This is a simple implementation of the paper "TopicSketch: Real-Time Bursty Topic Detection from Twitter" http://ieeexplore.ieee.org/document/7457245/?reload=true&arnumber=7457245&filter%3DAND(p_IS_Number:7505473)

This project is mainly coded using python, while for the reason of efficiency, some core parts are implemented using C and Cython.

The data flow is as follows. (See main.py.)

stream ==> preprocessor ==> detection component ==> topicsketch ==> detected topic

In order to run TopicSketch, a stream class is needed. It should extend topicsketch.stream.ItemStream . Everytime, TopicSketch will call stream.next() to get the next tweet. Each tweet is represented as a tuple (timestamp, user_id, tweet_text). experiment.tweet_stream is an example, which get tweets from a MySQL database.

Notice: make sure the tweet stream in the order of time !!!

In this project, I use twokenize (https://github.com/myleott/ark-twokenize-py) for tokenization. For different languages, other tokenizer may be used.

For hashing

hash.c is an implementation of the multi-linear family hash function.

cd c; gcc -fPIC -shared -o mlh.so hash.c

For acceleration and significance score

acceleration is defined in the above paper "TopicSketch".

significance score is defined in the paper "SigniTrend: scalable detection of emerging topics in textual streams by hashed significance thresholds".

install Cython first: pip install cython

cd cython

cd fast_smoother; python setup.py build_ext --inplace

cd fast_signi; python setup.py build_ext --inplace

After compiling, copy the libraries into folder TopicSketch.

Put stop words in twitter-stopwords.txt. Define your own stop words file according to your data set.

Set parameters in file "parameters.cnf". At least set start_t and end_t under section [detection].

It is suggested to use a machine which has enough memory (>=16GB).

It is suggested to set high threshold for reliable results.

It is suggested to remove spam accounts from the data set before running the detection algorithm.

Update scipy to the latest version for best computational performance.

Once you implement your stream class, replace tweet_stream in main.py by your stream. Then run main.py. Debugging information will appear in the CLI. Detected topics will be saved in file results.txt.