Product data comparison method and system based on deep learning

Abstract

The invention discloses a product data comparison method and system based on deep learning, and relates to the technical field of electric digital data processing. The method comprises the following steps of comparison performance evaluation, comparison effect optimization and comparison result monitoring. According to the method, the product data comparison condition of the to-be-compared platform is monitored in real time in the comparison period so as to perform comparison performance judgment, so that whether product data newly-added sample quality judgment is performed or not is determined, and if the product data newly-added sample quality judgment is performed, whether deep learning model updating optimization is performed or not is determined; otherwise, product data sample change monitoring is performed, and finally, continuous monitoring is performed on a product data comparison result after deep learning model optimization, so that triggering of deep learning model optimization is limited, the accuracy of the product data comparison result is improved, and the accuracy of the product data comparison result is improved. The problem that the accuracy of a product data comparison result is reduced due to updating lag of a deep learning model in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to a product data comparison method and system based on deep learning. Background Technology

[0002] Currently, e-commerce platforms, inventory management, price monitoring and other fields often require automated processing of large amounts of product data to achieve efficient matching, classification and identification. Therefore, there is an urgent need for a method to compare product data. The implementation process of existing technologies usually includes multiple steps such as data preparation and preprocessing, feature extraction, similarity calculation, deep learning model training, comparison and matching and result evaluation.

[0003] Data preparation and preprocessing includes multiple stages such as data collection, data cleaning, text preprocessing, and image preprocessing, primarily targeting product data from various sources, which may include product names, descriptions, specifications, prices, and images. Feature extraction mainly involves text feature extraction, such as using word embedding models to convert text data into vector form, or image feature extraction, such as using convolutional neural networks to extract deep features from images. Commonly used models include VGG (Visual Geometry Group Network) and ResNet (Residual Geometry Group Network). Networks (including residual networks) and structured data processing are used. For structured data (such as price, size, weight, etc.), standardization and normalization methods can be used. Similarity calculation mainly includes text similarity, image similarity, and joint similarity. Deep learning model training first designs the neural network and loss function, and then trains the model. Comparison and matching are used for products to be compared. The similarity between the comparison and the target product is calculated and ranked. A threshold is set according to the similarity value to determine which products can be considered as matches. If the data contains multiple modalities (such as text and images), a multimodal learning model can be used to combine text and image information for more accurate comparison. Result evaluation compares the comparison results with the actual results to evaluate the accuracy of the model. It is also necessary to evaluate the model's ability to identify positive examples. Especially for high-precision tasks such as product comparison, it is usually necessary to focus on the balance between precision and recall.

[0004] For example, Chinese invention patent CN116450645A discloses a data comparison method, apparatus, and storage medium based on back-end products, which includes: acquiring first data; the first data being data to be compared exported from a database; determining second data based on the first data; the second data being data associated with the first data exported from a data dictionary; and comparing the first data and the second data.

[0005] For example, the spatial data comparison method and related products disclosed in Chinese invention patent CN110738640B include: collecting detection data of a spatial detection area, the detection data including video data; obtaining key information from the video data, and downloading the standard detection template corresponding to the detection data from the server based on the key information, the standard detection template including historical data or spatial data model; and generating detection results based on the comparison results of the detection data and the standard detection template.

[0006] The above-mentioned technology has at least the following technical problems: Although the system possesses incremental learning and self-learning capabilities, deep learning models often update slowly, especially when new products are introduced or product data undergoes large-scale changes, making the model update cycle insufficiently timely. For e-commerce platforms, product data is updated frequently and is often dynamically changing (e.g., price adjustments, new product launches), so the system may lag behind actual market changes. This impacts the results, leading to a decrease in matching accuracy. For example, older models may fail to recognize subtle differences in new versions (e.g., new color schemes, new specifications), potentially misclassifying different products as identical, resulting in incorrect merging when the comparison system integrates cross-platform or cross-brand data. Summary of the Invention

[0007] To address the technical problem of reduced accuracy in product data comparison results due to lagging deep learning model updates in existing technologies, this invention provides a product data comparison method and system based on deep learning. The technical solution is as follows: On the one hand, a product data comparison method based on deep learning is provided. This method includes: real-time monitoring of the product data comparison status of the comparison platform during the comparison period to determine the comparison performance, thereby determining whether to perform quality assessment of newly added product data samples to identify the influencing factors of the product data comparison results; if quality assessment of newly added product data samples is performed, determining whether to update and optimize the deep learning model to improve the timeliness of the deep learning model for product data comparison; otherwise, monitoring changes in product data samples to determine whether changes in product data affect the comparison results; and continuously monitoring the product data comparison results after the deep learning model is optimized to limit the triggering of deep learning model optimization in subsequent comparison periods, thereby preventing the deep learning model from engaging in cyclical optimization.

[0008] On the other hand, a product data comparison system based on deep learning is provided. This system includes: a comparison performance evaluation module, a comparison effect optimization module, and a comparison result monitoring module. The comparison performance evaluation module monitors the product data comparison status of the platform to be compared in real time during the comparison period to determine the comparison performance and whether to perform a quality assessment of newly added product data samples to identify influencing factors on the product data comparison results. The comparison effect optimization module determines whether to update and optimize the deep learning model if a quality assessment of newly added product data samples is performed, thereby improving the timeliness of the deep learning model for product data comparison. Otherwise, it monitors changes in product data samples to determine whether these changes affect the comparison results. The comparison result monitoring module continuously monitors the product data comparison results after the deep learning model optimization, limiting the triggering of deep learning model optimization in subsequent comparison periods to prevent the deep learning model from undergoing cyclical optimization.

[0009] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. Real-time monitoring of product data comparison across different platforms during the comparison period helps to assess comparison performance and avoids the problem of reduced timeliness and accuracy of comparison results due to continuous changes in existing technology product data. It also allows for rapid response to changes in product data and models, ensuring consistently high quality and accuracy of comparison results. Next, it determines whether to assess the quality of newly added product data samples, helping to identify factors influencing the comparison results and promptly detect potential quality issues with new product data. If a quality assessment is conducted, it determines whether to update and optimize the deep learning model, thereby improving the timeliness of the deep learning model for product data comparison and ensuring that deep learning is only performed when data quality meets requirements. Model optimization ensures that each optimization is based on high-quality training data, reducing over-optimization caused by substandard data. Otherwise, product data sample changes are monitored to determine whether these changes affect the results. Finally, the product data comparison results after deep learning model optimization are continuously monitored, enabling real-time optimization of the comparison results and ensuring the accuracy of the deep learning model output. This limits the triggering of deep learning model optimization in subsequent comparison cycles, preventing the problem of cyclical optimization. It also prevents the traditional problem of over-optimization or model degradation caused by frequent deep learning model updates and optimizations, making the deep learning model less susceptible to excessive optimization, ensuring the stability of comparison performance, and avoiding performance fluctuations over long periods.

[0010] 2. The newly added sample set within the current comparison period is re-compared to obtain the corresponding comparison performance value, which is recorded as the new sample comparison performance value. This solves the technical problem in existing product data comparison processes where new samples are often directly incorporated into the training or comparison set without separately evaluating their impact on overall comparison performance. Furthermore, by independently comparing the new sample set and calculating the comparison performance value, it helps to separately assess the actual impact of new data, enabling quantitative monitoring of system performance changes after the introduction of new data. This provides a data foundation for subsequent quality checks and model optimization. Next, the difference between the comparison performance value of the previous comparison period and the new sample comparison performance value is quantified to obtain the new comparison performance performance value difference value. This overcomes the deficiency of traditional comparison systems that only monitor performance in a single period, establishing a cross-period performance difference quantification mechanism. This is beneficial for detecting the performance impact of new samples. Then, the new comparison performance value is... The performance difference value is compared with the comparison effect difference threshold. If the performance difference value of the newly added comparison effect is greater than the comparison effect difference threshold, a data quality check is performed on the newly added sample set. The data quality check includes one or more of the following: label quality check, feature quality check, distribution drift check, text data quality check, and image data quality check. Triggering the data quality check through performance difference detection can accurately identify the root cause at the data level before model optimization, avoiding blindly optimizing the model and causing further performance fluctuations. If the performance difference value of the newly added comparison effect is not greater than the comparison effect difference threshold, the deep learning model is updated and optimized. This prevents the traditional model update mechanism, which is usually executed periodically or manually, from adaptively judging whether optimization is needed based on performance changes, resulting in update lag or resource waste. Through this adaptive optimization mechanism, unnecessary training computational resource consumption is avoided, and the efficiency of the comparison system is improved.

[0011] 3. By comparing the product data change parameters obtained within the current comparison period with the corresponding extracted thresholds for the proportion of newly added samples and the distribution drift threshold, this approach helps address the lack of clear optimization trigger conditions in traditional model optimization mechanisms. These mechanisms often rely on manual experience or fixed-period triggering, leading to premature or delayed optimization. This not only standardizes the quantitative judgment of optimization triggers but also improves the accuracy and stability of deep learning model optimization triggers. If only the proportion of newly added samples exceeds the threshold, model learning intensity optimization is performed, avoiding the problem of insufficient model learning caused by traditional models using a fixed learning rate for updates. This allows the model to learn more fully in scenarios with rapidly changing data, thereby improving the accuracy of new sample comparisons. If only the distribution drift exceeds the distribution drift threshold, model learning rate optimization is performed, solving the problem of traditional optimization methods struggling to adjust model learning in a timely manner. Convergence rate can lead to slow model convergence or getting stuck in local optima, affecting real-time comparison performance. By optimizing the learning rate when distribution drift exceeds the limit, the convergence speed of the model under the new distribution is accelerated. If both the proportion of new samples and the distribution drift exceed the corresponding thresholds for the proportion of new samples and the distribution drift, comprehensive optimization of model learning is performed. This ensures that the deep learning model can fully absorb the information of new samples and quickly adapt to the new data distribution. If neither the proportion of new samples nor the distribution drift exceeds the corresponding thresholds for the proportion of new samples and the distribution drift, the performance of product data comparison in the next comparison period is monitored, and a comparison result analysis report for the current comparison period is output. This prevents the traditional system from performing periodic model updates when the performance is stable, which would waste computing resources. This improves the system's operating efficiency and resource utilization, and forms a periodic feedback closed-loop mechanism to support continuous optimization and decision support. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating a product data comparison method based on deep learning provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the process for determining the quality of newly added product data samples provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the process for updating and optimizing a deep learning model provided in an embodiment of the present invention; Figure 4This is a schematic diagram of the structure of a product data comparison system based on deep learning provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the architecture of the convolutional neural network provided in an embodiment of the present invention. Detailed Implementation

[0014] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0015] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0016] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0017] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0018] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0019] This invention provides a product data comparison method and system based on deep learning. For example... Figure 1 The illustrated embodiment of the present invention provides a flowchart of a product data comparison method based on deep learning. The processing flow of this method may include the following steps: During the comparison period, the product data comparison status of the platform to be compared is monitored in real time to determine the comparison performance, thereby determining whether to add new product data samples for quality assessment, and thus identifying the influencing factors of the product data comparison results.

[0020] It should be noted that the comparison platform usually refers to a platform that needs to automatically process large amounts of product data, such as e-commerce platforms, inventory management, and price monitoring. Furthermore, the deep learning models used vary depending on the product description text. For product description text comparison, deep learning models learn the semantic features of the text to determine whether the two match, such as convolutional neural networks or recurrent neural networks. For image comparison, especially the similarity comparison between product images, deep learning provides many effective methods, such as convolutional neural networks or Siamese networks.

[0021] If the quality of newly added product data samples is assessed, it will be determined whether to update and optimize the deep learning model to improve the timeliness of the deep learning model's comparison of product data. Otherwise, changes in product data samples will be monitored to determine whether the changes in product data will affect the comparison results.

[0022] Continuously monitor the comparison results of product data after deep learning model optimization to limit the triggering of deep learning model optimization in subsequent comparison cycles, thereby preventing the deep learning model from engaging in cyclical optimization.

[0023] In this embodiment, by performing real-time performance monitoring of the comparison results during the comparison period, the system can promptly capture the trend of model performance changes. This helps to quickly identify whether the decline in comparison performance is due to data quality issues or model aging, thereby achieving accurate determination of the cause of performance changes. Compared with traditional periodic detection methods, this significantly improves the timeliness and response speed of comparison performance monitoring. Furthermore, when the system detects abnormal comparison performance, it prioritizes determining whether it is a problem with the quality of newly added sample data, thus implementing a hierarchical processing mechanism of "data first, model later." If it is confirmed that there are quality problems with the newly added sample data, data quality correction or abnormal samples can be performed directly before model optimization, avoiding erroneous optimization caused by blindly updating the model. If the data quality is good, the system can automatically trigger model update optimization to improve the model's adaptability and generalization performance in new data scenarios. This helps to distinguish between data problems and model problems, avoids unnecessary model retraining, and improves the overall system stability and optimization efficiency. Moreover, when model update optimization is not required, the system will switch to monitoring changes in product data samples, improving the robustness of the comparison system, ensuring the scientific nature of model updates, and thus improving the accuracy of the comparison results.

[0024] Further, determine whether to conduct a quality assessment of the new product data samples. The specific process is as follows: Within a pre-set comparison period, the performance of the product data of the platform to be compared is monitored in real time to obtain the corresponding comparison performance value. The difference between the comparison performance value of the current comparison period and the comparison performance value of the previous comparison period is quantified to obtain the change in the comparison performance value. Specifically, the comparison performance value of the current comparison period is subtracted from the comparison performance value of the previous comparison period to obtain the change in the comparison performance value. When the comparison period is the first comparison period, the change in the comparison performance value is obtained by subtracting the average comparison performance value over the historical time period.

[0025] It should be explained that the comparison period refers to a fixed time period length, which is preset by professional technicians; the performance value of the comparison is usually represented by the F1 score, which is the harmonic mean of precision and recall. It is a comprehensive indicator that measures the precision and recall of a model and directly reflects the performance of the deep learning model when comparing product data. If the F1 score drops by more than a certain threshold (e.g., 3%), it indicates that the deep learning model has significantly degraded in accurately matching product data. This is usually the primary signal to update the model. Calculating the F1 score requires knowing the labels predicted by the model and the actual true labels, which is usually automatically generated by the deep learning model evaluation system.

[0026] The obtained change in the comparison performance value is compared with the set threshold for the change in the comparison performance value. The threshold for the change in the comparison performance value is extracted from the preset database and is obtained by professional technicians by averaging the comparison performance values ​​of the comparison results that are not qualified in the historical data. It is recorded as the threshold for the change in the comparison performance value and stored in the preset database for direct reading when needed.

[0027] If the change in the comparison performance value exceeds the threshold for the change in the comparison performance value used to limit the decline in the comparison performance value, the quality of newly added product data samples will be judged to determine whether to update and optimize the deep learning model; otherwise, the change in product data samples will be monitored.

[0028] In this embodiment, by monitoring the performance of product data comparison in real time within each preset comparison period and calculating the comparison effect performance value, the system can perform periodic, quantitative, and traceable performance evaluation of the comparison effect. Furthermore, by calculating the difference between the performance value and the previous comparison period, the resulting performance change can more intuitively reflect the model's performance change trend across different periods, realizing the transformation of comparison performance from "qualitative judgment" to "quantitative evaluation." This also improves the accuracy and interpretability of comparison performance monitoring, providing reliable numerical basis for subsequent optimization decisions. Moreover, comparing the change in the comparison effect performance value with a preset performance change threshold enables automated judgment of performance fluctuations. Additionally, when the performance change exceeds the threshold, the system automatically identifies it as an "abnormal fluctuation" state, triggering corresponding data quality judgment or model optimization decisions. This mechanism helps avoid the instability and lag problems caused by relying on manual judgment or fixed-period updates in traditional solutions. By constructing a quantitative definition and automatic identification mechanism for performance fluctuations, manual intervention is reduced, and the system's self-decision-making capability is improved.

[0029] like Figure 2 The diagram illustrates the process for determining the quality of newly added product data samples according to an embodiment of the present invention. The specific logic is as follows: The newly added sample set within the current comparison period is re-compared to obtain the corresponding comparison performance value, which is recorded as the new sample comparison performance value. The difference between the comparison performance value of the previous comparison period and the new sample comparison performance value is quantified to obtain the new comparison performance value difference value. This new comparison performance value difference value is compared with a comparison performance difference threshold: if the new comparison performance value difference value is greater than the comparison performance difference threshold, a data quality check is performed on the new sample set; if the new comparison performance value difference value is not greater than the comparison performance difference threshold, a deep learning model update and optimization are performed. This process helps improve the stability and reliability of the comparison system, as well as the accuracy, real-time performance, and robustness of the comparison results.

[0030] Furthermore, the specific process for determining the quality of newly added product data samples is as follows: The newly added sample set in the current comparison period is re-compared to obtain the corresponding comparison performance value, which is recorded as the comparison performance value of the new sample. The difference between the performance value of the previous comparison period and the performance value of the new sample comparison is quantified to obtain the difference value of the new comparison performance; that is, the difference value of the new comparison performance is obtained by performing a difference calculation between the performance value of the previous comparison period and the performance value of the new sample comparison.

[0031] The newly added performance difference value is compared with the performance difference threshold value, which represents the minimum difference that limits the change in the performance value. The performance difference threshold value is preset by professional technicians based on experience rules and stored in a preset database.

[0032] If the difference in the performance value of the newly added comparison results is greater than the threshold value for the difference in comparison results, it indicates that there may be a problem with the data quality of the newly added sample set. In this case, a data quality check is performed on the newly added sample set. The data quality check includes one or more of the following: label quality check, feature quality check, distribution drift check, text data quality check, and image data quality check.

[0033] If the difference in the performance value of the newly added comparison is not greater than the threshold value for the difference in the comparison, then the deep learning model will be updated and optimized.

[0034] In this embodiment, by independently comparing the newly added sample set and calculating the performance value of the comparison effect, not only is the actual impact of the new data assessed separately, but also the performance changes of the system after the introduction of the new data are quantitatively monitored. This provides a data foundation for subsequent quality checks and model optimization, and helps to achieve dynamic data adaptation at the model level, improving the robustness of the model under new samples. At the same time, a cross-period performance difference quantification mechanism is established, which is beneficial to detecting the performance impact of the new samples. The quantification results of the difference value not only provide an objective basis for subsequent data quality judgment, but also reduce subjective judgment or reliance on experience. In addition, by tracking the difference over continuous periods, a performance trend curve of the comparison system can be formed, which helps to provide early warning of model performance degradation. By introducing a difference definition threshold, a more objective standard for judging performance changes is established, realizing the quantitative management of performance fluctuations and avoiding misjudgments caused by accidental fluctuations or slight data noise, thus improving the stability and reliability of the comparison system.

[0035] like Figure 3The diagram illustrates the process for updating and optimizing a deep learning model according to an embodiment of the present invention. The specific logic is as follows: Obtain the product data change parameters within the current comparison period, and compare these parameters with the corresponding extracted thresholds for the proportion of newly added samples and distribution drift. If only the proportion of newly added samples exceeds the threshold, optimize the model learning intensity; if only the distribution drift exceeds the threshold, optimize the model learning rate; if both the proportion of newly added samples and the distribution drift exceed the corresponding thresholds, perform comprehensive optimization of model learning; if neither the proportion of newly added samples nor the distribution drift exceeds the corresponding thresholds, continue monitoring the performance of the product data comparison in the next comparison period and output an analysis report of the comparison results for the current comparison period. Through this process, not only is refined monitoring of the comparison data status achieved, but the accuracy of comparing product data is also improved.

[0036] Furthermore, the specific process for updating and optimizing the deep learning model is as follows: Obtain the product data change parameters within the current comparison period. The product data change parameters include the proportion of new samples and distribution drift.

[0037] Specifically, the percentage of new samples can be calculated through product databases and product update logs. Usually, the comparison platform will have incremental data records (such as the number of products per month or new product categories), which can be calculated through the update records of the data table. Through statistical analysis of product distribution, changes in product distribution can be obtained. It is usually necessary to compare the historical distribution with the current distribution and use distribution testing methods in data science to assess whether there is drift and obtain the corresponding distribution drift.

[0038] The product data change parameters are compared with the new sample proportion threshold and distribution drift threshold extracted from the preset database. The new sample proportion threshold and distribution drift threshold are extracted from the preset database, and both are preset data by professional technicians. Professional technicians usually set them in advance based on experience rules and store them in the corresponding preset database.

[0039] If only the proportion of newly added samples exceeds the threshold for the proportion of newly added samples, then the model learning intensity is optimized; if only the distribution drift exceeds the threshold for the distribution drift, then the model learning rate is optimized; if both the proportion of newly added samples and the distribution drift exceed their respective thresholds for the proportion of newly added samples and the distribution drift, then a comprehensive optimization of model learning is performed; if neither the proportion of newly added samples nor the distribution drift exceeds their respective thresholds for the proportion of newly added samples and the distribution drift, then the performance of the product data comparison in the next comparison period continues to be monitored, and a comparison result analysis report for the current comparison period is output.

[0040] In this embodiment, by introducing quantitative indicators of product data change parameters (proportion of new samples and distribution drift), refined monitoring of the data status of the comparison is achieved, providing interpretable data basis for subsequent model optimization. This helps to distinguish whether the root cause of performance fluctuations is data change or model problems. At the same time, it also realizes real-time visualization and quantifiable management of data dynamics within the comparison period, thereby improving the system's sensitivity and response speed to environmental changes. In addition, by setting and comparing thresholds (threshold for proportion of new samples and threshold for distribution drift), the quantitative judgment standardization of optimization triggers is also realized. Moreover, by separately monitoring the proportion of new samples, when a significant increase in the number of new samples is detected, the system automatically optimizes the learning intensity (e.g., increasing the learning rate, extending the number of training rounds, or expanding the training batch) to improve the model's ability to absorb new data. This allows the model to learn more fully in scenarios with rapidly changing data volume, thereby improving the comparison accuracy of new samples and the overall generalization ability. This avoids the "under-learning" problem caused by fixed learning parameters and enhances the model's adaptability in dynamic data environments.

[0041] It should be added that the model learning intensity optimization process is as follows: By comparing the change in the performance value with the threshold of the change in the performance value, and the deviation between the proportion of new samples and the threshold of the proportion of new samples, the corresponding optimization amount of model learning intensity is obtained.

[0042] Specifically, the comparison performance deviation value is obtained by averaging the difference between the change in the comparison performance value and the threshold for the change in the comparison performance value. The new sample deviation value is obtained by averaging the difference between the proportion of new samples and the threshold for the proportion of new samples. Finally, the model learning intensity optimization amount is obtained by averaging the comparison performance deviation value and the new sample deviation value.

[0043] The optimization amount of the model learning intensity is input into the learning intensity optimization function after training to obtain the corresponding learning intensity optimization value.

[0044] It should be added that the learning intensity optimization function represents the mapping relationship between the learning intensity optimization amount and the learning intensity optimization value of the fitted model. It is constructed based on the logistic regression algorithm, uses the cross-entropy loss function as the optimization criterion, and is trained through the scikit-learn framework. It is trained based on the learning intensity optimization training data, which includes the model learning intensity optimization amount within historical time periods, as well as the learning intensity optimization value set by professional technicians based on empirical rules.

[0045] The learning intensity optimization ratio and the learning intensity optimization value are adjusted based on the set learning intensity optimization ratio and the learning intensity optimization value. This means that the learning intensity optimization ratio and the learning intensity optimization value are multiplied to obtain the training batch optimization amount and the training cycle optimization amount, respectively. The learning intensity optimization ratio includes the batch optimization ratio and the cycle optimization ratio.

[0046] The training batch size and training cycle size of the deep learning model are increased by optimizing the training batch size and the training cycle size, respectively; that is, by multiplying the optimization size of the training batch size and the optimization size of the training cycle size with the training batch size and the training cycle size, respectively.

[0047] The increased training batch size and training period are compared with the comparison processing limits set according to the product data comparison rules. The comparison processing limits include the training batch size and training period size.

[0048] If the increased training batch size and training period length both meet the comparison processing constraints, then the corresponding model learning intensity optimization is performed based on the increased training batch size and training period length; otherwise, the corresponding model learning intensity optimization is performed based on the comparison processing constraints.

[0049] In this embodiment, the changes in performance values, sample proportions, and threshold deviations are smoothed to comprehensively calculate the optimized learning intensity of the model. This method smooths multi-dimensional performance change data, reducing over- or under-optimization caused by fluctuations in a single indicator. Furthermore, smoothing effectively reduces noise impact and prevents the model from over-responding to single changes. This not only ensures the stability and rationality of the optimization process but also improves the accuracy of optimization decisions and reduces overreaction to sudden data changes, guaranteeing the stability of the optimization effect. By inputting the calculated optimized learning intensity into the optimized learning intensity function, a precise optimized learning intensity value is obtained. This process can... The system dynamically adjusts the learning intensity of the deep learning model based on performance comparison and sample quality, adaptively adjusting the training strategy during the learning process. Furthermore, by adjusting the training intensity of batches and periods, the system ensures that the model maintains optimal efficiency in learning new data, avoiding the risks of instability or overtraining in traditional training. This ensures that the model learns efficiently and accurately under different data conditions, improving resource utilization and training effectiveness in the deep learning process. Moreover, the optimization ratio of the model can be automatically adjusted according to actual needs, optimizing the batch and period size to ensure efficient training and meet performance optimization requirements. Through refined optimization ratio settings, the system ensures both efficiency and flexibility in the training process.

[0050] Furthermore, the specific process for optimizing the model learning rate is as follows: By comparing the change in the performance value with the threshold of the change in the performance value, and by smoothing the deviation between the distribution drift and the distribution drift threshold, the corresponding optimization amount of the model learning rate is obtained.

[0051] Similarly, the comparison performance deviation value is obtained by comparing the difference between the change in the comparison performance value and the threshold value of the comparison performance value, and the comparison performance value change threshold value. The distribution drift deviation value is obtained by comparing the difference between the distribution drift and the distribution drift threshold value, and the distribution drift deviation value is obtained by averaging the comparison performance deviation value and the distribution drift deviation value.

[0052] By substituting the optimized learning rate of the model into the optimized learning rate function after training, the corresponding optimized learning rate value is obtained.

[0053] It should be explained that the learning rate optimization function represents the mapping relationship between the learning rate optimization amount and the learning rate optimization value of the fitted model. It is trained based on the learning rate optimization training data, which includes the model learning rate optimization amount within a historical time period, as well as the learning rate optimization value set by professional technicians based on empirical rules.

[0054] The optimized learning rate is input into an adaptive optimizer (such as Adam, Adagrad, RMSprop, etc.) to dynamically adjust the learning rate. That is, the learning rate is adaptively adjusted according to the changes in the gradient, so that the model has different learning strategies at different stages.

[0055] In this embodiment, by smoothing the change in the comparison performance value with the threshold of the change in the comparison performance value, as well as the deviation between the distribution drift and the distribution drift threshold, the comparison performance and data distribution changes can be comprehensively evaluated to obtain an accurate learning rate optimization amount. This smoothing process can reduce the impact of data fluctuations on the model training strategy, making the optimization process more stable and less susceptible to interference from single abnormal data. It not only improves the accuracy and stability of optimization decisions but also reduces misjudgments caused by abnormal data or local fluctuations, ensuring the long-term effectiveness of the optimization strategy. Furthermore, by substituting the obtained learning rate optimization amount into the already trained learning rate optimization function, an accurate learning rate optimization value can be calculated. This allows for adaptive adjustments to the current training stage of the model. By dynamically optimizing the learning rate, the model can adopt the most suitable learning strategy at different data stages, thereby improving training efficiency and accuracy. This achieves precise control over the model's learning rate, enabling the model to flexibly adjust the learning rate at different training stages and avoiding the problem of learning too fast or too slow caused by a fixed learning rate in traditional training.

[0056] Furthermore, the specific process of model learning and comprehensive optimization is as follows: After normalizing the differences between the change in the comparison performance value and the threshold for the change in the comparison performance value, the ratio of new samples to the threshold for the ratio of new samples, and the distribution drift to the threshold for the distribution drift, a weighted fusion operation is performed to obtain the representative value of the model learning comprehensive optimization.

[0057] It should be added that the differences between the change in the performance value of the comparison effect and the threshold of the change in the performance value of the comparison effect, the difference between the proportion of new samples and the threshold of the proportion of new samples, and the difference between the distribution drift and the distribution drift threshold are obtained by subtraction. After normalizing the above differences, they are weighted and fused with the assigned difference weights to obtain the representative quantity of the model learning comprehensive optimization. The difference weights include the comparison effect weight, the proportion of new samples weight, and the distribution drift weight. The above difference weights are all read from the preset database and are preset fixed weights. The sum of the three is 1, which is preset for professional technicians.

[0058] Based on the model learning comprehensive optimization representative quantity, the corresponding training data quantity adjustment value and learning rate influence factor are mapped in the trained model learning comprehensive optimization mapping model.

[0059] It should be noted that the model learning comprehensive optimization mapping model is built based on the logistic regression algorithm and relies on the least squares criterion. The training process is completed through the statsmodels framework. The model learning comprehensive optimization mapping model is used to reflect the mapping relationship between the comprehensive optimization representative quantity of model learning, the training data volume adjustment value, and the learning rate influence factor. At the same time, it is trained based on comprehensive optimization training data, which includes the comprehensive optimization representative quantity of model learning within a historical time period, as well as the training data volume adjustment value and the learning rate influence factor set by professional technicians according to empirical rules.

[0060] The model learning intensity and model learning efficiency are enhanced by adjusting the training data volume and the learning rate influence factor, respectively. Specifically, the training data volume adjustment value is multiplied by the training batch and training period of the model learning intensity, and the learning rate influence factor is multiplied by the learning rate optimization value of the model learning efficiency.

[0061] In this embodiment, the scheme normalizes the differences between the performance value change and its threshold, the proportion of new samples and the threshold, and the distribution drift and the threshold, respectively. This achieves a unified numerical scale among different indicators, making the influence of each parameter on the model optimization decision more balanced. Moreover, through normalization, the bias caused by the difference in the dimensions of different indicators is eliminated, avoiding the problem that a certain parameter may dominate the optimization direction due to its excessively large or small numerical magnitude. This not only improves the comparability and processing accuracy of the model optimization input data, but also achieves reasonable fusion of multi-dimensional indicators under a unified scale, ensuring the accuracy of the optimization calculation results. At the same time, this scheme constructs a model learning comprehensive optimization mapping model during the training phase, enabling the system to map the comprehensive optimization representative quantity to the training data volume adjustment value and the learning rate influence factor. This mapping model can automatically deduce the optimal training data volume adjustment strategy and learning rate adjustment strategy based on different optimization representative quantities, realizing intelligent and automated parameter control. Through the mapping relationship, intelligent decision-making of complex parameters is achieved, reducing manual intervention; improving the response speed and decision accuracy of the optimization process, and realizing self-learning model optimization control.

[0062] Furthermore, the specific steps for monitoring changes in product data samples are as follows: Step 1: Obtain the proportion of newly added samples within the comparison period and compare it with the set threshold for the proportion of newly added samples. If the proportion of newly added samples is greater than the threshold, then the intensity label representative value is recorded as the first sample representative value; otherwise, the intensity label representative value is recorded as the second sample representative value. The intensity label representative value represents the degree of model learning intensity optimization that needs to be performed, and the degree of model learning intensity optimization gradually decreases from the first sample representative value to the second sample representative value.

[0063] Step 2: Compare the distribution drift within the comparison period with the distribution drift threshold: If the distribution drift is greater than the distribution drift threshold, then the rate marker representative value is recorded as the first representative value of the distribution drift; otherwise, the rate marker representative value is recorded as the second representative value of the distribution drift. The rate marker representative value represents the degree of optimization of the model learning rate that needs to be performed, and the degree of optimization of the model learning rate gradually decreases from the first representative value of the distribution drift to the second representative value of the distribution drift.

[0064] Step 3: Combine the obtained intensity and velocity representative values ​​to obtain the corresponding labeling results, which include four types of labeling results: The first approach involves adjusting the model learning process based on the preset initial values ​​for model learning intensity and model learning rate, if the labeling results include the first representative value of the first sample and the first representative value of the distribution drift.

[0065] The second approach is to adjust the model learning intensity based on the preset initial value of the model learning intensity adjustment if the labeling result includes the first representative value and the second representative value of the distribution drift.

[0066] The third method involves obtaining a set number of labeling results within a specified comparison period if the labeling results include the second representative sample quantity and the first representative quantity of distribution drift. If the ratio of the labeling results to the second representative sample quantity and the first representative quantity of distribution drift exceeds the set ratio, the model learning rate is adjusted based on the initial value of the model learning rate adjustment. Otherwise, the product data comparison results are output.

[0067] Fourthly, if the labeling result includes the second representative quantity of the sample and the second representative quantity of the distribution drift, then the corresponding product data comparison result will be output.

[0068] In summary, the initial values ​​for adjusting the model learning intensity, the initial values ​​for adjusting the model learning rate, the set quantity, and the set ratio are all read and used from the preset database. These data are all preset and stored in the preset database by professional technicians based on experience rules.

[0069] In this embodiment, by comparing the proportion of newly added samples with a set threshold for the proportion of newly added samples, it can be determined whether the model needs to be optimized for learning intensity within the current comparison period. If the proportion of newly added samples exceeds the threshold, an intensity marker representative is used as the first sample representative, thereby increasing the model's learning intensity; conversely, a second sample representative is used to gradually reduce the optimization requirement for model learning intensity. Simultaneously, for distribution drift, the optimization requirement for model learning rate is determined based on the comparison result with a set distribution drift threshold. Rate optimization is achieved by setting a rate marker representative, providing a mechanism for automatically adjusting learning intensity and learning rate based on changes in data features, reducing human intervention, and improving the model's adaptability and intelligence. Moreover, the model learning strategy is dynamically adjusted according to the specific circumstances of changes in the proportion of newly added samples and distribution drift, avoiding a fixed approach. Overfitting or low training efficiency caused by optimization strategies can be mitigated. By combining strength and rate representative values, the model can make more accurate optimization decisions in different training scenarios. If the labeling result is the first representative value and the second representative value of distribution drift, only the learning strength adjustment is performed; if it is the second representative value and the first representative value of distribution drift, then the learning rate adjustment is performed. This flexible combination based on multi-dimensional labeling results ensures that the optimal optimization strategy is adopted in different situations, avoiding optimization imbalance caused by a single factor. By combining multiple representative values, the system can make more accurate comprehensive adjustments for different comparison periods. Furthermore, adopting a comprehensive optimization strategy is beneficial to improving the model's performance and adaptability in different data environments, avoiding overtraining or inefficiency that may result from relying solely on a single optimization strategy.

[0070] Furthermore, the specific details regarding the restrictions on triggering deep learning model optimization are as follows: First, monitor the performance values ​​of the comparison results in each comparison period after optimization in real time, and detect whether the deep learning model is triggered for updating and optimizing. If it is triggered, the deep learning model is updated and optimized.

[0071] Next, if deep learning model updates and optimizations are triggered for more than the preset number of comparison cycles, then deep learning model updates and optimizations will not be performed within a preset time period. The preset number of comparison cycles and the preset time period are set by professional technicians. They are pre-set and stored in a preset database based on comparison requirements and experience rules for retrieval and use.

[0072] Finally, the trigger frequency for deep learning model updates and optimizations within a preset time period is obtained and denoted as the trigger update frequency. The trigger update frequency is then matched with a frequency range, which is a data range set by technical personnel to determine whether the trigger update frequency is abnormal.

[0073] If the trigger update frequency is lower than the minimum value of the frequency range, the deep learning model will still not be updated or optimized, and the corresponding product data comparison results will not be output.

[0074] If the trigger update frequency is higher than the maximum value of the frequency range, the current model update parameters will be fed back to the manual management end, and a request will be made as to whether to continue the model update mechanism.

[0075] If the deep learning model is not updated or optimized within the preset time period, the product data comparison results will be output.

[0076] In this embodiment, the optimized comparison performance value is monitored in real time during each comparison period. This allows for timely detection of model performance fluctuations and dynamic determination of whether model updates and optimizations need to be triggered. This not only improves the accuracy and real-time performance of deep learning model optimization triggers but also enables fine-grained monitoring of model performance changes. It ensures timely optimization when performance declines or data distribution changes, while avoiding the problem of reduced accuracy of comparison results due to delayed optimization. When model updates and optimizations are detected to be triggered in multiple consecutive comparison periods, the system automatically determines that there is an over-optimization trend. Therefore, it sets a time limit for no further deep learning model updates and optimizations. This helps prevent the model from being repeatedly trained or cyclically optimized in a short period of time, avoiding model oscillations and performance instability caused by frequent updates. Furthermore, controlling the model update frequency makes the optimization behavior more stable, ensuring that the model is updated only after sufficient sample accumulation and performance monitoring, thus improving the effectiveness and stability of optimization.

[0077] Figure 4This is a schematic diagram of the structure of a product data comparison system based on deep learning provided in an embodiment of the present invention. (Refer to...) Figure 4 The system includes a comparison performance evaluation module, a comparison effect optimization module, and a comparison result monitoring module; The comparison performance evaluation module is used to monitor the comparison of product data on the comparison platform in real time during the comparison period to determine the comparison performance and thus determine whether to make a quality judgment on the new product data samples, so as to determine the influencing factors of the product data comparison results.

[0078] The comparison effect optimization module is used to determine whether to update and optimize the deep learning model if the quality of newly added product data samples is judged, so as to improve the timeliness of the deep learning model for product data comparison. Otherwise, it monitors changes in product data samples to determine whether changes in product data affect the comparison results.

[0079] The comparison result monitoring module is used to continuously monitor the comparison results of product data after deep learning model optimization, so as to limit the triggering of deep learning model optimization in subsequent comparison cycles, thereby preventing the deep learning model from having a problem of cyclic optimization.

[0080] In this embodiment, a convolutional neural network is used as an example. Assume the system is used to compare the appearance consistency of different batches of products, such as image detection tasks for mobile phone cases, bottle printing, and metal parts. The input is product image samples collected within the current comparison period. Each image can be an RGB channel matrix, for example, with a size of 224×224×3. The output is the classification probability (e.g., qualified, abnormal) or similarity score of the product comparison result. Figure 5The diagram shows the architecture of a convolutional neural network provided in an embodiment of the present invention. The input layer takes a 224×224×3 RGB image as input, which is a three-channel color image. Next are convolutional layers (C1, C2) and pooling layers (S1, S2). Layers C1 and C2 use 5×5 convolutional kernels to extract image features. These convolutional layers help the network extract low-level features (such as edges, textures, and shapes) from the input image. Layers S1 and S2 use 2×2 subsampling to reduce the size of the feature map while retaining important information. Then, feature extraction is performed. After convolution and pooling operations, the extracted features are passed to a fully connected layer for the final classification task. This stage is used to extract key information from the product image, providing useful features for subsequent classification. The final output layer outputs the "classification probability of the product comparison result", that is, the probability of being qualified or abnormal. This means that the network will eventually output 0 and 1, where 0 represents "abnormal" and 1 represents "qualified". In addition, there is a fully connected layer to fuse the features extracted by convolution and pooling, and finally generate the classification probability of qualified or abnormal. Through this architecture, not only is the comparison efficiency and consistency of product data improved, but the accuracy and reliability of product comparison are also improved.

[0081] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0082] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0083] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0084] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0085] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0086] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0087] In the embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0088] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0089] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0090] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0091] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A product data comparison method based on deep learning, characterized in that, The method includes: During the comparison period, the product data comparison status of the platform to be compared is monitored in real time to determine the comparison performance, thereby determining whether to add new product data samples for quality assessment, and thus determining the influencing factors of the product data comparison results. If the quality of newly added product data samples is assessed, it will be determined whether to update and optimize the deep learning model to improve the timeliness of the deep learning model in comparing product data. Otherwise, changes in product data samples will be monitored to determine whether changes in product data will affect the results. Continuously monitor the comparison results of product data after deep learning model optimization to limit the triggering of deep learning model optimization in subsequent comparison cycles, thereby preventing the deep learning model from engaging in cyclical optimization.

2. The product data comparison method based on deep learning according to claim 1, characterized in that, The specific process for determining whether to perform quality assessment on newly added product data samples is as follows: The performance of the product data of the comparison platform is monitored in real time within a pre-set comparison period to obtain the corresponding comparison performance value, and the difference between the comparison performance value and the comparison performance value of the previous comparison period is quantified to obtain the change in the comparison performance value. The obtained change in the comparison performance value is compared with the set threshold for the change in the comparison performance value: If the change in the comparison performance value exceeds the threshold for the change in the comparison performance value used to limit the decline in the comparison performance value, the quality of newly added product data samples will be judged to determine whether to update and optimize the deep learning model; otherwise, the change in product data samples will be monitored.

3. The product data comparison method based on deep learning according to claim 1, characterized in that, The specific process for determining the quality of newly added product data samples is as follows: The newly added sample set in the current comparison period is re-compared to obtain the corresponding comparison performance value, which is recorded as the comparison performance value of the new sample. The difference between the comparison performance value of the previous comparison period and the comparison performance value of the new sample is quantified to obtain the difference value of the comparison performance value of the new sample. The newly added comparison performance performance value difference value is compared with the comparison performance difference threshold value, which represents the minimum difference amount that limits the change of the comparison performance value. If the difference in the performance value of the newly added comparison effect is greater than the comparison effect difference threshold, then a data quality check is performed on the newly added sample set. The data quality check includes label quality check, feature quality check, distribution drift check, text data quality check, and image data quality check. If the difference in the performance value of the newly added comparison is not greater than the threshold value for the difference in the comparison, then the deep learning model will be updated and optimized.

4. The product data comparison method based on deep learning according to claim 3, characterized in that, The specific process for updating and optimizing the deep learning model is as follows: Obtain product data change parameters within the current comparison period, including the proportion of new samples and distribution drift; The product data change parameters are compared with the corresponding extracted thresholds for the proportion of newly added samples and the distribution drift threshold: If the proportion of newly added samples exceeds the threshold for the proportion of newly added samples, then the model learning intensity is optimized. If only the distribution drift exceeds the distribution drift threshold, then optimize the model learning rate. If the proportion of new samples and the distribution drift both exceed the corresponding thresholds for the proportion of new samples and the distribution drift, then model learning and comprehensive optimization will be performed. If the proportion of new samples and the distribution drift do not exceed the corresponding thresholds for the proportion of new samples and the distribution drift, the performance of the product data comparison in the next comparison period will continue to be monitored, and a comparison result analysis report for the current comparison period will be output.

5. The product data comparison method based on deep learning according to claim 4, characterized in that, The specific process for optimizing the model learning intensity is as follows: By comparing the change in the performance value with the threshold of the change in the performance value, and the deviation between the proportion of new samples and the threshold of the proportion of new samples, the corresponding optimization amount of model learning intensity is obtained. The learning intensity optimization amount of the model is input into the learning intensity optimization function after training to obtain the corresponding learning intensity optimization value; Based on the set learning intensity optimization ratio and learning intensity optimization value, the training batch optimization amount and training cycle optimization amount are obtained by making corrections. The learning intensity optimization ratio includes the batch optimization ratio and the cycle optimization ratio. The training batch size and training cycle size of the deep learning model are increased by optimizing the training batch size and the training cycle size, respectively. The increased training batch and training period are compared with the comparison processing limit range set according to the product data comparison rules. The comparison processing limit range includes the training batch setting range and the training period setting range. If the increased training batch size and training period length both meet the comparison processing constraints, then the corresponding model learning intensity optimization is performed based on the increased training batch size and training period length; otherwise, the corresponding model learning intensity optimization is performed based on the comparison processing constraints.

6. The product data comparison method based on deep learning according to claim 4, characterized in that, The specific process for optimizing the model learning rate is as follows: By comparing the change in the performance value with the threshold of the change in the performance value, and by smoothing the deviation between the distribution drift and the distribution drift threshold, the corresponding optimization amount of the model learning rate is obtained. By substituting the model learning rate optimization amount into the learning rate optimization function after training, the corresponding learning rate optimization value is obtained; The optimized learning rate is input into the adaptive optimizer to dynamically adjust the learning rate, allowing the model to employ different learning strategies at different stages.

7. The product data comparison method based on deep learning according to claim 4, characterized in that, The specific process of model learning and comprehensive optimization is as follows: After normalizing the differences between the change in the comparison performance value and the threshold for the change in the comparison performance value, the ratio of new samples and the threshold for the ratio of new samples, and the distribution drift and the distribution drift threshold, a weighted fusion operation is performed to obtain the representative value of the model learning comprehensive optimization. Based on the representative quantity of the model learning comprehensive optimization, the corresponding training data adjustment value and learning rate influence factor are mapped into the trained model learning comprehensive optimization mapping model. The model's learning intensity and efficiency are enhanced by adjusting the amount of training data and the learning rate influence factor, respectively.

8. The product data comparison method based on deep learning according to claim 1, characterized in that, The specific steps for monitoring changes in product data samples are as follows: The percentage of newly added samples within the comparison period is compared with a set threshold for the percentage of newly added samples. If the proportion of newly added samples is greater than the threshold for the proportion of newly added samples, then the intensity label representative value is recorded as the first sample representative value; otherwise, the intensity label representative value is recorded as the second sample representative value. The intensity label representative value represents the degree of model learning intensity optimization that needs to be performed, and the degree of model learning intensity optimization gradually decreases from the first sample representative value to the second sample representative value. The distribution drift within the comparison period is compared with the distribution drift threshold: If the distribution drift is greater than the distribution drift threshold, the rate marker is recorded as the first representative value of the distribution drift; otherwise, the rate marker is recorded as the second representative value of the distribution drift. The rate marker represents the degree of optimization of the model learning rate, and the degree of optimization of the model learning rate gradually decreases from the first representative value of the distribution drift to the second representative value of the distribution drift. The results of the intensity and velocity representative values ​​are combined to obtain the corresponding labeling results: If the labeling results include the first representative sample and the first representative distribution drift, then the model learning is comprehensively adjusted based on the preset initial values ​​of the model learning intensity adjustment and the model learning rate adjustment. If the labeling results include the first representative value and the second representative value of the distribution drift, then the model learning intensity adjustment is performed based on the preset model learning intensity adjustment initial value; If the labeling result includes the second representative sample and the first representative sample of distribution drift, then obtain the labeling result within a set number of comparison periods. If the ratio of the second representative sample to the first representative sample of distribution drift exceeds the set ratio, then adjust the model learning rate based on the initial value of the model learning rate adjustment. Otherwise, output the product data comparison result. If the labeling results include the second representative sample quantity and the second representative quantity of distribution drift, then the corresponding product data comparison results will be output.

9. The product data comparison method based on deep learning according to claim 1, characterized in that, The specific details of restricting the triggering of deep learning model optimization are as follows: Real-time monitoring of the alignment performance values ​​in each alignment period after optimization, and detection of whether deep learning model update and optimization are triggered. If triggered, deep learning model update and optimization are performed. If deep learning model updates and optimizations are triggered for more than a preset number of comparison cycles, then deep learning model updates and optimizations will not be performed within a preset time period. Obtain the trigger frequency for deep learning model updates and optimizations within a preset time period, denoted as the trigger update frequency, and match the trigger update frequency with the frequency range: If the trigger update frequency is lower than the minimum value of the frequency range, the deep learning model will still not be updated or optimized and the corresponding product data comparison results will not be output. If the trigger update frequency is higher than the maximum value of the frequency range, the current model update parameters will be fed back to the manual management end, and a request will be made as to whether to continue the model update mechanism. If the deep learning model is not updated or optimized within the preset time period, the product data comparison results will be output.

10. A product data comparison system based on deep learning, employing the product data comparison method based on deep learning as described in any one of claims 1-9, characterized in that, It includes a comparison performance evaluation module, a comparison effect optimization module, and a comparison result monitoring module; The comparison performance evaluation module is used to monitor the comparison of product data of the platform to be compared in real time during the comparison period to make a comparison performance judgment, thereby determining whether to make a quality judgment on the new product data samples, so as to determine the influencing factors of the product data comparison results. The comparison effect optimization module is used to determine whether to update and optimize the deep learning model if the quality of newly added product data samples is judged, so as to improve the timeliness of the deep learning model for product data comparison; otherwise, it monitors changes in product data samples to determine whether changes in product data affect the comparison results. The comparison result monitoring module is used to continuously monitor the comparison results of product data after deep learning model optimization. It can limit the triggering of deep learning model optimization in subsequent comparison cycles, thereby preventing the deep learning model from having a cyclical optimization problem.

Images