Task association induction team establishment method combining global cohesion and local decoupling

Abstract

The invention belongs to the technical field of team establishment, and discloses a task association induction team establishment method combining global cohesion and local decoupling. Team division based on a heterogeneous collaborative network is provided, cooperation between experts is described more finely, and implicit semantics in a text and dominant structure relations between the experts can be captured at the same time so as to meet specific task requirements. The cross-team cooperation potential is captured from the global perspective by using the hypergraph and the weighted folded graph, and the global representations are mapped to team internal expert representations through local decoupling. According to the method, the internal structure of the team is fully utilized, and the problem of range limitation is effectively avoided. According to the method, a team matching strategy driven by task correlation is introduced. By analyzing the correlation between the new task and the historical task, the strategy can select a proper candidate team, thereby effectively avoiding the false failure phenomenon in the matching process. The strategy aims at improving the team matching accuracy and efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of team building, and in particular to a task association induction team building method combining global cohesion and local decoupling. Background Art

[0002] The requirement of team formation can be regarded as a task, which specifies the professional skills and academic background of the required experts. The main challenge is to select experts who meet the task requirements and maintain close cooperation. In order to study the team formation problem, for a given task requirement, an expert library and its historical cooperation relationship, the goal of this invention is to automatically identify a group of experts who not only meet the task requirements but also show strong cooperation potential.

[0003] Existing team building methods are mainly divided into three types: topic-based, text-based, and graph-based methods. Topic-based methods achieve accurate matching of task requirements and expert teams by analyzing the topic characteristics of tasks and the research directions of experts. This type of method usually uses statistical language models and semantic embedding technology to combine task topics with expert research interests for modeling and optimization. Specifically, topic-based methods extract topic information from task descriptions and expert research content through topic modeling, reveal the correlation between tasks and experts on different topics, embed tasks and expert research content into high-dimensional semantic space, find potentially relevant experts through semantic similarity analysis, and use probability models to calculate the matching degree between experts and task topics to form teams with complementary skills and strong cooperation potential. Text-based methods aim to mine the potential relationship between experts and tasks by analyzing text data in task descriptions and expert background information. They usually rely on natural language processing technology to extract features from task requirements and expert research results or social network data, use these features to rank experts, and identify expert teams that are highly matched with task requirements. In addition, text-based methods can further optimize the selection of team members by analyzing social information such as cooperative relationships and emotional tendencies between experts. Graph-based methods model the expert network as nodes and edges, and optimize team formation by finding subgraphs that meet the task skill requirements. This method usually combines subgraph search with optimization algorithms to find a subset of experts that meet the task skill requirements and optimize team cohesion, and at the same time calculates task relevance based on weighted nodes and edges to screen team members that meet the conditions.

[0004] The paper "A variational neural architecture for skill-based teamformation" ACM Transactions on Information Systems 42.1(2023):1-28" improves on the traditional topic-based method and proposes a skill-matching team formation method based on variational neural networks, aiming to efficiently form teams that meet specific task requirements by learning the implicit associations between expert skills and task requirements. Expert skill data is mapped to the latent space to capture the complex nonlinear relationships between skills and optimize the team's skill complementarity and collaboration potential. After inputting expert profiles and collaboration history data, the model generates latent representations through training and evaluates different team configurations to match the best combination under new task requirements. The technical method described is an improvement on the topic-based team formation method. In real scenarios, the types of expert collaboration are diverse. This method is highly dependent on the quality and completeness of expert profiles and collaboration history data. The model may not accurately capture the real skill relationships and complex collaboration relationships between experts. At the same time, although this method can learn the implicit relationship between skill nodes, it may not be able to fully capture the deep semantic associations between different fields. The generated latent representations are often difficult to interpret directly and tend to limit member selection to the existing team, which may have certain limitations when collaborating across teams.

[0005] The paper "Adeep learning-based expert finding method to retrieve agilesoftware teams from CQAs" Information Processing&Management 60.2(2023):103144" proposes a text-based deep learning expert discovery method, which aims to optimize team building by analyzing the knowledge and skills of candidates. First, potential candidates are screened out using a recommendation decision model, and then the candidates' expertise in the main skill field and general knowledge in other related fields are estimated in parallel through a deep learning model, while considering the time factor and dynamically estimating their knowledge evolution. Finally, an integer linear programming model is used to select the most suitable team members from the screened candidates. The technical method can be regarded as a text-based team building method. The core idea is to estimate the professional level and general knowledge in a specific skill field by analyzing the text content published by experts, thereby providing expert selection for team building. However, this method mainly relies on text content to evaluate expert skills, which may not fully capture implicit semantics. For example, some candidates may have cross-skills in multiple fields, but if these skills are not clearly reflected in the text, the model may not be able to accurately identify these "implicit" skills. Therefore, the text content may not fully reflect the true professional level of the expert. At the same time, this simple reliance on text content cannot capture the explicit structural relationship between experts, and may ignore the experts' abilities in teamwork, communication, etc. Summary of the invention

[0006] Previous team formation methods have the following problems: they ignore the diversity of expert collaboration types, tend to limit member selection to existing teams, and fail to fully consider the correlation between tasks. To address these problems, we propose a task correlation-induced team formation method that combines global cohesion and local decoupling.

[0007] The technical solution of the present invention is as follows: a task association induction team formation method combining global cohesion and local decoupling, based on the heterogeneous collaborative network HCN for team division; based on expert data, historical task data and new task t new Construct a heterogeneous collaborative network graph G and select a special path Divide the expert nodes into teams and form a hypergraph H;

[0008] A weighted folding graph is constructed based on the hypergraph H, and team representation learning based on global cohesion is performed to form a team embedding vector, which is placed in a D-dimensional vector space. At the same time, team representation learning based on local decoupling obtains the relationship between tasks and experts in each divided team, and obtains task embedding vectors and expert embedding vectors.

[0009] The team matching strategy based on task relevance induction is used to new , calculate the task relevance, find the k most relevant historical tasks; according to the k most relevant historical tasks, t new Map to vector space and find the best matching team set.

[0010] The specific steps of performing team division based on the heterogeneous collaborative network HCN are as follows:

[0011] (1) Input expert data and historical task data, construct a heterogeneous collaborative network graph G and select the dominant node v s and special paths

[0012] (2) Initialization: Queue Q = {v s}, used to save nodes that can be further searched; set T = {v s}, records all nodes in a team; array R[v], records all explored neighbor nodes of node v; set Visited = {v s}, record all visited nodes, set Record nodes that do not satisfy constraints and should be deleted;

[0013] (3) Initial division of the team: Take out the node v in Q and find its node in the special path Find all the neighbor nodes μ that have not been visited, and add μ to R[v] and Visited, and add R[v] to T; the constraint is that the nodes in R[v] have n neighbor nodes. If the constraint is met, the nodes that meet the constraint are added to Q and further searched, otherwise, the nodes that do not meet the constraint are added to DL; repeat the above steps until Q is empty;

[0014] (4) Constraint-based pruning and optimization: Take a node from DL, delete it from T, iteratively delete all nodes that do not meet the constraints from T, and after deleting the node that does not meet the constraints, check whether each of its neighboring nodes still meets the constraints; if its neighboring nodes also do not meet the constraints, then delete the neighboring nodes from T; repeat the above steps until DL is empty; finally, remove v s All neighbor nodes that do not meet the constraints are added to T to relax the strict constraints;

[0015] (5) Constructing the hypergraph H: Based on the divided teams, each team is regarded as a hyperedge to construct the final hypergraph H.

[0016] In the heterogeneous collaboration network diagram, tasks and experts are represented as nodes, and the relationship between tasks and experts is represented as edges; the special path is expert-task-expert.

[0017] Each hyperedge in the hypergraph H is a team, and a weighted folding graph is constructed based on the hypergraph H; in the weighted folding graph, each node corresponds to a team, and the weight of the edge reflects the degree of association between the teams, and the weight is the ratio of experts shared by two hyperedges.

[0018] The team representation learning based on global cohesion is as follows: each team is regarded as a region in a D-dimensional vector space, represented by a position vector Pos and a range vector Ran, defined as Tm i =(Pos (i) ,Ran (i) );Team Tm i Satisfy: Tm i = {v∈R d |dis(Pos (i) ,v)≤Ran (i)}; where dis(Pos (i) ,v) is calculated as:

[0019]

[0020] The range vector is a constant that is dynamically adjusted according to the team size. Based on the range vector and the edge weights of the folding graph, the overlap constant of each team is obtained, which is defined as: d ij =(1-w ij )(Ran (i) +Ran (j) );

[0021] The position vector is obtained by combining Skip-gram negative sampling and overlap constant training on the folded graph; finally, the team embedding vector is obtained, and the team embedding loss function is as follows:

[0022]

[0023] The team representation learning based on local decoupling is specifically as follows: for each expert node and task node contained in the team embedding vector, their D-dimensional potential representations are learned to capture the semantic and structural relationships between each other, and the expert embedding vector and the task embedding vector are obtained. The loss function is defined as follows:

[0024]

[0025] Among them, T′v ={experts,questions},N t (v) represents the neighboring nodes of a node.

[0026] The P(v t |v;θ) is the softmax function: is the node type mapping for each node u, x u is the embedding representation of node u.

[0027] The specific steps of the team matching strategy induced by task relevance are as follows:

[0028] (1) Task relevance calculation: Input new task t new and historical tasks h , extract keywords respectively and Use the pre-trained word embedding model to map each keyword to an L-dimensional vector representation; construct a translation matrix Ts, where each element in Ts represents t new The words and t in h The embedding similarity between words in , Ts is defined as follows: The word-to-word interactions captured in the translation matrix Ts are transformed into task-solution ranking features φ(Ts):

[0029]

[0030] K(Ts i )={K1(Ts i ),...,K k (Ts i )}

[0031]

[0032] These ranking features are combined to generate the final ranking score:

[0033] F(t new ,t h )=tanh(w T φ(Ts)+b) Select the top k historical task sets T that are most relevant to the new task h ;

[0034] (2) Team matching: assign new tasks to new Mapped to T h In the same vector space;

[0035]

[0036] t newThe embedding vector of is matched with each team embedding vector; by calculating each potential team position vector and t new The distance between dis(Pos (i) ,G embeded (t new )), select the best matching t new The top m teams.

[0037] Beneficial effects of the present invention: a heterogeneous collaborative network HCN is designed to describe the cooperation between experts in a more detailed manner, and can simultaneously capture the implicit semantics in the text and the explicit structural relationship between experts. In addition, a team division method based on the heterogeneous collaborative network HCN is proposed to meet specific task requirements.

[0038] The present invention uses hypergraphs and weighted folding graphs to capture the potential of cross-team collaboration from a global perspective, and maps these global representations into team internal expert representations through local decoupling. This approach not only fully utilizes the team's internal structure, but also effectively avoids the scope limitation problem.

[0039] The present invention introduces a task relevance-driven team matching strategy. By analyzing the relevance between new tasks and historical tasks, the strategy can select suitable candidate teams, thereby effectively avoiding pseudo-failures in the matching process. The strategy aims to improve the accuracy and efficiency of team matching. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is the overall framework diagram of the method of the present invention;

[0041] Figure 2 It is the main flow chart of the method of the present invention;

[0042] Figure 3 A flow chart of the HCN-based team division method;

[0043] Figure 4 Representing learning flow charts for teams based on global cohesion and local decoupling;

[0044] Figure 5 Flowchart of team matching strategy induced by task relevance;

[0045] Figure 6 This is a schematic diagram of the task relevance calculation model framework;

[0046] Figure 7 The overall performance comparison of different methods under different datasets; (a) is the Physics dataset; (b) is the DBA dataset; (c) is the History dataset; and (d) is the Android dataset.

[0047] Figure 8Ablation experiments of different methods under different datasets; (a) is the Andriod-CL dataset; (b) is the Andriod-GM dataset; (c) is the Andriod-SC dataset; (d) is the Andriod-EL dataset; (e) is the History-CL dataset; (f) is the History-EL dataset; (g) is the History-SC dataset; (h) is the History-GM dataset; (i) is the DBA-CL dataset; (j) is the DBA-EL dataset; (k) is the DBA-SC dataset; (l) is the DBA-GM dataset; (m) is the Physics-EL dataset; (n) is the Physics-SC dataset; (o) is the Physics-CL dataset; (p) is the Physics-GM dataset;

[0048] Fig. 9 Comparative experiments of GM on different datasets. DETAILED DESCRIPTION

[0049] This method takes into account the heterogeneity of expert collaboration types, the potential for cross-team collaboration, and task relevance, thereby ensuring the accuracy, diversity, and efficiency of team formation. Specifically, this method captures the implicit semantics in the text and the explicit structural relationships between experts; it also considers the internal composition of the team and the potential for cross-team collaboration to alleviate the problem of limited member selection range; and it introduces task relevance factors to avoid pseudo-failures in the matching process.

[0050] The overall framework of this method is as follows Figure 1 As shown:

[0051] 1. Team division method based on heterogeneous collaborative network (HCN). Team formation needs to capture the complex relationships between experts more comprehensively. Relying only on the cooperative relationship between experts without distinguishing the type of cooperation may select experts who do not meet the requirements into the team. Based on the heterogeneous collaborative network, the team is initially divided by identifying special paths, and the initial division is pruned and refined according to the constraints, finally forming a refined division result, which will serve as the basis for constructing the hypergraph.

[0052] 2. Team representation learning method based on global cohesion and local decoupling. When forming a new team, it is necessary to avoid limiting the selection of members to existing teams. A weighted folding graph is constructed based on a hypergraph to capture the potential for cross-team collaboration from a global perspective. Subsequently, these overall representations are transformed into internal expert representations through local decoupling.

[0053] 3. Team matching strategy induced by task relevance. Task requirements and expert skills do not match in description, but there may be potential semantic associations between them. If the team formation process only relies on the matching of expert skills and task requirements in description, the deep connection hidden by the difference in expression may be ignored. By calculating the relevance between the new task and the historical tasks, the appropriate candidate team is selected and the participating expert members are further determined.

[0054] Figure 2 The main flow chart of the technical solution of the present invention is shown as follows: Figure 2 As shown, the task-relatedness-driven team building method proposed by the present invention that combines global cohesion and local decoupling includes the following steps:

[0055] Input expert data, historical task data and new task t new . Team division based on heterogeneous collaboration network (HCN). Constructing heterogeneous collaboration network diagram Select a special path The expert nodes are divided into teams and form a hypergraph H.

[0056] A weighted folding graph is constructed based on the hypergraph H, and team representation learning based on global cohesion is performed to form a team embedding vector, which is placed in a D-dimensional vector space. At the same time, team representation learning based on local decoupling obtains the relationship between tasks and experts in each divided team, and obtains task embedding vectors and expert embedding vectors.

[0057] The team matching strategy based on task relevance induction is used to new , calculate the task relevance, find the k most relevant historical tasks; according to the k most relevant historical tasks, t new Map to vector space and find the best matching team set.

[0058] Figure 3 The flowchart of HCN-based team division is shown. The specific steps are as follows:

[0059] Input expert data and historical task data, construct a heterogeneous collaborative network graph G, and select the dominant node v s and special paths

[0060] Initialization: queue Q = {v s}, used to save nodes that can be further searched; set T = {v s}, records all nodes in a team; array R[v], records all explored neighbor nodes of node v; set Visited = {v s}, record all visited nodes; set Record nodes that do not satisfy the constraints and should be deleted.

[0061] Initial division of the team: Take out the node v in Q and find its path Find all the neighbor nodes μ that have not been visited, and add μ to R[v] and Visited, and add R[v] to T; determine whether the nodes in R[v] have n neighbor nodes. If the constraint is met, add the nodes that meet the constraint to Q and search further. Otherwise, add the nodes that do not meet the constraint to DL. Repeat the above steps until Q is empty.

[0062] Constraint-based pruning and optimization: Iteratively delete all nodes that do not meet the constraints from T. Considering that deleting a node may affect its neighboring nodes, it is necessary to check whether each of its neighboring nodes still meets the constraints after deleting the node. If its neighboring nodes also do not meet the constraints, then the neighboring nodes also need to be deleted from T. Repeat the above steps until DL is empty. Finally, v s All neighbor nodes that do not satisfy the constraints are added to T to relax the strict constraints.

[0063] Construct the hypergraph H: According to the divided teams, each team is regarded as a hyperedge to construct the final hypergraph H.

[0064] Figure 4 The flowchart of team representation learning based on global cohesion and local decoupling is shown. The specific steps are as follows:

[0065] Construct a weighted folding graph: Each hyperedge in the hypergraph H represents a team. A weighted folding graph is constructed based on H. In this folding graph, each node corresponds to a team, and the weight of the edge reflects the degree of association between the teams. This weight is the ratio of experts shared by two hyperedges.

[0066] Team representation learning based on global cohesion: Each team is regarded as a region in a D-dimensional vector space, represented by a position vector Pos and a range vector Ran, defined as Tm i =(Pos (i) ,Ran (i) ). Team Tm i Satisfaction: Tm i = {v∈R d |dis(Pos (i) ,v)≤Ran (i)}. Among them, dis(Pos (i) ,v) is calculated as:

[0067]

[0068] The range vector is a constant that is dynamically adjusted according to the team size. Based on the range vector and the edge weights of the folding graph, the overlap constant of each team can be obtained, which is defined as: ij =(1-w ij )(Ran (i) +Ran (j) ). The position vector is obtained by combining Skip-gram negative sampling and overlap constant training on the folded graph. Finally, the team embedding vector can be obtained, and the team embedding loss function is as follows:

[0069]

[0070] Team representation learning based on local decoupling: For each team’s expert nodes and task nodes, learn their D-dimensional potential representations, capture their semantic and structural relationships, and obtain expert embedding vectors and task embedding vectors. The loss function is defined as follows:

[0071]

[0072] Among them, T′ v ={experts,questions},N t (v) represents the neighboring nodes of a node. t |v;θ) is considered as a softmax function, and we have the following definition:

[0073]

[0074] is the node type mapping for each node v, x v is the embedding representation of node v.

[0075] Figure 5 The flow chart of the team matching strategy induced by task relevance is shown. The specific steps are as follows:

[0076] Task dependency calculation: Input new task t new and historical tasks h , extract keywords respectively and Use a pre-trained word embedding model to map each keyword to an L-dimensional vector representation; the word embedding model can be any existing word embedding model, such as word2vec.

[0077] Construct the translation matrix Ts, each element in Ts represents t new The words and t in h The embedding similarity between words in , Ts is defined as follows: The word-to-word interactions captured in the translation matrix Ts are transformed into task-solution ranking features φ(Ts):

[0078]

[0079] K(Ts i )={K1(Ts i ),...,K k (Ts i )}

[0080]

[0081] And these ranking features are combined to generate the final ranking score:

[0082] F(t new ,t h )=tanh(w T φ(Ts)+b)

[0083] The framework for task relevance calculation is as follows Figure 6 As shown. With this sorting mechanism, the new task t can be calculated new And each historical task t h The correlation between them can identify the top k historical task sets T that are most relevant to the new task. h .

[0084] Team Matching: After identifying T h Then, the new task t new Mapped to T h In the same vector space:

[0085]

[0086] t new The embedding of is matched with each team embedding. By calculating each latent team and t new The distance between them can be used to select the best match t new The top m teams.

[0087] The present invention is implemented in Python and runs on an Intel(R) Xeon(R) Silver 4310CPU 2.10GHz, Nvidia A5000 GPU and 251GB RAM. We set the embedding dimension to 128 and the batch size to 256. The datasets used include four real datasets from Stack Exchange: Android, History, DBA and Physics.

[0088] The technical solution (E2T) proposed in this invention is compared with the following team building methods: CCR, SA-CA-CC, NCCO, TOSA, and TAPG. The main goal of CCR is to form a team that meets the needs and maximizes the potential for cooperation among members. NCCO and SA-CA-CC give priority to enhancing the professional level of team members and reducing communication costs. TOSA selects experts with high professional levels based on subject sensitivity, while TAPG ensures efficient identification of expert teams in large-scale text data. The experimental results show that E2T not only achieves significant advantages in skill coverage and accuracy, but also shows unique advantages in optimizing team communication costs. The key to its success lies in fully tapping the interaction and collaboration potential between experts by combining the historical collaborative relationship and cross-team cooperation potential of experts to ensure more efficient and accurate team building.

[0089] Figure 7 The experimental results are shown. The following is a detailed analysis of the experiment:

[0090] (1) CCR, NCCO, and SCA (i.e., SA-CA-CC) have obvious deficiencies in terms of skill coverage (SC). This is because most of these methods rely on graph search techniques to gradually add team members until all required skill requirements are met. However, this strategy is prone to falling into local optimal solutions, resulting in the algorithm not being able to fully cover all necessary skills when expanding the team. In particular, when certain skills are selected early on, the selection space for remaining members may be limited, thus affecting the final skill coverage effect.

[0091] (2) In terms of team communication cost (CL), TOSA performs significantly worse than other methods. The main goal of TOSA is to select the most suitable experts based on subject sensitivity, rather than optimizing the combination of team members to reduce communication costs. Therefore, TOSA has certain advantages in assessing whether experts have the required skills and can achieve better skill coverage, but it ignores the collaboration history and communication costs between team members, resulting in low communication efficiency when multiple experts collaborate.

[0092] (3) In terms of accuracy (PN), E2T and TAPG performed similarly, as both combined text information and considered the cooperative relationship between experts during the team formation process. However, the core difference between the two is that the method E2T of the present invention strengthens the cooperative potential between team members by directly integrating historical expert cooperation records and cross-team relationships, while TAPG indirectly infers the cooperative relationship between experts through the work content of experts and lacks direct support from the history of cooperation. Therefore, the technical solution of the present invention makes the matching of team members more accurate through this refined expert cooperation analysis, especially in optimizing communication costs.

[0093] The embedding component of E2T was replaced with advanced node embedding methods (NeRank and Seq) and the heterogeneous graph embedding method metapath2vec (M2V), and ablation experiments were conducted to evaluate expert teams of different sizes formed using expert search methods to verify the effectiveness of E2T's embedding method. The NeRank and Seq methods utilize the structural information in the heterogeneous network constructed based on the relationship between existing questions, questioners, and experts, and focus on learning how to represent the nodes of the network in the embedding space. Both use the text content of existing questions and the structural information of network nodes to predict the most suitable experts. The difference between them lies in the type of text information used: NeRank combines the title and body of the question, while Seq relies on the label of the question. Through comparative experiments, E2T performs well on most datasets. The experimental results are shown in Figure 8 As shown, the following is our analysis and summary of the experimental results:

[0094] (1) In terms of skill coverage (SC), E2T shows better performance than other methods. This is because it not only relies on the description of the problem, but also considers the content of the expert's answer to the question. In contrast, the NeRank and Seq methods do not use the expert's answer to the existing questions, resulting in their lack of skill coverage. By focusing on both the task content and the expert's answer, E2T can more comprehensively explore the potential connections between experts, questions and answers, thereby improving the accuracy of team member selection and the efficiency of expert matching. This enables E2T to accurately identify relevant experts for new tasks, thereby improving the effectiveness of team building.

[0095] (2) In terms of expert level (EL), E2T performs significantly better than other methods. Although NeRank and Seq also model the relationship between experts and questions as a heterogeneous network, E2T uses the quality score of expert answers as edge weights when constructing the network, which enhances the network's characterization of the expert's professional level. In contrast, NeRank and Seq do not consider the quality of expert answers and only use a simple network structure, which makes them unable to accurately reflect the actual level of experts. By assigning weights to expert answers, E2T can better evaluate the professional ability of experts, thereby improving the overall expert level of the team. As the size of the team increases, the expert level (EL) decreases. This may be because larger teams contain more candidate experts who may not fully meet the needs of the team, thus introducing some irrelevant skills, which in turn affects the expert level.

[0096] (3) In terms of team communication cost (CL), E2T performs well. CL is used to measure the frequency of joint participation in problem solving among team members. A higher CL means that team members have cooperated more in problem solving in the past, which helps reduce communication costs and improve team efficiency. Unlike other methods that only consider direct cooperation relationships, E2T can more comprehensively explore potential collaboration networks by considering direct and indirect cooperation relationships between experts. By adopting the HCN-based team division method, E2T can fully consider the historical cooperation records between experts when building a team, thereby effectively reducing the communication cost of the team and improving the team's collaboration efficiency.

[0097] (4) In terms of standard team matching (GM), E2T also performs better. GM is an indicator that measures the matching degree between the retrieved team and the real team. From the experimental results, it can be seen that E2T exceeds other methods in GM score, indicating that it can more accurately match a team that meets the actual needs for a given task. This result shows that E2T can better grasp the collaborative relationship and task requirements of experts when building a team, thereby achieving more accurate team matching.

[0098] To further illustrate the superiority of E2T in standard team matching (GM), we conducted additional experiments to compare different methods in retrieving teams of the same size as the standard team. For a test question, these methods identified teams with the same number of people as the actual answerer of the question. We calculated the average overlap percentage between the teams identified by each method and the standard team. The experimental results are shown in Figure 2. Fig. 9 shown.

Claims

1. A task association induced team formation method combining global cohesion and local decoupling, characterized in that: Team division based on heterogeneous collaborative network HCN; based on expert data, historical task data and new task t new Construct a heterogeneous collaborative network graph G and select a special path Divide the expert nodes into teams and form a hypergraph H; A weighted folding graph is constructed based on the hypergraph H, and team representation learning based on global cohesion is performed to form a team embedding vector, which is placed in a D-dimensional vector space. At the same time, team representation learning based on local decoupling obtains the relationship between tasks and experts in each divided team, and obtains task embedding vectors and expert embedding vectors. The team matching strategy based on task relevance induction is used to new , calculate the task relevance, find the k most relevant historical tasks; according to the k most relevant historical tasks, t new Map to vector space and find the best matching team set.

2. The task association induction team formation method combining global cohesion and local decoupling according to claim 1 is characterized in that: The specific steps of performing team division based on the heterogeneous collaborative network HCN are as follows: (1) Input expert data and historical task data, construct a heterogeneous collaborative network graph G and select the dominant node v s and special paths (2) Initialization: Queue Q = {v s }, used to save nodes that can be further searched; set T = {v s }, records all nodes in a team; array R[v], records all explored neighbor nodes of node v; set Visited = {v s }, record all visited nodes, set Record nodes that do not satisfy constraints and should be deleted; (3) Initial division of the team: Take out the node v in Q and find its node in the special path Find all the neighbor nodes μ that have not been visited, and add μ to R[v] and Visited, and add R[v] to T; the constraint is that the nodes in R[v] have n neighbor nodes. If the constraint is met, the nodes that meet the constraint are added to Q and further searched, otherwise, the nodes that do not meet the constraint are added to DL; repeat the above steps until Q is empty; (4) Constraint-based pruning and optimization: Take a node from DL, delete it from T, iteratively delete all nodes that do not meet the constraints from T, and after deleting the node that does not meet the constraints, check whether each of its neighboring nodes still meets the constraints; if its neighboring nodes also do not meet the constraints, then delete the neighboring nodes from T; repeat the above steps until DL is empty; finally, remove v s All neighbor nodes that do not meet the constraints are added to T to relax the strict constraints; (5) Constructing the hypergraph H: Based on the divided teams, each team is regarded as a hyperedge to construct the final hypergraph H.

3. The task association induction team building method combining global cohesion and local decoupling according to claim 2 is characterized in that: In the heterogeneous collaboration network diagram, tasks and experts are represented as nodes, and the relationship between tasks and experts is represented as edges; the special path is expert-task-expert.

4. The task association induction team building method combining global cohesion and local decoupling according to claim 1 is characterized in that: Each hyperedge in the hypergraph H is a team, and a weighted folding graph is constructed based on the hypergraph H; in the weighted folding graph, each node corresponds to a team, and the weight of the edge reflects the degree of association between the teams, and the weight is the ratio of experts shared by two hyperedges.

5. The task association induction team building method combining global cohesion and local decoupling according to claim 4 is characterized in that: The team representation learning based on global cohesion is as follows: each team is regarded as a region in a D-dimensional vector space, represented by a position vector Pos and a range vector Ran, defined as Tm i =(Pos (i) , Ran (i) );Team Tm i Satisfy: Tm i = {v∈R d |dis(Pos (i) ,v)≤Ran (i) }; where dis(Pos (i) , v) is calculated as: The range vector is a constant that is dynamically adjusted according to the team size. Based on the range vector and the edge weights of the folding graph, the overlap constant of each team is obtained, which is defined as: d ij =(1-w ij )(Ran (i) +Ran (j) ); The position vector is obtained by combining Skip-gram negative sampling and overlap constant training on the folded graph; finally, the team embedding vector is obtained, and the team embedding loss function is as follows:

6. The task association induction team building method combining global cohesion and local decoupling according to claim 5 is characterized in that: The team representation learning based on local decoupling is specifically as follows: for each expert node and task node contained in the team embedding vector, their D-dimensional potential representations are learned to capture the semantic and structural relationships between each other, and the expert embedding vector and the task embedding vector are obtained. The loss function is defined as follows: Among them, T′ v ={experts,questions},N t (v) represents the neighboring nodes of a node.

7. The task association induction team building method combining global cohesion and local decoupling according to claim 6 is characterized in that: The P(v t |v;θ) is the softmax function: is the node type mapping for each node u, x u is the embedding representation of node u.

8. The task association induction team building method combining global cohesion and local decoupling according to claim 1 is characterized in that: The specific steps of the team matching strategy induced by task relevance are as follows: (1) Task relevance calculation: Input new task t new and historical tasks h , extract keywords respectively and Use the pre-trained word embedding model to map each keyword to an L-dimensional vector representation; construct a translation matrix Ts, where each element in Ts represents t new The words and t in h The embedding similarity between words in , Ts is defined as follows: The word-to-word interactions captured in the translation matrix Ts are transformed into task-solution ranking features φ(Ts): K(Ts i )={K1(Ts i ),...,K k (Ts i )} These ranking features are combined to generate the final ranking score: F(t new , t h )=tanh(w T φ(Ts)+b) Select the top k historical task sets T that are most relevant to the new task h ; (2) Team matching: assign new tasks to new Mapped to T h In the same vector space; t new The embedding vector of is matched with each team embedding vector; by calculating each potential team position vector and t new The distance between dis(Pos (i) ,G embeded (t new )), select the best matching t new The top m teams.

Images