However, if for another prefix a false positive occurs in the first, but not in the second Bloom filter, or if there are plenty of values stored in the searched hash table buckets, then this case of matching will likely be slower than the former.
All the intermediate key–value pairs are sorted and hashed into buckets.
This means that we can perform an initial bucketing without inverting the hash functions; the buckets can then be processed independently, using multiple threads if required.
If the number of l-mers hash to the same bucket exceeds a threshold, they are considered as good candidates for motifs.
This can be realized using approximate nearest neighbor search (ANNS) methods such as locality sensitive hashing (LSH) [2], fast library for approximate nearest neighbors (FLANN) [3], and bucket distance hashing (BDH) [4].
For each table considered, it compares q with the sub-fingerprints that are hashed into the same bucket as q.
Two 'similar' points in the original space can be hashed into a same bucket with high probability, which makes LSH appropriate to perform indexing in the retrieval task.
In each table, the segments hashed into the same bucket as the forecast segment are returned, and the (N_mathrm{s}) more similar ones, which have a smaller distance from the forecast segment, are selected to be used in local forecasting.
where k is the width parameter, and {h j,j=1,…k} are LSH functions satisfying the LSH property, i.e., sub-fingerprints that are close to each other have a higher probability to be hashed into the same bucket than sub-fingerprints that are far apart.
