Data and code for "Public Self-Descriptions Encode Field Position in a French Twitter Core"

# Data Availability Statement *For submission to the Journal of Computational Social Science* --- ## Statement for manuscript review The data underlying this study were collected from the X (formerly Twitter)public API in 2022 under a standard Academic Research access agreement. Theoriginal corpus comprises 1,694,375 French-language user profiles stored asparquet-partitioned node files, and 229,078,863 directed follower ties storedas parquet-partitioned edge files. The reduced analytic core used in allreported analyses consists of 10,000 accounts and 3,418,570 directed edges. All data used in this study consist of publicly posted content on the Xplatform. No private messages, non-public accounts, or off-platform datasources were used. The public release contains: | File | Description ||------|-------------|| `user_ids.txt` | Dehydrated list of 10,000 numeric user IDs (one per line). No profile content. || `community_labels.csv` | User ID, community integer, and community label for each of the 10,000 core users. || `edge_list.csv` | All directed edges in the core as (source\_id, target\_id) pairs. No content columns. || `aggregates/cluster_summary.csv` | Community-level counts, degree statistics, and verified shares. || `aggregates/cluster_top_terms_tfidf.csv` | Top-6 TF-IDF terms per community. || `aggregates/deep_bio_feature_prevalence_by_cluster.csv` | Mean rate of each bio-feature family by community. || `aggregates/deep_bio_style_scores_by_cluster.csv` | Composite register scores (credentialized publicity, expressive ideology, brand/entrepreneurial) per community. || `aggregates/deep_bio_verified_comparison.csv` | Feature means split by verified status. || `aggregates/deep_bio_interpretable_model_coefficients.csv` | Multinomial logistic regression coefficients for the interpretable-features classifier. || `aggregates/bio_prediction_metrics.json` | TF-IDF classifier performance (accuracy, F1, confusion matrix). || `aggregates/deep_bio_interpretable_model_metrics.json` | Interpretable-features classifier performance. || `aggregates/network_overview.json` | Graph-level statistics (edges, modularity, component count). || `code/replicate_analysis.py` | Full end-to-end replication script (Python). || `code/prepare_public_data.py` | Script to regenerate the public release from the source DuckDB. || `code/requirements.txt` | Python dependency list with pinned versions. | All aggregate outputs contain only group-level or vocabulary-level statistics.No individual user content (bio text, screen names, display names, or locations)is included in the public deposit. --- ## What is not publicly available and why Raw profile content — including user bio text, screen names, and display names —cannot be redistributed under the X Developer Agreement (§ II.C and § II.D),which prohibits re-hosting or re-distributing Content (as defined by theAgreement) outside the X platform. This restriction applies regardless ofwhether the underlying content is publicly visible on the platform. Specifically, the following materials are held locally and not deposited: - The source DuckDB file (`2022 France Tweets.duckdb`) containing full user profile records and tweet text.- Per-user feature tables that map a user ID to bio text or profile metadata. These files are available to qualified researchers upon reasonable request,subject to the execution of a data sharing agreement consistent with the XDeveloper Agreement. Requests should be addressed to the corresponding author. --- ## How to rehydrate the data Researchers who wish to recover the profile text for the 10,000 core userscan do so using the X API's User Lookup endpoint: ```GET https://api.twitter.com/2/users ?ids=<comma-separated-ids> &user.fields=description,public_metrics,verified``` The IDs required are listed in `user_ids.txt` in the public deposit.Access requires an X Developer account (Basic tier or above). A helperscript is included in the deposit (`code/rehydrate_profiles.py`). Note that accounts that have been suspended, deactivated, or made privatesince the 2022 collection date will not be returned by the API. Therehydrated corpus may therefore differ slightly from the original. --- ## Ethical and legal considerations This study analyzes publicly posted user profiles on a social media platform.No private, restricted-access, or sensitive personal data were used. Profilebiographies are public-facing texts deliberately written for a general audience.All analysis is conducted at the group level or at the level of vocabularypatterns. No individual is identified or profiled in the reported results. The study did not require ethics board review under our institution's currentguidelines for analysis of publicly available social media data. However, wefollowed community best practices for computational social science, including: - Reporting results at the community level rather than highlighting individuals.- Withholding the mapping of user IDs to bio text from the public deposit.- Framing findings in terms of public self-presentation rather than private identity or psychological attributes.- Clearly stating the scope and limitations of the analytic sample. --- ## Software and computational environment All analyses were conducted in Python 3.11. Key dependencies: | Package | Version | Purpose ||---------|---------|---------|| duckdb | ≥ 0.10 | Source data access || igraph | ≥ 0.11 | Community detection (multilevel Louvain) || pandas | ≥ 2.0 | Data manipulation || scikit-learn | ≥ 1.4 | TF-IDF vectorization and classification || pyarrow | ≥ 14.0 | Parquet I/O | Full dependency list with pinned versions is in `code/requirements.txt`.Code was run on a single workstation (Apple Silicon, 32 GB RAM) and requiresno specialist hardware. Runtime for the full pipeline on the 10,000-node coreis approximately 8–12 minutes. --- *Last updated: April 2026*