CN111311115B - Group task allocation method based on space crowdsourcing social influence preference - Google Patents
Group task allocation method based on space crowdsourcing social influence preference Download PDFInfo
- Publication number
- CN111311115B CN111311115B CN202010170604.6A CN202010170604A CN111311115B CN 111311115 B CN111311115 B CN 111311115B CN 202010170604 A CN202010170604 A CN 202010170604A CN 111311115 B CN111311115 B CN 111311115B
- Authority
- CN
- China
- Prior art keywords
- task
- worker
- workers
- group
- tasks
- 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.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 29
- 230000007246 mechanism Effects 0.000 claims abstract description 14
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 8
- 230000003993 interaction Effects 0.000 claims description 64
- 239000013598 vector Substances 0.000 claims description 42
- 230000006870 function Effects 0.000 claims description 14
- 239000000126 substance Substances 0.000 claims description 8
- 230000008569 process Effects 0.000 claims description 6
- 239000013604 expression vector Substances 0.000 claims 2
- 230000000694 effects Effects 0.000 description 9
- 238000002474 experimental method Methods 0.000 description 5
- 238000005457 optimization Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000012358 sourcing Methods 0.000 description 3
- 241000220225 Malus Species 0.000 description 2
- 235000021016 apples Nutrition 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 230000002085 persistent effect Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
Images
Classifications
-
- 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
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0631—Resource planning, allocation, distributing or scheduling for enterprises or organisations
- G06Q10/06311—Scheduling, planning or task assignment for a person or group
- G06Q10/063118—Staff planning in a project environment
Landscapes
- Business, Economics & Management (AREA)
- Human Resources & Organizations (AREA)
- Engineering & Computer Science (AREA)
- Strategic Management (AREA)
- Entrepreneurship & Innovation (AREA)
- Economics (AREA)
- Operations Research (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Marketing (AREA)
- Educational Administration (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- Physics & Mathematics (AREA)
- General Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Factory Administration (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a group task allocation method based on spatial crowdsourcing social influence preference, which comprises the following steps of (S1) analyzing social influence preference values of each worker to different task types by using a bipartite graph embedding model BGEM and an attention mechanism, namely direct weights of each worker and the task types; (S2) the task assignment algorithm introduced in tree decomposition assigns a task to a plurality of workers, i.e., a worker group, according to the social influence preference value of each worker for different task categories. Through the scheme, the spatial tasks are allocated by using the Bipartite Graph Embedding Model (BGEM) and the attention mechanism learning worker group to the preference value of the task types based on the social influence on the basis of the preference value, so that the purpose of maximum task allocation is achieved, and the method has high practical value and popularization value.
Description
Technical Field
The invention belongs to the technical field of computers, and particularly relates to a group task allocation method based on space crowdsourcing social influence preference.
Background
With the widespread deployment of wireless networks and mobile devices (e.g., smart phones), Spatial Crowdsourcing (SC) is an emerging paradigm for monitoring various phenomena of human activity with distributed mobile devices, which has attracted widespread attention in both academic and industrial areas. The main idea of spatial crowdsourcing is to recruit a set of available workers to perform a task at a particular location by actually moving to those locations, which is referred to as task assignment.
Most of the existing SC studies are focused on single task allocation, which assumes that tasks are simple and that each task can only be allocated to a single worker. For example, Tong Yongxin et al designs some effective greedy algorithms to solve the Global Online Micro-task Allocation problem (GOMA) proposed in spatial crowdsourcing. Deng dingxinong et al, which considers task allocation and scheduling simultaneously, invented an approximation algorithm, iteratively improved allocation and scheduling to achieve more complex tasks. In reality, however, a single worker may not be able to independently perform a complex task (e.g., monitoring traffic flow or carrying heavy objects in an area) because performing this task alone exceeds the worker's ability. In this case, each Task should be assigned to a Group of workers, which is called Group Task Assignment (GTA).
Group task assignment requires a group of workers to perform the task by actually heading to the location of the task within a particular time. Some previous studies have explored the problem of group task assignment in spatial crowdsourcing. For example, Gao Dawei et al presents a Team-Oriented Task Planning (TOTP) problem that performs a feasible plan for a worker and meets the skill requirements of different tasks on the worker. Gao Dawei et al also proposed a Top-k team recommendation framework in spatial crowdsourcing, in which a team leader is appointed in each recommended team of workers to facilitate coordination of the different workers. Cheng Peng et al consider collaboration in task assignment, where workers are required to collaborate and complete a task, resulting in a higher overall collaboration quality score. However, they cannot efficiently integrate group preferences, which is an important factor in improving the quality of group task assignment in spatial crowd sourcing, because group members may be reluctant to perform tasks assigned to them when they are not interested in them.
Disclosure of Invention
In order to overcome the defects in the prior art, the invention provides a group task allocation method based on space crowdsourcing social influence preference, a Bipartite Graph Embedded Model (BGEM) and a attention mechanism learning worker group are used for allocating space tasks on the basis of a preference value of task types based on social influence, and thus, the maximum task allocation is achieved.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows:
a group task allocation method based on spatial crowdsourcing social influence preference comprises the following steps:
(S1) analyzing the social influence preference value of each worker on different task categories, namely the direct weight of each worker and task category, by using a bipartite graph embedding model BGEM and an attention mechanism;
(S2) the task assignment algorithm introduced in tree decomposition assigns a task to a plurality of workers, i.e., a worker group, according to the social influence preference value of each worker for different task categories.
Further, the judged types of the social influence preference value in the step (S1) include a personal interaction model and a worker group interaction model.
Further, the personal interaction model is the interaction between a given worker and a task, i.e. worker-task interaction data, using BGEM modeling personal worker-task interactions:
wherein W represents a worker set, C represents a task category, and W ≦ C represents a task categorySet of nodes in (E)WCRepresenting the set of edges between the worker and the task category.
Further, the worker group interaction model is for interactions between a given worker group and task category, using BGEM modeling worker group-task interactions:
wherein G represents a set of worker groups, and G { [ U ] C representsSet of nodes in (E)GCRepresenting the set of edges between the worker group and the task category.
Further, the assigning of the group tasks according to the social influence preference value of each worker for different task categories in the step (S2) may obtain available workers for each task or obtain a set of available worker groups for each task.
Further, the available workers who obtain each task obtain a set of workers who can complete the task without violating constraints, wherein the reachable set of workers for task s uses RWsIs shown, i.e.Simultaneously, the following conditions are met:
(1) the worker arrives at the task place before the task s failsThe location of affairs, i.e. tnow+t(w.l,s.l)≤s.e;
(2) The task s is within the reach of workers, namely d (w.l, s.l) is less than or equal to w.r;
(3) the worker arrives at the place of task s, i.e. t, within his own effective timenow+t(w.l,s.l)≤w.off;
Wherein, tnowRepresents the current time, t (w.l, s.l) represents the travel time from worker position w.l to position s.l, d (w.l, s.l) represents the travel distance between a given position w.l and position s.l; the three conditions described above ensure that worker w can travel directly from its location w.l to the location of task s before task s fails and within the time available to worker w.
Specifically, the obtaining of the set of available groups of workers for each task finds the set of available groups of workers given the reachable workers for each task s, under constraints of time available for workers in the group and the number of workers allowed to be allocated to perform the task s, and represents the set of available workers by awg(s), while awg(s) satisfies the following condition:
(1)|AWG(s)|=s.numW;
where | awg(s) | denotes the number of workers in awg(s), and the two conditions described above ensure that workers in a group can reach the location of task s without affecting the mutual available time.
Experimental setup:
the present invention conducted experiments using the check-in dataset in Twitter, which provides check-in data ranging from 9 months 2010 to 1 month 2011 in the united states except hawaii and ariska, including 62462 sites and 61412 user locations. The data set is widely applied to evaluation of crowdsourcing platforms. Since the category information of the place is missing in the data set, the present invention generates category information (i.e., task category information) for each place by means of the API information of Foursquare. When using this data set in experimental studies, it is assumed that users in the data set are workers in a crowd-sourced platform, and the worker location is the latest check-in point, since users checking in at different places may be good candidates to perform spatial tasks near those places. Assuming that these points are tasks of the crowdsourcing platform, the positions of the tasks are used to represent the positions of the tasks, and the earliest check-in time in the day represents the release time of the tasks. The invention extracts 20 kinds of check-in categories to simulate task categories (namely check-in categories). Checking in at a place is equivalent to accepting this task. Because Twitter does not contain explicit group information, implicit group task completion activities are extracted as follows: assume that a group of users visiting the same site or different sites of the same kind that are close to each other within one hour (e.g., a distance of less than 10km between any two sites in an experiment) is considered a member within a group.
The performance of the following methods were compared and evaluated:
1) OGTA: optimal Group Task Assignment (Optimal Group Task Assignment) based on a tree decomposition algorithm does not take into account preference values of worker groups.
2) AVG-OGTA: optimal Group Task Assignment with average worker Group preference (Optimal Group Task Assignment with average worker Group's preference), where the average preference value for Group g is set toWhereinThe class representing the group g has already executed is c number of tasks, NgRepresenting the total number of tasks performed by group g.
3) SIP-GGTA: greedy Group Task Assignment of worker groups based on Preference values for Social influence (Greedy Group Task Assignment with Social Impact-based Preference of worker groups). To improve efficiency, a basic greedy task assignment algorithm is introduced to greedily assign each task to the worker group with the greatest preference until all tasks are assigned or all worker groups are exhausted.
4) SIP-OGTA: optimal Group Task Assignment for a worker Group based on Preference values for Social influence (Optimal Group Task Assignment with Social Impact-based Preference of worker groups). I.e. the algorithm we propose.
In the above algorithm we compare three indicators:
1) cost of the CPU: the CPU time cost of the task allocation is looked up at a certain moment.
2) ASR: the allocation Success Rate (Assignment Success Rate) is the ratio of the number of successful allocations to the total number of allocations for all workers at a time. Note that this assignment of tasks is considered a successful assignment once all members of the group actually perform (check-in) the same category of tasks (places) that are close to each other within one hour (e.g., the distance between tasks in the experiment is less than 10 km).
3) The number of tasks allocated.
Wherein the experimental parameters are shown in table 1:
TABLE 1 Experimental parameters
Parameter(s) | Default value |
Effective time e-p of task | 1.5h |
Number of workers per |
2 |
Accessible radius r of the worker | 2km |
Off-on time available to the worker | 3h |
Number of tasks | S | 3000 |
The experimental results are as follows:
influence of e-p: the effect of the task validity time e-p was first investigated. As shown in fig. 2(a), this results in more CPU cost for all algorithms since we need to search for more available worker groups with the dead time side. As will be appreciated, as the task activity time increases, the accuracy of all algorithms except the OGTA increases, since as the task activity time increases, there is more opportunity for the worker group to be assigned to the task of interest (as shown in fig. 2 (b)). In terms of ASR, both SIP-OGTA and SIP-GGTA perform better than AVG-OGTA, which demonstrates the advantage of taking social impact into account the preferences of the worker group. This approach remains almost unchanged since the OGTA algorithm does not take into account the preferences of the worker group. Although SIP-GGTA is the fastest of all methods and has a similar ASR to SIP-OGTA, the SIP-GGTA algorithm allocates a lower number of tasks (as shown in fig. 2 (c)) than other methods (i.e., OGTA, AVG-OGTA, SIP-OGTA).
Influence of off-on: in this part of the experiment, the effect of worker uptime was evaluated. As can be seen from FIG. 3(a), the CPU cost of all algorithms increases with increasing off-on, as the available set of workers per task increases. As shown in FIG. 3(b), ASR of SIP-OGTA and SIP-GGTA are always superior to other algorithms, with SIP-GGTA being slightly lower than SIP-OGTA. The allocation success rate for all methods tends to be similar to that of e-p as off-on increases, for a similar reason that the worker groups have more opportunity to get the task they are interested in as off-on increases. And the task allocation number increases rapidly with increasing off-on, almost linearly (see fig. 3 (c)). The inherent reason for this is that the available worker groups per task are increasing as the worker's available time is longer.
Influence of r: i.e., the range of radii that the worker can reach. Obviously, as r increases, the CPU cost of all methods gradually increases (refer to fig. 4 (a)). For all methods that take into account the preferences of the worker group, the assignment success rate increases with increasing r, because the larger the worker reachable area, the greater the chance that the worker group is assigned the task of interest (fig. 4 (b)). As shown in fig. 4(c), it can be seen that the SIP-GGTA algorithm has a performance inferior to that of other algorithms, which also indicates the superiority of the optimal task allocation strategy.
Influence of numW: fig. 5(a) illustrates that the CPU cost gradually decreases as the number of workers per group (i.e., numW) becomes larger. The reason behind this is that as numW gets larger, there are fewer and fewer groups of workers available per task, thereby reducing the search space. In terms of allocation success rate, all algorithms trend downward as shown in fig. 5 (b). Due to the reduction in the available worker groups, tasks cannot be assigned to the appropriate groups. However, the SIP-OGTA method still has a higher advantage than other algorithms. In addition, fig. 5(c) demonstrates that the task allocation number of the SIP-GGTA algorithm has no advantage over other methods.
Influence of | S |: in the final part of the experiment, the scalability of the proposed algorithm was evaluated by varying the number of tasks | S | from 1k to 5 k. As expected, although the CPU cost increases as | S | increases, the SIP-OGTA performs well in promoting the allocation accuracy and the number of task allocations as shown in fig. 6(b) and 6 (c). As shown in fig. 6(a), SIP-GGTA is the most time-saving algorithm, while other OGTA-based algorithms run much slower, mainly because of the additional time cost required to build and search the tree to be searched. In terms of the allocation success rate, as can be seen from fig. 6(b), as the number of | S | increases, the accuracy of SIP-OGTA is still slightly higher than that of SIP-GGTA, and AVG-OGTA increases slowly. As can be seen from fig. 6(c), similar to the previous results, the performance of the OGTA-related algorithm is superior to the SIP-GGTA method for all values of | S |, in terms of the number of task assignments.
Compared with the prior art, the invention has the following beneficial effects:
(1) the invention uses a bipartite graph embedding model BGEM and an attention mechanism learning worker group to assign space tasks based on the preference value of the task types based on social influence, thereby achieving the maximum task assignment.
(2) The invention provides a group task allocation framework based on worker group preference. The frame is made up of two main parts: a preference model SIPM based on social influence and a group task assignment PGTA based on preference. Firstly, in the SIPM part, a bipartite graph embedded model and an attention mechanism are utilized to learn task types and representation of a worker group in a low-dimensional space from worker group-task interaction data, and in order to overcome the limitation of a data sparseness problem, the worker-task interaction data and social network structure information are integrated in a preference modeling process. Then in the PGTA section, an optimal task allocation algorithm based on a tree decomposition strategy is used to allocate tasks to corresponding worker groups by giving higher priority to the worker groups that are more interested in the tasks, thereby achieving maximized task allocation.
Drawings
FIG. 1 is a schematic diagram of the system of the present invention.
FIG. 2 is a graph of the impact of the group task allocation performance of the present invention on e-p.
FIG. 3 is a graph of the impact of the invention group assignment performance on off-on.
FIG. 4 is a graph of the impact of the group task assignment performance of the present invention with respect to r.
FIG. 5 is a graph of the impact of the group assignment performance of the present invention on numW.
FIG. 6 is a graph of the impact of the performance of the task allocation of the present invention on | S |.
FIG. 7 is a diagram of a model framework of the present invention.
Fig. 8 is a diagram illustrating an exemplary operation of the present invention.
Detailed Description
The present invention is further illustrated by the following figures and examples, which include, but are not limited to, the following examples.
Examples
As shown in fig. 1, a group task allocation method based on spatial crowdsourcing social influence preference includes the following steps:
(S1) analyzing the social influence preference value of each worker on different task categories, namely the direct weight of each worker and task category, by using a bipartite graph embedding model BGEM and an attention mechanism; the judging type of the social influence preference value comprises a personal interaction model and a worker group interaction model.
(S2) a task assignment algorithm incorporating tree splitting assigns a task to a plurality of workers, i.e., a worker group, according to each worker' S social influence preference value for different task categories; wherein, assigning group tasks according to the social influence preference value of each worker for different task categories may obtain available workers for each task or obtain a set of available worker groups for each task.
As shown in fig. 7, the present invention includes two main parts, one is Social-Impact-based Preference model SIPM (Social-Impact-based Preference model) of the worker Group, and the other is Preference-based Group Task Assignment PGTA (Preference-based Group Task Assignment) based on the worker Group Preference value.
In the SIPM part, the invention uses the worker-task interaction data and the worker group-task interaction data at the same time, and obtains the preference value of each worker group to different task types by using a Bipartite Graph Embedding Model (BGEM) and an attention mechanism. It is noted that the present invention indicates that a worker is interactive with a task if the worker has performed the task. More specifically, we model individual interactions (i.e., worker-task interactions) and worker-group interactions (i.e., worker-task interactions) using BGEM to learn vector representations of the classes of workers and tasks in the low-dimensional space, respectively. Since the groups of workers in spatial crowdsourcing are usually temporarily organized (called casual groups), and these groups and tasks do not have any interaction, which means that the group interaction data is sparse, the present invention cannot directly and effectively learn the vector representation of the group. To address this issue, the present invention introduces a worker's social influence representing the weight of the worker in the team when making a task selection decision. Specifically, worker-task interaction data and worker group-task interaction data are fused to construct a social network, and social network information is extracted based on the social network. In order to reduce sparseness of worker group-task interaction data, a joint optimization method is used for combining the worker group-task interaction data and the worker-task interaction data, and a representation vector of workers and task categories and weights of the workers (namely social influence of the workers) can be obtained. At the same time, a worker group vector is calculated using an attention mechanism that assigns different weights to different workers. Finally, the invention performs dot product on the group vector and the task category vector to obtain the preference value of the worker group to the task category.
In the PGTA section, given the workers and tasks to be assigned, the present invention first obtains an Available set of workers (AWGs) by considering the travel limits (i.e., Worker reachable radius, Worker Available time, and Task dead time), then assigns the tasks to the appropriate Worker Groups using an Optimal Task Assignment (OTA) algorithm based on tree decomposition to maximize the overall Task Assignment, and gives higher priority to the Worker Groups with higher preference values for the tasks.
Personal interaction model:
given the interaction between workers and tasks, i.e., worker-task interaction data, a bipartite graph is first constructed,wherein W represents a worker set, C represents a task category, and W ≦ C represents a task categorySet of nodes in (E)WCRepresenting the set of edges between the worker and the task category. When worker wiC is the executed class ofjWhen the task belongs to C, an edge e exists between the two tasksij∈EWC. Edge eijWeight h ofijIs arranged asWherein the content of the first and second substances,indicates worker wiThe executed category is cjThe number of tasks of (a) is,indicates worker wiThe total number of tasks performed.
Because of the success of BGEM in learning heterogeneous interaction entity embedded representations, the present invention uses BGEM to model individual worker-task interactions, for a given worker wi,wiAnd the kind is cjThe probability of interaction of the task(s) can be calculated as shown in equation (1):
wherein, wiIndicates worker wiRepresents a vector, which represents the worker's preference, cjIndicates the task type cjRepresents a vector.
Next, an objective function for BGEM is defined, the objective for BGEM being intended for each worker wiE.g. W, minimum empirical distributionAnd estimated neighbor probability distribution p (· | w)i) KL divergence in between. Using diRepresenting worker node wiThe output can be usedIs calculated to obtain, wherein hijRepresents an edge eijThe weight of (2). Defining an empirical distribution asThus, the objective function can be defined as equation (2):
worker group interaction model:
similarly, constructing a bipartite graph representing interactions between worker groups and task categories may be usedRepresents, wherein G represents a set of worker groups, and G ℃ @ C represents a set of worker groupsSet of nodes in (E)GCRepresenting the set of edges between the worker group and the task category. When worker group giC is the class executed by e GjWhen the task belongs to C, an edge e exists between the two tasksij∈EGC. At the same time, the edge eijWeight h ofijCan be simply set asWherein the content of the first and second substances,represents a worker group giThe executed category is cjThe number of tasks of (a) is,represents a worker group giThe total number of tasks performed. Use of g in the inventioniRepresents a worker group giIs a representative vector of cjIndicates the task type cjRepresents a vector. The object of the invention is to obtain a representation vector for each worker group to estimate the preference for all task categories.
Thus, similar to the worker-task interaction data, the objective function of the worker group-task interaction data can be calculated by equation (3):
however, in practice, there are few persistent groups (persistent groups) in spatial crowdsourcing, and there are a large number of occasional groups (occupational groups) formed temporarily to perform tasks. Thus, due to the cold start nature of the casual group (i.e., no or only few worker group-task interactions), the worker group-task interaction data is very sparse, which makes it difficult to directly learn casualThe representation vector of the group. To address data sparseness and cold start issues, vector representations of all members within a group are aggregated from worker group-task interaction data. It may be noted that in decisions such as task selection, some members of the group may outweigh others in expressing their preferences (due to reputation, authority, or other personal factors), and thus these people are more influential in group selection of tasks. Therefore, the present invention introduces a coefficient α (k, i) to represent the group giMiddle worker wkThe weight of (a) represents the group giWorker w in selecting a taskkGroup-aware personal social impact (group-aware personal social impact). Specifically, let us give chance group giThe vector g is represented by the definition set of formula (4)i:
However, occasional teams temporarily aggregate together at some point to perform a task. Due to the extreme data sparseness problem, it is difficult to learn the coefficients α (k, i) directly from the workgroup-task interaction data. The invention is therefore per worker wkIntroducing an additional positive value λkThis value represents a global personal social impact (global personal social impact) that is not dependent on any particular group. Using exp (λ)k) Indicating the relative influence of a worker on the task selected within the group. Thus, according to the attention mechanism, the coefficient α (k, i) can be calculated by the equation (5):
obviously, once the representation has been obtained for each worker wkλ of global personal social influence ofkThe social influence α (k, i) of the individual who has group awareness in a worker group can be easily obtained. But if a worker engages in a group with little activity, then overfitting problems may be encountered. Furthermore, if a worker never engages in any team activity,then the global personal social impact cannot be learned. Thus, the present invention does not learn satisfactory social influence from just the worker group-task interaction data.
In order to improve the accuracy of the global individual social influence estimation, the method constructs the social network of the worker based on the worker-task and worker group-task interaction data, extracts the social network information on the basis of the social network, and is favorable for the estimation of the global social influence of the worker. In a social network, each worker maps to a node, and there is an edge between two workers in the same group if they have worked on each other. The weight of the edge is set as the number of cooperations between workers. Each worker (node) is associated with the number of tasks he has completed. And then extracting the social network structure information by using different measures (such as degree center and middle center) and integrating the social network structure information into the learning process of the global social influence of workers, thereby effectively lightening the cold start problem of the worker group-task interaction data.
Specifically, each worker w is calculated from social network structure informationkSocial network feature vector betakThe feature selection vector h is used to assign different weights to different structural features. Normalizing all eigenvalues to [0, 1 ]]Within the range. Next, social network feature vector β is computedkAnd the feature selection vector h as a dot product, which is a Gaussian prior of the worker's global personal social influence, i.e.Where b is a deviation term. To learn more robust global personal social influence, λ is assumed, since global personal social influence may be influenced by other unknown factorskObeying a normal distribution with a mean value of betak·h+b。
In terms of the objective function, the parameter λ is influenced by socializing the personkIntroduces Gaussian prior, so a corresponding regularization term R should be added to the objective functionVI.e. byWherein the hyperparameterThe weight of the regularization term (i.e., the variance) may be controlled. Thus, the new objective function is shown in equation (6):
OVGC=OGC+RV (6)
the worker-task interaction data and the worker group-task interaction data are combined together in an optimization process to account for the cold start problem of the worker group-task interaction data. More specifically, a joint optimization method is designed, and the method can simultaneously learn the representation vectors of workers and task categories from the worker-task interaction data and the worker group-task interaction data. In addition, the global social influence of the worker can also be learned during the optimization process. Thus, mixing OVGCAnd OWCCombining to form a joint objective function, a simple definition can be given as shown in equation (7):
OGWC=OVGC+OWC (7)
minimizing the objective function O in equation (7) using a standard Stochastic Gradient Descent (SGD) strategyGWCIn this process, a worker representative vector w, a task class representative vector c, and a model parameter (i.e., λ) can be learnedkH). From equation (5), a coefficient α (k, i) representing the social influence of the individual of group awareness can be calculated. Then, the representative vector g for each group can be calculated accordingly according to equation (4). Finally, the representation vector of each group and the representation vector of each task category are subjected to dot product to obtain the preference of each group to each task category.
Generation of a set of available worker groups:
available workers for each task are obtained. Each task can only be completed by a small percentage of workers at a time due to constraints on worker reach, worker uptime, and task downtime. Thus, a set of workers capable of completing the task is first obtained without violating constraints. Reachable subset of workers for task s, available RWsIs shown, andthe following conditions should be satisfied:
(1) the worker arrives at the location of task s, i.e. t, before task s failsnow+t(w.l,s.l)≤s.e;
(2) The task s is within the reach of workers, namely d (w.l, s.l) is less than or equal to w.r;
(3) the worker arrives at the place of task s, i.e. t, within his own effective timenow+t(w.l,s.l)≤w.off;
Wherein, tnowRepresenting the current time, t (w.l, s.l) represents the travel time from worker location w.l to location s.l, and d (w.l, s.l) is the travel distance (e.g., euclidean distance) between a given location w.l and location s.l. The three conditions described above ensure that worker w can travel directly from its location w.l to the location of task s (which is within the reach of the worker) before task s fails and within the time available to worker w.
And RWsAre readable workers sets, meaning that each task can only be done by some workers due to space-time constraints (e.g., Worker reach, effective time, and task dead time), and RWsIs a monolithic body. The aforementioned W does not specify a task, i.e., crowd-sourcing all workers in space. For example, a total of 5 apples (equivalent to W) and child A can only eat one of the apples (equivalent to RW of child A)s)。
A set of available worker groups for each task is obtained. Given the reachable workers per task s, the set of available worker groups, denoted awg(s), is next found, given the constraints of the time available for workers in the group and the number of workers allowed to be allocated to perform task s, and the available worker groups in awg(s) should satisfy the following conditions:
(1)|AWG(s)|=s.numW;
wherein, | awg(s) | represents the number of workers in awg(s). The two conditions described above ensure that workers in a group can reach the location of task s without affecting the time available for each other.
The invention introduces a strategy based on tree decomposition to obtain the optimal task allocation with the maximum preference value. More specifically, a task dependency graph is first constructed based on dependencies between tasks (i.e., if two tasks share available workers they are interdependent on each other, otherwise they are independent of each other). Subsequently, all tasks are divided into separate clusters using a tree decomposition strategy (i.e., tasks in different clusters do not share the same available workers) and organized into a balanced tree structure such that tasks in sibling nodes of the tree do not share the same available workers. Through the tree structure, the invention can independently solve the problem of the optimal allocation sub-tree on each brother node, and then carries out depth-first search on the tree to find the optimal allocation.
Specific examples are as follows:
FIG. 8 shows an example of a group task assignment problem, where each task needs to be assigned to two workers, for a total of five workers (w)1,...,w5) And two tasks(s)1,s2). Each worker is associated with his current location and its range of reachable distances. Each task is marked with the location where it is executed. In addition, FIG. 8 also shows the preferences of each task of the different reachable task teams. The problem assigns the task to the appropriate worker group to maximize the overall task assignment. In spatial crowd-sourcing, assigning nearby tasks to workers without violating the spatio-temporal constraints (i.e., the assigned tasks should be within reach of the respective workers, and the workers should arrive at the assigned task's location before the task deadline) is an intuitive activity. Thus, we can obtain a task allocation<s1,{w1,w2}>,<s2,{w4,w5}>Its total set preference is 0.33. However, when we will task s2Assigned to a group of workers { w4,w5When the worker is paired with s2There is little interest (i.e., in s)2Is 0.04), the worker group may not perform task s2This will result in s2Cannot be completed. If we give higher priority to a group of workers who are more interested in a task when assigning the task, we may obtain a task assignment result<s1,{w2,w3}>,<s2,{w1,w4}>Its total set preference is 0.78.
The above embodiments are only preferred embodiments of the present invention, and are not intended to limit the scope of the present invention, but all changes that can be made by applying the principles of the present invention and performing non-inventive work on the basis of the principles shall fall within the scope of the present invention.
Claims (1)
1. A group task allocation method based on spatial crowdsourcing social influence preference is characterized by comprising a preference model SIPM based on social influence of a worker group and a group task allocation PGTA based on preference of the worker group preference value; in the SIPM part, worker-task interaction data and worker group-task interaction data are used at the same time, and a bipartite graph embedding model BGEM and an attention mechanism are adopted to obtain preference values of each worker group for different task types; modeling the worker-task interaction data and the worker group-task interaction data by adopting BGEM, thereby respectively learning vector representations of workers and task types in a low-dimensional space; introducing social influence of workers, the social influence representing the weight of workers in a team when making a task selection decision, specifically: fusing the worker-task interaction data and the worker group-task interaction data to construct a social network, and extracting social network information based on the social network; combining the worker group-task interaction data and the worker-task interaction data to obtain an expression vector of workers and task categories and weights of the workers, namely social influence of the workers; meanwhile, calculating a worker group vector by using an attention mechanism, wherein the attention mechanism allocates different weights to different workers, and finally, performing dot product on a representation vector of the worker group and a task category vector to obtain a preference value of the worker group to the task category;
in the PGTA, given the workers and tasks to be assigned, the set of available workers is first obtained by considering the trip limits, then the tasks are assigned to the appropriate worker groups using an optimal task assignment algorithm based on tree decomposition to maximize the overall task assignment, and the worker groups with higher preference values for the tasks are given higher priority;
personal interaction model:
given the interaction between workers and tasks, i.e., worker-task interaction data, a bipartite graph is first constructed,wherein the content of the first and second substances,a set of workers is represented as a set of workers,the type of the task is represented and,to representThe set of nodes in (1) is,representing a set of edges between workers and task categories; when the worker worksIs executed in the category ofWhen the task (2) is executed, an edge exists between the twoEdgeWeight of (2)Is arranged asWherein the content of the first and second substances,indicating the workerThe executed kind isThe number of tasks of (a) is,indicating the workerThe total number of tasks performed;
modeling personal worker-task interactions using BGEM for a given workerAnd is of the kindThe probability of interaction of the task(s) can be calculated as shown in equation (1):
wherein the content of the first and second substances,indicating the workerRepresents a vector, which represents the worker's preference,indicating the kind of taskA representative vector of (a);
defining an objective function for BGEM, the objective of BGEM being intended for each workerMinimizing empirical distributionAnd estimated neighbor probability distributionKL divergence in between; use ofRepresenting worker nodesOut of service ofAnd calculating to obtain the result that, wherein,representing edgesThe weight of (2); distributing the experienceIs defined asThus, the objective function can be defined as equation (2):
worker group interaction model:
constructing a bipartite graph representing interactions between worker groups and task classesIt is shown that, among others,a set of worker groups is represented as,to representAnd a set of nodes of the sum,representing a set of edges between a worker group and a task category; working groupIs executed in the category ofWhen the task (2) is executed, an edge exists between the twoAt the same time, the edgeWeight of (2)Can be simply set asWherein the content of the first and second substances,indicating groups of workersThe executed kind isThe number of tasks of (a) is,indicating groups of workersThe total number of tasks performed; use ofIndicating groups of workersIs used to represent a vector of (a) a,indicating the kind of taskA representative vector of (a); the goal is to obtain a representation vector for each worker group to estimate the preference values for all task categories;
the objective function of the workgroup-task interaction data can be calculated by equation (3):
introducing a coefficientIndicating groups of workersMiddle workerThe weight representing the groupWorkers in selecting tasksGroup-aware personal social influence of (a); the method specifically comprises the following steps: given worker groupThe expression vector for the group of workers is defined as equation (4)
For each workerIntroducing an additional positive valueThe value indicates independenceGlobal personal social influence on any particular group; use ofRepresenting the relative influence of a worker on a selected task within a group; thus, in the attention mechanism, the coefficientCan be calculated from equation (5):
computing each worker based on social network structure informationSocial network feature vector ofSelecting vectors using featuresAssigning different weights to different structural features; normalizing all eigenvalues to [0, 1 ]]Within the range; next, social network feature vectors are appliedAnd feature selection vectorsDot product as a Gaussian prior of the worker's global personal social influence, i.e.WhereinIs a deviation term; to learn more robust global personal social influence, the global personal social influence may be influenced by other unknown factorsObeying a normal distribution with a mean value of
In terms of objective function, parameters are influenced by socializing the individualGaussian priors are introduced, so that corresponding regularization terms are added to an objective functionNamely, it isWherein the hyperparameterThe weight of the regularization term may be controlled; thus, the new objective function is shown in equation (6):
will be provided withAndcombining to form a joint objective function, defined as shown in equation (7):
minimizing the objective function in equation (7) using a standard stochastic gradient descent strategyIn this process, worker representation vectors are learnedTask class representation vectorAnd model parameters,Calculating a coefficient representing the social influence of the individual on group consciousness according to equation (5)Then, a representative vector of each group is calculated according to equation (4)Finally, performing dot product on the representation vector of each group and the representation vector of each task type to obtain a preference value of each group to each task type;
generation of a set of available worker groups:
obtaining available workers for each task; due to the constraints of the reach of workers, the available time of workers and the failure time of tasks, each task can be completed by only a small part of workers at a certain moment; thus, a set of workers capable of completing the task is first obtained without violating constraints; taskReachable subset of workers, availableIs shown, andthe following conditions should be satisfied:
Wherein the content of the first and second substances,which is indicative of the current time of day,indicating slave worker positionAnd positionA travel time therebetween, andis a given positionAnd positionThe distance traveled in between; the three conditions ensure the workersCan be at taskBefore failure and before workersDirectly from its location within the available timeGo to taskThe position of (a);
and isAre Reachable workers sets, i.e., each task can only be completed by some workers due to space-time constraints,is a monolithic body; as mentioned aboveNot specifying a task, i.e., crowdsourcing all workers in the space;
obtaining a set of available worker groups for each task; given the reachable workers per task s, the time available for workers in the group and the performance tasks allowed to be assignedNext find a set of available worker groups, subject to the constraint of the number of workersBy usingIt is shown that,the available worker group in (1) should satisfy the following condition:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010170604.6A CN111311115B (en) | 2020-03-12 | 2020-03-12 | Group task allocation method based on space crowdsourcing social influence preference |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010170604.6A CN111311115B (en) | 2020-03-12 | 2020-03-12 | Group task allocation method based on space crowdsourcing social influence preference |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111311115A CN111311115A (en) | 2020-06-19 |
CN111311115B true CN111311115B (en) | 2021-04-23 |
Family
ID=71162247
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010170604.6A Expired - Fee Related CN111311115B (en) | 2020-03-12 | 2020-03-12 | Group task allocation method based on space crowdsourcing social influence preference |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111311115B (en) |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113839774B (en) * | 2021-10-12 | 2022-07-01 | 西南大学 | Group time base one-time password method and equipment |
CN114282645B (en) * | 2021-11-24 | 2023-04-21 | 杭州电子科技大学 | Space-time crowdsourcing task allocation method based on DQN |
CN114372680A (en) * | 2021-12-23 | 2022-04-19 | 电子科技大学(深圳)高等研究院 | Spatial crowdsourcing task allocation method based on worker loss prediction |
CN114020650B (en) * | 2022-01-10 | 2022-04-12 | 广东拓思软件科学园有限公司 | Crowd test task allocation method and device, electronic equipment and storage medium |
CN114519510A (en) * | 2022-01-27 | 2022-05-20 | 电子科技大学(深圳)高等研究院 | Spatial crowdsourcing task allocation method based on mutual information and preference perception |
CN115694877B (en) * | 2022-08-30 | 2023-08-15 | 电子科技大学长三角研究院(衢州) | Space crowdsourcing task allocation method based on federal preference learning |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106339852A (en) * | 2016-08-10 | 2017-01-18 | 南京邮电大学 | Crowdsourcing task allocation method considering preferences |
CN107886221B (en) * | 2017-10-24 | 2021-08-31 | 佛山科学技术学院 | Task allocation method for worker-oriented natural groups in crowdsourcing system |
CN107895220B (en) * | 2017-10-24 | 2021-09-28 | 佛山科学技术学院 | Autonomous team forming method for batch tasks in crowdsourcing system |
CN110400128B (en) * | 2019-07-29 | 2020-06-23 | 电子科技大学 | Spatial crowdsourcing task allocation method based on worker preference perception |
CN110390490A (en) * | 2019-07-29 | 2019-10-29 | 电子科技大学 | Space crowdsourcing method for allocating tasks based on profit |
-
2020
- 2020-03-12 CN CN202010170604.6A patent/CN111311115B/en not_active Expired - Fee Related
Also Published As
Publication number | Publication date |
---|---|
CN111311115A (en) | 2020-06-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111311115B (en) | Group task allocation method based on space crowdsourcing social influence preference | |
US11393341B2 (en) | Joint order dispatching and fleet management for online ride-sharing platforms | |
Liu et al. | A reinforcement learning-based resource allocation scheme for cloud robotics | |
Agussurja et al. | A state aggregation approach for stochastic multiperiod last-mile ride-sharing problems | |
Ray et al. | The multi-depot split-delivery vehicle routing problem: Model and solution algorithm | |
Lin et al. | Friend-as-learner: Socially-driven trustworthy and efficient wireless federated edge learning | |
US11567512B1 (en) | Unmanned aerial vehicle (UAV) task cooperation method based on overlapping coalition formation (OCF) game | |
Zhang et al. | Balancing energy consumption and reputation gain of UAV scheduling in edge computing | |
Dotoli et al. | A multi-agent advanced traveler information system for optimal trip planning in a co-modal framework | |
Hamrouni et al. | Low-complexity recruitment for collaborative mobile crowdsourcing using graph neural networks | |
Hossain et al. | Fuzzy decision-based efficient task offloading management scheme in multi-tier MEC-enabled networks | |
Li et al. | Research on QoS service composition based on coevolutionary genetic algorithm | |
Khanfor et al. | A trustworthy recruitment process for spatial mobile crowdsourcing in large-scale social IoT | |
Zhang et al. | Reinforcement learning assisted bandwidth aware virtual network resource allocation | |
Xie et al. | Loosely coupled cloud robotic framework for QoS-driven resource allocation-based Web service composition | |
Wu et al. | Delay-aware edge-terminal collaboration in green internet of vehicles: A multiagent soft actor-critic approach | |
Pilloni et al. | Task allocation among connected devices: Requirements, approaches, and challenges | |
Souri et al. | A trust-aware and authentication-based collaborative method for resource management of cloud-edge computing in social internet of things | |
Zhang et al. | Dynamic user recruitment in edge-aided mobile crowdsensing | |
Gu et al. | TERL: Two-stage ensemble reinforcement learning paradigm for large-scale decentralized decision making in transportation simulation | |
Wu et al. | Knowledge and Behavior‐Driven Fruit Fly Optimization Algorithm for Field Service Scheduling Problem with Customer Satisfaction | |
Zhang et al. | Online resource allocation for reusable resources | |
Zhang et al. | MPTO-MT: A multi-period vehicular task offloading method in 5G HetNets | |
Shen et al. | A dynamic task assignment model for aviation emergency rescue based on multi-agent reinforcement learning | |
Ma et al. | Mixed priority queue scheduling based on spectral clustering in spatial crowdsourcing |
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 | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20210423 |
|
CF01 | Termination of patent right due to non-payment of annual fee |