Search Tweets by Deep Reinforcement Learning

Search tweets referring to news, by Deep Reinforcement Learning

Tweet Search Problem

Search Process

We model search process by Markov Decision Process (MDP) in Reinforcement Learning.
Please refer to some open materials.

In MDP-based search process, we define

Immediate Reward:
Number of relevant tweets in current search results of tweets.
(Note that search results are returned by one-time query to a set of tweets)

State:
Difference between current search results (tweets) and previous search results (tweets).
E.g., one difference value is the difference value between textual contents, modeled by Word2Vec model.

Reinforcement Learning
We learn by Q-learning.
The loss function for learning:

R: immediate reward
Q(): predicted long-term return
s_t: state at time t
a_t: action (query strategy) used at time t

Model

Predict Long-term returns of search process by RNN Model

We use LSTM layer and FC layer

Search Results