KEFE is an approach that exploits the information of app description and user reviews (written in Chinese) to identify key features that have a significant relationship with app rating scores.

The application of KEFE involves three main steps: 1) applying a textual pattern-based approach and a deep machine learning classifier to extract features from app description; 2) applying another classifier to match features with their relevant user reviews; and 3) applying regression analysis to identify key features.

More details of KEFE can be found in the following paper:

Huayao Wu, Wenjun Deng, Xintao Niu, and Changhai Nie. Identifying Key Features from App User Reviews. International Conference on Software Engineering (ICSE), pp. 922-932, 2021

KEFE is developed and tested using Python 3, pyltp and tensorflow. Please install the following packages of specific versions:

pip install pyltp=0.2.1
pip install tensorflow=1.15.0

More instructions for installing pyltp and tensorflow can be found in their respective websites: pyltp, tensorflow.

1. Download the model files, which include:

   • pyltp model files: ltp-model
   • pre-trained BERT model: chinese_L-12_H-768_A-12
   • classification model of feature extraction: model-extract
   • classification model of user review matching: model-match

   The ltp-model should be put into the pyltp-resource directory, and the other three should be put into the bert-master directory.

2. To extract feature-describing phrases from a given app description, run:

   python feature_extraction.py -i [app_description].csv
   # for example
   # python feature_extraction.py -i example/description.csv

3. To identify key features of a given app, run:

   python feature_identification.py -f [features].txt -r [reviews].txt
   # for example
   # python feature_identification.py -f example/alipay_features.txt -r example/alipay_reviews.txt

   The above command will first apply the classification model to match features and user reviews (this will take a long time if there is a large volume of user reviews), and then identify key features. If an existing file of matching between features and user reviews is available, run:

   python feature_identification.py -f [features].txt -m [matching].txt
   # for example
   # python feature_identification.py -f example/alipay_features.txt -m example/alipay_matching.txt

   Format of Files

   • The [reviews].txt file should be organised as the following format per line: [review_text]-*-[review_date]-*-[rating_score]
   • The [matching].txt file should be organised as the following format per line: [feature]-*-[review_text]-*-[review_date]-*-[rating_score]-*-[label], where label = 0 and 1 indicate non-matching and matching pairs, respectively.

Dataset (app descriptions and raw user reviews collected) and replication package can be downlowded from the following links:

• dataset
• replication package