2026-10-13T03:55:12Z https://www.scidb.cn/oai

10.57760/sciencedb.j00133.00645 2026-06-29T08:54:00Z

2026-06-29 Dataset for Research on Self-Organizing Collaboration Pattern Recognition and Its Impact for Innovative Problem Solving doi:10.57760/sciencedb.j00133.00645 en 1. Data Source and Collection ProcessThis dataset originates from the TrailBlazer online innovation community operated by Salesforce (https://trailblazers.salesforce.com/), supporting the research on "Recognition of Self-Organizing Collaboration Patterns for Innovative Problem Solving and Their Impact." Data collection was conducted in 2023, selecting 24 user innovation proposal discussion threads with a final status of "completed" and top-ranked by votes on the main posts, yielding a total of 8,052 community member comments. Through identifying interaction relationships among comments, 6,044 argumentative collaboration chains were constructed as research samples.The data collection process consisted of three stages. First, Python programming was used to call the Salesforce community's public API to batch-retrieve the main post content, comment content, and metadata information (including posting time, participant information, vote counts, etc.) of target discussion threads. Second, collaboration chains were constructed based on reply relationships among comments, combining comment sets with interaction relationships into chain structures according to the chronological order of comments. Finally, non-argumentative member comments (such as pure emojis, pure likes, etc.) were manually removed, retaining 6,044 valid collaboration chains.2. Data Processing Methods and ToolsData processing primarily comprised three stages: argumentation annotation, pattern mining, and variable construction. In the argumentation annotation stage, the research team conducted two-level argumentation coding (micro and macro levels) for activity units in collaboration chains based on a two-layer argumentation framework. Micro-level annotation adopted a simplified conversational argument coding scheme (Canary & Seibold, 2010), encoding argumentation activities into four categories: claim (B), evidence (C), agreement (D), and opposition (E); macro-level annotation, based on micro-level results, identified three types of activities: positive integration, negative integration, and consensus integration. The coding process employed single-coder multiple-round coding, ensuring Cohen's Kappa coefficient reached above 90% through repeated measurements; coding ambiguities were resolved through triple-coder coding and expert discussion to reach consensus.In the pattern mining stage, frequent substructure mining algorithms were employed to identify collaboration patterns. First, collaboration chains were converted into Activity Dependency Structures (ADS), with 10% as the support threshold to filter frequent candidate structures; second, subgraph isomorphism relationships were introduced to group candidate structures; finally, independent support filtering ensured pattern independence. The entire processing used Python 3.9, primarily relying on open-source libraries such as pandas and networkx for data analysis and graph structure processing.In the variable construction stage, the density of each collaboration pattern in collaboration chains was calculated as the independent variable, with platform response as the proxy variable for innovation performance, introducing knowledge construction as a mediating variable, and controlling for collaboration sentiment, time interval, and participant experience. Knowledge construction level was measured using the ReaderBench tool based on the cohesion network method; sentiment tendency was calculated using the VADER sentiment analysis model.3. Dataset Content and StructureThis dataset is provided in Microsoft Excel format (.xlsx), containing 8 worksheets divided into two dimensions: micro-level and macro-level, with each dimension containing three types of content: raw data, pattern mining results, and regression model data.Micro-level data contains 4 worksheets: (1) "微观层面原始数据" (Micro-level Raw Data) worksheet, containing three fields: interaction sequence number, comment argumentation coding, and simplified coding, recording the specific composition of argumentation elements in single-speaker comment sequences, where coding B represents claim types (including ASRT assertion, PROP proposition), C represents evidence types (including ELAB elaboration, AMPL amplification, JUST justification, FRAM framework, etc.), D represents agreement types (including AGMT agreement, ACKN acknowledgment), and E represents opposition types (including OBJC objection, CHAL challenge, RESP response); (2) "微观模式挖掘子序列计数" (Micro Pattern Mining Subsequence Count) worksheet, with 173 records, counting the frequency of each subsequence in the dataset; (3) "微观频繁数据集支持度计数" (Micro Frequent Dataset Support Count) worksheet, with 60 records, counting sequences meeting the support threshold and their unique support counts; (4) "微观层面回归模型数据" (Micro-level Regression Model Data) worksheet, with 5,378 records, containing 16 variable fields: sequence identifier, original sequence, simplified coding, official response (0/1 variable), non-development pattern density, weak development pattern density and its three subcategory densities (claim-evidence, agreement-evidence, opposition-evidence), strong development pattern density and its two subcategory densities (repetitive development, sequential development), participant experience (number of participations), time interval (days), sentiment tendency (mean score), and knowledge construction (standardized score). Density indicators are proportional values with a range of 0-1; participant experience is an integer count; sentiment tendency and knowledge construction are continuous standardized scores.Macro-level data similarly contains 4 worksheets: (1) "宏观层面原始数据" (Macro-level Raw Data) worksheet, containing three fields: interaction sequence number, complete argumentation coding, and simplified coding; (2) "宏观层面模式挖掘子序列计数" (Macro Pattern Mining Subsequence Count) worksheet, with 10,959 records, counting the frequency of each subsequence at the macro level; (3) "宏观频繁数据集支持度计数" (Macro Frequent Dataset Support Count) worksheet, with 449 records, counting the support of frequent sequences at the macro level; (4) "宏观层面回归模型数据" (Macro-level Regression Model Data) worksheet, with 666 records, containing 17 variable fields: sequence number, original sequence, simplified coding, official intervention (0/1 variable), M1 positive development pattern density, M2 positive-negative development pattern density, M3 consensus development pattern density, and subcategory densities of each pattern type (weak evidence, strong evidence, weak opposition, strong opposition, strong evidence consensus, weak evidence consensus), participant experience, time interval, sentiment tendency, and knowledge construction.4. Data Integrity and Quality AssuranceThis dataset has no missing value issues; all fields are completely filled. The "备注" (Remarks) column in the raw data contains explanatory text displayed only in the first row, with remaining rows being empty, which is normal. Data quality is ensured through: (1) the argumentation coding process employing multiple-round coding and expert review mechanisms to ensure coding reliability reaches acceptable levels; (2) pattern mining algorithm parameters (support threshold of 10%) set with reference to existing research, with identification results determined for optimal pattern count through Sum of Squared Errors (SSE) criterion; (3) in regression model data, the dependent variable (official response/intervention) is an objectively recorded variable, pattern density is calculated through formulas, and mediating and control variables all use mature measurement methods.Possible sources of error in the data include: (1) the manual coding process may have subjective judgment differences, minimized through Kappa coefficient testing and expert discussion; (2) sentiment analysis and knowledge construction measurement depend on algorithmic models with certain measurement errors, but the VADER model and ReaderBench tool employed are both validated mature tools.5. Data Files and Usage InstructionsThis dataset is provided in Microsoft Excel format (.xlsx) and can be opened using mainstream office software such as Microsoft Excel, WPS Office, and LibreOffice Calc, or read through programming tools such as Python (pandas, openpyxl libraries) and R (readxl package). The dataset structure is clear with standardized field naming, and logical correspondence exists among worksheets, facilitating secondary analysis and verification by researchers. It is recommended to first read the raw data worksheets to understand the basic data composition, then combine pattern mining results and regression model data for analysis. Online Innovation Community; Collaboration Pattern; Argumentation Coding; Frequent Pattern Mining; Knowledge Construction; Innovation Performance zhang jie EMBARGO https://opendatacommons.org/licenses/by/1-0/ dataset Science Data Bank