How short lived topic peaks are detected AI
A local bump is a period with a higher topic rate than the years immediately before and after it. Requiring both comparisons helps distinguish a temporary peak from a sustained rise that simply reaches its highest value near the end of the record.
Compare the window with both sides
The screen considers windows of one, two or three consecutive years. Each window has a two-year flank immediately before it and another immediately after it. Counts are divided by the number of sampled papers in each period, so periods of different lengths remain comparable.
For each flank, a one-sided conditional test asks whether the window contains unusually many topic papers. Conditional on the combined topic count, the hypergeometric distribution supplies the reference distribution under equal topic shares in the two periods.
A bump must exceed both flanks. The larger of the two comparison p-values is therefore used for that window. A strong comparison on only one side is insufficient.
Account for searching across windows
There are 18 eligible windows in the 2010–2020 history with these settings. The window with the smallest comparison p-value is selected, and that value is multiplied by 18, capped at one. This Bonferroni adjustment accounts for the search across possible locations and durations. The result then enters the correction across topics and test types.
The selected window must also contain at least 15 topic papers and have a rate at least 1.5 times each flank. A half-count stabilizes effect ratios, while the tests use unsmoothed counts. Ranking combines excess papers above the higher flank’s expected count with relative effect size.
Read the boundaries carefully
The first and last years cannot be evaluated as local peaks because they lack observations on one side. A final-year surge may be visible in the rise ranking, but this record cannot establish whether it subsequently subsides.
A displayed bump candidate can fail the statistical screen. Its chart still identifies a period worth reading more closely; it does not establish a historical episode or explain why the count changed.