TBB flow graph would make this a very flexible set of producers and consumers.

Switch to follow and jemalloc to replace std::string and std::vector. Might be faster.

Introduce string and vector types in the lmwtree namespace.

Use likely and unlikely where useful.

Sparsifying storage of bitmap lookups can make them more efficient.

See https://github.com/cmdevries/LMW-tree/blob/master/src/lmw/BitMapList8.h#L53 and https://github.com/cmdevries/LMW-tree/blob/master/src/lmw/BitMapList16.h#L55.

Also can make it a single class while we are at it.

Also compare performance before and after.

This can just be done by switching the ACCUMULATOR vector type to an atomic version. Note that whole vector operations do NOT need to be atomic, just updating a single dimension in the accumulator vector.

After indexing compress document vectors so they can be written out in a compressed binary format.

Delta encode and variable byte sparse vectors.

Use https://github.com/lemire/FastPFOR.

It would be ideal for the indexer to output integer valued document vectors with term frequencies. These can be optionally written to disk in a compressed format using https://github.com/lemire/FastPFOR to allow for easy experimentation with different representation approaches. It would also output term collection statistics.

This would allow quick processing to convert the vectors to weighted vectors with TF-IDF, BM25, etc, or conversion to signatures.

It would be great to do the same with bi-grams and invent or reuse a weighting scheme that uses pointwise mutual information (http://nlpwp.org/book/chap-ngrams.xhtml#chap-ngrams-bigrams) in the weighting calculation.

GCC 4.9 brings useful new features like template lambdas and declaring templates using auto. Refactor the code where it makes the code cleaner.

Such as SVM-light and Matlab formats.

https://code.google.com/p/semanticvectors/wiki/ReflectiveRandomIndexing

However, no binary vector version exists yet.

http://arma.sourceforge.net/

There is a TODO in the code for this.

Currently the CMake build is specific for GCC.

To make it cross platform it needs to use the find_package() function instead of manually setting GCC compile and link flags.

With a large enough mini-batch size, parallelism can be exploited within the mini-batch of say 1 million signatures.

Hopefully this will converge in 1 iteration or less for excessively large datasets like ClueWeb.

This also works in a distributed setting by simply batching and broadcasting updates to the tree as the parallel mini-batches proceed.

Fast randomization of the input signature file is also useful.

Unrolling the loops for Hamming distance may give further performance improvements by keeping all integer execution units busy inside newer processors. This basically unrolls the uint64_t chunk1, chunk2; result = chunk1 ^ chunk2; POPCNT result; operations inside the Hamming distance calculation.

Use compiler intrinsics to try different SIMD implementations of Hamming distance for both exlusive or and population count

change non-performance critical concepts to interfaces instead of concepts; i.e. virtual functions instead of templates

There is some memory overhead on vectors.

std::string is around 32 bytes and we use this for a vector ID, but vectors are usually 512 bytes at most. So make the ID a template parameter and switch to char* for strings.

Some lower level unit tests for vector types and other concepts would also be useful.

Compress transmission of integer accumulators between machines vectors using https://github.com/lemire/FastPFOR.

Hadoop + HDFS (just get hadoop to hand over the bytes, or use HDFS directly).

ZeroMQ + GlusterFS.

Apache Spark might work well with python bindings for library, https://github.com/apache/spark.

HDFS + Erlang scheduler (gascheduler) + C++ code as a simple TCP server.

The K-tree implementation was broken somewhere along the way. It needs to be fixed. Maybe it never worked completely in the first place.

Using smart pointers for cases where the pointer overhead is not critical such as tree nodes.

Switch to move semantics for vectors.

Use memory pools, custom allocators, and, allocation of nearby vectors in contiguous memory to reduce allocation overhead and improve locality of reference.

Create a set of standard test datasets and their expected quality in terms of internal and external measures. This can be used to test for any regressions modifications may introduce.

It also serves as documentation.

Get some non-document data sets for UCI.

TF-IDF
BM25 - probably quite useful with reflexive random indexing because it preserves the inner product space where BM25 works well
Log Likelihood from TopSig paper

Use Lance's idea for cyclic generation of random index vectors. Very cache friendly.