Online news article metadata and hourly monitoring windows for crisis detection in digital media monitoring, Indonesia, 21 October to 2 November 2025
Description
This dataset contains the data of one case used to demonstrate a streak-based hysteresis architecture for crisis detection in digital media monitoring. The case covers online news coverage of a fast-moving consumer goods brand in Indonesia over 13 days, from 21 October 2025 to 2 November 2025, divided into 312 hourly monitoring windows. All data come from online news portals. The dataset contains two tables. The article table lists the 2,283 articles published in the observation window, with the article identifier, URL, news portal, publication and ingestion times, the result of the relevance filter, sentiment, source tier, and story identifier. Of these articles, 2,063 were marked as relevant and 220 as irrelevant. The window table lists the 312 hourly windows with the scope state at the end of each window, the spike signal, the scope-level risk level, the quality gate, the number of stories and relevant articles, and whether an alert was dispatched. The article text is not included, to respect the copyright of the publishers. Sentiment and story labels were produced by a language model and were not validated or audited by hand. Times are given in epoch seconds UTC and in Western Indonesia Time (WIB, UTC+7). The analysis scripts are not included. A README file describes every column.
Files
Steps to reproduce
The data were produced by a commercial digital media monitoring platform during its operation and exported after the run. The steps below describe how the two tables were produced and how the counts reported in the associated research paper can be checked. 1. Article collection. The platform collected news articles from online news portals. The article table keeps every article with a publication time from 21 October 2025 00:00 WIB to 3 November 2025 00:00 WIB (publish_ts from 1760979600 to 1762102800, both ends included), which gives 2,283 rows. 2. Article attributes. A relevance filter marked each article as relevant (2,063) or irrelevant (220). Sentiment was assigned by a language model. Source tier comes from the media catalog of the platform. Relevant articles were grouped into stories by a commercial large language model API, and articles that received the same story title share one story_id. Irrelevant articles have no story. 3. Monitoring windows. The observation window is divided into 312 hourly windows. Window k (k = 1 to 312) starts at 1760979600 + 3600 x (k - 1) and ends one hour later. A relevant article belongs to a window when its ingestion time is after the window start and no later than the window end, and its publication time is no earlier than three hours before the window end. Counting relevant articles per window with this rule reproduces article_now in the window table, 2,062 in total. 4. Window signals. story_now, sr_star, risk_level, qg_pass, final_state, and the alert flag were recorded by the platform for each window. In the evaluated run the expected baseline was fixed at 3 stories, so sr_star equals story_now divided by 3. 5. Threshold Baseline. The comparison rule in the paper assigns a relevant article to a window by ingestion time only and raises an alert when a window contains at least five negative articles. Applying this rule to the article table gives 101 windows. 6. Checks. 2,283 articles, 2,063 relevant, 220 irrelevant (all neutral), 1,531 relevant negative, 179 relevant negative from tier 1 sources, 312 windows, 70 windows with qg_pass of 1, 98 windows with sr_star of 2 or more, and 62 windows with an alert. The analysis scripts are not included. Any tool that reads CSV files, such as Python with pandas or R, can reproduce these counts. The article text is not included and can be read at the URLs, subject to availability. Epoch timestamps are in seconds UTC, and columns ending in _wib are in Western Indonesia Time (UTC+7).