ShortText-Fasttext

ShortText classification is an intractability problem in Machine Learning. Based on Facebook's Fasttext, this project optimizes its performance in ShortText and introduces a mechanism named category dropout. This mechanism works very well in Weibo's ads classification. Details about category dropout can read Blog。

Building

git clone https://github.com/AckClinkz/ShortText-Fasttext
cd ShortText-Fasttext
make

Usage

This section assumes that you are somewhat familiar with Fasttext. You can try to learn the basics of Fasttext from officical website.

Category Dropout only supports mode of supervised. Train model with category dropout, just set -cate_dropout.

In pratice, setting minCount can improve model performance. But it will lost infomation in low-frequency word. Settting vocabulary is an efficient method to balance them. If word in vocabulary, don't consider minCount.

Summarize the above points ShortText-Fasttext's usage as follows:

fasttext supervised \
  -input your_input.txt \
  -output your_model \
  -dim 120 \
  -lr 0.3 \
  -wordNgrams 3 \
  -minCount 30 \
  -bucket 10000000 \
  -epoch 100 \
  -thread 4 \
  -cate_dropout \
  -vocab ./your_vocabulary.txt

Updated 2018.11.6

According to 《Item2Vec: Neural Item Embedding for Collaborative Filtering 》, realize item2vec, usage as follow:

fasttext skipgram_item2vec \
  -input ./your_input.txt \
  -output "./your_model" \
  -dim 120 \
  -lr 0.01 \
  -wordNgrams 1 \
  -minCount 2 \
  -bucket 1000000 \
  -epoch 5 \
  -thread 8