GraphJet is a real-time graph processing library written in Java that maintains a full graph index over a sliding time window in memory on a single server. This index supports a variety of graph algorithms including personalized recommendation algorithms based on collaborative filtering. These algorithms power a variety of real-time recommendation services within Twitter, notably content (tweets/URLs) recommendations that require collaborative filtering over a heterogeneous, rapidly evolving graph.
GraphJet is able to support rapid ingestion of edges in an evolving graph while concurrently serving lookup queries through a combination of compact edge encoding and a dynamic memory allocation scheme. Each GraphJet server can ingest up to one million graph edges per second, and in steady state, computes up to 500 recommendations per second, which translates into several million edge read operations per second. More information about the internals of GraphJet can be found in the VLDB'16 paper.
After cloning the repo, build as follows (for the impatient, use option -DskipTests to skip tests):
$ mvn package install
GraphJet includes a demo that reads from the Twitter public sample stream using the Twitter4j library and maintains two separate in-memory bipartite graphs:
- A bipartite graph of user-tweet interactions. The left-hand side vertices represent users, the right-hand side vertices represent tweets, and the edges represent tweet posts and retweets.
- A bipartite graph of tweet-hashtag contents. The left-hand side vertices represent tweets, the right-hand side vertices represent hashtags, and the edges represent content association (e.g., a tweet contains a hashtag).
To run the demo, create a file called twitter4j.properties in the GraphJet base directory with your Twitter credentials (replace xxxx with actual credentials):
oauth.consumerKey=xxxx
oauth.consumerSecret=xxxx
oauth.accessToken=xxxx
oauth.accessTokenSecret=xxxx
For obtaining the credentials, see documentation on obtaining Twitter OAuth tokens. The public sample stream is available to registered users, see documentation about Twitter streaming APIs for more details.
Once you've built GraphJet, start the demo as follows:
$ mvn exec:java -pl graphjet-demo -Dexec.mainClass=com.twitter.graphjet.demo.TwitterStreamReader
Once the demo starts up, it begins ingesting the Twitter public sample stream. The program will print out a sequence of status messages indicating the internal state of the user-tweet graph and the tweet-hashtag graph.
You can interact with the graph via a REST API, running on port 8888 by default; use -Dexec.args="-port xxxx" to specify a different port.
The following calls are available to query the state of the in-memory bipartite graph of user-tweet interactions:
userTweetGraph/topTweets: queries for the top tweets in terms of interactions (retweets). Use parameterkto specify number of results to return (default ten). Sample invocation:
curl http://localhost:8888/userTweetGraph/topTweets?k=5
userTweetGraph/topUsers: queries for the top users in terms of interactions (retweets). Use parameterkto specify number of results to return (default ten). Sample invocation:
curl http://localhost:8888/userTweetGraph/topUsers?k=5
userTweetGraphEdges/tweets: queries for the edges incident to a particular tweet in the user-tweet graph, i.e., users who have interacted with the tweet. Use parameteridto specify tweetId (e.g., fromuserTweetGraph/topTweetsabove). Sample invocation:
curl http://localhost:8888/userTweetGraphEdges/tweets?id=xxx
userTweetGraphEdges/users: queries for the edges incident to a particular user in the user-tweet graph, i.e., tweets the user interacted with. Use parameteridto specify userId (e.g., fromuserTweetGraph/topUsersabove). Sample invocation:
curl http://localhost:8888/userTweetGraphEdges/users?id=xxx
The following calls are available to query the state of the in-memory bipartite graph of tweet-hashtag contents:
tweetHashtagGraph/topTweets: queries for the top tweets in terms of hashtags. Use parameterkto specify number of results to return (default ten). Sample invocation:
curl http://localhost:8888/tweetHashtagGraph/topTweets?k=5
tweetHashtagGraph/topHashtags: queries for the top hashtags in terms of tweets. Use parameterkto specify number of results to return (default ten). Sample invocation:
curl http://localhost:8888/tweetHashtagGraph/topHashtags?k=5
tweetHashtagGraphEdges/tweets: queries for the edges incident to a particular tweet in the tweet-hashtag graph, i.e., hashtags contained in the tweet. Use parameteridto specify tweetId (e.g., fromtweetHashtagGraph/topTweetsabove). Sample invocation:
curl http://localhost:8888/tweetHashtagGraphEdges/tweets?id=xxx
tweetHashtagGraphEdges/hashtags: queries for the edges incident to a particular hashtag in the tweet-hashtag graph, i.e., tweets the given hashtag is contained in. Use parameteridto specify hashtagId (e.g., fromtweetHashtagGraph/topHashtagsabove). Sample invocation:
curl http://localhost:8888/tweetHashtagGraphEdges/hashtags?id=xxx
The demo program illustrates collaborative filtering via similarity queries on the tweet-hashtag graph. Note that the demo does not offer personalized recommendation algorithms on the user-tweet graph (as is deployed inside Twitter) because the public sample stream API is too sparse in terms of interactions to give good results. The following endpoint for similarity queries offers related hashtags given an input hashtag:
similarHashtags: computes similar hashtag to the input hashtag based on real time data. Use parameterhashtagto specify hashtag (e.g., fromtweetHashtagGraph/topHashtagsabove). Sample invocation:
curl http://localhost:8888/similarHashtags?hashtag=trump&k=10
Copyright 2016 Twitter, Inc.
Licensed under the Apache License, Version 2.0