CN115495648B - Space crowdsourcing method for recommending task release time - Google Patents

Space crowdsourcing method for recommending task release time Download PDF

Info

Publication number
CN115495648B
CN115495648B CN202211046356.XA CN202211046356A CN115495648B CN 115495648 B CN115495648 B CN 115495648B CN 202211046356 A CN202211046356 A CN 202211046356A CN 115495648 B CN115495648 B CN 115495648B
Authority
CN
China
Prior art keywords
task
release time
sequence
representing
matrix
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202211046356.XA
Other languages
Chinese (zh)
Other versions
CN115495648A (en
Inventor
郑凯
赵艳
苏涵
陈轩磊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Yangtze River Delta Research Institute of UESTC Huzhou
Original Assignee
Yangtze River Delta Research Institute of UESTC Huzhou
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Yangtze River Delta Research Institute of UESTC Huzhou filed Critical Yangtze River Delta Research Institute of UESTC Huzhou
Priority to CN202211046356.XA priority Critical patent/CN115495648B/en
Publication of CN115495648A publication Critical patent/CN115495648A/en
Application granted granted Critical
Publication of CN115495648B publication Critical patent/CN115495648B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9536—Search customisation based on social or collaborative filtering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/01—Social networking

Abstract

The invention discloses a space crowdsourcing method for recommending task release time. The lightweight graph convolution neural network is used, the self-connection of the graph and the weight matrix of feature transformation are canceled, and the training speed of the model is accelerated while high-efficiency recommendation is ensured. In the multi-view relation learning of the task release sequence, firstly, multi-view attributes of the tasks are fused, and factors which possibly influence the task release time are comprehensively considered. The importance of other tasks in the sequence to the current task is then calculated using the transform's multi-head attention mechanism, thereby allowing further mining of the context information in the sequence. Finally, the problem of sparsity of data can be solved to a certain extent by using a time period issued by a full-connection layer prediction task, and the prediction accuracy is improved.

Description

Space crowdsourcing method for recommending task release time
Technical Field
The invention relates to the technical field of information, in particular to a space crowdsourcing method for recommending task release time.
Background
In recent years, with the development and widespread use of GPS equipped smart devices and wireless mobile networks (e.g., 5G networks), people can move as sensors and participate in some location-based tasks such as monitoring traffic conditions and reporting local hot spots. Spatial Crowdsourcing (SC) is a recently proposed concept and framework that has found widespread use in many applications. In SC, the platform gathers the space tasks and requires the staff member to physically move to a specific location to complete the assigned tasks. Task allocation techniques in different application scenarios of SC have been widely studied at present, and these techniques are generally based on the assumption that task release time is specified by task requesters. However, in practice, improper task publication time may cause a task requester to wait a long time on the SC platform or cause task allocation/completion failure. The recommendation problem in the SC mainly is to recommend a proper task for a worker or a proper route for completing the task for the worker, and basically no recommendation related to task release time is generated. Conventional task recommendation methods mostly use static features (e.g., custom worker scoring of tasks, worker skills, and task categories) to capture worker preferences. A more typical approach is to learn the worker's preferences for tasks based on some or several static features, using matrix decomposition (MF) or Collaborative Filtering (CF), to implement task recommendations. Because CF has poor expandability on the data set, probability matrix decomposition (PMF) is introduced to replace CF, and PMF can linearly expand along with the data amount and has good performance on large, sparse and unbalanced data sets. However, under the same conditions, the worker can preferentially finish the favorite tasks, so that the sequence of the tasks performed by the worker can also reflect the preference of the worker to a certain extent, and the sequential mode of the moving track of the worker is ignored in the methods. Furthermore, the worker's preference for task execution time is also an important factor in modeling the worker, and existing methods are not substantially concerned with this factor. For the route recommendation method, mainly, route recommendation (dynamic change of tasks and workers) in dynamic scenes is adopted, wherein each task has a corresponding starting point and a corresponding end point, and workers must start from the starting point of the task to the end point of the task when executing the task. All of the above methods are deficient in considering the task requester's recommendation from the worker's perspective.
Disclosure of Invention
The invention aims to provide a space crowdsourcing method for recommending task release time.
In order to achieve the above purpose, the invention is implemented according to the following technical scheme:
the method comprises task related expression learning, multi-view relation learning and task release time prediction, wherein the task related expression learning part eliminates self-connection of the graph and a characteristic transformation weight matrix, and accelerates training speed of a model while ensuring efficient recommendation; the multi-view relation learning part is used for carrying out aiming at a task release sequence, firstly fusing multi-view attributes of each task in the sequence, then using a multi-head attention mechanism of a transducer to further mine context information contained in the sequence, the task release time prediction part is used for predicting task release time by using a full-connection layer to obtain recommended release time of each task, then respectively carrying out dynamic task allocation by using a greedy algorithm and a maximum flow algorithm based on minimum cost, and finally calculating the maximum flow in a network flow diagram by using the maximum flow algorithm of minimum cost, wherein the maximum flow corresponds to the maximum task completion number at the current moment.
The invention can effectively solve the problem of overlong waiting time of task requesters in the SC. Firstly, constructing a cross graph neural network to extract characteristics of task requesters, tasks and task release time; then, carrying out multi-attribute fusion (task, task release time, position of task in sequence) on each task in the task release sequence to obtain a new entity and replace the original task; then utilizing a multi-head self-attention mechanism of a transducer to mine context information in a historical release task sequence of a task requester; and finally, connecting the task release sequence with the characteristic data of the task requester, and inputting a full connection layer (FC) to predict the final task release time. The invention can obviously improve the task completion rate in the task allocation process and reduce the average waiting time of task requesters.
The beneficial effects of the invention are as follows:
compared with the prior art, the space crowdsourcing method for recommending task release time utilizes the cross-map neural network to extract the characteristics of task requesters, tasks and task release time, so that the characteristics of the task requesters, the tasks and the task release time comprise information in different semantic spaces, and the characteristic expression is enriched. The lightweight graph convolution neural network (Light GCN) is used, the self-connection of the graph and the weight matrix of feature transformation are canceled, and the training speed of the model is accelerated while high-efficiency recommendation is ensured. In the multi-view relation learning of the task release sequence, firstly, multi-view attributes of the tasks are fused, and factors which possibly influence the task release time are comprehensively considered. The importance of other tasks in the sequence to the current task is then calculated using the transform's multi-head attention mechanism, thereby allowing further mining of the context information in the sequence. Finally, a full-connection layer is used for predicting the task release time period instead of the task release time, so that the problem of sparseness of data can be solved to a certain extent, and the prediction accuracy is improved.
Drawings
FIG. 1 is a general architecture diagram of the present invention;
fig. 2 is an example of a minimum cost maximum flow network flow diagram of the present invention.
Detailed Description
The invention will be further described with reference to the accompanying drawings and specific embodiments, wherein the exemplary embodiments and descriptions of the invention are for purposes of illustration, but are not intended to be limiting.
As shown in fig. 1: the invention comprises three sub-parts: task related expression learning, multi-view relation learning and task release time prediction. The input data of the whole model is interaction data of a task requester and a task, wherein the task requester and the task release time, and the task requester q releases the task s in a time period t, so that the task requester q and the task s are interacted, and the task release time t is generated. The network output is the recommended task publication time period.
The task related expression learning part of the invention utilizes a cross-map neural network. The graph neural network is not a general graph roll-up neural network (GCN) but a lightweight graph roll-up neural network (Light GCN), eliminates self-connection of graphs and a characteristic transformation weight matrix, and accelerates training speed of a model while guaranteeing efficient recommendation. Firstly, based on the interaction data of a task requester and a task, respectively constructing two graph convolution neural networks to extract the characteristics of the task requester, the task and the task release time. The propagation process of each layer of the graph roll-up neural network of the task requester and the task is as follows:
Figure BDA0003822532730000041
/>
Figure BDA0003822532730000042
wherein Q represents the feature vector of the task requester, S represents the feature vector of the task, Q (i) And S is (i) Feature vectors representing the task requester and the task of the i-th layer, respectively, l 1 Indicating the number of layers of the neural network,
Figure BDA0003822532730000043
A qs adjacency matrix representing interaction matrix of task requester and task +.>
Figure BDA0003822532730000044
Is A qs Degree matrix, Q qs And S is qs A feature matrix representing the task requester and the task that the network ultimately outputs. The output result of the last layer of the graphic neural network is the average value of the characteristic matrix of each layer, and the +.>
Figure BDA0003822532730000045
The propagation process of each layer of the graph convolution neural network corresponding to the task requester and the task release time is similar, and finally a feature matrix Q of the task requester is obtained qt And a feature matrix T of task release time qt 。
After the above operation, the feature matrix of two task requesters can be obtained. In order to control the flow of information in two different semantic spaces (task requester and task, task requester and task release time), a door mechanism is introduced to fuse feature matrices in the two different semantic spaces of the task requester, and the expression mode is as follows:
Q * ＝gate*Q qt +(1-gate)*Q qs
gate＝σ(W g (Concat(Q qs ,Q qt ))+b g )
sigma represents a sigmoid activation function, W g And b g Representing a trainable parameter matrix, Q * Representing the resulting feature matrix of the task requester.
The multi-view relation learning part in the invention is carried out for the task release sequence, and the multi-view attribute of each task in the sequence is fused as follows:
e i ＝s i +t i +pos i
wherein s is i Feature matrix, t, representing task i Feature matrix, pos, representing task release time i Position coding matrix, e, representing task in sequence i Representing the new entity obtained finally, containing more task related information. By e i Instead of the original s i The expression of the new task release sequence can be obtained. Then, using a multi-head attention mechanism of a transducer, the context information contained in the sequence is further mined, and the propagation process of each layer is as follows:
Z l ＝LN(MHA(H l-1 )+H l-1 )
H l ＝LN(TFN(Z l )+Z l )
wherein MHA represents a multi-head attention mechanism, LN represents a normalization layer, TFN represents a two-layer feed forward network, H l Representing the output matrix of the i-th layer. For the multi-head attention mechanism, firstly, the query matrixes are respectively mapped from the output matrixes of the previous layer
Figure BDA0003822532730000051
Key->
Figure BDA0003822532730000052
Value->
Figure BDA0003822532730000053
The following is shown:
Figure BDA0003822532730000054
wherein the method comprises the steps of
Figure BDA0003822532730000055
Representing a weight matrix. ThenThe output of each attention head is calculated,
Figure BDA0003822532730000056
d k representing the dimension of the attention head. Finally, all attention heads are spliced to obtain a final output +.>
Figure BDA0003822532730000057
/>
Figure BDA0003822532730000058
Wherein the method comprises the steps of
Figure BDA0003822532730000061
Is an expression of a task release sequence containing context semantic information. Averaging the expression of each task in the sequence to obtain the final expression of the task issuing sequence +.>
Figure BDA0003822532730000062
The task release time prediction part predicts the task release time by using the full connection layer. Since the task release sequence data is sparse, a time period (a period of 15 minutes is taken as one time period, and one day is divided into 96 time periods) for predicting task release is selected instead of a specific time. The input of the full connection layer is the concatenation of the task requester feature vector and the vector of the task release sequence as follows:
Figure BDA0003822532730000063
Figure BDA0003822532730000064
representing the predicted time period for task release, FC represents fully connected layers, ρ represents the softmax activation function.
After the recommended release time period of each task is obtained, a specified number of workers are generated according to uniform distribution. Dynamic task allocation (the number of workers and tasks is changing at each moment) is then performed using greedy algorithm and minimum cost maximum flow algorithm, respectively. The core of the greedy algorithm is that the worker closest to the task is selected from the available worker set of each task to execute the task, so that the algorithm can only reach the local optimal solution. And constructing a network flow diagram (shown in fig. 2) according to the worker and the task at the current moment based on the minimum cost maximum flow algorithm, and calculating the maximum flow in the network flow diagram by utilizing the minimum cost maximum flow algorithm, wherein the maximum flow corresponds to the maximum task completion number at the current moment.
The technical scheme of the invention is not limited to the specific embodiment, and all technical modifications made according to the technical scheme of the invention fall within the protection scope of the invention.

Claims (1)

1. The space crowdsourcing method for recommending task release time is characterized by comprising task related expression learning, multi-view relation learning and task release time prediction, wherein the task related expression learning part eliminates self-connection of a graph and a characteristic transformation weight matrix, and the training speed of a model is accelerated while high-efficiency recommendation is ensured; the multi-view relation learning part is used for carrying out aiming at a task release sequence, firstly fusing multi-view attributes of each task in the sequence, then using a multi-head attention mechanism of a transducer to further mine context information contained in the sequence, predicting task release time by using a full-connection layer to obtain recommended release time of each task, then respectively carrying out dynamic task allocation by using a greedy algorithm and a maximum flow algorithm based on minimum cost, and finally calculating the maximum flow in a network flow graph by using the maximum flow algorithm of the minimum cost, wherein the maximum flow corresponds to the maximum task completion number at the current moment;
the task related expression learning part firstly respectively builds two graph roll-up neural networks to extract characteristics of task requesters, tasks and task release time based on interaction data of the task requesters and task release time, and the propagation process of each layer of the graph roll-up neural networks of the task requesters and the tasks is as follows:
Figure FDA0004207041130000011
Figure FDA0004207041130000012
wherein Q represents the feature vector of the task requester, S represents the feature vector of the task, Q (i) And S is (i) Feature vectors representing the task requester and the task of the i-th layer, respectively, l 1 Indicating the number of layers of the neural network,
Figure FDA0004207041130000013
A qs adjacency matrix representing interaction matrix of task requester and task +.>
Figure FDA0004207041130000014
Is A qs Degree matrix, Q qs And S is qs A feature matrix representing the task requester and the task which are finally output by the network;
the output result of the last layer of the graphic neural network is the average value of the characteristic matrix of each layer,
Figure FDA0004207041130000015
the propagation process of each layer of the graph convolution neural network corresponding to the task requester and the task release time is similar, and finally the graph convolution neural network is obtained
Feature matrix Q to task requester qt And a feature matrix T of task release time qt ；
The multi-view relation learning part is carried out for a task release sequence, and firstly, multi-view attributes of each task in the sequence are fused, wherein the multi-view attributes are represented by the following formula:
e i ＝s i +t i +pos i
wherein s is i Feature matrix, t, representing task i Feature matrix, pos, representing task release time i Position coding matrix, e, representing task in sequence i Representing the new entity finally obtained, including more task related information; by e i Instead of the original s i The expression of a new task release sequence can be obtained; then, using a multi-head attention mechanism of a transducer, the context information contained in the sequence is further mined, and the propagation process of each layer is as follows:
Z l ＝LN(MHA(H l-1 )+H l-1 )
H l ＝LN(TFN(Z l )+Z l )
wherein MHA represents a multi-head attention mechanism, LN represents a normalization layer, TFN represents a two-layer feed forward network, H l An output matrix representing an i-th layer; the multi-head attention mechanism firstly maps the query matrix from the output matrix of the previous layer
Figure FDA0004207041130000021
Key->
Figure FDA0004207041130000022
Value->
Figure FDA0004207041130000023
The formula is as follows: />
Figure FDA0004207041130000024
Wherein the method comprises the steps of
Figure FDA0004207041130000025
Representing a weight matrix; the output of each attention head is then calculated,
Figure FDA0004207041130000026
d k representing the dimension of the attention head; finally, all the notes are injectedSplicing the force heads to obtain final output +.>
Figure FDA0004207041130000027
Figure FDA0004207041130000028
Wherein the method comprises the steps of
Figure FDA0004207041130000029
Is an expression of a task release sequence containing context semantic information; averaging the expression of each task in the sequence to obtain the final expression of the task issuing sequence +.>
Figure FDA00042070411300000210
The task release time prediction part predicts task release time by using a full-connection layer; the input of the full connection layer is the concatenation of the task requester feature vector and the vector of the task release sequence, as follows:
Figure FDA0004207041130000031
Figure FDA0004207041130000032
representing a predicted task release time period, FC representing a fully connected layer, ρ representing a softmax activation function;
after the recommended release time period of each task is obtained, generating a specified number of workers according to uniform distribution; then respectively utilizing a greedy algorithm and a maximum flow algorithm based on minimum cost to carry out dynamic task allocation;
in order to control the flow of information in two different semantic spaces, a door mechanism is introduced to fuse feature matrices in the two different semantic spaces of a task requester, and the expression mode is as follows:
Q * ＝gate*Q qt +(1-gate)*Q qs
gate＝σ(W g (Concat(Q qs ,Q qt ))+b g )
sigma represents a sigmoid activation function, W g And b g Representing a trainable parameter matrix, Q * Representing the feature matrix of the task requester finally obtained;
the two different semantic spaces are task requesters and tasks, task requesters and task publication times.
CN202211046356.XA 2022-08-30 2022-08-30 Space crowdsourcing method for recommending task release time Active CN115495648B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211046356.XA CN115495648B (en) 2022-08-30 2022-08-30 Space crowdsourcing method for recommending task release time

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211046356.XA CN115495648B (en) 2022-08-30 2022-08-30 Space crowdsourcing method for recommending task release time

Publications (2)

Publication Number Publication Date
CN115495648A CN115495648A (en) 2022-12-20
CN115495648B true CN115495648B (en) 2023-06-09

Family

ID=84466983

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211046356.XA Active CN115495648B (en) 2022-08-30 2022-08-30 Space crowdsourcing method for recommending task release time

Country Status (1)

Country Link
CN (1) CN115495648B (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113852492A (en) * 2021-09-01 2021-12-28 南京信息工程大学 Network flow prediction method based on attention mechanism and graph convolution neural network
CN114038200A (en) * 2021-11-29 2022-02-11 东北大学 Attention mechanism-based time-space synchronization map convolutional network traffic flow prediction method
CN114817663A (en) * 2022-05-05 2022-07-29 杭州电子科技大学 Service modeling and recommendation method based on class perception graph neural network

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20220129790A1 (en) * 2020-10-28 2022-04-28 Verizon Media Inc. System and method for deep enriched neural networks for time series forecasting

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113852492A (en) * 2021-09-01 2021-12-28 南京信息工程大学 Network flow prediction method based on attention mechanism and graph convolution neural network
CN114038200A (en) * 2021-11-29 2022-02-11 东北大学 Attention mechanism-based time-space synchronization map convolutional network traffic flow prediction method
CN114817663A (en) * 2022-05-05 2022-07-29 杭州电子科技大学 Service modeling and recommendation method based on class perception graph neural network

Also Published As

Publication number Publication date
CN115495648A (en) 2022-12-20

Similar Documents

Publication Publication Date Title
CN110032782B (en) City-level intelligent traffic signal control system and method
Liang et al. Applying genetic algorithm and ant colony optimization algorithm into marine investigation path planning model
CN109887284B (en) Smart city traffic signal control recommendation method, system and device
Kong et al. Big data‐driven machine learning‐enabled traffic flow prediction
CN114139637B (en) Multi-agent information fusion method and device, electronic equipment and readable storage medium
CN114415735B (en) Dynamic environment-oriented multi-unmanned aerial vehicle distributed intelligent task allocation method
CN112686428A (en) Subway passenger flow prediction method and device based on subway network station similarity
CN112990485A (en) Knowledge strategy selection method and device based on reinforcement learning
Dai et al. Spatio-temporal deep learning framework for traffic speed forecasting in IoT
Huo et al. Cooperative control for multi-intersection traffic signal based on deep reinforcement learning and imitation learning
CN114780739A (en) Time sequence knowledge graph completion method and system based on time graph convolution network
Xiong et al. DCAST: a spatiotemporal model with DenseNet and GRU based on attention mechanism
CN115495648B (en) Space crowdsourcing method for recommending task release time
CN113205220A (en) Unmanned aerial vehicle logistics distribution global planning method facing real-time order data
CN111507499B (en) Method, device and system for constructing model for prediction and testing method
CN107491841A (en) Nonlinear optimization method and storage medium
CN117131979A (en) Traffic flow speed prediction method and system based on directed hypergraph and attention mechanism
KR102477717B1 (en) Big data-based carbon emission calculation for each situation and automatic trading method, device and system based on artificial intelligence
CN116643877A (en) Computing power resource scheduling method, training method and system of computing power resource scheduling model
CN116258253A (en) Vehicle OD prediction method based on Bayesian neural network
CN116596105A (en) Charging station load prediction method considering power distribution network development
CN116523187A (en) Engineering progress monitoring method and system based on BIM
CN111596668B (en) Mobile robot anthropomorphic path planning method based on reverse reinforcement learning
Chen et al. Q‐EANet: Implicit social modeling for trajectory prediction via experience‐anchored queries
Wang et al. Hybrid neural network modeling for multiple intersections along signalized arterials-current situation and some new results

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant