Ease Of Creating Intelligent Systems With Azure Cloud
Makode, Sunny (2024)
Makode, Sunny
2024
Master's Programme in Computing Sciences
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2024-10-22
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202410149259
https://urn.fi/URN:NBN:fi:tuni-202410149259
Tiivistelmä
It is often a growing challenge to accommodate new technical skills from time to time in both professional and academic settings. It takes lots of effort to develop technical skills like machine learning which has its own learning curve, alongside it also needs some IT infrastructure in case to try out new things. This thesis addresses such issues using Azure cloud to provide an indication whether it's good to invest in cloud and leveraging it for machine learning/AI operations that could shorten the gap between skills development and making useful products. It also focuses on how much flexibility cloud-based solutions could offer when it comes to developing systems with machine learning. It also emphasized on checking if it is possible to make machine learning models through Azure ML services irrespective of knowledge background.
With machine learning, the conventional approach is to use some programming language like python with frameworks like Anaconda or Google Colab etc. On top of that, it also needs knowledge of certain concepts ranging from statistics, mathematics, feature/data engineering, algorithms, programming and so on. However, it can be compensated with the low code approach in Azure Cloud using Designer and Automated ML. The thesis also determines performance of models developed through Azure ML in comparison with model generated locally.
As part of the thesis, Azure ML architecture is studied. Designer pipeline and Automated ML provided by Azure ML is explored and experimented for predicting single family housing prices as well as electricity consumption per municipality in Finland. The housing and electricity data is collected/requested from various sources like Tilastokeskus Finland and Energiateollisuus ry. Through experimentation, it was found that Azure ML does not allow multi-target regression, so the regression operation is performed for each target value. On the other hand, making use of scikit learn library locally provided way to do multi-label/multi-target regression through MultiOut putRegressor. In terms of performance, Models developed with Automated ML performed better in comparison with Designer and locally developed models. With Automated ML, it was observed that the dependency on repetitive tasks like data cleaning, data/feature engineering, model selection and hyperparameter tuning are not required and thus in retrospect it saved time in comparison to Designer and conventional method of doing machine learning. Also, an anonymous survey was conducted to determine what people in academics and professional domain think about cloud usage. It was observed that most of the participant find flexibility and ease of using Azure cloud services due to its user friendliness and on-demand availability of services. On the contrary the challenging aspect of Azure cloud is reflected on Cost management and limited control over its services.
With machine learning, the conventional approach is to use some programming language like python with frameworks like Anaconda or Google Colab etc. On top of that, it also needs knowledge of certain concepts ranging from statistics, mathematics, feature/data engineering, algorithms, programming and so on. However, it can be compensated with the low code approach in Azure Cloud using Designer and Automated ML. The thesis also determines performance of models developed through Azure ML in comparison with model generated locally.
As part of the thesis, Azure ML architecture is studied. Designer pipeline and Automated ML provided by Azure ML is explored and experimented for predicting single family housing prices as well as electricity consumption per municipality in Finland. The housing and electricity data is collected/requested from various sources like Tilastokeskus Finland and Energiateollisuus ry. Through experimentation, it was found that Azure ML does not allow multi-target regression, so the regression operation is performed for each target value. On the other hand, making use of scikit learn library locally provided way to do multi-label/multi-target regression through MultiOut putRegressor. In terms of performance, Models developed with Automated ML performed better in comparison with Designer and locally developed models. With Automated ML, it was observed that the dependency on repetitive tasks like data cleaning, data/feature engineering, model selection and hyperparameter tuning are not required and thus in retrospect it saved time in comparison to Designer and conventional method of doing machine learning. Also, an anonymous survey was conducted to determine what people in academics and professional domain think about cloud usage. It was observed that most of the participant find flexibility and ease of using Azure cloud services due to its user friendliness and on-demand availability of services. On the contrary the challenging aspect of Azure cloud is reflected on Cost management and limited control over its services.