A major change in this maintenance release is the support for TF-IDF computation.

• Major changes

  • Added a support for TF-IDF computation [876b335]
  • Changed the default eta0 (initial learning rate) to 0.2 [aa92317]
  • Added a DDL to create hivemall UDFs as permanent functions [5cc77c9]

• Minor changes

  • Added a server throughput monitoring functionality using JMX [744bc8d]
  • Fixed MixServerTest to pass a TravisCI's build

A major change in this maintenance release is the support for AdaGrad/AdaGradRDA/AdaDelta.

For the usage, see the following examples:

• https://github.com/myui/hivemall/wiki/news20-binary-classification-AdaGradRDA-AdaGrad-AdaDelta
• https://github.com/myui/hivemall/wiki/KDDCup-2012-track-2-CTR-prediction-AdaGrad-AdaDelta

• Enhancement

  • Supported AdaGrad/AdaDelta
  • Supported AdaGradRDA classification

• Major changes

  • logress_iter() is no more supported [ad10c33]

• Minor changes

  • Fixed not to use String#split because of its performance [9e0a95c]

We are pleased to announce the initial beta release (beta1) of Hivemall v0.3.

The major enhancement in this release is the support for model mixing.
See this page to know how to use the new feature.

You can find a brief explanation of the internal design of MIX protocol in this slide.

Note this is a pre-release and for alpha users. Any feedback is welcome!

The major changes in this release is the support for a dense model using "-densemodel" option.
When feature dimension is large (greater than 2^24), SpaceEfficientDenseModel is used.

Note this release contains destructive changes that "-fh" and "-b" options are removed.
Implicit feature hashing and bias clause through training options are no more supported.
Use explicit mhash() or add_bias() instead of them.

• Major Changes

  • Added a support for dense model and removed bias option [84e1e0b]
  • Added SpaceEfficientDenseModel [a9076b5]

• Minor Changes

  • Changed the default value of confidence parameter (-phi) [98077b4]
  • The default bias feature dimension is changed to "0" [dc46f73]
  • Fixed mhash() and sha() to return a hashed value starting from 1 [0b0ca68]
  • Added function aliases [3c7c868]
  • Removed output_touched option [bf99be3]
  • Added a support to use a fixed seed in rand_amplify() [731275c]

A major change in this maintenance release is argmin_kld().
See the usage in this example.

• Enhancement

  • Added extract_weight(string featureVectors)::weights UDF [88af7b4]
  • Added "-outputs_touched" option for model loading [4cd59f2]
  • Add a function lr_datagen() used to generate regression datasets [273c486, cdbebbd]
  • Added generate_series() UDTF [5b21017]
  • Added normalize() UDF [707a16f]

• Major Changes

  • Added argmin_kld(mean, covar) to replace voted_avg(mean) [0731913]

• Minor Changes

  • Modified to reuse a decompressor [5c9c2ff]
  • Added a signature addBias(array features, int bias>) [2fae08c]
  • Reduced object allocations in rand_amplify() [753e123]

• Bugfix

  • Fixed a bug in distcache_get() [946fe10]
  • Fixed a bug in logress_iter() [667c3fd]

Important: Hive v0.11 or later is required (or recommended) from this release.

A major enhancement for this pre-release is the support for iterative training using distributed cache.
Find examples in our wiki pages [1] [2].

Also, the returning value of cw/arow/scw is changed to return covariance as well.
See examples (e.g., this one) for details.

• Enhancement

  • Supported a feature to load a prediction model ("-loadmodel" option) [e7ae2f1, e63a9fc]
  • Added rowid() and taskid() UDF [f0a688b]
  • Added distcache_gets UDF [3810923] [PR#64, PR#65]
  • Added extract_feature UDF [a984536]

• Major Changes

  • Modified to return covarinace in cw/arow/scw [e7ae2f1]
  • Added aliases for UDFs [8ee0586]
  • Changed hive version dependency to 0.11 or later [a856753, 96d7593]

• Minor Changes

  • Fixed UDFs to return Hive values instead of Java primitive values [71a3415]
  • Fixed '==' to 'equals' for HivemallConstants [c9c0291, 8d34ce5]
  • Fixed mhash UDF for arrays inputs [fa8efcb]

• Bugfix

  • Fixed a bug in multiclass PA [46da974]

• Enhancement
  • kNN search using Minhash
  • kNN search using b-Bit Minhash
  • kNN search using cosine similarity
  • Introduced a recommendation scheme using Minhash
  • Modified to use OpenHashTable instead of HashMap for reducing memory consumption [9698972]
• Bugfix
  • Fixed a bug in rand_amplify() [0962c07]

• Enhancement
  • Added rand_amplify() [538d866, 23b4f3b]
  • Added EMR bootstrap for hivemall [2a834e8]
  • Added AROW with a hinge loss [ae5c100]
• Bugfix
  • Fixed a bug in maven pom.xml [0d33ed7]

• Enhancement

  • Added a function prefixed_hash_values() [0a6ffb3]

• Bugfix

  • Fixed recursion in OnlineVariance#mean() [c5b8c5b]
  • Fixed score calculation w.r.t bias values in predict() [f895269, 0130cfe]

• Enhancement

  • Add new classifiers (Confidence Weighted, AROW, Soft Confidence Weighted)

• Bugfix

  • fixed a bug in PA1a and PA2a that stddev was not calculated correctly [75ccdd3]
  • fixed option handle for aggressive parameter C [095d939]
  • changed the default power_t from 0.25 to 0.1 [7081291]

• removed a dependency to serde.Constants for Hive 0.10 and later.
• fixed a serious bug in logress().