| As Internet technology continues to develop,crowdsourcing is a new humanmachine based solution that facilitates the rapid completion of various tasks by different workers in an organisation.Among the various types of crowdsourcing tasks,knowledge-based crowdsourcing is particularly important as it requires a high level of expertise,skills and experience and requires high-quality deliverables.However,crowdsourcing platforms for such tasks often act as a mere middleman.Crowdsourcing task publishers spend a lot of time waiting for enough workers to sign up for crowdsourcing tasks,and then filtering them to find the right ones.Even if the selected workers are able to complete the task,there is no guarantee that they are the best person for a particular task,which can affect the quality and efficiency of the final deliverables.Therefore,recommending suitable and reliable workers for knowledge-based crowdsourcing tasks is a valuable and rewarding task.Knowledge-based crowdsourcing tasks are different from other types of tasks in that they require the target worker to have a corresponding professional background and there is a natural correspondence between the two,so in order to recommend suitable workers for such tasks,this thesis uses multi-label classification techniques to handle the process,i.e.the workers to be recommended are regarded as labels,so that the relevant labels can be automatically matched for a given crowdsourcing task,thus achieving the purpose of recommending the corresponding worker.Although some progress has been made in the research of classification techniques,the following challenges exist in the crowdsourcing scenario: on the one hand,insufficient mining of the textual content of crowdsourcing tasks,data sparsity exhibited by the historical data of crowdsourcing and category imbalance caused by the huge difference in the number of target and non-target labels for crowdsouring tasks,which affect the recommendation accuracy.On the other hand,due to the nature of crowdsourcing,workers participate in the crowdsourcing process on a free and voluntary basis,which leads to a huge number of workers(labels)in some crowdsourcing platforms.Finding a subset of labels relevant to a crowdsourcing task in a large label space is a challenging task,which can affect the performance and scalability of the approach.To address these issues,this thesis proposes a crowdsourcing recommendation method based Capsule Network by combining Capsule Network,a loss function designed based on Circle Loss and hierarchical recommendation based on a tree structure.The specific research work is as follows:(1)A crowdsourcing recommendation model based on Capsule Network is proposed.Firstly,in order to solve the problem of inadequate feature mining of crowdsourcing task texts,this thesis adopts a Capsule Network model for feature extraction.This model has a dynamic routing mechanism,which can better learn the relationship between the parts of the crowdsourcing text and the whole,and thus better characterise the crowdsourcing task.Secondly,this thesis applies a loss function designed according to Circle Loss to alleviate the category imbalance problem caused by the large difference between the number of target labels(workers)and the number of non-target labels(workers)for crowdsourcing tasks.With the proposed method,the recommendation model can recommend the right crowdsourcing workers more accurately and is more suitable for practical use.(2)A hierarchical crowdsourcing recommendation model based on a tree structure is proposed.To cope with the scalability limitations of Capsule Network due to the large number of workers(labels)in a crowdsourcing platform,this thesis constructs a label tree by clustering labels based on the task features corresponding to each label.Each layer of the tree will train a Capsule Network model corresponding to it separately,which will only train or predict on the set of candidate labels predicted by the previous layer,thus reducing the size of the model output space.At the same time,this thesis proposes a candidate routing mechanism,i.e.the Capsule Network will only route to part of a given output space instead of all of it,thus reducing the computational resource overhead of the Capsule Network in training or prediction.Furthermore,in order to better optimise the convergence of the model,a routing algorithm that can adaptively determine the number of iterations is applied in the routing process.Through these optimisations,the proposed model is able to scale and have better accuracy on datasets with a large number of labels.(3)This thesis collects data from the real crowdsourcing platform Topcoder and divides it into four datasets of different sizes: Topcoder-1701,Topcoder-453,Topcoder-255 and Topcoder-166,and introduces the public dataset EUR-Lex in order to better measure the performance of the model in the case of large-scale crowdsourcing workers(labels).In this thesis,full experimental analysis and argumentation are conducted on these datasets,and experimental evaluation is carried out in terms of recommendation performance and parameter sensitivity.The experimental results show that the crowdsourcing recommendation model based Capsule Network can effectively mine the textual information of crowdsourcing tasks and improve the recommendation performance compared to conventional methods.In the face of large-scale labels,the hierarchical crowdsourcing recommendation model combining tree structure and Capsule Network proposed in this thesis effectively breaks the application bottleneck of Capsule Network in terms of scalability.The model shows better recommendation performance compared with methods under the same domain. |
