(1) Overview
Repository location
Collection: https://nakala.fr/collection/10.34847/nkl.7dbf6k3o
README DOI: https://doi.org/10.34847/nkl.c5767c8o
Context
The dataset was created as part of a doctoral research project on user browsing practices in OpenEdition. It supports the analysis of navigation patterns and information-seeking behaviors in an open-access digital library environment, extending previous work based on session datasets in information retrieval research (Mayr, 2016; Buscher, 2010). To promote reproducibility and data reuse, the dataset includes documented preprocessing procedures, notably session reconstruction and bot detection.
(2) Method
Session identification is a key challenge in web usage analysis. Existing methods typically rely on inactivity thresholds (Halfaker et al., 2015; He et al., 2002) or navigation-based heuristics exploiting website structures (Bayir & Toroslu, 2022). Because modern browsing behaviors limit the effectiveness of purely structural approaches, session reconstruction often combines temporal and contextual signals.
In this work, sessions are identified using a probabilistic soft-threshold model that combines the inter-request time interval (Δt) with referrer continuity (ref). The inactivity threshold (τ) is determined empirically from the distribution of inter-request intervals, while the severity parameter (α) is selected as the first value for which the number of reconstructed sessions stabilizes. This approach balances session cohesion and distinctness in the absence of ground-truth labels. The probability model is defined as follows:
where Δt denotes the inter-request time interval, τ the inactivity threshold, ref the referrer continuity indicator, and α a severity parameter controlling the decay rate of the function.
After session reconstruction, automated sessions (bots) are identified. Existing bot detection methods typically rely on user-agent signatures, behavioral patterns, or machine learning models (Doran & Gokhale, 2011). In this work, bot detection combines syntax-based and pattern-based techniques. Known crawlers are first identified using the CrawlDetect Python library. Additional behavioral heuristics are then applied, and a session is classified as automated if it contains more than 100 requests, exceeds a browsing speed of two requests per second, or exhibits perfectly regular inter-request intervals (standard deviation equal to zero).
Steps
Raw Nginx server log files were automatically generated on a daily basis from OpenEdition’s production servers. Each platform generates separate access logs (for instance, journal logs record interactions specific to academic journals). These raw log files were collected from the OpenEdition Data team. To capture complete user activity across the entire ecosystem (covering Journals, Books, Calenda, Hypothèses, and the Search engine) the logs from all individual platforms were aggregated into a single integrated dataset.
The logs were parsed and pre-processed to remove server-generated requests for static assets such as scripts, stylesheets, and images.
User sessions were reconstructed using a probabilistic soft-threshold strategy that evaluates the likelihood of two consecutive requests belonging to the same session based on inter-request time intervals and referrer continuity (as defined in mathematical formula above).
Automated traffic was filtered using a combination of signature-based detection (CrawlDetect) and behavioral heuristics, including request volume, browsing speed, and regularity of inter-activity times.
IP addresses were geolocated with the GeoIP2 database corresponding to each month.
Document metadata were extracted from OpenEdition’s Solr instance and normalized into a corpus table, while access modalities were compiled into separate CSV files.
The final dataset was organized into weekly Parquet files for sessions and CSV files for corpus and access-modality information.
Sampling strategy
The dataset is not based on random sampling. It was built as a time-bounded snapshot covering two comparable periods: December 2024 and December 2025. December was selected because it represents a period of sustained activity in annual server logs. Each period is split into weekly files so that navigation behavior can be compared across years and across the five calendar weeks represented in the December traces.
Quality control
Server-generated requests were removed before session reconstruction to avoid inflating activity counts.
Potential bot traffic was filtered using both user-agent signatures and behavioral thresholds.
Session boundaries were estimated using a probabilistic rule designed to balance continuity and separation of browsing episodes.
Metadata were translated into English for consistency, except for the disciplinary index, which was preserved in its original form because accurate translation would require domain-specific normalization.
Access-modality information was assembled with archived records so that changes in embargo status and access policy during the study periods could be represented.
(3) Dataset Description
Repository name
Nakala.
Object name
A collection of files containing session data, corpus metadata, and access-modality tables.
Format names and versions
Parquet and CSV.
The dataset is divided into three data families: pre-processed session data (user traffic) stored in Parquet format for December 2024 and December 2025; one table containing document metadata; and two tables describing access modalities for both periods, stored in CSV format.
| FILE | DESCRIPTION |
|---|---|
| final_sessions_week_1_year_2024_month_12.parquet | User sessions during the specified period. |
| final_sessions_week_2_year_2024_month_12.parquet | |
| final_sessions_week_3_year_2024_month_12.parquet | |
| final_sessions_week_4_year_2024_month_12.parquet | |
| final_sessions_week_5_year_2024_month_12.parquet | |
| final_sessions_week_1_year_2025_month_12.parquet | |
| final_sessions_week_2_year_2025_month_12.parquet | |
| final_sessions_week_3_year_2025_month_12.parquet | |
| final_sessions_week_4_year_2025_month_12.parquet | |
| final_sessions_week_5_year_2025_month_12.parquet | |
| final_corpus.csv | Metadata of all digital resources in the library |
| diffusion_modality_2024.csv | Access modality per resource for both December 2024 and 2025. |
| diffusion_modality_2025.csv |
Sessions files (Parquet)
The dataset contains ten Parquet files (five per year), covering the entire 31-day period of December for either 2024 or 2025. Each file corresponds to a calendar week (or partial week) spanning the full month. Each file is derived from the raw server logs of the digital library platform and has been pre-processed to resolve IP addresses to countries and to assign session identifiers (see the methods section). Each row in a sessions file corresponds to one HTTP request.
| COLUMN | DESCRIPTION |
|---|---|
| id | Unique identifier of the digital resource; links to the id column in the modality access files |
| type | Type of the resource (e.g., article, book chapter, facsimile) |
| publication_date | Date and time (UTC) when the resource was published on the platform |
| url | Canonical URL of the resource; can be matched against the request column in the sessions’ files |
| title | Title of the document |
| language | ISO 639-1 language code of the document |
| publisher_country | Country of the publisher |
| discipline_index | Comma-separated list of disciplines assigned to the document based on the OpenEdition ontologies see link The index was left into its original language as we don’t have the knowledge to translate to their equivalent disciplines in English. |
Corpus file (CSV)
The file final_corpus.csv contains metadata for all digital resources available in the library as of 31 December 2025. Each row describes one resource. The id column serves as the primary key of this file and is referenced by the modality access files. The “url” column can be joined with the request column of the sessions files to link user traffic to document metadata.
| COLUMN | DESCRIPTION |
|---|---|
| session_id | Unique identifier that groups all HTTP requests belonging to the same user session |
| time_local | Timestamp (with timezone) at which the server received the request from the user |
| request | Full URL of the requested resource on the platform |
| http_referer | Referrer URL indicating the previous page the user visited before the current request; ‘-’ if not available |
| http_x_forwarded_for | Forwarded IP address from the proxy chain; None if not available |
| country | Country of origin of the request, inferred from the IP address |
Modality access files (CSV)
Access to resources on OpenEdition is governed by diverse and time-varying policies. Depending on the document’s specific status at the time of the request, access modalities include:
Open access: users can freely read the HTML version and download the PDF or EPUB files.
Freemium: users can freely read the HTML version, but downloading files requires institutional access.
Barrière mobile (moving wall): content becomes freely accessible after an embargo period (typically 6, 12, 24, or 36 months).
Barrière mobile via commercial access: the document is under a moving wall on OpenEdition, but remains accessible to users who purchased access through Cairn.
The two modality access files record the access model associated with each resource for December 2024 and 2025. The id column is a foreign key to the id column in final_corpus.csv.
This table describes the columns of the modality access.
General information about the data collection
Creation dates 2024-12-01to 2024-12-31 and 2025-12-01 to 2025-12-31 (December 2024 and December 2025 collections).
Dataset creators Mohsine Aabid (OpenEdition and Aix Marseille University, Marseille, France): Investigation, Data curation, Formal analysis, Validation. Simon Dumas Primbault (OpenEdition, Marseille, France): Supervision. Patrice Bellot (Aix Marseille University, Marseille, France): Supervision.
Language English for labels and documentation, with one original French field preserved in the disciplinary index.
License The dataset is distributed under the Etalab Open License 2.0 (etalab-2.0). Choosing this license ensures full legal alignment with French public research frameworks while guaranteeing unrestricted reuse and re-analysis of the data, provided proper attribution is given.
Publication date 2026-05-12
(4) Reuse Potential
The dataset can be reused to study scholarly navigation, search behavior, and patterns of consultation in open-access digital libraries. Within information science, humanities computing, and human-computer interaction, it can support analyses of session structure, browsing transitions, and content discovery. In computer science, it can be used to benchmark session-based recommendation methods, ranking approaches, and user-behavior models. The longitudinal design also makes it suitable for comparing reading and research practices across two years. Beyond research, the dataset can be used for teaching, methodological validation, and interface evaluation in digital library design. Reuse is limited by the fact that the data come from a single platform, include only platform-recorded interactions, and exclude direct personal identifiers and most user-level contextual variables.
Note
[1] Available at https://www.openedition.org/19362?lang=en.
AI Declaration
During the preparation of this work the authors used AI in order to rephrase. After using this tool, the authors reviewed and edited the content as needed and takes full responsibility for the content of the published article.
Ethical Considerations and Data Anonymization
Compliance with personal data protection regulations, including the GDPR, was rigorously maintained throughout the creation of this dataset. OpenEdition’s official privacy policy1 explicitly informs users that connection logs may be reused for scientific research purposes, and prior to data extraction, the entire processing workflow underwent formal review and validation by the CNRS Data Protection Officer (DPO). Furthermore, strict anonymization measures were applied: raw IP addresses and user agents were stripped or transformed during pre-processing, and interactions were aggregated into isolated session identifiers based solely on temporal gaps and referrer continuity, ensuring that users cannot be traced or re-identified.
Acknowledgements
The authors thank the OpenEdition team, the Pôle Universitaire de Données (PUD) of Aix-Marseille University, and the CNRS Ethics Committee for their valuable support, guidance, legal advice and assistance throughout this work.
Author Contributions
Mohsine Aabid: Conceptualization, Methodology, Software, Data Curation, Formal Analysis, Writing – original draft.
Simon Dumas Primbault: Supervision, Writing – review & editing.
Patrice Bellot: Supervision, Writing – review & editing.