An adaptive crowdsourcing task recommendation method based on graph neural network

Abstract

The invention discloses an adaptive crowdsourcing task recommendation method based on graph neural network, obtains a crowdsourcing task object set and a recommended population set in a real application scenario, constructs a crowdsourcing network graph structure and a recommended population network graph structure according to the crowdsourcing task object and the recommended population according to the classic KNN method; uses the graph neural network method GCN to perform multiple rounds of information dissemination on the crowdsourcing network graph structure and the recommended population network graph structure, respectively, and performs task embedding representation and recommender embedding representation learning on the crowdsourcing task and the recommender; recursively performs embedding representation learning operation, and after multiple rounds of learning, the obtained crowdsourcing embedding representation and recommender embedding representation are mapped to the same space using t-SNE, and according to the proximity principle, the corresponding task is recommended to a specific individual. The invention can not only accurately recommend crowdsourcing tasks to individuals, but also is more efficient and accurate than traditional crowdsourcing task recommendation methods, and is more scientific and reasonable.

Description

Technical Field

[0001] The present invention relates to the fields of task recommendation and graph neural network, and is mainly used in crowdsourcing scenarios, and specifically to adaptive crowdsourcing task recommendation based on graph neural network. Background Art

[0002] The use and development of information and communication technology has enabled knowledge to be codified, standardized and digitized, and has also enabled service products to be split or decomposed into small parts. The application of information technology in the business field has enabled services that were traditionally provided within the company to be externalized, and consultations between service providers and customers can be conducted remotely rather than face to face. There are many types of service products that can be decomposed, whether it is low-value data (such as data entered into a computer) or high-value data (such as architectural design, sophisticated financial data analysis, X-rays, movies, software development and advertising clips).

[0003] Crowdsourcing, on the other hand, benefits from the innovation potential brought by social differentiation and diversification. It is a practice in which a company or organization outsources work tasks to non-specific volunteers on a voluntary basis. Crowdsourcing tasks are usually undertaken by individuals, but if they involve tasks that require collaboration among multiple people, they may also appear in the form of individual production based on open source. According to data from the China Internet Network Information Center, as of January 2007, there were 1.1 billion Internet users worldwide. Among the numerous Internet users, there are many talented people with their own characteristics, providing a steady stream of solvers for the various problems faced by enterprises. The latest incarnation of the Internet, Web 2.0, encourages people to develop like-minded communities, not just websites; the development of broadband has accelerated the two-way flow of information, making it easier for enterprises to operate in a crowdsourcing manner.

[0004] There are connections between crowdsourcing tasks, but the correlation between tasks is not taken into account when recommending tasks. In addition, the correlation between recommenders is not taken into account. Graphs have a natural advantage in modeling relationships. Graphs are objects with broad significance. In mathematics, they are the main research object of graph theory; in the computer field, they are a common data structure; and in data science, they are widely used to describe various types of relational data. In fact, in actual data scenarios, we usually call graphs networks. Since the network data form can naturally express the connection between objects, many application scenarios in reality can be modeled as graphs (networks), and graph data is everywhere. Usually, all deep learning methods on graph data are collectively referred to as "graph neural networks."

[0005] Since 2005, the study of graph neural networks has begun. The concept of graph neural networks was first proposed by Gori et al., who drew on the research results in the field of neural networks and designed a model for processing graph structured data. Since then, new graph neural network models and application research have been proposed one after another. In recent years, with the rise of deep learning and the powerful performance of convolutional neural networks, researchers have begun to pay attention to introducing convolution operations into the field of graph data. At the beginning of modeling graph convolutional neural networks, how to use deep learning methods to construct two important operations similar to convolutional neural networks on graph data: convolution operators and pooling operators, or how to construct deep learning models on graphs has attracted the attention of researchers. Summary of the invention

[0006] Purpose of the invention: In the face of a large amount of emerging crowdsourcing task data, the use of traditional crowdsourcing task allocation methods to recommend and allocate tasks is often time-consuming and inefficient, and fails to capture and model the relationship between crowdsourcing tasks, and can no longer meet the growing demand for crowdsourcing task recommendations. The present invention proposes an adaptive crowdsourcing task recommendation method based on a graph neural network, which can construct a graph structure for crowdsourcing tasks and recommended people, and use a graph neural network algorithm to embed the graph structure. It can accurately recommend crowdsourcing tasks to individuals, which is more efficient and accurate than traditional crowdsourcing task recommendation methods. In addition, by constructing a graph structure, the correlation between tasks and the correlation between people can be taken into account, which is more scientific and reasonable.

[0007] Technical solution: To achieve the above purpose, the technical solution adopted by the present invention is:

[0008] An adaptive crowdsourcing task recommendation method based on graph neural network includes the following steps:

[0009] Step 1, data collection: obtain the set of crowdsourcing task objects and the set of recommended people in real application scenarios.

[0010] Step 2, data preprocessing stage: According to the crowdsourcing task object and the recommended population, the crowdsourcing network graph structure and the recommended population network graph structure are constructed based on the classic KNN method.

[0011] Step 3, data analysis phase: Use the graph neural network method GCN to perform multiple rounds of information propagation on the crowdsourcing network graph structure and the recommendation crowd network graph structure, and learn task embedding representation and recommender embedding representation for crowdsourcing tasks and recommenders.

[0012] Step 4, task recommendation stage: recursively perform the embedding representation learning operation in step 3. After multiple rounds of learning, the crowdsourcing embedding representation and the recommender embedding representation are mapped to the same space using t-SNE. According to the proximity principle, the corresponding task is recommended to a specific individual.

[0013] Preferably: the characteristic information of the task object in step 1 includes the start time, end time, category, and task description. The attribute information of the recommended group includes gender, age, occupation, region, and personal information description.

[0014] Preferred: In step 2, the method of constructing the crowdsourcing network graph structure and the recommended population network graph structure according to the crowdsourcing task object and the recommended population according to the classic KNN method:

[0015] Step 2a), for the crowdsourcing task objects, the crowdsourcing task objects are cleaned according to their own characteristic information, and then the k1 other crowdsourcing tasks that are most similar to each crowdsourcing task are found according to the classic KNN method, and edges are established between them and the K1 tasks, thereby constructing a crowdsourcing network graph structure composed of all crowdsourcing tasks, where k1 = 1, 2, 3, ..., K1, K1 represents the number of similar crowdsourcing tasks that need to be found.

[0016] Step 2b), for the recommended groups, the recommended groups are cleaned according to their attribute information, and then the k2 other recommenders that are most similar to each recommended group are found based on the classic KNN method, and edges are established between them and the K2 recommenders, thereby constructing a recommended group network graph structure composed of recommenders, where k2=1,2,3,…,K2, K2 represents the number of similar recommenders that need to be found.

[0017] Preferably: In step 3, the graph neural network method GCN is used to perform multiple rounds of information propagation on the crowdsourcing network graph structure and the crowd network graph structure, and the method for learning the task embedding representation and recommender embedding representation of the crowdsourcing task and the recommended individual is as follows:

[0018] In step 3a), multiple rounds of information propagation are performed on the crowdsourcing network graph structure based on the graph neural network method GCN, and the corresponding task embedding representation is learned for each crowdsourcing task to facilitate its application to downstream tasks.

[0019] Step 3b): Based on the graph neural network method GCN, multiple rounds of information propagation are performed on the recommender network graph structure, and the corresponding recommender embedding representation is learned for each recommender, which is convenient for application to downstream tasks.

[0020] Preferably: in step 4, the embedding representation learning operation of step 3 is recursively performed. After multiple rounds of learning, the crowdsourcing embedding representation and the recommender embedding representation are mapped to the same space using t-SNE. According to the proximity principle, the corresponding task is recommended to a specific individual in the following method:

[0021] In step 4a), the corresponding task embedding representation learned for each crowdsourcing task is mapped to the corresponding two-dimensional space using t-SNE.

[0022] In step 4b), the corresponding recommender embedding representation learned for each recommender is mapped into the corresponding two-dimensional space using t-SNE.

[0023] Step 4c), according to the proximity principle, the corresponding tasks are recommended to specific individuals, multiple tasks can be assigned to one person at the same time, and one task can be assigned to multiple people at the same time.

[0024] Preferably, the number K1 of similar crowdsourcing tasks to be found is 5.

[0025] Preferred: The number of similar recommenders K2 to be found is 5

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] The present invention constructs a graph structure for crowdsourcing tasks and recommended groups, and adopts a graph neural network algorithm to embed and represent the graph structure, which can accurately recommend crowdsourcing tasks to individuals. It is more efficient and accurate than traditional crowdsourcing task recommendation methods. Moreover, by constructing a graph structure, the correlation between tasks and the correlation between people can be taken into account, which is more scientific and reasonable. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The figure is an overall flow chart of the method of the present invention. DETAILED DESCRIPTION

[0029] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0030] An adaptive crowdsourcing task recommendation method based on graph neural network, such as Figure 1 As shown, the following steps are included:

[0031] Step 1, data collection: (1) Obtain a sufficient set of crowdsourcing task objects in real application scenarios, as well as the data generated by these objects with multiple features such as start time, end time, category, task description, etc.; (2) Obtain a set of task recommendation people, as well as the attribute information corresponding to these objects, such as gender, age, occupation, region, and personal information description.

[0032] Step 2, data preprocessing stage: clean the crowdsourcing task objects according to their own multiple feature information, and then find the k other crowdsourcing tasks that are most similar to each crowdsourcing task according to the classic KNN algorithm (K defaults to 5), and establish edges between them and the K tasks, thereby constructing a crowdsourcing network graph structure composed of all crowdsourcing tasks; in addition, for the recommended crowd, the recommended crowd network graph structure is constructed according to the same strategy. The specific steps are as follows:

[0033] Step 2a), for crowdsourcing task objects, the data of the crowdsourcing task objects are cleaned according to their own characteristic information, mainly including data completion, data normalization, and removal of useless fields. Then, according to the classic KNN method, the k1 other crowdsourcing tasks that are most similar to each crowdsourcing task are found, and edges are established between them and the K1 tasks, thereby constructing a crowdsourcing network graph structure composed of all crowdsourcing tasks, where k1 = 1, 2, 3, ..., K1, K1 represents the number of similar crowdsourcing tasks that need to be found. The number of similar crowdsourcing tasks that need to be found, K1, is 5.

[0034] Step 2b), for the recommended groups, the recommended groups are cleaned according to their attribute information, and then the k2 other recommenders that are most similar to each recommended group are found according to the classic KNN method, and edges are established between them and the K2 recommenders, thereby constructing a recommended group network graph structure composed of recommenders, where k2 = 1, 2, 3, ..., K2, K2 represents the number of similar recommenders to be found. The number of similar recommenders to be found K2 is 5.

[0035] Step 3, data analysis phase: Use the graph neural network method GCN to perform multiple rounds of information propagation on the crowdsourcing network graph structure and the recommendation crowd network graph structure, and learn the task embedding representation and recommender embedding representation of the crowdsourcing task and recommender, as follows:

[0036] In step 3a), multiple rounds of information propagation are performed on the crowdsourcing network graph structure based on the graph neural network method GCN, and the corresponding task embedding representation is learned for each crowdsourcing task to facilitate its application to downstream tasks.

[0037] Step 3b): Based on the graph neural network method GCN, multiple rounds of information propagation are performed on the recommender network graph structure, and the corresponding recommender embedding representation is learned for each recommender, which is convenient for application to downstream tasks.

[0038] Step 4, task recommendation stage: recursively perform the embedding representation learning operation in step 3. After multiple rounds of learning, the crowdsourcing embedding representation and the recommender embedding representation are mapped to the same space using t-SNE. According to the proximity principle, the corresponding task is recommended to a specific individual. The specific method is as follows:

[0039] In step 4a), the corresponding task embedding representation learned for each crowdsourcing task is mapped to the corresponding two-dimensional space using t-SNE.

[0040] In step 4b), the corresponding recommender embedding representation learned for each recommender is mapped into the corresponding two-dimensional space using t-SNE.

[0041] Step 4c), according to the proximity principle, the corresponding tasks are recommended to specific individuals, multiple tasks can be assigned to one person at the same time, and one task can be assigned to multiple people at the same time.

[0042] In actual deployment, the choice of K in KNN and the specific graph neural network algorithm cannot be a one-time solution. They need to be changed in experiments according to specific needs, and the embedded representation also needs to be periodically updated.

[0043] In summary, the present invention proposes an adaptive crowdsourcing task recommendation method based on graph neural network, which can construct a graph structure for crowdsourcing tasks and recommended crowds, and use a graph neural network algorithm to embed and represent the graph structure, so as to accurately recommend crowdsourcing tasks to individuals. It is more efficient and accurate than traditional crowdsourcing task recommendation methods. Moreover, by constructing a graph structure, the correlation between tasks and the correlation between people can be taken into account, which is more scientific and reasonable.

[0044] The present invention constructs a graph structure for crowdsourcing tasks and recommended groups, and adopts a graph neural network algorithm to embed and represent the graph structure, which can accurately recommend crowdsourcing tasks to individuals. It is more efficient and accurate than traditional crowdsourcing task recommendation methods. Moreover, by constructing a graph structure, the correlation between tasks and the correlation between people can be taken into account, which is more scientific and reasonable.

[0045] The present invention can complete accurate task recommendations for a large number of crowdsourcing tasks and people waiting to claim tasks on the Internet.

[0046] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An adaptive crowdsourcing task recommendation method based on graph neural network, characterized in that: The steps include: Step 1, data collection: obtain the set of crowdsourcing task objects and the set of recommended people in real application scenarios; Step 2, data preprocessing stage: construct the crowdsourcing network graph structure and the recommendation group network graph structure according to the crowdsourcing task object and the recommendation group based on the classic KNN method; According to the crowdsourcing task object and the recommended population, the method of constructing the crowdsourcing network graph structure and the recommended population network graph structure based on the classic KNN method: Step 2a), for crowdsourcing task objects, the crowdsourcing task objects are cleaned according to their own characteristic information, and then the k1 other crowdsourcing tasks that are most similar to each crowdsourcing task are found according to the classic KNN method, and edges are established between them and the K1 tasks, thereby constructing a crowdsourcing network graph structure composed of all crowdsourcing tasks, where k1 = 1, 2, 3, ..., K1, K1 represents the number of similar crowdsourcing tasks that need to be found; Step 2b), for the recommended groups, the recommended groups are cleaned according to their attribute information, and then the k2 other recommenders that are most similar to each recommended group are found according to the classic KNN method, and edges are established between them and the K2 recommenders, thereby constructing a recommended group network graph structure composed of recommenders, where k2 = 1, 2, 3, ..., K2, K2 represents the number of similar recommenders that need to be found; Step 3, data analysis phase: Use the graph neural network method GCN to perform multiple rounds of information propagation on the crowdsourcing network graph structure and the recommendation crowd network graph structure, and learn task embedding representation and recommender embedding representation for crowdsourcing tasks and recommenders; Using the graph neural network method GCN, multiple rounds of information propagation are performed on the crowdsourcing network graph structure and the crowd network graph structure respectively. The method of learning the task embedding representation and recommender embedding representation of the crowdsourcing task and the recommended individual is as follows: Step 3a), based on the graph neural network method GCN, multiple rounds of information propagation are performed on the crowdsourcing network graph structure, and the corresponding task embedding representation is learned for each crowdsourcing task, which is convenient for application to downstream tasks; Step 3b), based on the graph neural network method GCN, multiple rounds of information propagation are performed on the recommender network graph structure, and the corresponding recommender embedding representation is learned for each recommender, which is convenient for application to downstream tasks; Step 4, task recommendation stage: recursively perform the embedding representation learning operation in step 3. After multiple rounds of learning, the crowdsourcing embedding representation and the recommender embedding representation are mapped to the same space using t-SNE. According to the proximity principle, the corresponding task is recommended to a specific individual.

2. The method for adaptive crowdsourcing task recommendation based on graph neural network according to claim 1, characterized in that: The characteristic information of the task object in step 1 includes the start time, end time, category, and task description; the attribute information of the recommended group includes gender, age, occupation, region, and personal information description.

3. The method for adaptive crowdsourcing task recommendation based on graph neural network according to claim 2 is characterized in that: In step 4, the embedding representation learning operation of step 3 is recursively performed. After multiple rounds of learning, the crowdsourcing embedding representation and the recommender embedding representation are mapped to the same space using t-SNE. According to the proximity principle, the corresponding task is recommended to a specific individual in the following way: Step 4a), map the corresponding task embedding representation learned for each crowdsourcing task into the corresponding two-dimensional space using t-SNE; Step 4b), map the corresponding recommender embedding representation learned by each recommender into the corresponding two-dimensional space using t-SNE; Step 4c), according to the proximity principle, the corresponding tasks are recommended to specific individuals, multiple tasks can be assigned to one person at the same time, and one task can be assigned to multiple people at the same time.

4. The method for adaptive crowdsourcing task recommendation based on graph neural network according to claim 3 is characterized by: The number K1 of similar crowdsourcing tasks that need to be found is 5.

5. The method for adaptive crowdsourcing task recommendation based on graph neural network according to claim 4 is characterized in that: The number of similar recommenders K2 that needs to be found is 5.

Images