# Analysis of Software Engineering Practices in General Software and Machine Learning Startups

**Authors:** Bishal Lakha, Kalyan Bhetwal, Nasir U. Eisty  
**arXiv:** [2304.01523](https://arxiv.org/abs/2304.01523)  
**Date:** 2023-04-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.01523v1/

## Key points

- Objective: Understand software engineering practices followed by machine learning (ML) startups and identify any additional needs
- Method: Conducted a systematic literature review on 37 papers published in the last 21 years, focusing on general software startups and ML startups
- Phases of software development life-cycle studied: requirement engineering, design, development, quality assurance, and deployment
- Database search performed in IEEE Xplore and ACM Digital Library using terms related to ML startups; alternate terminologies like "Deep learning" and "Artificial Intelligence" used to enrich the database for ML startups
- Metadata collected from database search including title, author names, abstracts, published year, URL citations, etc.
- Web scraper developed using BeautifulSoup library in Python to extract metadata from ACM Digital Library search results
- Deduplication performed using Pandas package in Python; duplicate papers dropped based on metadata; RegEx used to drop papers without required keywords in abstracts
- Snowballing techniques employed to increase number of papers for ML startups; relevant paper citations checked for additional relevant papers based on inclusion-exclusion criteria; added twenty-one papers to the database for ML startups
- Manual selection process conducted to finalize papers for analysis
- Total of 92 papers collected: 72 belonged to general software startups while 20 belonged to ML startups distributed among different phases of software development life cycle.

## Summary

The objective of this study was to understand the software engineering practices followed by machine learning (ML) startups and identify any additional needs. To achieve this goal, a systematic literature review was conducted on 37 papers published in the last 21 years, focusing on both general software startups and ML startups. The data collected aimed to understand software engineering practices in five phases of the software development life-cycle: requirement engineering, design, development, quality assurance, and deployment. Initially, a database search was performed in IEEE Xplore and ACM Digital Library using terms related to ML startups. As an alternative approach, alternate terminologies for machine learning such as "Deep learning" and "Artificial Intelligence" were used to enrich the database for ML startups. This resulted in a few additional useful papers. Metadata including title, author names, abstracts, published year, URL citations and other details were collected from the database search and stored in a CSV file. This allowed for deduplication and validation of papers as well as manual filtering based on abstracts. While IEEE Xplore had built-in features for exporting metadata as a CSV file; ACM Digital Library and Google Scholar did not have these features. Therefore; a web scraper was developed using BeautifulSoup library in Python to extract metadata from ACM Digital Library search results. After collecting metadata from various sources for both general software startups and ML startups; deduplication was performed using Pandas package in Python. Duplicate papers were dropped from the list based on their metadata; additionally RegEx (regular expressions) was used to identify papers where the required keywords were not present in the abstracts; these papers were also dropped from the list. To increase the number of papers for ML startups since there were very few initially found through database search alone; snowballing techniques were employed. Citations of relevant papers were checked to identify any additional relevant papers based on inclusion-exclusion criteria; this resulted in the addition of twenty-one papers to the database for ML startups. After de-duplication; validation; and snowballing; a manual selection process was conducted to finalize the papers for analysis. Out of 92 total papers collected 72 belonged to general software startups while 20 belonged to ML startups distributed among different phases of software development life cycle.