Automatic task pushing method based on crowdsourcing human platform

Abstract

The invention belongs to the technical field of Internet crowdsourcing platforms, and particularly relates to an automatic task pushing method based on a crowdsourcing human platform, which comprises the following steps: S1, data collection: collecting executor information and task information; s2, data analysis and modeling: models constructed in the step comprise an executor portrait model, a task model and a matching algorithm model; s3, automatic pushing: based on the task information and the comprehensive matching score, automatically pushing a corresponding task to an executor according to a pushing rule; and S4, feedback and optimization: collecting feedback information and optimizing the executor portrait model, the task model and the matching algorithm model according to the feedback information. According to the method, through multi-dimensional data collection and analysis, an accurate executor portrait and a task model are constructed, high matching of the task and the executor is realized through a multi-dimensional matching algorithm, and the task completion quality and efficiency are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of Internet crowdsourcing platforms, and particularly relates to a method for automatically pushing tasks based on a crowdsourcing human platform. Background Art

[0002] In a crowdsourcing human platform, how to efficiently push tasks to suitable executors is the key to improving the platform operation efficiency and user satisfaction. Traditional task pushing methods often have many deficiencies. For example, the tasks pushed do not match well with the skills, interests, and geographical locations of the executors, resulting in a low task completion rate and poor execution effects. At the same time, the method of manually screening and assigning tasks not only consumes a large amount of manpower and time, but also is prone to subjective biases. With the continuous expansion of the crowdsourcing business scale, there is an urgent need for a method that can automatically and accurately push tasks to improve the matching degree between tasks and executors and enhance the overall operation efficiency of the platform. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for automatically pushing tasks based on a crowdsourcing human platform, which is used to solve the technical problem that the tasks pushed by the system in the prior art do not match well with the skills, interests, and geographical locations of the executors.

[0004] The described method for automatically pushing tasks based on a crowdsourcing human platform includes the following steps: S1. Data collection: Collect executor information and task information;

[0005] S2. Data analysis and modeling: The model constructed in this step includes an executor portrait model, a task model, and a matching algorithm model;

[0006] S3. Automatic push: Based on the task information and the comprehensive matching score, automatically push the corresponding tasks to the executors according to the push rules;

[0007] S4. Feedback and optimization: Collect feedback information and optimize the executor portrait model, task model, and matching algorithm model accordingly.

[0008] Preferably, step S1 includes:

[0009] 1) Collection of executor information: The executor information includes professional skills, hobbies, and geographical location information;

[0010] 2) Collection of task information: The task information includes task type, task requirements, and task location, and the task requirements include skill requirements and completion time.

[0011] Preferably, step S2 includes:

[0012] S2.1. Establish an executor portrait model: This model constructs an executor portrait based on the collected executor information by using data analysis algorithms;

[0013] S2.2. Build a task model: Analyze task information, extract key features and requirements, and build a task model;

[0014] S2.3. Matching algorithm model design: Design a matching algorithm model based on multi-dimensional features to achieve matching between tasks and executors;

[0015] S2.4. Train each model to adjust parameters.

[0016] Preferably, step 2.3 includes:

[0017] S2.3.1. Calculate the matching degree of each factor feature;

[0018] S2.3.2. Assign basic weights based on the importance of each factor requirement;

[0019] S2.3.3. Adjust the weight of each factor based on the basic weight based on the task situation, task type and task feedback information;

[0020] S3.3.4. After the weight adjustment is completed, the comprehensive matching score is calculated based on the final weight determined after the adjustment using the comprehensive matching score calculation formula; the calculation formula for the comprehensive matching score after the weight adjustment is: Comprehensive matching score = w1\timesC1+w2\timesC2+…+w n \timesC n ,

[0021] Among them, C n Indicates the matching degree of the nth factor, w n represents the weight corresponding to the nth factor, w2\timesC2 represents the product of the matching degree of the nth factor and the corresponding weight. Specifically, the matching degree includes at least skill matching degree and position matching degree.

[0022] Preferably, the step S2.3.3 includes:

[0023] 1) Adjust weights according to the task situation: For ordinary tasks, appropriately balance the weights of various factors in the matching degree calculation; for urgent tasks, since the urgency is high and a large amount of data needs to be collected in a short time, the weights of key factors such as skill matching degree and location matching degree should be increased on the basis of the basic weights; in long-term tasks, the stability of the executor is important, and the weights of factors related to stability should be appropriately increased. When the interests and hobbies are related to stability, the weight of interest matching degree should be increased, and the executor stability factor can be added. The stability factor extracts corresponding feature vectors by quantifying the completion coherence and reputation of the executor's past tasks, and quantifies the information on the task continuation time in the task information to extract corresponding feature vectors, thereby calculating the stability matching degree and giving a certain weight to the stability factor;

[0024] 2) Adjust weights according to the task type: For professional skill-based tasks that require highly professional skills, the skill matching degree weight should be dominant; for service-based tasks, the location matching degree weight and the skill matching degree weight are equally important; for creative tasks, the interest matching degree weight and the skill matching degree weight are relatively high;

[0025] 3) Adjust the weights of each factor based on feedback information: Collect feedback information from the executor on the task push for weight adjustment; obtain feedback on the task completion quality and executor matching degree from the task publisher and adjust the weights; through the platform's data monitoring system, observe indicators including the task completion rate, average completion time, and executor satisfaction, and optimize the weight settings based on the indicator changes.

[0026] Preferably, the step S2.4 includes:

[0027] S2.4.1, Data preparation, used to form the training set required for training, including data integration and cleaning, data annotation, and data partitioning;

[0028] S2.4.2, Executor portrait model training, using the executor feature data in the training set as input and the classification label of the executor as output to train the random forest model to obtain the executor portrait model;

[0029] S2.4.3, Task model training, using the bag-of-words model, TF-IDF, or word embedding to convert the task description text into a numerical vector, converting the time requirement into a numerical form of time, using the feature vector of the task as input and the corresponding label of the feature vector as output to train the SVM model to obtain the task model;

[0030] S2.4.4. Matching algorithm training: For each task, a certain number of executor samples with different degrees of matching are selected from the executor dataset to form task-executor pairs, and their true degrees of matching are labeled. Using the generated task-executor pair data, the weight parameters of the matching algorithm are adjusted through the gradient descent optimization algorithm to optimize the matching algorithm model.

[0031] Preferably, the push rule includes: determining the urgency level and task location of the task based on the task information, classifying the tasks according to the urgency level of the task, and determining the position range for preferentially pushing the task according to the task location and the urgency level of the task. A screening threshold for the comprehensive matching score of the executor for the preferentially pushed task is determined based on the urgency level. For tasks with a low urgency level, no position range and screening threshold for preferentially pushing the task are set. When no position range and screening threshold for preferentially pushing the task are set, the system pushes to all executors who meet the basic skill requirements in the order of the comprehensive matching score from high to low.

[0032] Preferably, in step S2.1, the data analysis performed by the executor portrait model includes: quantitatively evaluating the skill level of the executor; classifying and labeling the description of the executor's interests and hobbies using natural language processing technology; analyzing the geographical location information of the executor to determine their main activity area. After quantifying the executor information, the data analysis extracts the feature vectors of the corresponding factors. The feature vectors include skill vectors representing the skill level, position vectors representing the geographical location of the executor, and interest vectors representing the interests and hobbies of the executor.

[0033] Preferably, step S2.2 includes:

[0034] 1) Standardized coding of task types for identifying and distinguishing tasks of different types and at different times;

[0035] 2) Quantitative processing of skill requirements: For the skill requirements required for the task, an accurate representation is made in a quantitative manner. According to the skill level standard, the corresponding skill requirements are quantified into a specific value. Considering the importance differences of different skills in the task, corresponding weights are assigned to each skill, and after quantifying the skill requirements, they are transformed into skill requirement vectors. The skill requirement vectors contain the values of the skill requirements of each category of skills and the corresponding weights;

[0036] 3) Determination of geographical range features: According to the task location information, accurate geographical coordinates or a geographical area object containing boundary coordinate information are obtained.

[0037] The advantages of the present invention are as follows: improving task matching degree: through multi-dimensional data collection and analysis, accurate executor portraits and task models are constructed. Through the data processing of these models, on the one hand, a multi-dimensional portrait is established for each executor to more accurately describe their characteristics and capabilities; on the other hand, task information is standardized and quantitatively represented to form a feature vector of the task, so as to facilitate the subsequent matching of tasks and executors.

[0038] The present invention designs a matching algorithm for tasks and executors based on multi-dimensional features. This algorithm comprehensively considers various features of the executor portrait and the task model to achieve a high degree of matching between tasks and executors, improving the quality and efficiency of task completion. At the same time, this method also provides a method for adjusting the corresponding feature weights of the matching algorithm according to task types, task situations, and feedback information, so that the matching algorithm can more specifically adapt to the various needs of tasks, thereby achieving a better matching effect.

[0039] The present invention sets corresponding automatic push rules based on the matching results and key factors. The automatic push of tasks reduces manual intervention and saves a large amount of labor and time costs, enabling the platform to quickly and efficiently process task allocation and improving the overall operation efficiency. At the same time, the push method also maximally ensures that the executors suitable for tasks can obtain task pushes more preferentially. This precise task push enables executors to receive tasks that are more in line with their own capabilities and interests, improving the satisfaction and participation of executors in the platform; at the same time, task publishers can also find suitable executors faster, improving the success rate and effect of task publishing. Description of the Drawings

[0040] Figure 1 It is the basic flowchart of an automatic task push method based on a crowdsourcing manpower platform in the present invention.

[0041] Figure 2 It is the basic flowchart of step S2 data analysis and modeling in the present invention. Detailed Embodiments

[0042] The following further details the specific embodiments of the present invention by describing the embodiments with reference to the drawings, so as to help those skilled in the art have a more complete, accurate, and in-depth understanding of the inventive concept and technical solution of the present invention.

[0043] As Figure 1 、 Figure 2 shown, the present invention provides an automatic task push method based on a crowdsourcing manpower platform, including the following steps.

[0044] S1. Data collection: Collect executor information and task information.

[0045] 1) Collection of executor information: The crowdsourcing human resources platform collects the information of registered executors, including but not limited to name, contact information, age, gender, professional skills (such as copywriting, design, programming and other skill levels and relevant certification information), work experience (types of crowdsourcing tasks participated in in the past and completion status), interests and hobbies, geographic location information (real-time location or usual work location), etc.

[0046] The specific method is to set up detailed information filling fields on the platform registration page to guide performers to fill in their personal information truthfully. For skill level and certification information, performers are required to upload relevant supporting documents, which the platform will manually review or verify through data connection with relevant certification bodies.

[0047] To obtain the real-time geographic location information of performers, a positioning function is integrated into the platform's mobile application. After the performers' authorization, their location information is obtained regularly. At the same time, performers are allowed to set their frequently used work locations to serve as a supplement when positioning information is unavailable.

[0048] 2) Task information collection: For tasks posted on the platform, the crowdsourcing human resources platform collects detailed information on the tasks, including but not limited to task type (such as data entry, market research, delivery services, etc.), task requirements (skill requirements, completion time, quality standards, etc.), task remuneration, task location (if there is a need for on-site execution), etc.

[0049] The specific method is as follows: When a task publisher posts a task on the platform, the system provides a detailed task information template, requiring the publisher to clearly define the task type, describe the task requirements in detail, set a reasonable task reward, and accurately determine the task location (if on-site execution is required). For the skill requirements of the task, the publisher can select from the platform's preset skill categories and indicate the required skill level.

[0050] S2. Data analysis and modeling: The models constructed in this step include the executor portrait model, task model, and matching algorithm model. The specific construction and training process of each model is as follows.

[0051] S2.1. Establish an executor portrait model: This model uses data analysis algorithms to construct executor portraits based on the collected executor information.

[0052] The data analysis performed by the executor portrait model includes: quantitatively evaluating the executor's skill level. For example, the copywriting skill is divided into primary, intermediate, and advanced levels, corresponding to different score ranges. Based on the certificates and work experience provided by the executor, their skill level is scored and calibrated. Using natural language processing technology, the descriptions of the executor's interests and hobbies are classified and labeled. For example, "likes reading and writing" is labeled as an interest in "writing creation". By analyzing the executor's geographical location information, their main activity area is determined. For example, with their usual work location as the center, an activity range with a certain radius is set. After quantification, the executor's skill level forms a skill vector of the executor's skill level, and the skill vector contains the quantified values of various categories of skills mastered by the executor. Through the data analysis of the executor portrait model, the executor information is quantified, and the feature vectors of the corresponding factors are extracted. The feature vectors involved include the skill vector representing the skill level, the position vector representing the executor's geographical location, the interest vector representing the executor's interests and hobbies, etc.

[0053] S2.2. Construct a task model: Analyze the task information, extract key features and requirements, and construct a task model.

[0054] The task model standardizes and quantifies information such as task type, skill requirements, and task location, forming a feature vector of the task. It specifically includes the following content.

[0055] 1) Standardized coding of task types: For the convenience of computer processing and analysis, various different types of tasks are uniformly coded. For example, the data entry task is coded as F+yyyyMMdd+4-digit random number, and the design task is coded as D+yyyyMMdd+4-digit random number; this step can also continuously update the coding system according to the expansion of the platform's business. For example, if a video editing task is newly added to the platform, it is coded as V+yyyyMMdd+4-digit random number. In the above examples, F, D, and V respectively represent different task types, and yyyyMMdd represents the date of year, month, and day. Through this standardized coding, different types of tasks at different times can be quickly identified and distinguished in subsequent algorithm calculations and model processing.

[0056] 2) Quantification processing of skill requirements: For the skill requirements of a task, use a quantification method for accurate representation. Taking programming skills as an example, if a task requires intermediate Java programming skills, according to the industry-wide common skill level standard, the intermediate Java programming skills can be quantified into a specific value, such as 80 points (assuming a full score of 100 points). At the same time, considering the differences in the importance of different skills in the task, assign corresponding weights to each skill. For example, in a software development project task, the weight of the core programming skill may be set to 0.6, while the weight of the auxiliary document writing skill is set to 0.2. In this way, the skill requirement vector A of the task is obtained. The skill requirement vector A contains the values of the skill requirements for each category of skills, and different categories of skills have corresponding weights.

[0057] 3) Determination of geographical scope characteristics: According to the task location information, convert it into accurate geographical coordinates, such as longitude and latitude; if the task has specific execution scope requirements, use Geographic Information System (GIS) technology to calculate the geographical scope centered on the geographical coordinates of the task location and represent it as a geographical area object containing information such as boundary coordinates. For example, if the task requires that the executor within a 5-kilometer radius around the task execution location be given priority, then calculate a geographical scope with a radius of 5 kilometers centered on the geographical coordinates of the task location and form a geographical area object. This geographical area object will be an important part of the geographical-related characteristics in the task model and will be used for subsequent matching calculations with the executor's geographical location.

[0058] Quantify the task information through data analysis of the task model and extract the feature vectors of the corresponding factors. The feature vectors involved include the skill requirement vector representing the task skill requirements, the location requirement vector representing the task's geographical location requirements, the interest requirement vector representing the task's requirements for the executor's interests, etc.

[0059] S2.3. Design of the matching algorithm model: Design a matching algorithm model based on multi-dimensional features to achieve the matching between tasks and executors.

[0060] The matching algorithm model comprehensively considers the characteristics of the executor portrait and the task model and calculates the matching degree score between the task and each executor. For example, for a copywriting task, the algorithm will focus on considering the copywriting skill level of the executor, past relevant work experience, and the part related to writing creation in the interests and hobbies, and at the same time combine the matching degree between the task location and the executor's geographical location to give a comprehensive matching degree score. The specific construction method of the matching algorithm model includes the following content.

[0061] S2.3.1. Calculate the matching degree of each factor feature: The cosine similarity algorithm is used to measure the similarity between two vectors. In the scenario of matching tasks and executors, the cosine similarity between the feature vector of the task and the feature vector of the executor is calculated to evaluate the matching degree of the two in terms of factors such as skills and interests. As mentioned above, if the skill requirement vector of the task is A and the skill vector of the executor is B, the skill matching degree score between the two is calculated through the cosine similarity formula: [cos(A,B) = \frac{A\cdot B}{||A||\times||B||}]; the numerator A\cdot B of this formula is the dot product of the two vectors, reflecting the sum of their similarities in each dimension; while the denominator ||A||\times||B|| is the product of the norms of the two vectors, \frac{}{} represents a fraction, the content in the previous {} is the numerator, the content in the latter {} is the denominator, and cos(A,B) represents the calculation result of the cosine similarity, which is also the matching degree of the executor and the task in this factor feature. This calculation formula is used to normalize the result, making the similarity score between -1 and 1. The closer the value is to 1, the more similar the two vectors are. It should be noted that the skill requirements include basic skill requirements that must be possessed. If the executor's skills do not meet the basic skill requirements, the skill matching degree is 0.

[0062] S2.3.2. Assign basic weights based on the importance of each factor requirement: In practical applications, different features have different importance for the matching of tasks and executors. Therefore, the comprehensive matching degree is adjusted by assigning different weights to the matching degrees of different factor features. In the initial allocation stage, the system can achieve the allocation of basic weights according to the initial settings. The calculation formula for the comprehensive matching degree after weight adjustment is: Comprehensive matching degree score = w1×C1 + w2×C2 + … + w n ×C n , where, C n represents the matching degree of the nth factor, w n represents the weight corresponding to the nth factor, and w2×C2 represents the product of the matching degree of the nth factor and the corresponding weight. Specifically, the matching degree includes at least the skill matching degree and the location matching degree. The skill matching degree and the location matching degree belong to the matching degrees of key factors. In addition, it can also include the matching degrees of less important factors such as the interest matching degree.

[0063] For example, whether the skills match is a key factor in the successful completion of a task. Therefore, when calculating the comprehensive matching score, a relatively high weight, such as 0.5, is assigned to the skill matching part. For location matching, if the task has high requirements for the location (such as local delivery tasks), a relatively high weight, such as 0.3, is assigned; if the task has low requirements for the location (such as online copywriting tasks), a lower weight, such as 0.1, is assigned. For the matching factor of hobbies and interests, since its impact on task completion is relatively small, a lower weight, such as 0.1, can be assigned. Suppose the skill matching degree is calculated as 0.8, the location matching degree is 0.6, and the hobby and interest matching degree is 0.7 through the cosine similarity algorithm. According to the above weight settings, the comprehensive matching score can be calculated as follows:

[0064] Comprehensive matching score = 0.8×0.5 + 0.6×0.3 + 0.7×0.1 = 0.4 + 0.18 + 0.07 = 0.65. The above weights are directly obtained through initial allocation based on the general requirements of the task, so they are the basic weights.

[0065] S2.3.3. Adjust the weights of each factor on the basis of the basic weights according to the task situation, task type, and task feedback information. Specifically, it includes the following aspects.

[0066] 1) After the basic weights are assigned, this step can also adjust the weights according to the task situation; the task situation includes ordinary tasks, urgent tasks, and long-term tasks.

[0067] Ordinary tasks: For ordinary tasks, such as general data entry tasks, the time requirement is relatively loose, and the weights of various factors are appropriately balanced in the matching degree calculation. Calculate according to the basic weights determined above, or fine-tune according to the platform's operation strategy. For example, keep the skill matching weight at 0.5, the location matching weight at 0.3, and the interest matching weight at 0.2.

[0068] Urgent tasks: When the task urgency is relatively high, such as a market research task for a time-limited promotion activity, a large amount of data needs to be collected in a short time. At this time, the weights of key factors such as skill matching degree and location matching degree should be increased on the basis of the basic weights. For example, adjust the skill matching weight to 0.6 and the location matching weight to 0.4 to ensure that executors with high matching degrees and timely responses can be quickly pushed. Because in an emergency, finding an executor with appropriate skills and who can quickly reach the task location is the primary task, and the priority of factors such as hobbies and interests is relatively reduced.

[0069] Long-term tasks: For some long-term and complex project tasks, such as the development of large software systems, in addition to skill requirements, hobbies and the stability of the executor are also very important. In this case, the weights of factors related to stability should be appropriately increased. If the hobbies are related to stability, the weight of the interest matching degree can be increased to 0.2 - 0.3. The factor of executor stability can also be added at the same time. The stability factor extracts corresponding feature vectors by quantifying aspects such as the completion coherence and reputation of the executor's past tasks, and quantifies the information on the task continuation time in the task information to extract corresponding feature vectors, thereby calculating the stability matching degree and giving a certain weight to the stability factor, such as 0.1 - 0.2. Calculating the comprehensive matching degree score in this way is conducive to matching an executor who not only has the skills, but also has enthusiasm for the task and can be committed to the task in the long term.

[0070] 2) After the basic weights are assigned, this step can also adjust the weights according to the task type; the task types include three types of tasks: professional skill-based, service-based, and creative-based tasks.

[0071] Professional skill-based tasks: For tasks that require high professional skills, such as professional design and advanced programming, the weight of skill matching degree should be dominant. For example, for an advanced UI design task, the weight of skill matching degree can be set to 0.7 - 0.8, the weight of location matching degree is reduced to 0.1 - 0.2, and the weight of interest matching degree is maintained at about 0.1. Because the key to the success of such tasks lies in the professional skills of the executor, and other factors are relatively less important.

[0072] Service-based tasks: For service-based tasks, such as customer service support and on-site repair, the weights of location matching degree and skill matching degree are equally important. Taking the customer service support task as an example, the weights of skill requirements (such as language communication ability and product knowledge) can be set to 0.4 - 0.5, the weight of location matching degree is 0.3 - 0.4, and the weight of interest matching degree is 0.1 - 0.2. This is because service-based tasks require the executor to be able to reach the service location in a timely manner and have good service skills.

[0073] Creative-based tasks: For creative-based tasks, such as advertising creative design and copywriting planning, the weights of interest matching degree and skill matching degree are relatively high. For example, in an advertising creative design task, the weight of skill matching degree is 0.5 - 0.6, the weight of interest matching degree is 0.3 - 0.4, and the weight of location matching degree is 0.1 - 0.2. Because creative-based tasks require the executor to have relevant interests and creative abilities to provide high-quality creative works.

[0074] 3) Finally, when the platform system has obtained the feedback information, adjust the weights of each factor based on the feedback information.

[0075] Executor Feedback: This section collects feedback from executors on task pushes. For example, executors often report that received tasks are highly compatible with their skills, but are geographically inappropriate, preventing them from taking them. If this type of feedback reaches a certain percentage, such as over 30%, the location match weighting will need to be adjusted. This adjustment will be based on the needs of platform administrators. If the platform wants to encourage executors to expand their work scope, the location match weighting can be appropriately lowered; if it wants to prioritize executors' location convenience needs, the location match weighting can be appropriately increased.

[0076] Tasker Feedback: Receive feedback from taskers regarding task completion quality and executor compatibility. If taskers frequently report that the executors they receive have matching skills but lack enthusiasm, resulting in slow progress, consider increasing the weighting of interest and hobbies. Additionally, based on the importance of different factors assigned by taskers (e.g., some prioritize skills over others, while others prioritize location), we can adjust the weighting based on the tasker's information.

[0077] Platform Data Monitoring: The platform's data monitoring system monitors indicators such as task completion rate, average completion time, and executor satisfaction. If a certain task type has a low completion rate, analysis may indicate that improper weighting of a specific factor is leading to poor matching. For example, if a task requiring on-site work has a low completion rate, this could indicate that the position match weight is too low, resulting in insufficient executors being assigned suitable positions. In this case, the position match weight can be appropriately increased, and the weight settings can be continuously optimized based on the changes in the indicators.

[0078] S3.3.4. After the weight adjustment is completed, the comprehensive matching score is calculated based on the final weight determined after the adjustment using the comprehensive matching score calculation formula.

[0079] In addition to adjusting the weights, if it is necessary to add other factor characteristics to the matching calculation according to the actual situation, such as work experience matching, etc., then the corresponding factor characteristics should be extracted in step 1), and the algorithm for the comprehensive matching score with the addition of new factor characteristics should be updated. For example, when considering work experience matching, if the task requires an executor with relevant project experience, the work experience matching score can be calculated by comparing the types of crowdsourcing tasks that the executor has participated in the past and their completion status with the similarity of the current task, and included in the calculation of the comprehensive matching score. Through this matching algorithm that integrates multi-dimensional features and weight adjustment, the matching degree between tasks and executors can be evaluated more comprehensively and accurately, providing strong support for accurate task push.

[0080] S2.4. Train each model to adjust parameters. The training process includes the following steps.

[0081] S2.4.1. Data Preparation. Data preparation is used to form the training set required for training, and specifically includes the following steps.

[0082] (1) Data Integration and Cleaning: Integrate the collected executor information and task information to form a complete data set. Clean and process duplicate data, error data, and missing values in the data set. For example, for duplicate executor registration information, only keep the latest or most complete record; for records in task information that lack key skill requirements or payment information, delete or supplement and improve them.

[0083] (2) Data Annotation: For some data, manual annotation is required to assist model training. For example, for some complex task descriptions, manually annotate the core skills and key requirements involved to ensure that the model can accurately learn the key features of the task. For the hobbies of the executor, further refine and annotate the specific domain labels to which they belong to improve the accuracy of classification.

[0084] Manual annotation is divided into three modules: annotation management terminal, annotation work terminal, and annotation review terminal.

[0085] The main work of the annotation management terminal is task allocation, progress monitoring, setting annotation rules, and managing annotators.

[0086] The annotation work terminal is used to receive the tasks assigned by the annotation management terminal, provide annotation tools such as bounding boxes, labels, punctuation, lines, polygons, text, etc., annotate the data, and finally submit the annotation results to the annotation management terminal and the annotation review terminal.

[0087] The annotation review terminal mainly reviews the annotation results submitted by the annotation work terminal to ensure the accuracy and consistency of the annotation. Check the annotation quality through sampling, double annotation, etc., and give feedback and correction opinions. Statistically review the results, including annotation error rate, annotation speed, etc., to provide a basis for the performance of annotators.

[0088] The main work process includes:

[0089] 1. Determine the annotation requirements, including data type, annotation type, etc.

[0090] 2. Allocate annotation tasks and assign the tasks to suitable annotators.

[0091] 3. The annotator logs in to the annotation work terminal and receives the tasks.

[0092] 4. Use the annotation tools to annotate the data according to the annotation specifications.

[0093] 5. After the annotation is completed, submit the annotation results. It is also possible to modify and improve the annotation results.

[0094] 6. The auditor logs in to the annotation review terminal and receives the annotation tasks to be reviewed.

[0095] 7. Review the annotation results one by one to ensure the accuracy and consistency of the annotations. If any annotation errors or non-compliant places are found, give feedback and corrective opinions.

[0096] 8. After the review is completed, submit the review results to the annotation management terminal.

[0097] 9. Conduct quality assessment and rewards for the annotators according to the review results.

[0098] Data annotation quality assessment rules:

[0099] Completion quality score = qualified annotation score (full score 50) + annotation quantity score (full score 50)

[0100] Qualified annotation score = number of qualified sampled data / total number of sampled data.

[0101] Annotation quantity score = (standard number of fields / number of completed tasks) / 10% * 50

[0102] Overall quality: Grade A: Completion quality score > 80; Grade B: Completion quality score 60 - 80; Grade C: Completion quality score < 60;

[0103] (3) Data division: Divide the processed dataset according to the ratio of 70% training set, 20% validation set, and 10% test set. The training set is used for model training, the validation set is used to evaluate the model's performance during training, adjust model parameters, and prevent overfitting, and the test set is used to finally evaluate the generalization ability of the trained model.

[0104] S2.4.2. Training of the executor portrait model.

[0105] (1) Feature engineering: Further engineering processing of various features of the executor. For example, subdivide the executor's work experience according to dimensions such as task type, completion time, and completion quality to form multiple feature dimensions. Discretize continuous features such as age and income to better integrate into the model.

[0106] (2) Model selection and training: Select a suitable machine learning model to construct the executor portrait, such as decision tree, random forest, or neural network, etc. Here we use random forest, take the executor feature data in the training set as input, and the executor's classification labels (such as skill level classification, interest field classification, etc.) as output, and train the random forest model. During the training process, adjust parameters such as the number of trees, maximum depth, and node splitting conditions to improve the accuracy and stability of the model.

[0107] (3) Model evaluation and optimization: Evaluate the trained executor portrait model using the validation set, and calculate evaluation metrics such as accuracy, recall, and F1 value. If the model performance is poor, optimize the executor portrait model by adjusting the feature engineering method, model parameters, or trying different models. For example, if the classification accuracy of a certain skill area is found to be low, the amount of training data in this area can be increased, or the weight of the model for the features in this area can be adjusted.

[0108] S2.4.3. Task model training.

[0109] (1) Feature extraction and transformation: Extract and transform the features of the task more deeply. For example, for the task description text, use techniques such as the bag-of-words model, TF-IDF, or word embeddings to convert it into a numerical vector for model processing. Convert the time requirements of the task (such as completion time) into numerical forms such as timestamps or relative time intervals to enhance the model's understanding of time factors.

[0110] (2) Model construction and training: Similarly, select a suitable model to construct the task model, such as a support vector machine (SVM), logistic regression, or a deep learning model. Taking SVM as an example, use the feature vectors of the task as input and the labels such as the type and difficulty level of the task as output to train the SVM model. Optimize the model performance by adjusting the kernel function type, penalty parameter, etc.

[0111] (3) Model validation and improvement: Use the validation set to validate the task model and check the model's ability to extract and classify task features. If the model performs poorly on certain task types or features, analyze the reasons and make improvements. For example, if it is found that the classification accuracy of the model for tasks requiring specific professional knowledge is low, the feature extraction of relevant professional vocabulary can be increased, or the sensitivity of the model to these professional features can be adjusted.

[0112] S2.4.4. Matching algorithm training.

[0113] (1) Training data generation: Generate data for training the matching algorithm based on the training set and the validation set. For each task, select a certain number of executor samples with different degrees of matching from the executor dataset to form task-executor pairs and label their true degrees of matching (which can be determined through manual evaluation or existing business data).

[0114] (2) Algorithm optimization and training: Use the generated task-executor pair data to train and optimize the matching algorithm. For example, for a matching algorithm based on cosine similarity combined with weight adjustment, the weights of features such as skills, geographic location, and interests are adjusted through optimization algorithms such as gradient descent, so that the matching score calculated by the algorithm is closer to the matching score of the actual annotations. During the training process, the algorithm is continuously iterated and optimized to improve the accuracy of the matching score.

[0115] (3) Algorithm evaluation and adjustment: Use the test set to evaluate the trained matching algorithm and calculate indicators such as the mean square error (MSE) and mean absolute error (MAE) between the predicted and true matching degrees. Based on the evaluation results, further adjustments and optimizations are made to the algorithm. For example, if it is found that the matching error of the algorithm is large for certain task types or executor groups, the weight of this part of the data during the training process can be adjusted specifically, or the algorithm's calculation method can be improved.

[0116] S3. Automatic push: Based on task information and comprehensive matching scores, the corresponding tasks are automatically pushed to the executors according to the push rules.

[0117] The push rules include: determining the urgency and location of the task based on the task information, grading the tasks according to their urgency, determining the location range for priority push tasks based on the task location and urgency level, and determining the screening threshold for the comprehensive matching score of the executors based on the urgency level. For example, tasks are divided into three levels based on their urgency: high, medium, and low. For tasks with high urgency, the location range for priority push tasks is set to be within 5 kilometers of the task location, with a screening threshold of 80 points. This means that tasks are prioritized to executors within 5 kilometers of the task location and with a matching score of 80 points or above. For tasks with medium urgency, the location range for priority push tasks is set to be within 10 kilometers of the task location, with a screening threshold of 60 points. For tasks with low urgency, no location range or screening threshold is set for priority push tasks. When no location range or screening threshold is set for priority push tasks, the system pushes tasks to all executors who meet the basic skill requirements in descending order of comprehensive matching scores.

[0118] In this step, after the system completes the matching process, it sends the task push information to the selected performer via platform messaging, SMS, or push notifications based on the aforementioned push rules. This push information includes key task information, such as task type, reward, general task requirements, and task location, to facilitate the performer's quick decision-making.

[0119] S4. Feedback and optimization: Collect feedback information and optimize the executor portrait model, task model and matching algorithm model accordingly.

[0120] The collection of feedback information includes: setting up phased feedback entry points through which the executor and the task issuer can feedback problems encountered in the task, thereby obtaining feedback information. For example, the executor can report problems such as the task difficulty exceeding expectations or the task description being unclear through the feedback entry point. Or after the task is completed, the system sends evaluation and feedback questionnaires to the task issuer and the executor respectively. The task issuer evaluates the executor's work quality, completion time, communication ability, etc.; the executor gives feedback on aspects such as the reasonableness of the task reward, task difficulty, and task interestingness. In this way, the collection of feedback information is achieved. The collection of feedback information can also be carried out by the platform system monitoring the data. For example, through the platform's data monitoring system, indicators such as the task completion rate, average completion time, and executor satisfaction are observed.

[0121] After that, the team regularly analyzes the collected feedback information. For example, if it is found that the task matching effect in a certain area is generally poor, by analyzing the data of the executors and tasks in that area, possible problems can be identified, such as the mismatch between the task types and the executor skill distribution in that area, or the need to optimize the processing method of geographical location information. According to the analysis results, targeted optimization is carried out on the executor portrait, task model, and matching algorithm. For example, if it is found that the weight setting of a certain skill type in the matching algorithm is unreasonable, resulting in inaccurate matching between tasks and executors, the weight of that skill type is readjusted, and then the historical data is recalculated and verified to ensure that the optimized model can improve the accuracy of task push.

[0122] The present invention has been described exemplarily above in conjunction with the accompanying drawings. Obviously, the specific implementation of the present invention is not limited by the above methods. As long as various non-substantive improvements are made using the inventive concept and technical solutions of the present invention, or the inventive concept and technical solutions of the present invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. An automatic task pushing method based on a crowdsourcing human platform, characterized in that: It includes the following steps: S1. Data collection: Collect executor information and task information; S2. Data analysis and modeling: The models constructed in this step include an executor portrait model, a task model, and a matching algorithm model; S3. Automatic push: Based on the task information and the comprehensive matching score, automatically push corresponding tasks to the executor according to the push rules; S4. Feedback and optimization: Collect feedback information and optimize the executor portrait model, task model, and matching algorithm model accordingly.

2. The automatic task pushing method based on a crowdsourcing human platform according to claim 1, wherein: Step S1 includes: 1) Collection of executor information: The executor information includes professional skills, hobbies, and geographical location information; 2) Collection of task information: The task information includes task type, task requirements, and task location. The task requirements include skill requirements and completion time.

3. The automatic task pushing method based on a crowdsourcing manpower platform according to claim 1, wherein: Step S2 includes: S2.

1. Establish an executor portrait model: This model constructs an executor portrait based on the collected executor information using data analysis algorithms; S2.

2. Construct a task model: Analyze the task information, extract key features and requirements, and construct a task model; S2.

3. Design of a matching algorithm model: Design a matching algorithm model based on multi-dimensional features to achieve the matching between tasks and executors; S2.

4. Train each model to implement parameter adjustment.

4. The automatic task pushing method based on a crowdsourcing human platform according to claim 3, wherein: Step 2.3 includes: S2.3.

1. Calculate the matching degree of each factor feature; S2.3.

2. Assign basic weights based on the importance of each factor requirement; S2.3.

3. Adjust the weights of each factor on the basis of the basic weights according to the task situation, task type, and task feedback information; S3.3.

4. After the weight adjustment is completed, calculate the comprehensive matching degree score through the calculation formula of the comprehensive matching degree score based on the finally determined adjusted weight; the calculation formula of the comprehensive matching degree after the weight adjustment is: Comprehensive matching degree score = w1×C1 + w2×C2 + … + w n ×C n , Among them, C n represents the matching degree of the nth factor, and w n represents the weight corresponding to the nth factor. w2×C2 represents the product of the matching degree of the nth factor and the corresponding weight. Specifically, the matching degree includes at least the skill matching degree and the location matching degree.

5. The automatic task pushing method based on a crowdsourcing human platform according to claim 4, wherein: The said step S2.3.3 includes: 1) Weight adjustment according to the task situation: For ordinary tasks, appropriately balance the weights of each factor in the matching degree calculation; the urgency of urgent tasks is relatively high, and a large amount of data needs to be collected in a short time. The weights of key factors such as skill matching degree and location matching degree should be increased on the basis of the basic weights; in long-term tasks, the stability of the executor is important, and the weights of factors related to stability should be appropriately increased. When the hobby is related to stability, the weight of the interest matching degree should be increased, and the executor stability factor can be added. The stability factor extracts corresponding feature vectors by quantifying the completion coherence and reputation of the executor's past tasks, and quantifies the information of the task continuation time in the task information to extract corresponding feature vectors, thereby calculating the stability matching degree and giving a certain weight to the stability factor; 2) Weight adjustment according to the task type: For professional skill-based tasks that require highly professional skills, the weight of the skill matching degree should be dominant; for service-based tasks, the weights of the location matching degree and the skill matching degree are equally important; for creative tasks, the weights of the interest matching degree and the skill matching degree are relatively high; 3) Adjust the weights of various factors based on feedback information: Collect the feedback information from the executors on the task push for weight adjustment; Obtain the feedback on the task completion quality and executor matching degree from the task publisher and adjust the weights; Through the platform's data monitoring system, observe indicators including the task completion rate, average completion time, and executor satisfaction, and optimize the weight settings based on the indicator changes.

6. The automatic task pushing method based on a crowdsourcing human platform according to claim 4, wherein: The step S2.4 includes: S2.4.

1. Data preparation, which is used to form the training set required for training, including data integration and cleaning, data annotation, and data partitioning; S2.4.

2. Training of the executor portrait model. Use the executor feature data in the training set as the input and the classification label of the executor as the output to train the random forest model to obtain the executor portrait model; S2.4.

3. Training of the task model. Use the bag-of-words model, TF-IDF, or word embedding to convert the task description text into a numerical vector, convert the time requirement into a numerical form of time, use the feature vector of the task as the input and the corresponding label of the feature vector as the output to train the SVM model to obtain the task model; S2.4.

4. Training of the matching algorithm. For each task, select a certain number of executor samples with different matching degrees from the executor dataset to form task-executor pairs and label their true matching degrees; Use the generated task-executor pair data to adjust the weight parameters of the matching algorithm through the gradient descent optimization algorithm to optimize the matching algorithm model.

7. The automatic task pushing method based on a crowdsourcing human platform according to claim 1, wherein: The push rules include: Determine the urgency level and task location of the task based on the task information, classify the tasks according to the urgency level of the task, and determine the location range of the priority push tasks according to the task location and urgency level of the task, and determine the screening threshold of the comprehensive matching score of the priority push tasks for the executors based on the urgency level; For tasks with a low urgency level, no location range and screening threshold for the priority push tasks are set. When no location range and screening threshold for the priority push tasks are set, the system pushes to all executors who meet the basic skill requirements in the order of the comprehensive matching score from high to low.

8. The automatic task pushing method based on a crowdsourcing human platform according to claim 3, characterized in that: In step S2.1, the data analysis performed by the executor portrait model includes: Quantitatively evaluating the skill level of the executor; Using natural language processing technology to classify and label the description of the executor's hobbies; Analyzing the geographical location information of the executor to determine their main activity area; After quantifying the executor information, the data analysis extracts the feature vectors of the corresponding factors. The feature vectors include the skill vector representing the skill level, the location vector representing the geographical location of the executor, and the interest vector representing the executor's hobbies.

9. The automatic task pushing method based on a crowdsourcing human platform according to claim 3, wherein: The step S2.2 includes: 1) Standardized coding of task types, which is used to identify and distinguish tasks of different types and at different times; 2) Quantify the skill requirements. For the skill requirements needed for the task, use a quantitative method for accurate representation. According to the skill level standard, quantify the corresponding skill requirements into a specific value. Considering the importance differences of different skills in the task, assign corresponding weights to each skill. After quantifying the skill requirements, convert them into a skill requirement vector, and the skill requirement vector contains the values of the skill requirements of each category of skills and the corresponding weights; 3) Determine the geographical range characteristics. According to the task location information, transform it to obtain accurate geographical coordinates or a geographical region object containing boundary coordinate information.

Images