# Do users talk about the software in my product? Analyzing user reviews on IoT products

## TL;DR

This study explores user reviews of IoT products to identify software-related information, demonstrating that machine learning models can effectively classify such data to aid requirements engineering.

## Contribution

It provides an analysis of IoT product reviews for software insights and evaluates machine learning methods for automatic classification of review content.

## Key findings

- Reviews contain significant software-related information.
- SVM and CNN models achieve high precision and recall in classification.
- Machine learning can automate extraction of software insights from reviews.

## Abstract

Consumer product reviews are an invaluable source of data because they contain a wide range of information that could help requirement engineers to meet user needs. Recent studies have shown that tweets about software applications and reviews on App Stores contain useful information, which enable a more responsive software requirements elicitation. However, all of these studies' subjects are merely software applications. Information on system software, such as embedded software, operating systems, and firmware, are overlooked, unless reviews of a product using them are investigated. Challenges in investigating these reviews could come from the fact that there is a huge volume of data available, as well as the fact that reviews of such products are diverse in nature, meaning that they may contain information mostly on hardware components or broadly on the product as a whole. Motivated by these observations, we conduct an exploratory study using a dataset of 7198 review sentences from 6 Internet of Things (IoT) products. Our qualitative analysis demonstrates that a sufficient quantity of software related information exists in these reviews. In addition, we investigate the performance of two supervised machine learning techniques (Support Vector Machines and Convolutional Neural Networks) for classification of information contained in the reviews. Our results suggest that, with a certain setup, these two techniques can be used to classify the information automatically with high precision and recall.