Task-related induction method for team formation combining global cohesion and local decoupling

Abstract

The application belongs to the technical field of team formation, and discloses a task-related induced team formation method combining global cohesion and local decoupling. A team division based on a heterogeneous collaboration network is proposed, which more finely describes the cooperation between experts and can simultaneously capture the implicit semantics in the text and the explicit structural relationship between experts to meet the specific task requirements. The hypergraph and the weighted folding graph are used to capture the potential of cross-team cooperation from a global perspective, and these global representations are mapped into the expert representation within the team through local decoupling. This method not only makes full use of the internal structure of the team, but also effectively avoids the scope limitation problem. The application introduces a task-related driven team matching strategy. By analyzing the correlation between the new task and the historical task, the strategy can select appropriate candidate teams, thereby effectively avoiding the false failure phenomenon in the matching process. The strategy aims to improve the accuracy and efficiency of team matching.

Description

Technical Field

[0001] This invention relates to the field of team building technology, and in particular to a task-related method for inducing team building that combines global cohesion and local decoupling. Background Technology

[0002] Team building requirements can be viewed as a task that details the professional skills and academic backgrounds of the required experts. The main challenge is selecting experts who meet the task requirements and maintain close collaboration. To study the team building problem, given task requirements, an expert pool, and their historical collaborations, the goal of this invention is to automatically identify a group of experts who not only meet the task requirements but also demonstrate strong collaborative potential.

[0003] Existing team-building methods are mainly divided into three types: topic-based, text-based, and graph-based methods. Topic-based methods achieve precise matching between task requirements and expert teams by analyzing the topical characteristics of the task and the research directions of experts. These methods typically employ statistical language models and semantic embedding techniques, combining task topics with expert research interests for modeling and optimization. Specifically, topic-based methods extract topical information from task descriptions and expert research content through topic modeling, revealing the correlation between tasks and experts on different topics. They embed the research content of tasks and experts into a high-dimensional semantic space, identify potentially relevant experts through semantic similarity analysis, and use probabilistic models to calculate the matching degree between experts and task topics, thus forming teams with complementary skills and strong collaborative potential. Text-based methods aim to uncover the potential relationship between experts and tasks by analyzing textual data in task descriptions and expert background information. They typically rely on natural language processing techniques to extract features from task requirements and expert research results or social network data. These features are used to rank experts and identify expert teams that highly match the task requirements. Furthermore, text-based methods can further optimize team member selection by analyzing social information such as collaborative relationships and sentiment tendencies among experts. Graph-based approaches model expert networks as nodes and edges, optimizing team composition by finding subgraphs that satisfy task skill requirements. This method typically combines subgraph search and optimization algorithms to find a subset of experts that meets task skill requirements and optimizes team cohesion. Simultaneously, it calculates task relevance based on weighted nodes and edges to select qualified team members.

[0004] The paper "A variational neural architecture for skill-based teamformation" (ACM Transactions on Information Systems 42.1(2023):1-28) improves upon traditional topic-based methods by proposing a skill-matching team-building method based on variational neural networks. This method aims to efficiently build teams that meet specific task requirements by learning the implicit correlation between expert skills and task demands. Expert skill data is mapped to a latent space to capture complex nonlinear relationships between skills and optimize the team's skill complementarity and collaborative potential. After inputting expert profiles and collaboration history data, the model generates latent representations through training and evaluates different team configurations to match the optimal combination under new task demands. The proposed technique improves upon topic-based team-building methods. However, in real-world scenarios, expert collaboration types are diverse, and this method heavily relies on the quality and completeness of expert profiles and collaboration history data. The model may not accurately capture the true skill relationships and complex collaborative relationships between experts. At the same time, although this method can learn the implicit relationships between skill nodes, it may not be able to fully capture the deep semantic connections between different domains. The generated potential 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 "A deep learning-based expert finding method to retrieve agile software teams from CQAs" (Information Processing & Management 60.2(2023):103144) proposes a text-based deep learning expert discovery method aimed at optimizing team building by analyzing candidates' knowledge and skills. First, a recommendation decision model is used to screen potential candidates. Then, a deep learning model is used in parallel to estimate candidates' expertise in their primary skill domain and general knowledge in other related domains, while dynamically estimating their knowledge evolution considering the time factor. Finally, an integer linear programming model is used to select the most suitable team members from the screened candidates. This technical method can be considered a text-based team building approach. Its core idea is to estimate the expertise and general knowledge of experts in specific skill domains by analyzing the text content they publish, thereby providing expert selection for team building. However, this method primarily relies on textual content to assess expert skills, which may not fully capture implicit semantics. For example, some candidates may possess cross-domain skills, but if these skills are not explicitly stated in the text, the model may not accurately identify these "implicit" skills. Therefore, the textual content may not fully reflect the expert's true professional level. Furthermore, this reliance solely on textual content fails to capture the explicit structural relationships between experts, potentially overlooking their abilities in areas such as teamwork and communication. Summary of the Invention

[0006] Previous team-building methods suffer from the following problems: they ignore the diversity of expert collaboration types, tend to limit member selection to existing teams, and fail to adequately consider the relevance between tasks. To address these issues, we propose a task-related team-building method that combines global cohesion and local decoupling.

[0007] The technical solution of this invention is as follows: a task-related team-building method combining global cohesion and local decoupling, which divides teams based on a heterogeneous collaborative network (HCN); and uses expert data, historical task data, and new task t. new Construct a heterogeneous collaborative network graph G and select a special path. The expert nodes are divided into teams and formed into a hypergraph H;

[0008] We construct a weighted folded graph based on the hypergraph H, perform team representation learning based on global cohesion to form team embedding vectors, and place them in a D-dimensional vector space. At the same time, we obtain the relationship between tasks and experts in each partitioned team based on local decoupling team representation learning, and obtain task embedding vectors and expert embedding vectors.

[0009] A team matching strategy based on task relevance is applied to a new task t. new Calculate task relevance and find the k most relevant historical tasks; based on the k most relevant historical tasks, assign task t... new Map to the vector space to find the best-matching set of teams.

[0010] The specific steps for team partitioning 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} is used to store 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}, records all visited nodes, set Record nodes that do not meet the constraints and should be deleted;

[0013] (3) Initial team division: Extract node v from Q and find its special path. Find all unvisited neighbor nodes μ and add μ to R[v] and Visited, and add R[v] to T; The constraint is that each node in R[v] has n neighbor nodes. If the constraint is satisfied, add the node that satisfies the constraint to Q and search further. Otherwise, add the node that does not satisfy the constraint to DL. Repeat the above steps until Q is empty.

[0014] (4) Constraint-based pruning and optimization: Remove a node from the DL and delete it from T. Iteratively delete all nodes from T that do not satisfy the constraints. After deleting a node that does not satisfy the constraints, check if each of its neighboring nodes still satisfies the constraints. If its neighboring nodes also do not satisfy the constraints, then that neighboring node is also deleted from T. Repeat the above steps until the DL is empty. Finally, set v... s All neighboring nodes that do not satisfy the constraints are added to T to relax the strict constraints;

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

[0016] In the heterogeneous collaborative network graph, 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 represents a team, and a weighted folded graph is constructed based on the hypergraph H. In the weighted folded graph, each node corresponds to a team, and the weight of the edge reflects the degree of association between teams. The weight is the proportion of experts shared between two hyperedges.

[0018] The team representation learning based on global cohesion specifically involves treating each team 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) The calculation method for v is as follows:

[0019]

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

[0021] The position vector is obtained by training on the folded graph using Skip-gram negative sampling and overlap constant; 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 specifically involves: for each team embedding vector containing expert nodes and task nodes, learning their D-dimensional latent representations to capture the semantic and structural relationships between them, resulting in expert embedding vectors and task embedding vectors. The loss function is defined as follows:

[0024]

[0025] Among them, T′v = {experts,questions}, N t (v) represents a node's neighboring nodes.

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

[0027] The specific steps of the task-related team matching strategy are as follows:

[0028] (1) Task relevance calculation: Input new task t new and historical missions h Extract keywords respectively and Each keyword is mapped to an L-dimensional vector representation using a pre-trained word embedding model; a translation matrix Ts is constructed, where each element in Ts represents t. new The word and t in h The embedding similarity between words in Ts is defined as follows: The 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: Matching new tasks with teams. new Mapping to T h In the same vector space;

[0035]

[0036] t newThe embedding vector is matched with the embedding vector of each team; by computing the position vector of each potential team and t new The distance between dis(Pos) (i) G embeded (t new Select the best match t new The first m teams.

[0037] The beneficial effects of this invention are as follows: A heterogeneous collaborative network (HCN) is designed to more precisely describe collaboration among experts, simultaneously capturing both implicit semantics in the text and explicit structural relationships between experts. Furthermore, a team partitioning method based on the HCN is proposed to meet specific task requirements.

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

[0039] This invention introduces a task-relevance-driven team matching strategy. By analyzing the correlation between new and historical tasks, this strategy can select suitable candidate teams, thereby effectively avoiding false failures during the matching process. This strategy aims to improve the accuracy and efficiency of team matching. Attached Figure Description

[0040] Figure 1 This is a general framework diagram of the method of the present invention;

[0041] Figure 2 This is the main flowchart of the method of the present invention;

[0042] Figure 3 A flowchart illustrating the HCN-based team partitioning method;

[0043] Figure 4 A flowchart for team representation learning 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 of different methods is compared across 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 using different methods on different datasets: (a) Android-CL dataset; (b) Android-GM dataset; (c) Android-SC dataset; (d) Android-EL dataset; (e) History-CL dataset; (f) History-EL dataset; (g) History-SC dataset; (h) History-GM dataset; (i) DBA-CL dataset; (j) DBA-EL dataset; (k) DBA-SC dataset; (l) DBA-GM dataset; (m) Physics-EL dataset; (n) Physics-SC dataset; (o) Physics-CL dataset; (p) Physics-GM dataset.

[0048] Figure 9 This is a comparative experiment of GM on different datasets. Detailed Implementation

[0049] This method fully considers 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 implicit semantics in the text and 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; and it introduces task relevance factors to avoid false failures in the matching process.

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

[0051] 1. Team Partitioning Method Based on Heterogeneous Collaborative Networks (HCNs). Team formation requires a more comprehensive capture of the complex relationships between experts. Relying solely on collaborative relationships without distinguishing between collaboration types may result in the selection of unsuitable experts into teams. Based on heterogeneous collaborative networks, teams are initially partitioned by identifying specific paths. The initial partitions are then pruned and refined according to constraints, ultimately forming a refined partitioning result. These results will serve as the basis for constructing a hypergraph.

[0052] 2. A team representation learning method based on global cohesion and local decoupling. When forming a new team, it's crucial to avoid limiting member selection to existing teams. This approach constructs a weighted folded graph based on a hypergraph to capture the potential for cross-team collaboration from a global perspective. Subsequently, these global representations are transformed into internal expert representations through local decoupling.

[0053] 3. Task-Relevance-Induced Team Matching Strategy. Task requirements and expert skills may not match in description, but there may be underlying semantic connections between them. Relying solely on the descriptive match between expert skills and task requirements during team formation may overlook deeper connections hidden by these differences in expression. By calculating the relevance between new and historical tasks, suitable candidate teams can be selected, and the participating expert members can be further identified.

[0054] Figure 2 The main flowchart of the technical solution of this invention is shown, as follows: Figure 2 As shown, the task-related team building method proposed in this invention, which combines global cohesion and local decoupling, includes the following steps:

[0055] Input expert data, historical task data, and new task t new Team partitioning based on Heterogeneous Collaboration Network (HCN). Construction of the heterogeneous collaboration network graph. Select a special path The expert nodes are divided into teams and formed into a hypergraph H.

[0056] We construct a weighted folded graph based on the hypergraph H, perform team representation learning based on global cohesion to form team embedding vectors, and place them in a D-dimensional vector space. At the same time, we obtain the relationship between tasks and experts in each partitioned team based on local decoupling team representation learning, and obtain task embedding vectors and expert embedding vectors.

[0057] A team matching strategy based on task relevance is applied to a new task t. new Calculate task relevance and find the k most relevant historical tasks; based on the k most relevant historical tasks, assign task t... new Map to the vector space to find the best-matching set of teams.

[0058] Figure 3 The flowchart for team partitioning based on HCN is shown, and the specific steps are as follows:

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

[0060] Initialization: Queue Q = {v s} is used to store 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}, records all visited nodes; set Record nodes that do not meet the constraints and should be deleted.

[0061] Initial team partitioning: Extract node v from Q and find its path Collect all unvisited neighbor nodes μ and add μ to R[v] and Visited, then add R[v] to T; determine if all nodes in R[v] have n neighbor nodes. If the constraint is satisfied, add the nodes that satisfy the constraint to Q and continue searching; otherwise, add the nodes that do not satisfy the constraint to DL. Repeat the above steps until Q is empty.

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

[0063] Constructing the hypergraph H: Based on the partitioned teams, each team is regarded as a hyperedge, and the final hypergraph H is constructed.

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

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

[0066] Team representation learning based on global cohesion: Each team is treated 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) The calculation method for v is as follows:

[0067]

[0068] The range vector is a constant that is dynamically adjusted based on the team size. Based on the range vector and the edge weights of the folded graph, the overlap constant for each team can be obtained, defined as: d ij =(1-w ij (Ran) (i) +Ran (j) The position vector is obtained by combining Skip-gram negative sampling on the folded graph and the overlap constant during training. 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 containing expert nodes and task nodes, learn their D-dimensional latent 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 given node. P(v) t |v;θ) is considered the softmax function, and we have the following definition:

[0073]

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

[0075] Figure 5 The flowchart illustrates a task-relevance-driven team matching strategy, with the specific steps as follows:

[0076] Task relevance calculation: Input new task t new and historical missions h Extract keywords respectively and Each keyword is mapped to an L-dimensional vector representation using a pre-trained word embedding model; the word embedding model can be any existing word embedding model, such as word2vec.

[0077] Construct a translation matrix Ts, where each element in Ts represents t. new The word and t in h The embedding similarity between words in Ts is defined as follows: The 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] These ranking features are then 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. Using this sorting mechanism, the new task t can be calculated. new And each historical task t h The correlation between them is used to identify the top k historical task sets T that are most relevant to the new task. h .

[0084] Team matching: After identifying T h After that, the new task t new Mapping to T h In the same vector space:

[0085]

[0086] t new The embeddings are matched with the embeddings of each team. This is done by calculating the embeddings of each potential team and t. new The distance between them can be used to select the best match t. new The first m teams.

[0087] This invention is implemented in Python and runs on an Intel(R) Xeon(R) Silver 4310 CPU 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-world datasets from Stack Exchange: Android, History, DBA, and Physics.

[0088] The proposed technical solution (E2T) was compared with the following team building methods: CCR, SA-CA-CC, NCCO, TOSA, and TAPG. The main goal of CCR is to build a team that meets needs and maximizes the collaborative potential among members. NCCO and SA-CA-CC prioritize 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. Experimental results show that E2T not only achieves significant advantages in skill coverage and accuracy but also demonstrates unique advantages in optimizing team communication costs. Its success lies in fully exploring the interaction and collaboration potential among experts by combining their historical collaborative relationships and cross-team collaboration potential, ensuring more efficient and accurate team building.

[0089] Figure 7 The experimental results are presented, and a detailed analysis of the experiment follows:

[0090] (1) CCR, NCCO, and SCA (i.e., SA-CA-CC) have significant shortcomings in terms of skill coverage (SC). This is because these methods mostly rely on graph search techniques to progressively add team members until all required skills are met. However, this strategy is prone to getting trapped in local optima, causing the algorithm to fail to fully cover all necessary skills when expanding the team. In particular, when some skills are selected early on, the selection space for the remaining members may be limited, thus affecting the final skill coverage effect.

[0091] (2) In terms of team communication costs (CL), TOSA performed significantly worse than other methods. The main goal of TOSA is to select the most suitable experts based on subject sensitivity, rather than to optimize the combination of team members to reduce communication costs. Therefore, TOSA has a certain advantage in assessing whether experts have the required skills and can achieve good skill coverage, but it ignores the collaboration history and communication costs among team members, resulting in lower communication efficiency in multi-expert collaboration.

[0092] (3) In terms of accuracy (PN), E2T and TAPG perform similarly, both incorporating textual information and considering the collaborative relationships between experts during team formation. However, the core difference lies in the fact that E2T, the method of this invention, enhances the collaborative potential among team members by directly integrating historical expert collaboration records and cross-team relationships, while TAPG indirectly infers expert collaboration relationships through expert work content, lacking direct support from collaborative history. Therefore, the technical solution of this invention, through this refined expert collaboration analysis, makes team member matching more accurate, especially in optimizing communication costs.

[0093] Ablation experiments were conducted to replace the embedding components of E2T with advanced node embedding methods (NeRank and Seq) and the heterogeneous graph embedding method metapath2vec (M2V), evaluating expert teams of different sizes formed using expert search methods to validate the effectiveness of E2T's embedding methods. NeRank and Seq methods leverage structural information in heterogeneous networks built based on existing questions, question proposers, and relationships between experts, focusing on learning how to represent the network's nodes in the embedding space. Both utilize the textual content of the existing question and the structural information of the network nodes to predict the most suitable expert. Their difference lies in the type of textual information used: NeRank combines the question title and body, while Seq relies on the question label. Comparative experiments show that E2T performs well on most datasets. Experimental results are as follows: Figure 8 As shown, the following is our analysis and summary of the experimental results:

[0094] (1) In terms of skill coverage (SC), E2T demonstrates superior performance compared to other methods. This is due to its reliance not only on the problem description but also on the content of expert responses. In contrast, NeRank and Seq methods do not utilize expert responses to existing problems, resulting in deficiencies in skill coverage. By simultaneously focusing on task content and expert responses, E2T can more comprehensively uncover the potential connections between experts, problems, 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, thus improving the effectiveness of team building.

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

[0096] (3) In terms of team communication cost (CL), E2T performs superiorly. CL measures the frequency with which team members collaborate in problem-solving. A higher CL indicates that team members have collaborated more in the past when solving problems, which helps reduce communication costs and improve team efficiency. Unlike other methods that only consider direct collaboration relationships, E2T can more comprehensively uncover potential collaboration networks by considering both direct and indirect collaboration relationships between experts. By adopting an HCN-based team partitioning method, E2T can fully consider the historical collaboration records between experts when building teams, thereby effectively reducing team communication costs and improving team collaboration efficiency.

[0097] (4) In terms of standard team matching (GM), E2T also performs better. GM is an indicator that measures the degree of matching between retrieved teams and real teams. The experimental results show that E2T outperforms other methods in GM score, indicating that it can more accurately match teams that match the actual needs of a given task. This result shows that E2T can better grasp the collaborative relationships of experts and task requirements when building teams, thus achieving higher accuracy in team matching.

[0098] To further illustrate the superiority of E2T in standard team matching (GM), we conducted additional experiments comparing 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 actually answering the question. We calculated the average percentage overlap between the teams identified by each method and the standard team. Experimental results are as follows: Figure 9 As shown.

Claims

1. A method for inducing team formation by combining global cohesion and local decoupling through task association, characterized in that, Team division is based on heterogeneous collaborative network (HCN); based on expert data, historical task data, and new tasks. Constructing a heterogeneous collaborative network graph Choose a special path The expert nodes are divided into teams and a hypergraph is formed. In the heterogeneous collaborative network graph, tasks and experts are represented as nodes, and the relationships between tasks and experts are represented as edges. The special path is Expert-Task-Expert; According to the Hypergraph A weighted folded graph is constructed, and team representation learning based on global cohesion is performed to form team embedding vectors, which are then placed in a D-dimensional vector space; the hypergraph Each hyperedge in the graph represents a team, based on the hypergraph. Construct a weighted folded graph; in the weighted folded graph, each node corresponds to a team, and the weight of the edge reflects the degree of association between teams, wherein the weight is the proportion of experts shared by two hyperedges; the team representation learning based on global cohesion specifically involves: treating each team as a region in a D-dimensional vector space, and using position vectors... and range vector It means, defined as );team satisfy: ;in, The calculation method is as follows: ； 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 folded graph, the overlap constant for each team is obtained and defined as: ; The position vector is obtained by training on the folded graph using Skip-gram negative sampling and overlap constant; finally, the team embedding vector is obtained, and the team embedding loss function is as follows: Simultaneously, based on locally decoupled team representation learning, the relationship between tasks and experts in each partitioned team is obtained, yielding task embedding vectors and expert embedding vectors. Specifically, the locally decoupled team representation learning involves learning the relationships between expert nodes and task nodes within each team embedding vector. The latent representation captures the semantic and structural relationships between the vectors, yielding expert embedding vectors and task embedding vectors. The loss function is defined as follows: ； in, Then it represents the neighboring nodes of a certain node; Team matching strategies based on task relevance are applicable to new tasks. Calculate task relevance and find the k most relevant historical tasks; based on the k most relevant historical tasks, Map to the vector space to find the best-matching set of teams.

2. The task-related team building method combining global cohesion and local decoupling as described in claim 1, characterized in that, The specific steps for team partitioning based on the heterogeneous collaborative network HCN are as follows: (1) Input expert data and historical task data to construct a heterogeneous collaborative network diagram. And select the dominant node and special paths ; (2) Initialization: queue ={ } is used to store nodes that can be further searched; set T = { }, records all nodes in a team; array Record nodes All explored neighboring nodes; set Record all visited nodes, set Record nodes that do not meet the constraints and should be deleted; (3) Initial team division: Nodes in Take it out and find it in a special path. All unvisited neighbor nodes and will Add to and In, and will join in The constraints are: All nodes in the middle have If a neighboring node satisfies the aforementioned constraints, then that node is added to the list of neighboring nodes. And further search, otherwise, in Add nodes that do not meet the constraints; repeat the above steps until... Empty; (4) Constraint-based pruning and optimization: From Retrieve the node, remove the node from Delete, iteratively from Delete all nodes that do not satisfy the constraints. After deleting a node that does not satisfy the constraints, check if each of its neighboring nodes still satisfies the constraints. If its neighboring nodes also do not satisfy the constraints, then that neighboring node is also removed from the list. Delete; repeat the above steps until... Empty; finally, Add all neighbor nodes that do not satisfy the constraints to China relaxes strict restrictions; (5) Constructing a hypergraph Based on the divided teams, each team is considered as a hyperedge, and the final hypergraph is constructed. .

3. The task-related team building method combining global cohesion and local decoupling as described in claim 1, characterized in that, The For the softmax function: ; For each node Node type mapping, It is a node Embedded representation.

4. The task-related team building method combining global cohesion and local decoupling as described in claim 1, characterized in that, The specific steps of the task-related team matching strategy are as follows: (1) Task relevance calculation: Input new task and historical mission Extract keywords respectively and Each keyword is mapped to an L-dimensional vector representation using a pre-trained word embedding model; a translation matrix is ​​then constructed. , Each element in represents words and Embedding similarity between words in the text The definition is as follows: = ;Translation matrix The word interactions captured are transformed into ranking features for task-solution analysis. : ； ； ； These ranking features are combined to generate the final ranking score: Select the top k historical task sets that are most relevant to the new task. ; (2) Team matching: matching new tasks Mapping to In the same vector space; ； Will The embedding vector is matched with the embedding vector of each team; by computing the position vector of each potential team and Distance between Select the best match The first m teams.

Images