This repository contains experiments for comparing the estimation accuracy and running times of the following weighted minwise hashing algorithms:

• DartMinHash https://arxiv.org/abs/2005.11547
• ICWS and FastICWS https://research.google/pubs/pub36928/
• BagMinHash https://arxiv.org/abs/1802.03914

For BagMinHash and ICWS we use the implementation from here https://github.com/oertl/bagminhash with the relevant files included in the /bagminhash folder.

See the DartMinHash paper https://arxiv.org/abs/2005.11547 for a description of the algorithm and further results of experiments.

The BagMinHash algorithm uses XXHash64 which must be installed:

1. Get xxhash from https://github.com/Cyan4973/xxHash, e.g. using git clone https://github.com/Cyan4973/xxHash.git
2. Build using make lib
3. Place xxhash.h and libxxhash.a into the directory /bagminhash/xxhash

The code compiles under GCC version 7.5.0 https://gcc.gnu.org/ with relevant commands in the makefile https://www.gnu.org/software/make/.

make run compiles and executes the main function in main.cpp.

make test compiles and run unit tests.

The different experiments are all placed in the main.cpp and write their output to stdout in CSV format.

1. time_performance: Times different algorithms on synthetic data for all combinations of sketch lengths, and L0 and L1 norms chosen.
2. time_performance_specific: Same as above, but only runs on specified tuples of parameters.
3. measure_similarity: Returns the estimated Jaccard similarity of different algorithms on synthetic pairs of weighted sets with a specific similarity.

By default make run will run time_performance_specific on a subset of the settings used in Table 1 in the paper.

In order to pipe the output to the file data.csv use command make run > data.csv.

Notation:

• t denotes the sketch length (usually k in the paper).
• ICWS is a simple and unoptimized version of ICWS using tabulation hashing.
• ICWS_xxhash is the implementation from the BagMinHash repository which uses the ziggurat algorithm for fast sampling: https://en.wikipedia.org/wiki/Ziggurat_algorithm
• FastICWS is our own highly optimized implementation of ICWS that tabulates expensive operations and only computes the logarithms of weights once.
• BagMinHash1 and BagMinHash2: BagMinHash variants described in the BagMinHash paper. BagMinHash2 is essentially always faster and is what we compare against.
• DartMinHash: Optimized implementation following the pseudocode in the paper.

We use Catch2 https://github.com/catchorg/Catch2 for unit testing.

To compile and run tests use the command: make test