CN113361928B - Crowd-sourced task recommendation method based on heterogram attention network - Google Patents

Abstract

The invention discloses a crowdsourcing task recommendation method based on an heterogram attention network, which comprises the steps of firstly collecting crowdsourcing data, sequentially carrying out feature cleaning, feature screening and feature completion processing to obtain attribute feature vectors of crowdsourcing participators and tasks, constructing a heterogeneous crowdsourcing network, sampling and pre-training according to a set element path, obtaining node feature representation of the personnel and the tasks, carrying out model training, and obtaining attention between node pairs; transmitting and aggregating neighbor node information according to the attention coefficient, and updating node characteristic representation of personnel and tasks; finally, calculating similarity scores of personnel and tasks according to the learned node characteristic representation, and recommending the tasks according to score ordering; aiming at the existing task information and personnel information in crowdsourcing, the crowdsourcing heterogeneous network is established, the attribute vector in the crowdsourcing network is updated by using the graph annotation meaning network, the accuracy of crowdsourcing task recommendation is improved, and the completion efficiency and the completion quality are further improved.

Description

Crowd-sourced task recommendation method based on heterogram attention network

Technical Field

The invention relates to the technical field of crowdsourcing task recommendation, in particular to a crowdsourcing task recommendation method based on an iso-composition attention network.

Background

With the development of economic globalization and network science technology, crowdsourcing is used as an internet-based distributed problem solving mechanism, which integrates computers and unknown masses on the internet to complete tasks which are difficult to be completed by the computers alone. Crowd sourcing has gained very wide attention in academia and industry in the computer field in recent years due to many application contexts, such as massive image annotation, online commodity evaluation, software testing, etc.

In the crowdsourcing mode, how to accurately recommend the task to the participators who are more likely to complete the task is one of the key problems to be solved. The accurate recommendation of the task directly influences the completion efficiency and the completion quality of the crowdsourcing task. However, the lack of information and limitations of recommendation methods make it difficult for existing tasks to be accurately recommended to the participants. Therefore, aiming at the crowdsourcing task recommendation problem, how to accurately recommend the task to the participators who are more likely to complete the task faces a great challenge.

When crowdsourcing task recommendation is performed, only task and personnel attributes are considered, and often an inaccurate recommendation result is obtained. Because the attribute data is an independent data structure, crowdsourcing interaction information of participators and historical tasks, similarity between people and task to task is also important for accurate task recommendation. Thus, modeling crowd-sourced data as a heterogeneous crowd-sourced network will provide rich interaction information for recommended tasks. By learning node representation of personnel and tasks in the crowdsourcing network, accurate task recommendation can be achieved by calculating node representation similarity. However, considering heterogeneous characteristics of nodes in the network, the traditional network representation learning method cannot be directly applied to crowdsourcing task recommendation, and a targeted method needs to be adopted to build a model aiming at the characteristics of crowdsourcing data.

The graph neural network (Graph Neural Networks, GNNs) is a neural network model that learns from a network structure data representation. The graph neural network iteratively learns and updates the characteristic representation of each node by combining the downstream task through the linear transformation of the node characteristics and the neighbor characteristic aggregation. The graph attention network is a graph neural network combined with an attention mechanism, can learn the weight of feature aggregation of different neighbor nodes, and realizes more effective node feature representation learning.

Disclosure of Invention

The invention aims to: most of the prior graph neural networks aim at isomorphic networks, but the crowdsourcing networks have the node isomerism of tasks and personnel, and the prior graph neural network model needs to be modified to realize the extraction of isomerism semantic information. The invention provides a crowdsourcing task recommendation method based on an heterogram attention network, which aims at the existing task information and personnel information in crowdsourcing and establishes a crowdsourcing heterogeneous network. And extracting network structure attributes based on the element paths, and updating feature vectors in the crowdsourcing network by using the graph annotation meaning network so as to improve the accuracy of crowdsourcing task recommendation and improve the completion efficiency and the completion quality of the crowdsourcing task.

The technical scheme is as follows: in order to achieve the above purpose, the invention adopts the following technical scheme:

a crowd-sourced task recommendation method based on an heterograph attention network comprises the following steps:

s1, data acquisition;

collecting crowdsourcing data generated by crowdsourcing participants and crowdsourcing task objects in a plurality of real application scenes; the crowdsourcing data comprises crowdsourcing participant names, sexes, ages, advantages, historical task information, task ids, task descriptions, modules to which tasks belong and task types;

s2, preprocessing data;

sequentially performing feature cleaning, feature screening and feature completion processing on the obtained crowdsourcing data to respectively obtain attribute feature vectors of crowdsourcing participants and crowdsourcing tasks; according to historical interaction information of connection between crowdsourcing participants and crowdsourcing tasks and characteristic information of the crowdsourcing participants and the crowdsourcing tasks, constructing a crowdsourcing network by combining a KNN algorithm, and sampling and pre-training according to a set element path to respectively obtain structural characteristic vectors of the personnel and the tasks based on structural information; finally, the personnel structure feature vector and the task structure feature vector based on the structure information are spliced to obtain node feature representation of personnel and tasks;

s3, training a model;

calculating the attention between node pairs according to the node characteristic representation of the personnel and the task obtained in the step S2, transmitting and aggregating neighbor node information according to the calculated attention coefficient, and updating the node characteristic representation of the personnel and the task;

s4, recommending tasks;

and (3) recursively updating node characteristic representations of the personnel and the tasks in the step (S3), calculating similarity scores of the personnel and the tasks according to the learned node characteristic representations, and recommending the tasks according to score sequences.

Further, in the step S2, a crowdsourcing network is constructed according to historical interaction information of the crowdsourcing participant and the crowdsourcing task, and feature information of the crowdsourcing participant and the crowdsourcing task, and by combining with a KNN algorithm, the method specifically includes:

step S2.1, performing feature cleaning, feature screening and feature completion on all data to obtain a node attribute feature vector H _att ；

And S2.2, selecting the personnel and tasks with the history interaction connection according to the history interaction information of the crowdsourcing participants and the establishment of connection with the crowdsourcing tasks, and establishing personnel-task connection edges. The KNN algorithm is adopted to respectively calculate Euclidean distance for the characteristic information representation of each crowdsourcing participant and other crowdsourcing participants, and a pair-by-pair similarity index set S is obtained _w Calculating Euclidean distance for the characteristic representation of each crowdsourcing task and other crowdsourcing tasks to obtain pair-by-pair similarity index set S _t . For each crowd-sourced participant, select S _w The first K similar nodes with the maximum similarity index values are used for establishing a person-person continuous edge; for each crowdsourcing task, S is selected _t The first K similar nodes with the maximum similarity index values are established as task-task edges, so that a crowdsourcing heterogeneous network G is obtained;

step S2.3, according to the set meta-path set ρ= { ρ ₁ ,ρ ₂ ,…,ρ _M Sampling and obtaining a node structure feature vector set H by using a meta 2vec algorithm _st ＝{H _st1 ,H _st2 ,…,H _stM -a }; the Metapath2vec algorithm is a heterogeneous graph structural feature extraction method based on a meta-path, sampling is carried out by random walk of a predefined meta-path, then node feature representation training is carried out by using a shallow neural network, and semantic information and structural information in a heterogeneous network can be effectively reserved.

Step S2.4, attribute feature vector and structure featureThe sign vectors are spliced to obtain corresponding node characteristic representation H= [ H ] _att ||H _st1 ||H _st2 ||…||H _stM ]。

Further, the model training specific step in the step S3 includes:

step S3.1, based on the input node characteristic H, performing linear transformation through a linear layer, multiplying by a weight matrix W, and obtaining a transformed node characteristic H';

step S3.2, calculating an attention coefficient alpha pair by pair for two nodes connected by each group of edges of the crowdsourcing network; specifically, first, the feature vectors of two nodes i and j are spliced to obtain [ H '' _i ||H’ _j ]Multiplying a learnable attention parameter vector a, outputting a result through a LeakyReLU, and finally carrying out attention coefficient normalization operation by using a softmax function to obtain alpha;

s3.3, correspondingly multiplying the attention coefficient alpha obtained by calculation with the transformed node characteristic H', calculating the aggregation of neighbor attention information for each node, and outputting through a sigmoid activation function to obtain an aggregated central node characteristic vector;

step S3.4, repeatedly executing the steps S3.1-S3.3 by adopting a multi-head attention mechanism to obtain multi-head node representation H _agg ＝{H _agg1 ,H _agg2 ,…,H _aggT And finally, splicing node representations calculated by each attention head to obtain a final node representation Z= [ H ] _agg1 ||H _agg2 ||…||H _aggT ]。

Further, the task recommending step in the step S4 specifically includes:

s4.1, acquiring node representation Z of personnel and tasks in a crowdsourcing network;

step S4.2, for each person p _i Acquiring a task set t= { t which does not have connection relation with the person in the network ₁ ,t ₂ ,…,t _N }；

Step S4.3, calculating personnel p _i And each task t _j The cosine similarity of the task to be recommended corresponding to each person is obtained to obtain a similarity score s= { s ₁ ,s ₂ ,…,s _N }；

Step S4.4, for each person p _i And sorting according to the similarity scores s of the tasks to be recommended, and selecting the L tasks with highest sorting to be recommended.

The beneficial effects are that:

the invention provides a crowdsourcing task recommendation method based on an heterogram attention network, which can effectively realize accurate task recommendation under crowdsourcing scene and improve the working efficiency of a crowdsourcing platform. The method for establishing the heterogeneous crowdsourcing network can fully utilize the interactive information in the data; the advantage of the graph neural network model is fully utilized, and the node representation can be effectively learned for task recommendation.

Drawings

FIG. 1 is a flowchart of a crowd-sourced task recommendation method based on an heterogenous composition attention network provided by the invention;

fig. 2 is a flowchart of a data preprocessing process in an embodiment of the present invention.

FIG. 3 is a flow chart of model training based on a graph attention network in accordance with the present invention.

Detailed Description

The invention will be further described with reference to the accompanying drawings.

The crowd-sourced task recommendation method based on the heterographical attention network shown in fig. 1 comprises the following steps:

and S1, data acquisition.

And collecting crowdsourcing data generated by crowdsourcing participants and crowdsourcing task objects in a plurality of real application scenes, wherein the crowdsourcing data comprises names, sexes, ages, advantages, historical task information, task ids, task descriptions, modules to which tasks belong and task types of the crowdsourcing participants.

And S2, preprocessing data.

The acquired data are subjected to processing such as feature cleaning, feature screening, feature complement and the like to obtain attribute feature vectors of personnel and tasks; meanwhile, a crowdsourcing network is built according to the historical interaction information and a KNN algorithm, a person-task connecting edge is built according to the historical interaction information, and the person-person connecting edge and the task-task connecting edge are built according to the KNN calculation attribute similarity, so that the crowdsourcing network is built. Then sampling and pre-training are carried out according to the set element path, and structural feature vectors of personnel and tasks based on structural information are obtained; and finally, splicing the two types of vectors to obtain the node characteristic representation of the personnel and the task. In particular, the method comprises the steps of,

step S2.1, performing feature cleaning, feature screening and feature completion on all data to obtain a node attribute feature vector H _att ；

And S2.2, selecting the personnel and tasks with the history interaction connection according to the history interaction information of the crowdsourcing participants and the establishment of connection with the crowdsourcing tasks, and establishing personnel-task connection edges. The KNN algorithm is adopted to respectively calculate Euclidean distance for the characteristic information representation of each crowdsourcing participant and other crowdsourcing participants, and a pair-by-pair similarity index set S is obtained _w Calculating Euclidean distance for the characteristic representation of each crowdsourcing task and other crowdsourcing tasks to obtain pair-by-pair similarity index set S _t . For each crowd-sourced participant, select S _w The first K similar nodes with the maximum similarity index values are used for establishing a person-person continuous edge; for each crowdsourcing task, S is selected _t The first K similar nodes with the maximum similarity index values are established as task-task edges, so that a crowdsourcing heterogeneous network G is obtained;

step S2.3, according to the set meta-path set ρ= { ρ ₁ ,ρ ₂ ,…,ρ _M Sampling and obtaining a node structure feature vector set H by using a meta 2vec algorithm _st ＝{H _st1 ,H _st2 ,…,H _stM }；

Step S2.4, splicing the obtained attribute feature vector and the structural feature vector to obtain a corresponding node feature representation H= [ H ] _att ||H _st1 ||H _st2 ||…||H _stM ]。

And S3, model training.

And (3) calculating the attention between the node pairs according to the node characteristic representation of the personnel and the task obtained in the step (S2), and transmitting and aggregating neighbor node information according to the calculated attention coefficient, and updating the node characteristic representation of the personnel and the task. In particular, the method comprises the steps of,

step S3.1, based on the input node characteristic H, performing linear transformation through a linear layer, multiplying by a weight matrix W, and obtaining a transformed node characteristic H';

step S3.2, calculating an attention coefficient alpha pair by pair for two nodes connected by each group of edges of the crowdsourcing network; specifically, first, the feature vectors of two nodes i and j are spliced to obtain [ H '' _i ||H’ _j ]Multiplying a learnable attention parameter vector a, outputting a result through a LeakyReLU, and finally carrying out attention coefficient normalization operation by using a softmax function to obtain alpha;

step S3.3, correspondingly multiplying the calculated attention coefficient alpha by the transformed node characteristic H ', and adding the vectors multiplied by the attention of the neighbor nodes to each node i to obtain an aggregated central node characteristic vector H' i, namely H 'i=Σαij H' j, j epsilon Ni;

step S3.4, the representation H 'of all nodes is used as the input of a sigmoid activation function, and a node characteristic vector Hagg after nonlinear activation is obtained, namely Hagg=sigmoid (H'); wherein sigmoid (x) =1/(1+e-x);

step S3.5, repeatedly executing the steps S3.1-S3.3 by adopting a multi-head attention mechanism to obtain multi-head node representation H _agg ＝{H _agg1 ,H _agg2 ,…,H _aggT And finally, splicing node representations calculated by each attention head to obtain a final node representation Z= [ H ] _agg1 ||H _agg2 ||…||H _aggT ]。

S4, recommending tasks;

and (3) recursively updating node characteristic representations of the personnel and the tasks in the step (S3), calculating similarity scores of the personnel and the tasks according to the learned node characteristic representations, and recommending the tasks according to score sequences. In particular, the method comprises the steps of,

s4.1, acquiring node representation Z of personnel and tasks in a crowdsourcing network;

step S4.2, for each person p _i Acquiring a task set t= { t which does not have connection relation with the person in the network ₁ ,t ₂ ,…,t _N }；

Step S4.3, calculating personnel p _i And each task t _j The cosine similarity of the task to be recommended corresponding to each person is obtained to obtain a similarity score s= { s ₁ ,s ₂ ,…,s _N }；

Step S4.4, for each person p _i And sorting according to the similarity scores s of the tasks to be recommended, and selecting the L tasks with highest sorting to be recommended.

The foregoing is only a preferred embodiment of the invention, it being noted that: it will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the principles of the present invention, and such modifications and adaptations are intended to be comprehended within the scope of the invention.

Claims (4)

1. The crowd-sourced task recommendation method based on the heterograms attention network is characterized by comprising the following steps of:

s1, data acquisition;

collecting crowdsourcing data generated by crowdsourcing participants and crowdsourcing tasks in a plurality of real application scenes; the crowdsourcing data comprises crowdsourcing participant names, sexes, ages, advantages, historical task information, task ids, task descriptions, modules to which tasks belong and task types;

s2, preprocessing data;

sequentially performing feature cleaning, feature screening and feature completion processing on the obtained crowdsourcing data to respectively obtain attribute feature vectors of crowdsourcing participants and tasks; according to historical interaction information of connection between crowdsourcing participants and crowdsourcing tasks and characteristic information of the crowdsourcing participants and the crowdsourcing tasks, a crowdsourcing network is constructed based on a KNN algorithm, sampling and pre-training are carried out according to a set element path, and structural characteristic vectors of the personnel and the tasks based on structural information are respectively obtained; finally, the personnel structure feature vector and the task structure feature vector based on the structure information are spliced to obtain node feature representation of personnel and tasks;

s3, training a model;

calculating the attention between node pairs according to the node characteristic representation of the personnel and the task obtained in the step S2, transmitting and aggregating neighbor node information according to the calculated attention coefficient, and updating the node characteristic representation of the personnel and the task;

s4, recommending tasks;

and (3) recursively updating node characteristic representations of the personnel and the tasks in the step (S3), calculating similarity scores of the personnel and the tasks according to the learned node characteristic representations, and recommending the tasks according to score sequences.

2. The crowd-sourced task recommendation method based on the iso-patterned attention network according to claim 1, wherein the constructing the crowd-sourced network based on the KNN algorithm in step S2 specifically includes:

step S2.1, performing feature cleaning, feature screening and feature completion on all data to obtain a node attribute feature vector H _att ；

Step S2.2, selecting the personnel and tasks with history interaction connection according to the history interaction information of the connection between the crowdsourcing participants and the crowdsourcing tasks, establishing personnel-task connection edges, and calculating Euclidean distances respectively for the characteristic information representation of each crowdsourcing participant and other crowdsourcing participants by adopting a KNN algorithm to obtain a pair-by-pair similarity index set S _w Calculating Euclidean distance for the characteristic representation of each crowdsourcing task and other crowdsourcing tasks to obtain pair-by-pair similarity index set S _t The method comprises the steps of carrying out a first treatment on the surface of the For each crowd-sourced participant, select S _w The first K similar nodes with the maximum similarity index values are used for establishing a person-person continuous edge; for each crowdsourcing task, S is selected _t The first K similar nodes with the maximum similarity index values are established as task-task edges, so that a crowdsourcing heterogeneous network G is obtained;

step S2.3, according to the set meta-path set ρ= { ρ ₁ ,ρ ₂ ,…,ρ _M Sampling and obtaining a node structure feature vector set H by using a meta 2vec algorithm _st ＝{H _st1 ,H _st2 ,…,H _stM }；

And S2, performing step S2.4. Splicing the obtained attribute feature vector and the structural feature vector to obtain a corresponding node feature representation H= [ H ] _att ||H _st1 ||H _st2 ||…||H _stM ]。

3. The crowd-sourced task recommendation method based on an heterographical attention network according to claim 1, wherein the model training specific step in step S3 comprises:

step S3.1, based on the input node characteristic H, performing linear transformation through a linear layer, multiplying by a weight matrix W, and obtaining a transformed node characteristic H';

step S3.2, calculating an attention coefficient alpha pair by pair for two nodes connected by each group of edges of the crowdsourcing network; specifically, first, the feature vectors of two nodes i and j are spliced to obtain [ H '' _i ||H’ _j ]Multiplying a learnable attention parameter vector a, outputting a result through a LeakyReLU, and finally carrying out attention coefficient normalization operation by using a softmax function to obtain alpha;

s3.3, correspondingly multiplying the attention coefficient alpha obtained by calculation with the transformed node characteristic H', calculating the aggregation of neighbor attention information for each node, and outputting through a sigmoid activation function to obtain an aggregated central node characteristic vector;

step S3.4, repeatedly executing the steps S3.1-S3.3 by adopting a multi-head attention mechanism to obtain multi-head node representation H _agg ＝{H _agg1 ,H _agg2 ,…,H _aggT And finally, splicing node representations calculated by each attention head to obtain a final node representation Z= [ H ] _agg1 ||H _agg2 ||…||H _aggT ]。

4. The crowd-sourced task recommendation method based on an iso-patterned attention network according to claim 1, wherein the task recommendation step in step S4 specifically includes:

s4.1, acquiring node representation Z of personnel and tasks in a crowdsourcing network;

step S4.2, for each person p _i Acquiring a task set t= { t which does not have connection relation with the person in the network ₁ ,t ₂ ,…,t _N }；

Step S4.3, calculating personnel p _i And each task t _j The cosine similarity of the task to be recommended corresponding to each person is obtained to obtain a similarity score s= { s ₁ ,s ₂ ,…,s _N }；

Step S4.4, for each person p _i And sorting according to the similarity scores s of the tasks to be recommended, and selecting the L tasks with highest sorting to be recommended.