Mètriques per a la detecció d'esdeveniments en app reviews

User reviews in app stores contain valuable user feedback useful for app developers and stakeholders. However, it is impractical to manually read this high-volume data as popular apps can receive thousands of reviews per day. This master's thesis proposes a machine learning pipeline for automat...

Descripción completa

Detalles Bibliográficos
Autor: Moonsammy, Leandra
Tipo de recurso: tesis de maestría
Fecha de publicación:2024
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/411655
Acceso en línea:https://hdl.handle.net/2117/411655
Access Level:acceso abierto
Palabra clave:Machine learning
Mobile apps
aplicaciones móviles
análisis de mercado
aprendizaje automático
reseñas de usuarios
detección de eventos
microblogging
mobile apps
market analysis
machine learning
user reviews
event detection
Aprenentatge automàtic
Aplicacions mòbils
Àrees temàtiques de la UPC::Informàtica::Sistemes d'informació
Descripción
Sumario:User reviews in app stores contain valuable user feedback useful for app developers and stakeholders. However, it is impractical to manually read this high-volume data as popular apps can receive thousands of reviews per day. This master's thesis proposes a machine learning pipeline for automatically detecting significant software-based events from user reviews of mobile apps. The pipeline involves preprocessing raw reviews, aggregating data weekly, extracting relevant metrics like review count, rating, polarity, and count and percentage of negative, neutral and positive reviews, as well as training machine learning classifiers on two manually annotated truth sets to identify events from 10 microblogging Android apps spanning one year. Each model was trained on two truth sets: reviews of all ratings and positively-rated reviews only. Out of four trained models: Multilayer Perceptron, Naive Bayes, Support Vector Machines and Random Forest, the Multilayer Perceptron and Random Forest models achieved high F1 scores of 97\% and 93\% on the all ratings truth set and 53\% and 87\% on the positive only truth set respectively. Through qualitative assessment via summarization, the detected events provided actionable insights for app developers about bugs, performance issues, feature requests, and user feedback. The findings demonstrate the effectiveness of leveraging machine learning techniques for extracting valuable information from large volumes of app reviews to support development efforts.