A Structured Approach to Understanding Recovery and Relapse in AA

This repository contains the code for the paper:

"A Structured Approach to Understanding Recovery and Relapse in AA." Yue Zhang, Arti Ramesh, Jennifer Golbeck, Dhanya Sridhar and Lise Getoor. WWW 2018.

This code includes two parts. 'aa-feature' includes code for extracting linguistic features, psycho-linguistic features, structural features from collected data set. And these features are used as input for PSL models.

Linguistic Features: include Term Frequency, Alcohol/Sober Word Usage, Topic Distribution from Seeded Topic Modeling, Sentiment Scores.

Psycho-linguistic Features: we extract psycho-linguistic features using LIWC[Linguistic Enquiry Word Count (LIWC). https://liwc.wpengine.com/]. Here we use affect and social categories.

Structural Features: Friends, Replies, Retweets, Similarity.

'aa-feature/recover3' includes code for extracting features in 3 months period.

'aa-feature/recover-1year' includes code for extracting features in 1 year period.

'PSL-aa' includes variations HL-MRF recovery prediction models. All these models are based on the Probabilistic Soft Logic framework[http://psl.linqs.org/].

PSL-Recovery-All: uses all the features.

PSL-Linguistic: only uses linguistic features drawn from AA users’ tweets and dependencies among them.

PSL-Relational: Combines structural features, but excludes psycho-linguistic features from LIWC.

PSL-Topic: Combines structural and topic features without sentiment; and combines structural and linguistic on friends’ tweets, but excludes psycho-linguistic features from LIWC.

PSL-Psychological(Affect): combines structural features with affect.

PSL-Psychological(Social): Combines structural features with social.

PSL-Sentiment: Combines structural and topic features with sentiment.

PSL-LIWCSimilarity: Uses collective Rules with LIWCSimilarity.