with the feature vector parameters.

[0152] In one possible implementation, the system constructs a rating loss function by minimizing the squared residual between the predicted and actual ratings, and continuously updates the parameter vectors of users and items during the training phase. The prediction module, as a closed-loop component in the rating modeling process, outputs results that can be used to construct a set of candidate rankings for user-item recommendations.

[0153] In some embodiments, to support multi-dimensional scoring scenarios, the system can independently model scores for different dimensions. For example, prediction functions can be constructed for sub-items such as "functional score", "appearance score", and "service score", and then a total score estimate can be formed by weighted or rule-based fusion. This structure can enhance the system's ability to interpret scoring semantics and is particularly suitable for modeling tasks in multi-label scoring scenarios.

[0154] Furthermore, the rating prediction module can also be linked with the subsequent recommendation system module to set threshold strategies based on the predicted rating value, dynamically control the triggering mechanism of the recommendation results, and support the output of recommendation level, confidence level or preference classification results based on the user's preference level.

[0155] Module 5: Scoring Modeling Module

[0156] In summary, the rating prediction module completes the structure of the rating data by modeling based on feature vectors. It is the terminal component of this system to complete the rating task. In the process of constructing the prediction function, the system considers both structural representation and rating trend information, which makes the rating prediction results highly consistent and expressive. This provides a calculable output data foundation for the final realization of user preference modeling and recommendation generation.

[0157] After the front-end module completes the collection, preliminary processing, and feature extraction of the raw data, and forms a structured, continuous, or sparse multidimensional feature representation during the intermediate processing, the system introduces a scoring modeling module to achieve score prediction of the input samples. This module is used to quantitatively model and predict the target score value. A direct logical connection is established between this module and the feature mapping module. The high-dimensional feature vectors provided by the latter constitute the input basis for the scoring modeling module, which is used for learning and dynamically adjusting the scoring model parameters.

[0158] In this embodiment, the scoring modeling module uses the sparse Bayesian regression method to construct a scoring prediction model. This method introduces a sparsity control mechanism on the basis of traditional Bayesian linear regression to achieve compressed selection of feature dimensions and improve generalization ability.

[0159] Generally, the scoring modeling module establishes predictive relationships based on the following linear model:

[0160] y = Φ(X)·w + ε;

[0161] in:

[0162] y is the score prediction output vector;

[0163] Φ(X) is the kernel matrix after eigenvalue transformation;

[0164] w is the vector of regression coefficients that the model is to learn;

[0165] ε is a zero-mean, Gaussian-distributed noise vector.

[0166] To control the sparsity of regression coefficients in the scoring modeling module, the model assigns an independent Gaussian prior distribution to the regression coefficients w, i.e.:

[0167] The parameters are defined as follows:

[0168] w represents the regression parameter vector to be learned in the scoring modeling module, with dimension m, and each component w j This represents the projected weights of the model at the j-th kernel basis;

[0169] α is a hyperparameter vector, called the precision control hyperparameter, and its dimension is also m;

[0170] The prior precision of each j-th regression coefficient is the reciprocal of the covariance in the Gaussian distribution.

[0171] The variance of the Gaussian prior distribution is used to control w. j The degree of sparsity;

[0172] This means that the mean is 0. The variance is a Gaussian distribution;

[0173] m represents the number of regression basis functions in the scoring model, which is usually related to the kernel dimension or feature dimension.

[0174] In general, by using each dimension w j Constrained by its corresponding independent prior accuracy α j This allows for the suppression of redundant parameters through penalties, thereby enabling the construction of sparse structure models.

[0175] Specifically, the scoring modeling module further includes a posterior inference submodule, used to calculate the mean and variance of the predicted distribution given the input. In one implementation, the predicted result is no longer a single-point score, but a score estimate with an uncertainty interval, suitable for complex scenarios with observation errors or feature drift.

[0176] As an alternative, the scoring modeling module can threshold the predicted score output based on a segmented activation mechanism to classify result levels or trigger conditional responses. For example, when the predicted score is within a certain high confidence interval, a system control response is triggered; when the score confidence intervals overlap significantly, a "fuzzy judgment" state is output.

[0177] In some embodiments, the scoring modeling module also integrates a self-updating mechanism. When the system detects a drift in the statistical characteristics of the scoring residual distribution, it can automatically update the kernel function parameters in the model or recalculate the posterior distribution based on the sliding window statistical features, thereby improving the system's continuous adaptability in dynamic task scenarios.

[0178] Furthermore, the scoring modeling module supports collaborative operation with the feature selection unit in module 3 to achieve reverse feature importance feedback driven by scoring residuals. This mechanism can reflect scoring errors back onto the feature space, assisting in optimizing feature extraction channels, thereby constructing a scoring system with weakly supervised self-feedback capabilities.

[0179] In summary, the scoring modeling module in this invention constructs a scoring prediction model by introducing sparse Bayesian regression, which balances modeling sparsity, interpretability, and prediction accuracy. It also forms a parameter interaction mechanism with the front-end feature mapping module. This module is technically innovative in both its structural design and parameter update strategy, and is an important component in realizing the system's continuous prediction capability and feedback optimization mechanism.

[0180] Module 6: Model Training Module

[0181] After the scoring modeling module completes the regression structure for predicting target scores, a model training module is introduced as the core training unit to effectively learn and optimize the model parameters. The main function of the model training module is to receive output data from the scoring modeling module, including feature inputs, score outputs, and corresponding true score labels. Based on the Maximum A posteriori Estimation (MAP) method, it estimates and updates the parameters involved in the regression model, thereby making the predicted scores more closely match the sample distribution and improving the overall model's accuracy and generalization ability.

[0182] In this embodiment, the model training module uses the maximum a posteriori probability estimation method to optimize the parameter vector in the scoring modeling module. These parameters include the regression coefficient vector and its set of hyperparameters controlling sparsity. In the scoring modeling module, the regression coefficients are assigned independent Gaussian prior distributions, and each dimension is constrained by corresponding precision-controlled hyperparameters. Therefore, during the model training phase, it is necessary to comprehensively consider the information from the likelihood function and the prior distribution to solve for the posterior distribution in order to obtain the most probable parameter estimates.

[0183] Generally, the objective function used by the model training module consists of two parts: one part is the loss term representing the prediction error, which is usually constructed based on the squared residuals between the input features and the true labels; the other part is the regularization term, which is used to introduce constraints on the parameter space, so that the parameters maintain a sparse structure and avoid the model from overfitting the training data. The above two parts together constitute the optimization objective of the posterior distribution in the logarithmic sense.

[0184] Specifically, during the parameter estimation process, the training module updates the hyperparameters that control sparsity synchronously. The larger the hyperparameter, the weaker or closer to zero the corresponding regression parameter is, thereby compressing the feature dimension. During the training convergence process, the model training module gradually removes redundant or insignificant feature channels, retaining only the parameters that significantly contribute to the prediction results.

[0185] In one possible implementation, the model training module includes a loss tracking mechanism and a learning rate adjustment strategy. This mechanism can record the change in the posterior estimate during each training iteration and adjust the update magnitude according to the error convergence, thereby controlling the stability of the training process. Once the objective function is detected to enter an oscillating or stable state, the system will automatically reduce the parameter update step size to prevent the risk of divergence.

[0186] As an alternative, the training module can also introduce a sample weighting mechanism to address the problems of uneven sample noise or unbalanced class distribution. By dynamically weighting the influence of different samples on the training objective function, the model can pay more attention to sample points with larger prediction errors, thereby improving the fitting ability under boundary sample or extreme feature conditions.

[0187] In some embodiments, the model training module also supports an incremental training strategy, which involves updating the existing model parameters online by combining new sample data instead of retraining the entire dataset. This strategy is particularly suitable for scenarios where data is gradually introduced or the system needs to run for a long time. The incremental update process achieves parameter fine-tuning by correcting the local posterior distribution, effectively saving computing resources while maintaining model performance.

[0188] Furthermore, to adapt to the differences in computing resources under different deployment environments, the training module also supports low-rank approximation and block structure update methods. When the feature dimension is high, the system can use a block matrix form to split the parameter update process, reducing the computational complexity in each training iteration.

[0189] In summary, the model training module achieves joint optimization of regression parameters and hyperparameters in the scoring modeling module through the maximum a posteriori estimation method, constructing a feedback closed-loop mechanism from score prediction to parameter update. This module not only has a direct impact on model performance, but also constitutes the core link of parameter adaptation and structural adjustment in the system, and is an indispensable component of the technical solution of this invention.

[0190] Module 7: Output Module

[0191] After the model training module completes the convergence and optimization of the rating modeling parameters, the system enters the final output stage to achieve product ranking and personalized recommendations based on the predicted rating results. This stage uses the output module to perform operations such as result filtering, content sorting, and user profile matching, transforming the modeled rating information into user-perceptible recommendation output. The output module is directly connected to the model training module, obtaining the latest rating results and combining them with group profile information for multi-dimensional fusion processing, ultimately completing the display or storage of product recommendation content. Simultaneously, the output module is also responsible for interfacing with the system's front-end interface or data storage unit to achieve final result delivery.

[0192] In this embodiment, the output module receives the predicted score results from the scoring modeling module and sorts the candidate product set in descending order based on the score values. The score results are the model output optimized by maximum a posteriori estimation, representing the predicted fit or preference value of each product sample under the current feature conditions.

[0193] Generally, the output module constructs a recommendation list based on the rating and ranking results, and further processes the recommendation results by combining them with the system's built-in user profile information. The user profile is a statistical representation based on a comprehensive model of multiple users' historical behaviors, preference trends, rating tags, and social aggregation information, and typically includes, but is not limited to, indicators such as age group, spending power, style preference, regional characteristics, and category popularity.

[0194] Specifically, the output module, based on the initial ranking, introduces a weighted correction coefficient for each product's score. This correction coefficient is derived from the tag distribution and similarity measurement results extracted from the user profile and is used to strengthen the weight of products that match the attributes of the current user group. This process constitutes a group preference guidance mechanism, improving the matching degree of the recommendation results within the target user range.

[0195] Alternatively, a rating smoothing mechanism based on content diversity can be embedded in the output module. When the rating concentration is too high in a multi-product rating set, the system will introduce rating distribution entropy or coverage metrics to re-rank the results. This mechanism is used to avoid the problem of overly singular recommendation results and ensure the breadth and explorability of content recommendations.

[0196] In one possible implementation, the output module further sub-clusters the group profile, decomposing large-scale profile information into several highly homogeneous groups to achieve finer-grained interest fusion and matching. During recommendation ranking, the system uses the distance between the current user or session features and the centers of each sub-cluster as a reference, selecting the preference tag set represented by the most matching group as the benchmark for product adjustment.

[0197] In some embodiments, the output module also supports a cross-time period recommendation history fusion processing mechanism. That is, based on the current rating output, the system will combine high-frequency keywords or significant feedback generated in recent interaction behaviors to perform time-series weighting processing on the original ranking results, so that the recommendation results take into account both short-term interest changes and long-term preference characteristics.

[0198] Furthermore, by interacting with the data storage unit, the output module can structurally archive each recommendation result and its corresponding ranking criteria, forming a traceable recommendation log, which is convenient for subsequent model retraining or preference evolution analysis. Simultaneously, when connected to the system's front-end interface, the output module can provide real-time visualization of the recommendation results and support interactive operations such as multi-dimensional filtering, label switching, or dynamic adjustments.

[0199] In summary, the output module, as the final carrier and display node of the scoring results in the system, not only realizes the closed-loop connection from model prediction to result application, but also constructs a multi-dimensional recommendation mechanism with adjustability and interpretability through technical means such as group profiling guidance, multi-dimensional weight fusion and diversity control. This ensures the adaptability and stability of the recommendation results in different user groups and provides the final personalized recommendation output capability for the system described in this invention.

[0200] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A product appraisal system for e-commerce, characterized in that, include: The rating data acquisition module is used to receive ratings from multiple users for multiple products across multiple dimensions, and to provide raw rating data for correction processing. The user bias correction module is used to model and correct individual biases in user ratings; the rating data acquisition module is connected to the user bias correction module and outputs the corrected rating data. The feature extraction module performs dimensionality reduction on the scoring data based on principal component analysis; the user bias correction module is connected to the feature extraction module and outputs the dimensionality-reduced feature representation. The feature mapping module is used to perform nonlinear mapping on the dimensionality reduction results to enhance feature representation; The feature extraction module is connected to the feature mapping module, providing mapped features for score prediction modeling. The scoring modeling module constructs a scoring prediction model based on sparse Bayesian regression; the feature mapping module is connected to the scoring modeling module and provides feature input to optimize model parameters. The model training module is used to optimize model parameters through maximum a posteriori estimation; the scoring modeling module is connected to the model training module and outputs the model's predicted scoring results. The output module is used to sort products based on the predicted rating results and make personalized recommendations in combination with group profile information. The model training module is connected to the output module, and the output module is directly connected to the system interface or data storage unit to display or save the final recommendation results. The system constructs a product appreciation perspective based on the group profile to realize group preference recommendation output under multi-dimensional integration.

2. The product appraisal system for e-commerce according to claim 1, characterized in that, The rating data acquisition module constructs a three-dimensional rating tensor, where the first dimension is the product identifier, the second dimension is the rating dimension, and the third dimension is the user ID. The rating tensor is used to support subsequent dimensionality reduction and modeling.

3. The product appraisal system for e-commerce according to claim 1, characterized in that, The user bias correction module centralizes the ratings of each user, removing individual rating biases. These biases are obtained by modeling the mean or statistical distribution of the user's historical ratings.

4. The product appraisal system for e-commerce according to claim 1, characterized in that, The feature extraction module uses principal component analysis to reduce the dimensionality of the two-dimensional rating matrix formed after expanding the rating tensor. The formula is: Z = RU q ; in: R represents a two-dimensional rating matrix obtained by expanding the rating tensor, with dimensions n×d, where n is the number of samples and d is the rating dimension; U q The eigenvector matrix formed by the directions of the first q principal components has a dimension of d×q and is obtained by the covariance decomposition of the scoring matrix. q represents the number of principal components selected, used to retain the main change information in the scoring data; Z represents the dimensionality-reduced rating representation matrix with dimensions n×q, which is used in the subsequent feature mapping and rating modeling process.

5. The product appraisal system for e-commerce according to claim 1, characterized in that, The feature mapping module uses a kernel function to perform a nonlinear transformation on the dimensionality-reduced score vector. The kernel function includes a polynomial kernel function or a radial basis function to improve the model's ability to fit nonlinear structures.

6. The product appraisal system for e-commerce according to claim 1, characterized in that, The scoring modeling module represents product scores as a weighted combination of features, where the weights of each feature follow a zero-mean Gaussian distribution, and sparsity constraints on the scoring model are achieved by controlling its variance parameter.

7. The product appraisal system for e-commerce according to claim 1, characterized in that, The model training module employs the expectation-maximization algorithm for iterative optimization. In the E-step, the posterior distribution of the model parameters is estimated, and in the M-step, the feature precision parameter and noise variance parameter are updated until convergence. The expectation-maximization algorithm formula is as follows: E-step is the desired step: in: w represents the feature weight vector in the scoring model, and represents the coefficient of each feature in the score prediction; q(w) represents the posterior distribution of the feature weight vector, which is calculated in the E-step; Represents a normal distribution; μ represents the mean vector of the posterior distribution of the feature weights, and μ represents the optimal estimate of each feature weight in the current iteration. Σ represents the covariance matrix of the posterior distribution of feature weights, and represents the uncertainty relationship between features; M-step is the maximization step: in: α j The precision parameter represents the j-th feature, which measures how well the feature is retained in the model; Σ jj The diagonal element corresponding to the j-th feature in the covariance matrix represents the uncertainty of the estimated value of that feature. μ j This represents the current estimate of the j-th feature in the posterior mean vector; σ 2 The noise variance parameter represents the overall fluctuation level of the scoring error; y represents the rating vector, which represents the actual rating value of a user or group for a product in the training samples; This represents the product of the feature matrix and its transpose, used to measure the joint variance of the feature space; Xμ represents the model's predicted score, i.e., the score result under the current feature estimation; ||y-Xμ|| 2 This represents the sum of squares of the score prediction error, and this represents the magnitude of the residuals from the model fit. Tr(·) represents the trace operator, used to calculate the sum of the diagonal elements of a matrix; n represents the number of training samples, used to normalize the overall error.

8. The product appraisal system for e-commerce according to claim 1, characterized in that, The output module sorts the products according to their predicted ratings and then adjusts the ranking by weighting the ratings based on user profiles to generate a group-based product review list with preference adjustment capabilities.

9. The product appraisal system for e-commerce according to claim 1, characterized in that, The user profile includes the user's preference distribution, rating volatility, similar group clustering labels, and historical behavior trajectory across various rating dimensions, which are used to guide the personalized reconstruction of rating output.

10. The product appraisal system for e-commerce according to claim 1, characterized in that, The generation method of the group evaluation perspective includes modeling the rating contribution of multiple similar profile groups, and adjusting the final evaluation ranking of products with the aggregated results as weights.