Bayesian network model for public platform user to participate in multiple tasks and classification method
Through the combination of the multitasking Bayesian neural network model and Swin Transformer, the problem of user labeling reliability assessment in the public science platform is solved, efficient image classification and user participation estimation are achieved, and data quality and scientific research reliability are improved.
Patent Information
- Application Number
- CN202510514801.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-23
- Publication Date
- 2025-08-26
AI Technical Summary
The prior art is difficult to effectively evaluate the reliability of user annotations in public science platforms, especially the lack of dynamic analysis of user participation and processing of uncertain annotations, resulting in a decline in the accuracy of data sets and the value of scientific research.
The multi-task Bayesian neural network model is used to combine Swin Transformer large-scale pre-trained visual model, and through image classification and user engagement estimation, user engagement is evaluated in real time and data allocation is optimized. The Bayesian neural network's multi-task learning framework is used for feature sharing and dynamic task allocation, and low-quality annotations are filtered based on user interaction behavior and classification uncertainty.
It significantly improves the quality of crowdsourcing data and the reliability of scientific research. Through dynamic evaluation and feedback mechanisms, the accuracy of image classification and user behavior analysis capabilities are improved, the data allocation strategy is optimized, and the reliability and efficiency of labeling are ensured.
Smart Images
Figure CN120543902A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image classification and user parameter estimation of public science platforms, and more specifically, to a multi-task Bayesian network model and classification method for public platform user participation. Background Art
[0002] Citizen science platforms, such as GalaxyZoo, allow non-expert participants to contribute to scientific data classification tasks, particularly in the classification of galaxy images. While these platforms have attracted significant public participation, they still face a key challenge: ensuring the reliability of user-generated annotations. Some users may lack scientific expertise or interest in the task, leading to random clicks or sloppy annotations. These low-quality annotations directly impact the accuracy of datasets and diminish their scientific value.
[0003] Existing methods typically rely on basic quality control mechanisms (such as majority voting) to filter out unreliable annotations. However, such methods have significant limitations. First, majority voting cannot distinguish between highly engaged and less engaged annotators, making it susceptible to random clicks and unable to effectively remove low-quality labels. Second, there is a lack of tools for dynamically analyzing user behavior, making it difficult to determine whether participants truly understand the task and are paying sufficient attention. Furthermore, when processing complex or ambiguous images, majority voting may overlook scientifically significant anomalous objects, thereby reducing the value of the dataset for subsequent research. Furthermore, current technologies have significant shortcomings in quantifying and analyzing user engagement, including the lack of comprehensive mechanisms for evaluating user behavior (such as annotation time, click patterns, and annotation consistency) and the lack of models that effectively integrate user engagement with classification tasks, resulting in a lack of intelligence in data allocation and quality control decisions. Furthermore, existing methods fail to account for annotation uncertainty, making it difficult to assess the reliability of ambiguous labels or anomalous data points in classification tasks. Summary of the Invention
[0004] In response to the shortcomings of the existing technology, the purpose of the present invention is to provide a multi-task Bayesian network model and classification method for public platform user participation. The invention improves the data quality and classification accuracy of crowdsourcing labeling tasks by integrating Bayesian neural networks and Swin Transformer large-scale pre-trained visual models, evaluates user participation in real time, and optimizes data allocation strategies. By combining user participation scores with classification uncertainty, unreliable annotations are effectively filtered out, significantly improving the overall quality of crowdsourcing data, and enhancing the reliability of scientific research and user behavior analysis capabilities.
[0005] To achieve the above object, the present invention provides the following technical solutions:
[0006] A multi-task Bayesian neural network model for user engagement on public science platforms uses a Bayesian neural network and a pre-trained visual model to classify celestial objects and simultaneously quantify user engagement levels. The model integrates a pre-trained visual model, a Bayesian neural network with a multi-task learning framework, and a public science platform.
[0007] The pre-trained model extracts key features of the celestial body image that has completed image preprocessing, and identifies the target celestial body based on the key features;
[0008] The Bayesian neural network has a multi-task learning framework. The Bayesian neural network simultaneously performs two tasks: image classification and user engagement estimation. In the image classification task, celestial objects are classified into predefined classification categories. In the user engagement estimation task, user interaction data is used to calculate the user's engagement score. The extracted features are shared between the two tasks, improving classification accuracy and enhancing the ability to filter low-quality user input. Real-time feedback ensures the stability of the Bayesian neural network output.
[0009] The public science platform integrates real-time interaction and data collection, records user interaction indicators when users perform classification tasks, and stores the interaction indicators in a database for iterative optimization.
[0010] Further, predefined classification categories include elliptical galaxies, spiral galaxies, edge-on galaxies, and anomalous objects.
[0011] Furthermore, in the user engagement estimation task, user effectiveness is evaluated by analyzing the user's interaction patterns on complex celestial images, and highly engaged users are used to handle tasks with uncertainty.
[0012] Furthermore, the task learning framework includes the marking and analysis of abnormal celestial bodies. The marked abnormal celestial bodies are verified by experts and fed back to the Bayesian neural network for iterative optimization to ensure the output accuracy of the Bayesian neural network.
[0013] A multi-task classification method for user participation in public science platforms is proposed. Based on the multi-task Bayesian neural network model described above, it uses a Bayesian neural network in a multi-task learning framework to classify images of target celestial bodies, quantitatively evaluate the level of user participation, dynamically assign tasks of different confidence levels based on the user's participation level, and improve the quality of target object classification by combining user analysis of target object images with expert verification. The method specifically includes the following steps:
[0014] Step 1. Image preprocessing of the target celestial object: The acquired image of the target celestial object is preprocessed to enhance the key features of the celestial object, including binary conversion and contour detection, bounding box calculation, image cropping and resizing, and generating a bounding box that accurately covers the main part of the celestial object;
[0015] Step 2. Target celestial object classification and user engagement estimation: The celestial object images pre-processed in Step 1 are classified using the citizen science platform and a multi-task learning framework, Bayesian neural network. Based on the user's interaction data in image classification, the citizen science platform and the multi-task learning framework, Bayesian neural network, calculate user engagement scores.
[0016] Step 3. Data annotation recommendation strategy: Combine the user engagement score obtained in step 2 with the data annotation recommendation strategy of the Bayesian neural network model classification uncertainty, give priority to rendering suitable image types to improve image classification efficiency; output multi-category posterior probability distribution and user engagement score data annotation,
[0017] Step 4. Utilization of the Citizen Science Platform: The image classification task and user engagement estimation are completed in the Citizen Science Platform. The Citizen Science Platform, which serves as an interactive interface for users to participate in celestial object annotation, collects user interaction data and renders images according to the recommended strategy in Step 3. The Citizen Science Platform supports expert review operations.
[0018] Step 5. Data Storage and Analysis: Image classification results, user engagement scores, and interaction data are stored in a MySQL database. The data is regularly exported to CSV format for subsequent analysis and optimization, and data backup is performed. During the data analysis phase, Python's Pandas library is used to perform statistical analysis on the exported data, including user interaction data, to provide data on user classification levels and engagement behavior. User classification permissions are adjusted based on the user engagement behavior data. The Bayesian neural network is evaluated and analyzed for low-confidence annotations or expert verification failures to identify weaknesses in the classification model when dealing with anomalous objects. Active learning is used to select high-value data points, re-annotate them, and add them to the training set to improve classification accuracy.
[0019] Step 6. Expert Verification and Optimization of the Bayesian Neural Network Classification Model: In image classification, based on real-time feedback on the classification of image data and user engagement scores, low-confidence labels or abnormal classification results are identified as potential candidates for abnormal celestial bodies. Candidates are prioritized for expert verification, and experts review and correct the classification results to generate highly reliable labeled data. The expert review results are stored in the MySQL database to form high-quality training data for optimizing the image classification model. The image classification model is iteratively updated based on the results of expert verification, and new weight parameters are recorded in each optimization cycle to continuously adapt to new celestial body types and user engagement discovered in astronomical research.
[0020] Furthermore, step 2 includes dynamic task allocation, which extracts classification uncertainty based on the probability distribution output of the image classification results of the Bayesian neural network model, and combines it with the engagement score generated by the user's interactive behavior to match and allocate tasks according to the complexity of the image task. High-engagement users are preferentially assigned image tasks with high classification uncertainty or possible abnormal celestial objects, and low-engagement users are assigned tasks with high classification confidence and clear structural features, so as to achieve dual optimization of user participation and classification accuracy.
[0021] Furthermore, in step 2, user engagement is evaluated through interactive behaviors such as user annotation time, click behavior and annotation consistency, image classification completion time, and error rate. The multi-task Bayesian neural network model synchronously outputs the classification results and updates the corresponding user engagement score.
[0022] In summary, the invention has the following beneficial effects:
[0023] This paper simultaneously performs image classification and user engagement estimation, utilizing the Swin Transformer (a large-scale pre-trained visual model) for feature extraction and combining it with a Bayesian neural network (BNN) to perform a dual task: classifying celestial images and assessing user engagement levels during the annotation process. The classification task predicts the most accurate label for each image, while the engagement estimation task assesses user focus by analyzing user interaction metrics such as annotation time and annotation consistency. This paper introduces a multi-task learning framework that dynamically assesses data quality and classification accuracy. The ability to quantify uncertainty through Bayesian inference represents a significant advancement for citizen science platforms, making the identification of anomalous celestial objects more reliable. A real-time engagement scoring mechanism introduces adaptability in task allocation, matching complex classification tasks with highly engaged users while providing simpler tasks for less engaged users, optimizing classification efficiency and user experience. This paper integrates engagement analysis into a feedback loop that considers classification time, click patterns, and annotation consistency, simultaneously improving individual user performance and overall data quality. By intelligently filtering unreliable annotations based on comprehensive participation scores and classification uncertainty, the present invention significantly improves the overall quality of crowdsourcing data and greatly improves the reliability of scientific research and user behavior analysis capabilities in public science platforms. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 is a flow chart of the present invention;
[0025] Figure 2 Visualization of the classification and user engagement model based on Bayesian neural networks. DETAILED DESCRIPTION
[0026] The present invention will be described in further detail below with reference to the accompanying drawings.
[0027] It should be noted that, for the sake of convenience, the directions described below are consistent with the directions of the drawings themselves, but do not limit the structure of the present invention.
[0028] like Figures 1-2 As shown, the present invention discloses a multi-task Bayesian neural network model for user participation in a public science platform, which uses a Bayesian neural network and a pre-trained visual model to classify celestial objects and quantify the level of user participation. The model includes a pre-trained visual model, a Bayesian neural network with a multi-task learning framework, and a public science platform integration, wherein: the pre-trained model extracts the key features of the celestial image that has completed image preprocessing, and identifies the target celestial body according to the key features; the Bayesian neural network has a multi-task learning framework, and the Bayesian neural network simultaneously performs two tasks: image classification and user participation estimation, and in the image classification task, celestial bodies are divided into predefined classification categories, and the predefined classification categories include elliptical galaxies, spiral stars, and celestial bodies. The system integrates user interaction data to calculate user engagement scores, including galaxies, side-on galaxies, and anomalous objects, ensuring comprehensive coverage of both common and rare objects. In the user engagement estimation task, user interaction data is used to calculate user engagement scores, and extracted features are shared between the two tasks to improve classification accuracy and enhance the ability to filter low-quality user input. Real-time feedback ensures the stability of the Bayesian neural network output. In the user engagement estimation task, user effectiveness is assessed by analyzing user interaction patterns on complex astronomical images, and highly engaged users are utilized to handle tasks with uncertainty. Public science platforms are integrated for real-time interaction and data collection. When users perform classification tasks, user interaction metrics are recorded and stored in a database for iterative optimization. The Bayesian neural network quantifies the uncertainty of the predictions. For anomalous objects with low confidence, quantified uncertainty prioritizes subsequent manual verification, significantly improving the reliability of the classification task.
[0029] The multi-task learning framework includes the labeling and analysis of abnormal celestial bodies. The labeled abnormal celestial bodies are verified by experts and fed back to the Bayesian neural network for iterative optimization to ensure the output accuracy of the Bayesian neural network.
[0030] The present invention also discloses a multi-task classification method for user participation on a public science platform. Based on the multi-task Bayesian neural network model described above, the Bayesian neural network of the multi-task learning framework is used to classify images of target celestial bodies, quantitatively evaluate the level of user participation, dynamically assign tasks of different confidence levels based on the user's participation level, and improve the quality of target celestial body classification by combining user analysis of target celestial body images with expert verification. The method specifically includes the following steps:
[0031] Step 1. Image preprocessing of the target celestial object: The acquired image of the target celestial object is preprocessed to enhance the key features of the celestial object, including binary conversion and contour detection, bounding box calculation, image cropping and resizing, and generating a bounding box that accurately covers the main part of the celestial object;
[0032] Step 2. Target Object Classification and User Engagement Estimation: The astronomical images pre-processed in Step 1 are classified using the Citizen Science Platform's multi-task learning framework, a Bayesian neural network. Based on user interaction data during image classification, the Citizen Science Platform and the multi-task learning framework, a Bayesian neural network, calculate user engagement scores. These scores are dynamically used to guide subsequent classification task assignments, thereby improving the reliability of crowdsourced labels. This method integrates classification confidence and user behavior data into a unified feedback mechanism, effectively improving image classification accuracy while eliminating low-quality annotations.
[0033] The system also includes dynamic task allocation. Based on the probability distribution output of the image classification results from a Bayesian neural network model, it extracts classification uncertainty and combines it with engagement scores generated by user interaction behavior to match task complexity. Highly engaged users are prioritized for tasks with images with high classification uncertainty or potential anomalous astronomical objects, while low-engaged users are assigned tasks with high classification confidence and clear structural features. This optimizes both user engagement and classification accuracy. User engagement is assessed through user annotation time, click behavior, annotation consistency, image classification completion time, and error rate. The multi-task Bayesian neural network model simultaneously outputs classification results and updates the corresponding user engagement score. User annotation time reflects the user's attention to the classified image, click behavior records the number and distribution of clicks in the user table, annotation consistency measures the match between user annotations and the model or expert opinions, image classification completion time assesses user efficiency, and error rate reflects the deviation between user annotations and verification data.
[0034] The collected interaction data is converted into numerical features, such as the mean and variance of dwell time and the statistical properties of click distribution. These features are then used as input data for the engagement assessment task in a Bayesian neural network. Before entering the Bayesian neural network model, all data is normalized to eliminate scale differences in different user behavior characteristics. Based on these input features, the Bayesian neural network calculates a probability distribution of engagement, capturing the dynamic changes and uncertainty of user engagement behavior. Compared to traditional rule-based scoring methods, this probabilistic calculation method is more flexible and accurate, and can incorporate behavioral characteristics of specific domains.
[0035] The engagement score is used to classify users as "active" or "inactive" based on whether their engagement probability is above a set threshold. For example, "active" users typically exhibit high annotation consistency, detailed click patterns, and long dwell times, resulting in engagement scores above 0.8. "Inactive" users, on the other hand, exhibit rapid clicks or very short dwell times, resulting in engagement scores below 0.5. The model outputs a confidence level for each user's engagement, which determines whether to accept the classification result or submit it for expert verification. This effectively filters out low-quality annotated data and improves overall data quality.
[0036] For image classification, this paper uses the Swin Transformer, a large visual model, as the backbone for feature extraction. The Swin Transformer processes astronomical images through hierarchical encoding to extract complex spatial patterns and optical features. After feature extraction, the resulting feature vectors are passed to two fully connected layers for further processing, ultimately using Bayesian inference to calculate the probability distribution of each astronomical object class.
[0037] Step 3. Data annotation recommendation strategy: Combine the user engagement score obtained in step 2 with the data annotation recommendation strategy of the Bayesian neural network model classification uncertainty, prioritize rendering suitable image types to improve image classification efficiency; output multi-category posterior probability distribution and data annotation of the user engagement score.
[0038] The multi-category posterior probability distribution is used to reflect the multi-task Bayesian neural network model's confidence and uncertainty in each category of results. By analyzing the maximum value of the multi-category posterior probability distribution and the differences between categories, the multi-task Bayesian neural network model can identify images with low classification confidence and prioritize them in the recommendation strategy. When assigning image tasks to users with different engagement scores, images suitable for their processing capabilities are selectively pushed to different users based on their engagement scores and the image classification uncertainty, enabling priority rendering of appropriate image types. Highly engaged users will receive low-confidence or abnormal images first to improve overall annotation accuracy; low-engagement or new users will be assigned images with higher confidence to ensure data quality and annotation efficiency.
[0039] This method prioritizes the allocation of astronomical images based on two dimensions: (1) classification uncertainty, which is the confidence level reflected by the probability distribution output by the Bayesian classifier; and (2) user engagement score, which reflects the user's focus and reliability in the task. Highly engaged users are preferentially assigned to images with low classification confidence or potential anomalies, while low-engagement or new users are assigned to images with high classification confidence. By matching task difficulty with user reliability, the efficiency and credibility of the crowdsourcing annotation process are optimized, thereby improving the quality of data used for scientific research analysis. Together with the user engagement score in step 2, this constitutes a dynamic feedback loop between user behavior and model optimization.
[0040] Based on the confidence level of the classification probability, images are divided into three categories: high-confidence labels (classification probability > 0.9), low-confidence labels (classification probability < 0.6), and true label images. High-confidence label images have clear classifications and are suitable for evaluating the annotation capabilities of new users. Low-confidence label images have high uncertainty or ambiguity and are therefore preferentially assigned to verified high-participation users to significantly improve overall classification accuracy. True label images are annotated by experts and serve as benchmark data for evaluating user annotation consistency and system performance. The recommendation strategy not only balances the workload among users, but also creates a supportive environment for less experienced participants, which is conducive to users gradually improving their professional capabilities.
[0041] Step 4. Utilize the Citizen Science Platform: The image classification task and user engagement estimation are completed within the Citizen Science Platform. Serving as an interactive interface for users to participate in celestial object annotation, the Citizen Science Platform collects user interaction data and renders images according to the recommended strategies from Step 3. The Citizen Science Platform supports expert review and enables real-time updates of classification models and user engagement profiles. This platform-level data integration ensures consistency between front-end user interaction and back-end model analysis and database storage, supporting the continuous optimization of classification reliability and the quality of user contributions.
[0042] Highly engaged users are preferentially assigned low-confidence images (classification probability <0.6) or images of potentially anomalous celestial objects to fully utilize their annotation skills and improve model performance. Meanwhile, new or unverified users are assigned high-confidence images (classification probability >0.9) to assess their annotation ability and consistency. This task assignment mechanism ensures that task difficulty matches user capabilities, effectively improving overall annotation quality. Users are categorized as "highly engaged" or "lowly engaged" based on their engagement scores. Highly engaged users with scores above a preset threshold (e.g., 0.8) indicate their reliability and initiative in the annotation task. Lowly engaged users with scores below the threshold (e.g., 0.5) may indicate random annotation behavior or incompatibility with the task requirements. Engagement scores not only influence image presentation decisions, such as assigning simpler tasks to low-engagement users, but are also stored in a MySQL database and used alongside user interaction data and classification results for model optimization and user behavior analysis.
[0043] Step 5. Data storage and analysis: The image classification results, user engagement scores and interaction data are stored in the MySQL database. The data is regularly exported to CSV format for subsequent analysis and optimization, and data backup is performed. In the data analysis stage, Python's Pandas library is used to perform statistical analysis on the exported data, covering the user's interaction data, and providing data on user classification level and participation behavior. User classification permissions are adjusted based on the data on user participation behavior. User participation is evaluated through indicators such as total classification volume, anomaly detection rate, average classification accuracy and completeness to provide user classification level and participation behavior. Comprehensive observation; evaluate and analyze low-confidence annotations or expert verification failure samples of the Bayesian neural network to determine the weaknesses of the classification model in handling abnormal celestial bodies. By analyzing user behavior patterns, the system can detect inefficient users, such as users with short stay times or high error rates, and adjust recommendation strategies or restrict their classification permissions. Performance evaluation focuses on analyzing low-confidence annotations or expert verification failure samples to determine the weaknesses of the classification model in handling abnormal celestial bodies; apply active learning to select high-value data points, such as confirmed abnormal celestial bodies or misclassified samples, re-label them, and add them to the training set to improve classification accuracy;
[0044] Step 6. Expert Verification and Bayesian Neural Network Classification Model Optimization: In image classification, low-confidence labels or anomalous classification results are identified as potential candidates for anomalous astronomical objects based on real-time feedback on image data classification and user engagement scores. These candidates are prioritized for expert verification, who review and revise the classification results to generate highly reliable annotated data. The expert review results are stored in a MySQL database, forming high-quality training data for optimizing the image classification model. The image classification model is iteratively updated based on the expert verification results, with new weight parameters recorded during each optimization cycle to continuously adapt to new astronomical object types discovered in astronomical research and user engagement. Real-world labeled images are annotated by experts and serve as benchmark data for evaluating user annotation consistency and system performance.
[0045] Figure 2 It includes three curve graphs, namely Figure 2 A. Figure 2 B and Figure 2 C, respectively illustrate the key analysis curves of Bayesian reasoning, user engagement analysis and classification confidence assessment in the present invention.
[0046] Figure 2 Figure A shows the Bayesian inference curve, which illustrates the evolution of the prior distribution, likelihood function, and posterior distribution for a Bayesian neural network (BNN) at different sample sizes (e.g., count_105 and count_1050). The horizontal axis represents the parameter values of the Bayesian neural network model (ranging from 0 to 1), and the vertical axis represents the corresponding probability density. As the sample size increases, the posterior distribution converges, indicating that the Bayesian neural network model's prediction uncertainty decreases and confidence increases.
[0047] Figure 2 Figure B shows the user engagement distribution curve, which illustrates the statistical distribution of user engagement scores on the platform. The horizontal axis represents the user engagement score (ranging from 16 to 80, calculated based on user behavior metrics), and the vertical axis represents the number of users within the corresponding score range. Based on the score range, engagement is categorized into three categories: "unengaged" (<50 points), "moderately engaged" (50–65 points), and "highly engaged" (>65 points). Figure 2 B is used to reflect the overall activity characteristics of users and provide a basis for task allocation and user modeling.
[0048] Figure 2 C is the classification confidence curve, which shows the density distribution of the probability output by the Bayesian neural network model, exhibiting a typical bell-shaped curve. The horizontal axis represents the probability value predicted by the Bayesian neural network model for a specific category (ranging from 0 to 1), and the vertical axis represents the statistical density of that probability. The shaded areas correspond to the 90%, 95%, and 99% confidence intervals, respectively, which measure the reliability and statistical significance of the classification results.
[0049] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.
Claims
1. A multi-task Bayesian neural network model for user participation in public science platforms, characterized by: Utilizes Bayesian neural networks and pre-trained visual models to classify celestial objects while quantifying user engagement levels. This includes pre-trained visual models, Bayesian neural networks with a multi-task learning framework, and public scientific platform integration, including: The pre-trained model extracts key features of the celestial body image that has completed image preprocessing, and identifies the target celestial body based on the key features; The Bayesian neural network has a multi-task learning framework. The Bayesian neural network simultaneously performs two tasks: image classification and user engagement estimation. In the image classification task, celestial objects are classified into predefined classification categories. In the user engagement estimation task, user interaction data is used to calculate the user's engagement score. The extracted features are shared between the two tasks, improving classification accuracy and enhancing the ability to filter low-quality user input. Real-time feedback ensures the stability of the Bayesian neural network output. The public science platform integrates real-time interaction and data collection, records user interaction indicators when users perform classification tasks, and stores the interaction indicators in a database for iterative optimization.
2. The multi-task Bayesian neural network model for user participation in a public science platform according to claim 1, characterized in that: The predefined classification categories include elliptical galaxies, spiral galaxies, edge-on galaxies, and anomalous objects.
3. The multi-task Bayesian neural network model for user participation in a public science platform according to claim 1, characterized in that: In the user engagement estimation task, user effectiveness is evaluated by analyzing the user's interaction pattern on complex celestial images, and highly engaged users are used to handle tasks with uncertainty.
4. The multi-task Bayesian neural network model for user participation in a public science platform according to claim 1, characterized in that: The multi-task learning framework includes the labeling and analysis of abnormal celestial bodies. The labeled abnormal celestial bodies are verified by experts and fed back to the Bayesian neural network for iterative optimization to ensure the output accuracy of the Bayesian neural network.
5. A multi-task classification method for user participation on a public science platform, based on the multi-task Bayesian neural network model of any one of claims 1 to 4, characterized in that: The Bayesian neural network of the multi-task learning framework is used to classify the images of the target celestial objects and quantitatively evaluate the level of user participation. Tasks with different confidence levels are dynamically assigned based on the user's participation level. The quality of the target celestial object classification is improved by combining user analysis of the target celestial object images with expert verification. The specific steps include: Step 1. Image preprocessing of the target celestial object: The acquired image of the target celestial object is preprocessed to enhance the key features of the celestial object, including binary conversion and contour detection, bounding box calculation, image cropping and resizing, and generating a bounding box that accurately covers the main part of the celestial object; Step 2. Target celestial object classification and user engagement estimation: The celestial object images pre-processed in Step 1 are classified using the citizen science platform and a multi-task learning framework, Bayesian neural network. Based on the user's interaction data in image classification, the citizen science platform and the multi-task learning framework, Bayesian neural network, calculate user engagement scores. Step 3. Data annotation recommendation strategy: Combine the user engagement score obtained in step 2 with the data annotation recommendation strategy of the Bayesian neural network model classification uncertainty, give priority to rendering suitable image types to improve image classification efficiency; output multi-category posterior probability distribution and user engagement score data annotation, Step 4. Utilization of the Citizen Science Platform: The image classification task and user engagement estimation are completed in the Citizen Science Platform. The Citizen Science Platform, which serves as an interactive interface for users to participate in celestial object annotation, collects user interaction data and renders images according to the recommended strategy in Step 3. The Citizen Science Platform supports expert review operations. Step 5. Data Storage and Analysis: Image classification results, user engagement scores, and interaction data are stored in a MySQL database. The data is regularly exported to CSV format for subsequent analysis and optimization, and data backup is performed. During the data analysis phase, Python's Pandas library is used to perform statistical analysis on the exported data, including user interaction data, to provide data on user classification levels and engagement behavior. User classification permissions are adjusted based on the user engagement behavior data. The Bayesian neural network is evaluated and analyzed for low-confidence annotations or expert verification failures to identify weaknesses in the classification model when dealing with anomalous objects. Active learning is used to select high-value data points, re-annotate them, and add them to the training set to improve classification accuracy. Step 6. Expert Verification and Optimization of the Bayesian Neural Network Classification Model: In image classification, based on real-time feedback on the classification of image data and user engagement scores, low-confidence labels or abnormal classification results are identified as potential candidates for abnormal celestial bodies. Candidates are prioritized for expert verification, and experts review and correct the classification results to generate highly reliable labeled data. The expert review results are stored in the MySQL database to form high-quality training data for optimizing the image classification model. The image classification model is iteratively updated based on the results of expert verification, and new weight parameters are recorded in each optimization cycle to continuously adapt to new celestial body types and user engagement discovered in astronomical research.
6. The multi-task classification method for user participation in a public science platform according to claim 5, characterized in that: The step 2 includes dynamic task allocation, which extracts classification uncertainty based on the probability distribution output of the image classification results of the Bayesian neural network model, and combines the engagement score generated by the user's interactive behavior to match and allocate tasks according to the complexity of the image task. High-engagement users are preferentially assigned image tasks with high classification uncertainty or possible abnormal celestial objects, and low-engagement users are assigned tasks with high classification confidence and clear structural features, so as to achieve dual optimization of user engagement and classification accuracy.
7. The multi-task classification method for user participation in a public science platform according to claim 6, characterized in that: In step 2, user engagement is evaluated through interactive behaviors such as user annotation time, click behavior and annotation consistency, image classification completion time, and error rate. The multi-task Bayesian neural network model synchronously outputs the classification results and updates the corresponding user engagement score.