Review Rating Prediction
A variety of mplementations of models used for predicting review
ratings. They are linear regression, sparse linear regression,
basic topic model, supervised topic model.
Author: Anthonylife
Date: 12/27/2012
- Introduction =============== This project concentrates on review rating prediction. The data set is consisted of 5000 documents. The first 2500 documents are used as traning data while the second used as test data. Each document owns a rating value which is an iterger from 1-5.
Many models or methods can be used to solve this problem. In this programme, we focus on using regression model to predict the rating values. What matters most for us are the features how they are generated and used. Note multi-class classification is also OK here.
A variety of mplementations of models used for predicting review ratings. They are linear regression, sparse linear regression, basic topic model, supervised topic model.
-
How to Run all Related Models Two steps are OK. After launching the MATLAB software in the commander,
- excute "initpath.m" to add temporal path to the current workspace;
- excute "main.m" to run the code. "main.m" is a function contains all the start codes to run all the models. Users should modify the code to set the model selection option. More detailes can be seen in the file.
-
Additional library source files 1.IT++: C++ library of mathmatical computation, with similar coding style like MATLAB. URL: http://itpp.sourceforge.net/stable/
-
Code files layout dataset/ :: Raw data files train_review.dat --> train data file test_review.dat --> test data file ...
evaluation/ :: evaluation scripts predictiveR2.m --> implementation of evaluation metric
linearReg/ :: multiple linear regression codes gen_tfidf.py --> generates TF-IDF value as features lr.m --> linear regression with closed-form solution when the number of features is samll, e.g., topic model lrsgd.m --> linear regression with all words as features lrlasso.m --> linear regreesion with L1-norm regularization
ldaGibbs/ :: standard topic model with Gibbs Sampling, Matlab. ldaGibbs.m --> main code file all other files are packed in "private" directory...
ldaVariation/ :: standard topic model with Variational Bayesian, Matlab. ldaVariation.m --> main code file all other files are packed in "private" directory...
ldaGibbs++/ :: modification of an existed C++ implemention of Gibbs Sampling
sLDA/ :: supervised topic model with Variational Bayesian Learning slda.m --> main code file vbe_step.m --> variational bayesian E-step vbm_step.m --> variational bayesian M-step ... discLDA/ :: supervised topic model with Gibbs Sampling EM algorithm src/lda.cpp --> start code src/model.cpp --> main code
5.Predictive results see file "Result.txt" for more information.
Note, currently, I implemente Supervised topic model by matlab, basic LDA with Gibbs sampling by matlab. I also give a little modification for basic LDA with variational inference written by other users. Moreover, in order to speed up the procedure of Gibbs sampling, I also modified LDAGIBBS++, the C++ version written by Xuan-Hieu Phan. More details can be seen in the sub-directory in each method.