Certifying similarity retrieval for incomplete sensor time series
New methods provide guarantees for nearest neighbour searches when sensor data has missing values.
Sensor time series often reach similarity search with missing values. The retrieved neighbours depend on how the gaps were filled.
Researchers provide a pairwise certificate that one candidate is strictly closer than another. They also give a guaranteed candidate set that contains every possible nearest neighbour.
A certificate can show that a nearest-neighbour label cannot change. This applies to the true query whenever it lies inside a specified box.
The work uses a family of window-minimum distances. This family contains the distance as its separable end.
On three traffic-sensor collections, linear interpolation changes the nearest neighbour in 15 to 33 percent of the queries. The fast layer resolves about 90 percent of all tested boundary pairs without dynamic programming.
The exact method is 82 to 110 times faster than the programme alone on paired network-scale queries. Band width trades forecasting accuracy and certification coverage for lower imputation sensitivity.
Key facts
- Linear interpolation changes the nearest neighbour in 15 to 33 percent of queries on three traffic-sensor collections.
- The exact method is 82 to 110 times faster than the programme alone for paired network-scale queries.
- The fast layer resolves about 90 percent of all tested boundary pairs without dynamic programming.