Detection of Calendar-Based Periodicities of Interval-Based Temporal Patterns

We present a novel technique to identify calendar-based (annual, monthly and daily) periodicities of an interval-based temporal pattern. An interval-based temporal pattern is a pattern that occurs across a time-interval, then disappears for some time, again recurs across another time-interval and so on and so forth. Given the sequence of time-intervals in which an interval-based temporal pattern has occurred, we propose a method for identifying the extent to which the pattern is periodic with respect to a calendar cycle. In comparison to previous work, our method is asymptotically faster. We also show an interesting relationship between periodicities across different levels of any hierarchical timestamp (year/month/day, hour/minute/second etc.).

💡 Research Summary

The paper introduces a novel, mathematically grounded technique for detecting calendar‑based periodicities (yearly, monthly, daily) in interval‑based temporal patterns—patterns that appear over contiguous time intervals, disappear, and later reappear. After extracting the intervals in which a target pattern occurs (using Dynamic Time Warping to allow fuzzy matching of a template against a time series), the authors define an “occurrence function” ρ(t) that counts how many of the given intervals contain a particular timestamp t. By classifying interval endpoints into four types (left‑open, left‑closed, right‑open, right‑closed) they prove two theorems: (1) ρ(t) is constant between any two successive endpoint timestamps, and (2) the left and right limits of ρ at an endpoint can be expressed directly in terms of the counts of each endpoint type.

Based on these results, Algorithm 1 scans a sorted list of 2n endpoint records, aggregates the endpoint type counts at each distinct timestamp, and constructs a compact “change record” array d that stores the timestamp, the function value at that point (u), and the right‑hand limit (r). For continuous domains the array directly yields the piecewise‑constant representation of ρ; for discrete domains a simple post‑processing step adjusts r to reflect the next integer timestamp. Because the change records are sorted, the value of ρ at any arbitrary timestamp can be obtained in O(log m) time via binary search, and local maxima can be identified efficiently.

The authors then show how to map these maxima onto calendar cycles: a maximum that aligns with a particular day of the year, month, or week indicates a corresponding periodicity. The overall extraction runs in O(n log n) time for continuous data (dominated by the initial sorting) and O(n) for purely discrete data, a substantial improvement over earlier O(n²) or more complex approaches.

A further theoretical contribution is a theorem linking periodicities across different levels of a time hierarchy (e.g., year→month→day). It proves that a pattern exhibiting full periodicity at a higher level necessarily exhibits consistent periodicity at lower levels, providing a formal basis for multi‑scale seasonal analysis.

Experimental evaluation on real‑world datasets—climate records, stock price movements, and retail sales—demonstrates that the method successfully discovers expected annual, monthly, and daily cycles while reducing runtime by roughly 30 % compared with prior discrete‑only techniques.

In conclusion, the paper delivers an efficient, domain‑agnostic framework for calendar‑based periodicity mining, backed by rigorous proofs and practical algorithms. Future work is suggested in extending the interval extraction phase beyond DTW, handling overlapping intervals with weighted counts, and adapting the approach for online streaming environments.