E-commerce platform commodity evaluation analysis method based on multi-source data mining

Abstract

The invention discloses an e-commerce platform commodity evaluation analysis method based on multi-source data mining. The method comprises the following steps: S1, collecting data of each platform; s2, performing data preprocessing on the text data of the platform; s3, carrying out data analysis on the preprocessed text data, splicing LDA features and BERT features, carrying out standardization, and providing suggestions for commodities through results of sentiment analysis, evaluation clustering analysis and user behavior analysis; and S4, continuously monitoring and optimizing the commodity, and establishing a user feedback mechanism to collect feedback output by the user to the model to guide the optimization direction. According to the method, deeper-level relations and laws among data are mined, and comprehensiveness and diversity of the data are ensured; potential negative evaluation and product problems are accurately found and processed, and targeted improvement is carried out; demands and problems of the user are responded in time, and customer service experience is improved; data analysis and learning are continuously carried out, and precious market and user knowledge is accumulated.

Description

Technical Field

[0001] The present invention relates to the field of commodity evaluation and analysis, and particularly to a method for analyzing commodity evaluations on an e-commerce platform based on multi-source data mining. Background Art

[0002] Commodity evaluation analysis is a process of collecting and analyzing consumers' evaluation information of commodities to gain insights into consumers' needs, preferences, and feedback. The purpose of commodity evaluation analysis is to help merchants better understand the voices of consumers, thereby improving product quality, optimizing the customer experience, and enhancing market competitiveness.

[0003] Traditional commodity evaluation analysis has limitations in data collection. When conducting commodity evaluation analysis on an e-commerce platform, it mainly relies on data collected within the platform and ignores other potential valuable data sources, such as social media and forums. This limits the comprehensiveness and depth of the analysis results. Moreover, sentiment analysis is an important part of evaluation analysis, but traditional sentiment analysis tools may not be able to accurately capture the subtle sentiment changes in user comments, resulting in deviations in the analysis results. Summary of the Invention

[0004] In order to solve the technical problems of the existing commodity evaluation analysis method, such as limited data collection and inaccurate sentiment analysis, the present invention proposes a method for analyzing commodity evaluations on an e-commerce platform based on multi-source data mining.

[0005] A method for analyzing commodity evaluations on an e-commerce platform based on multi-source data mining includes the following steps:

[0006] S1: Collect data from various platforms;

[0007] S2: Perform data preprocessing on the text data of the platforms, including data unification, duplicate removal and cleaning, word segmentation and annotation;

[0008] S3: Analyze the preprocessed text data, splice and standardize LDA features and BERT features, and provide suggestions for commodities based on the results of sentiment analysis, evaluation clustering analysis, and user behavior analysis;

[0009] S4: Continuously monitor and optimize the commodities, and establish a user feedback mechanism to collect users' feedback on the model output to guide the optimization direction.

[0010] Further, in a method for analyzing commodity evaluations on an e-commerce platform based on multi-source data mining, the step S1 includes the following sub-steps:

[0011] S11: Collect the ratings, review texts, review pictures, review videos, user behavior data, purchase history, and spatial data of the users on the e-commerce platform;

[0012] S12: Collect data from social media, blogs, and news websites.

[0013] Further, a method for analyzing product reviews on an e-commerce platform based on multi-source data mining, where the step S2 includes the following sub-steps:

[0014] S21: Data unification: Use optical character recognition (OCR) to extract text information from review pictures and other picture data. Divide review videos into image information and audio information. Use OCR to extract the image information as text information, and use an audio-to-text tool to convert the audio information into text information.

[0015] S22: Duplicate removal and cleaning: Perform duplicate removal and cleaning on all the collected text data to obtain high-quality and low-redundancy data.

[0016] S23: Word segmentation and annotation:

[0017] S231: Use Chinese word segmentation tools to perform word segmentation and part-of-speech annotation on the data after duplicate removal and cleaning. Split the continuous text string into a sequence of meaningful words, and add a custom dictionary.

[0018] The Chinese word segmentation tools include jieba, SnowNLP, and THULAC.

[0019] S232: Delete words that are not important for sentiment analysis and all punctuation marks.

[0020] S233: Create or use existing positive and negative sentiment lexicons, assign weights to each sentiment word. Usually, positive words are assigned positive weights, and negative words are assigned negative weights. Match the word segmentation results with the sentiment dictionary to assign sentiment weights to each word.

[0021] S234: Check whether there is a negation word before the sentiment word that affects the sentiment tendency. If there is a negation word, then readjust the weight of the sentiment word.

[0022] Further, a method for analyzing product reviews on an e-commerce platform based on multi-source data mining, where the S22 includes the following sub-steps:

[0023] For records that are exactly the same in the data, use the Pandas library in Python and remove duplicate review data through the drop_duplicates() function.

[0024] For records that are similar but not exactly the same, set a similarity threshold sml for judgment and processing. For data with a similarity exceeding the threshold s ml, directly delete it, and keep the data with a similarity lower than the threshold sml.

[0025] For the missing data, set the missing degree threshold def. Data exceeding the missing degree threshold def is directly deleted, and data below the threshold def is filled with missing values using the mean, median, and mode methods;

[0026] For incorrect or inconsistent values in the data, correct or delete them.

[0027] Furthermore, a method for analyzing product evaluations on an e-commerce platform based on multi-source data mining, where S3 includes the following sub-steps:

[0028] S31: Use the BERT model to extract sentence vectors from the preprocessed text data to obtain the vector representation of each text;

[0029] S32: Use the LDA model to perform topic modeling on the preprocessed text data to obtain topic feature vectors through training;

[0030] S33: Concatenate the LDA features and BERT features row by row. The dimension of the fused feature vector is k + d, and the fused feature vector Fused Feature is:

[0031] Fused Feature = [t1, t2,..., t k , b1, b2,..., b d ;

[0032] Among them, [t1, t2,..., t k represents the LDA features, [b1, b2,..., b d represents the BERT features, k represents the number of topics of the LDA features, and d represents the output dimension of the BERT features;

[0033] S34: Use the Z-score normalization method to normalize the concatenated features. The formula is:

[0034]

[0035] Among them, z represents the normalized concatenated features, x represents the concatenated feature value, μ represents the feature mean, and σ represents the feature standard deviation;

[0036] S35: Perform sentiment analysis, evaluation clustering analysis, and user behavior analysis on the normalized concatenated features;

[0037] S36: Based on the results of sentiment analysis, evaluation clustering analysis, and user behavior analysis, summarize the different focuses that consumers care about when purchasing different products, and provide guiding suggestions for product production and improvement.

[0038] Further, a method for analyzing product evaluations on an e-commerce platform based on multi-source data mining, where S35 includes the following sub-steps:

[0039] S351: Sentiment analysis: Use the BERT model to train the standardized concatenated features to obtain the positive or negative emotions of consumers towards different attention dimensions of each different category of products;

[0040] S352: Evaluation clustering analysis: Use the K-means algorithm to perform spatial clustering on the results of sentiment analysis in combination with the spatial information of users to mine the needs and preferences of consumer groups in different regions;

[0041] S353: User behavior analysis: Combine the results of sentiment analysis with user behavior data to construct user portraits and understand the psychological characteristics, demand characteristics, and behavior patterns of different user groups;

[0042] S36 includes the following sub-steps:

[0043] S361: Provide personalized product recommendations based on the sentiment analysis of users towards products, improve the user experience and conversion rate, identify target customer groups, formulate high-quality marketing strategies, and push specific products targeted;

[0044] Further, a method for analyzing product evaluations on an e-commerce platform based on multi-source data mining, where S4 includes the following sub-steps:

[0045] S41: Use the Apache Kafka tool to collect and analyze user evaluations in real time and visualize the analysis results;

[0046] S42: Real-time track the evaluation trend, set a monitoring threshold, and when the number of negative evaluations or the results of sentiment analysis exceed the threshold, the system automatically issues an alarm;

[0047] S43: Adjust products or services in a timely manner according to the real-time monitoring results, and feedback the effects of improvement measures to the monitoring system to form a continuous improvement cycle;

[0048] S44: Perform model optimization, regularly evaluate the model performance through weighted summation of precision, recall, and F1-score to obtain evaluation indicators, identify deficiencies, collect labeled data, and enhance the generalization ability of the model. The formula is:

[0049] N = w1 * Precision + w2 * Recall + w3 * F1 Score;

[0050] where N represents the evaluation indicator, Precision represents precision, and Recall represents recall;

[0051] Among them, w1 represents the precision weight, w2 represents the recall weight, w3 represents the F1-score weight, and the weight distribution where the sum of w1, w2, and w3 is 1 is adjusted according to the actual application scenario;

[0052] S45: Use data augmentation techniques to generate synthetic data to improve the robustness and performance of the model, fine-tune the existing model according to the new data, make adjustments for the details of specific product categories, establish a user feedback mechanism, and collect user feedback on the model output to guide the optimization direction.

[0053] Furthermore, an e-commerce platform product evaluation analysis method based on multi-source data mining, the model optimization includes the following sub-steps:

[0054] S441: Initialize the best parameter dictionary and the best metric record, set the iteration counter and the early stopping counter, create a training history record dictionary, randomly sample a set of parameter combinations from the parameter search space, and update the model with the current parameters;

[0055] S442: Train the model on the training set, make predictions on the validation set, calculate the evaluation metrics of the current parameter combination, and update the training history;

[0056] S443: Calculate the comprehensive score of the current parameter combination. If the comprehensive score exceeds the historical best score, update the best score, the best parameter combination, the best evaluation metrics, and reset the early stopping counter by adding a threshold. Otherwise, increase the early stopping count;

[0057] S444: After reaching the maximum number of iterations, output the optimal parameter combination, the corresponding evaluation metrics, and the training history.

[0058] The beneficial effects of the present invention are:

[0059] 1. Collect product evaluation data from different channels, mine deeper connections and patterns between the data, find potential connections between spatial locations and product evaluations, and ensure the comprehensiveness and diversity of the data;

[0060] 2. Accurately discover and handle potential negative evaluations and product problems, make targeted improvements, improve product quality, and reduce brand reputation risks;

[0061] 3. Respond to user needs and problems in a timely manner, enhance the customer service experience, increase customer satisfaction and product stickiness, and enhance the brand image and market competitiveness;

[0062] 4. Continuously conduct data analysis and learning, accumulate valuable market and user knowledge, and provide support for long-term development. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is a flowchart of this method. Detailed implementation manners

[0064] The present invention will be further described, but the protection scope of the present invention is not limited to the following description.

[0065] An e-commerce platform product evaluation analysis method based on multi-source data mining, comprising the following steps:

[0066] S1: Collect data from each platform;

[0067] S2: Perform data preprocessing on the text data of the platform, including data unification, duplicate removal and cleaning, word segmentation and annotation;

[0068] S3: Analyze the preprocessed text data, splice and standardize LDA features and BERT features, and provide suggestions for products based on the results of sentiment analysis, evaluation clustering analysis, and user behavior analysis;

[0069] S4: Continuously monitor and optimize the products, and establish a user feedback mechanism to collect user feedback on the model output to guide the optimization direction.

[0070] Furthermore, for an e-commerce platform product evaluation analysis method based on multi-source data mining, the step S1 includes the following sub-steps:

[0071] S11: Collect the ratings, review texts, review pictures, review videos, user behavior data, purchase history, and spatial data of e-commerce platform users;

[0072] S12: Collect data from social media, blogs, and news websites.

[0073] Among them, the data of social media includes, for example, user feedback, forum discussions, and customer service records;

[0074] Furthermore, for an e-commerce platform product evaluation analysis method based on multi-source data mining, the step S2 includes the following sub-steps:

[0075] S21: Data unification: Use optical character recognition (OCR) to extract text information from review pictures and other picture data, divide review videos into image information and audio information, use OCR to extract the image information as text information, and use an audio-to-text tool to convert the audio information into text information;

[0076] S22: Duplicate removal and cleaning: Perform duplicate removal and cleaning on all collected text data to obtain high-quality and low-redundancy data:

[0077] S23: Word segmentation and annotation:

[0078] S231: Use Chinese word segmentation tools to segment and tag the data after deduplication and cleaning, split the continuous text string into meaningful word sequences, and add a custom dictionary;

[0079] The Chinese word segmentation tools include jieba, SnowNLP, and THULAC;

[0080] S232: Remove words and all punctuation marks that are not important for sentiment analysis;

[0081] Among them, unimportant words such as "的", "是" and other stop words;

[0082] S233: Create or use existing positive and negative sentiment word libraries, assign weights to each sentiment word, positive words are usually assigned positive weights, and negative words are assigned negative weights, match the word segmentation results with the sentiment dictionary, and assign sentiment weights to each word;

[0083] S234: Check whether there are negative words before the sentiment word that affect the sentiment tendency, and if there are negative words, readjust the weight of the sentiment word;

[0084] Among them, negative words such as "not" in "bad" are negative words, so the sentiment weight of "good" should become negative.

[0085] Furthermore, in an e-commerce platform product evaluation analysis method based on multi-source data mining, the S22 includes the following sub-steps:

[0086] For identical records in the data, use Python's Pandas library to remove duplicate comment data through the drop_duplicates() function;

[0087] For records that are similar but not identical, a similarity threshold sml is set for judgment and processing. s ml of data will be deleted directly, and data below the threshold sml will be retained;

[0088] For missing data, set the missing degree threshold def. Data exceeding the missing degree threshold def will be directly deleted. Data below the threshold def will be filled with the mean, median, or mode method.

[0089] Correct or delete errors or inconsistent values ​​in the data.

[0090] Furthermore, in an e-commerce platform product evaluation analysis method based on multi-source data mining, S3 includes the following sub-steps:

[0091] S31: Use the BERT model to extract sentence vectors from the preprocessed text data to obtain the vector representation of each text;

[0092] S32: Use the LDA model to perform topic modeling on the preprocessed text data, and obtain the topic feature vector after training;

[0093] S33: Concatenate LDA features and BERT features row by row. The dimension of the fused feature vector is k+d. The fused feature vector Fused Feature is:

[0094] Fused Feature=[t1, t2,...,t k , b1, b2, …, b d ];

[0095] Among them, [t1, t2, ..., t k ] represents LDA features, [b1, b2, ..., b d ] represents BERT features, k represents the number of topics of LDA features, and d represents the output dimension of BERT features;

[0096] S34: Use the Z-score standardization method to standardize the splicing features. The formula is:

[0097]

[0098] Among them, Z represents the standardized splicing feature, x represents the splicing feature value, μ represents the feature mean, and σ represents the feature standard deviation;

[0099] S35: Perform sentiment analysis, evaluation cluster analysis, and user behavior analysis on the standardized splicing features;

[0100] S36: Based on the results of sentiment analysis, evaluation cluster analysis, and user behavior analysis, we summarize the different focuses that consumers care about when purchasing different products, and provide guiding suggestions for product production and improvement.

[0101] Furthermore, in an e-commerce platform product evaluation analysis method based on multi-source data mining, the S35 includes the following sub-steps:

[0102] S351: Sentiment analysis: The standardized concatenated features are trained using the BERT model to obtain consumers’ positive or negative emotions on different dimensions of concern for each category of products;

[0103] S352: Evaluation cluster analysis: The results of sentiment analysis are spatially clustered using the K-means algorithm combined with the user's spatial information to discover the needs and preferences of consumer groups in different regions;

[0104] S353: User Behavior Analysis: Combine the results of sentiment analysis with user behavior data to construct user portraits and understand the psychological characteristics, demand features, and behavior patterns of different user groups;

[0105] The above-mentioned S36 includes the following sub-steps:

[0106] S361: Based on the sentiment analysis of users towards products, provide personalized product recommendations, enhance user experience and conversion rate, identify target customer groups, formulate high-quality marketing strategies, and push specific products targeted.

[0107] Furthermore, an e-commerce platform product evaluation analysis method based on multi-source data mining, the above-mentioned S4 includes the following sub-steps:

[0108] S41: Use the Apache Kafka tool to collect and analyze user evaluations in real time and visualize the analysis results;

[0109] S42: Real-time track the evaluation trend, set a monitoring threshold, and when the number of negative evaluations or the results of sentiment analysis exceed the threshold, the system automatically issues an alarm;

[0110] The function of automatically issuing an alarm is for timely response;

[0111] S43: Adjust products or services in a timely manner according to real-time monitoring results, and feedback the effects of improvement measures to the monitoring system to form a continuous improvement cycle;

[0112] S44: Conduct model optimization, and obtain the evaluation index through weighted summation of precision, recall rate, and F1-scor e Regularly evaluate the model performance, identify deficiencies, collect labeled data, and enhance the generalization ability of the model. The formula is:

[0113] N = w1 * Precision + w2 * Recall + w3 * F1 Score;

[0114] Among them, N represents the evaluation index, Precision represents the precision, and Recall represents the recall rate;

[0115] Among them, w1 represents the precision weight, w2 represents the recall rate weight, w3 represents the F1-score weight, and the weight allocation where the sum of w1, w2, and w3 is 1 is adjusted according to the actual application scenario;

[0116] Among them, the labeled data is, for example, new product evaluations;

[0117] S45: Use data augmentation techniques to generate synthetic data to improve the robustness and performance of the model, fine-tune the existing model based on the new data, make adjustments for the details of specific product categories, establish a user feedback mechanism, and collect user feedback on the model output to guide the optimization direction.

[0118] Furthermore, an e-commerce platform product evaluation analysis method based on multi-source data mining, the model optimization includes the following sub-steps:

[0119] S441: Initialize the best parameter dictionary and the best metric record, set the iteration counter and the early stopping counter, create a training history record dictionary, randomly sample a set of parameter combinations from the parameter search space, and update the model using the current parameters;

[0120] S442: Train the model on the training set, make predictions on the validation set, calculate the evaluation metrics of the current parameter combination, and update the training history;

[0121] S443: Calculate the comprehensive score of the current parameter combination. If the comprehensive score exceeds the historical best score, update the best score, the best parameter combination, the best evaluation metric, and reset the early stopping counter by adding a threshold. Otherwise, increase the early stopping count;

[0122] S444: After reaching the maximum number of iterations, output the optimal parameter combination, the corresponding evaluation metrics, and the training history;

[0123] Among them, the training history is used to analyze the model optimization process.

[0124] This solution is an e-commerce platform product evaluation analysis method based on multi-source data mining, which collects product evaluation data from different channels, mines deeper connections and rules between the data, searches for potential connections between spatial locations and product evaluations, and ensures the comprehensiveness and diversity of the data; accurately discovers and processes potential negative evaluations and product problems, makes targeted improvements, improves product quality, and reduces brand reputation risks; responds to user needs and problems in a timely manner, enhances the customer service experience, increases customer satisfaction and product stickiness, enhances the brand image and market competitiveness; continuously conducts data analysis and learning, accumulates valuable market and user knowledge, and provides support for long-term development.

[0125] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for analyzing product evaluation on an e-commerce platform based on multi-source data mining, characterized in that: The following steps are involved: S1: Collect data from various platforms; S2: Data preprocessing of the platform's text data includes data unification, deduplication and cleaning, word segmentation and annotation; S3: Perform data analysis on the preprocessed text data, concatenate LDA features and BERT features and standardize them, and provide product recommendations based on the results of sentiment analysis, evaluation cluster analysis, and user behavior analysis; S4: Continuously monitor and optimize products, and establish a user feedback mechanism to collect user feedback on model output to guide optimization direction.

2. The method for analyzing product evaluation on an e-commerce platform based on multi-source data mining according to claim 1, characterized in that: The step S1 comprises the following sub-steps: S11: Collect ratings, review texts, review images, review videos, user behavior data, purchase history, and spatial data from users on the e-commerce platform; S12: Collect data from social media, blogs, and news websites.

3. The method for analyzing product evaluation on an e-commerce platform based on multi-source data mining according to claim 1, characterized in that: The step S2 comprises the following sub-steps: S21: Data unification: Use optical character recognition (OCR) to extract text information from comment images and other image data, divide comment videos into image information and audio information, use optical character recognition (OCR) to extract image information into text information, and use audio-to-text tools to convert audio information into text information; S22: Deduplication and cleaning: Deduplication and cleaning of all collected text data to obtain high-quality, low-redundancy data: S23: Word segmentation and tagging: S231: Use Chinese word segmentation tools to segment and tag the data after deduplication and cleaning, split the continuous text string into meaningful word sequences, and add a custom dictionary; The Chinese word segmentation tools include jieba, SnowNLP, and THULAC; S232: Remove words and all punctuation marks that are not important for sentiment analysis; S233: Create or use existing positive and negative sentiment word libraries, assign weights to each sentiment word, positive words are usually assigned positive weights, and negative words are assigned negative weights, match the word segmentation results with the sentiment dictionary, and assign sentiment weights to each word; S234: Check whether there are negative words before the sentiment word that affect the sentiment tendency. If there are negative words, readjust the weight of the sentiment word.

4. The method for analyzing product evaluation on an e-commerce platform based on multi-source data mining according to claim 3 is characterized in that: The S22 comprises the following sub-steps: For identical records in the data, use Python's Pandas library to remove duplicate comment data through the drop_duplicates() function; For records that are similar but not identical, a similarity threshold sml is set for judgment and processing. Data with similarity exceeding the threshold sml is directly deleted, and data with similarity below the threshold sml is retained. For missing data, set the missing degree threshold def. Data exceeding the missing degree threshold def will be directly deleted. Data below the threshold def will be filled with the mean, median, or mode method. Correct or delete errors or inconsistent values ​​in the data.

5. The method for analyzing product evaluation on an e-commerce platform based on multi-source data mining according to claim 1, characterized in that: The S3 comprises the following sub-steps: S31: Use the BERT model to extract sentence vectors from the preprocessed text data to obtain the vector representation of each text; S32: Use the LDA model to perform topic modeling on the preprocessed text data, and obtain the topic feature vector after training; S33: Concatenate LDA features and BERT features row by row. The dimension of the fused feature vector is , the fused feature vector for: , ,…, , , ,…, ； in, , ,…, represents the LDA feature, , ,…, represents BERT features, The number of topics representing LDA features, Represents the output dimension of BERT features; S34: Use the Z-score standardization method to standardize the splicing features. The formula is: ； in, represents the standardized splicing features, represents the concatenated feature value, represents the feature mean, represents the characteristic standard deviation; S35: Perform sentiment analysis, evaluation cluster analysis, and user behavior analysis on the standardized splicing features; S36: Based on the results of sentiment analysis, evaluation cluster analysis, and user behavior analysis, we summarize the different focuses that consumers care about when purchasing different products, and provide guiding suggestions for product production and improvement.

6. The method for analyzing product evaluation on an e-commerce platform based on multi-source data mining according to claim 5, characterized in that: The S35 comprises the following sub-steps: S351: Sentiment analysis: The standardized concatenated features are trained using the BERT model to obtain consumers’ positive or negative emotions on different dimensions of concern for each category of products; S352: Evaluation clustering analysis: The results of sentiment analysis are spatially clustered using the K-means algorithm combined with the user's spatial information to discover the needs and preferences of consumer groups in different regions; S353: User behavior analysis: Combine the results of sentiment analysis with user behavior data to build user portraits and understand the psychological characteristics, demand characteristics and behavior patterns of different user groups; The S36 comprises the following sub-steps: S361: Provide personalized product recommendations based on user sentiment analysis of products, improve user experience and conversion rate, identify target customer groups, develop high-quality marketing strategies, and push specific products in a targeted manner.

7. The method for analyzing product evaluation on an e-commerce platform based on multi-source data mining according to claim 1, characterized in that: The S4 comprises the following sub-steps: S41: Use Apache Kafka to collect and analyze user reviews in real time and visualize the analysis results; S42: Real-time tracking of evaluation trends, setting monitoring thresholds, and automatically issuing alarms when the number of negative reviews or sentiment analysis results exceeds the threshold; S43: Adjust products or services in a timely manner based on real-time monitoring results, and feed back the effects of improvement measures into the monitoring system to form a continuous improvement cycle; S44: Perform model optimization. Regularly evaluate model performance by obtaining evaluation indicators through weighted summation of precision, recall rate, and F1-score, identify deficiencies, collect labeled data, and enhance the generalization ability of the model. The formula is: ； in, N represents the evaluation indicator, Precision Represents accuracy, Recall represents the recall rate; in, represents the precision weight, represents the recall weight, represents the F1-score weight, The weight distribution of the sum of the three is 1 and is adjusted according to the actual application scenario; S45: Use data augmentation technology to generate synthetic data to improve the robustness and performance of the model, fine-tune the existing model based on the new data, adjust the details of specific product categories, establish a user feedback mechanism, and collect user feedback on the model output to guide the optimization direction.

8. The method for analyzing product evaluation on an e-commerce platform based on multi-source data mining according to claim 7, characterized in that: The model optimization includes the following sub-steps: S441: Initialize the best parameter dictionary and best indicator record, set the iteration counter and early stop counter, create a training history dictionary, randomly sample a set of parameter combinations from the parameter search space, and update the model using the current parameters; S442: Train the model on the training set, make predictions on the validation set, calculate the evaluation index of the current parameter combination, and update the training history record; S443: Calculate the comprehensive score of the current parameter combination. If the comprehensive score exceeds the best score in history, add the threshold to update the best score, best parameter combination, best evaluation index, and reset the early stopping counter. Otherwise, increase the early stopping count. S444: After reaching the maximum number of iterations, the optimal parameter combination, the corresponding evaluation index, and the training history record are output.

Images