yao-jz / rating-prediction-dl

Some deep learning methods for rating prediction

Rating-Prediction-DL

All the deep learning models are implemented using jittor.

Introduction

I have implemented LSTM, CNN, and BERT models, and ensemble learning.

All models use the embedding layer of the pretrained bert provided by jittor to encode the words of the comment text, token_dim = 768

Result

model	Accracy	F1	MSE	RMSE	MAE
LSTM	0.5868	0.1479	1.777	1.333	0.7440
CNN_BATCH_64_FEATURE_256	0.662	0.4369	0.8730	0.9343	0.4746
CNN_BATCH_32_FEATURE_256	0.653	0.4465	0.8595	0.9271	0.4802
CNN_BATCH_2_FEATURE_256	0.6447	0.4657	0.8365	0.9146	0.4801
CNN_BATCH_2_FEATURE_128	0.647	0.4654	0.8600	0.9274	0.4836
CNN_BATCH_2_FEATURE_64	0.6488	0.4566	0.8716	0.9336	0.4855
CNN_BATCH_2_FEATURE_256_ENSEMBLE	0.6301	0.317	1.111	1.054	0.559
BERT-base	0.6976	0.5339	0.5709	0.7556	0.3743

Overall, the bert model with the most parameters and using attentions performed best on this dataset and outperformed the other models in various metrics.