Please share with the community what you think needs improvement with Microsoft Azure Machine Learning Studio.
What are its weaknesses? What would you like to see changed in a future version?
It's the first software that I've used in terms of machine learning. Therefore, I don't have anything to compare it to, however, it was okay for me. I didn't have any problems or anything. Maybe it can be integrated with something else. For example, business analytics. That way, you could also give creative reports. It's possible it could be integrated with the Power BI, as it's also Microsoft. That said, I'm not really sure. It if isn't possible, it's something they could consider for a future release. Microsoft needs to be sure to monitor the security and ensure they are constantly updating it. It would be nice if the product offered more accessibility in general.
Every tool requires some improvement. They have already improved many things. They had added new features and a new pipeline. They should have an on-premise version, other than Python and R Studio, which is only good for cloud-based deployments. If they could have a copy of the on-premise version on Mac or Linux or Windows, it would be helpful. It should have the flexibility to work o the desktop. They should have a desktop version to work on the platform.
We've found that the solution runs at a high cost. It's not cheap to utilize it. Two additional items I would like to see added in future versions are software life cycle features and more security capabilities. There should be data access security, a role level security. Right now, they don't offer this.
In terms of improvement, I'd like to have more ability to understand the detailed impact of the variables on the model and their interactions. Their algorithms are very powerful and they explain overall the net contribution of each of the variables to the solution. In terms of being able to say to people "If you did this, you'll get this much more improvement" Azure (at least my understanding of it) doesn't provide readily accessible tools to assess from a management perspective the impact of their changing a sinimized, the better.gle value - for instance in closing a lead, decreasing response time by 10%. I recognize that the multivariate algorithms used from decision trees to neural nets do not readily provide the coefficients for each variable ala the older regression modeling approaches. My experience over my 50 years of developing and implementing predictive models has been that more than half the value of modeling lies in improving management's understanding of the process being modeled, often leading to major organization and operational structure changes. More ability to understand the variables impacting the end result being optimized would be very useful.
I really can't see where it needs much improvement. My experience is only half-matured and is still maturing. I don't think we have reached the stage where the customer has enough cohesion to really complain about anything. Also, a Microsoft team is personally involved which really simplifies the process. In the machine learning world, when you are defining the model, typically people go for an interesting library of algorithms that are available. It's an imperfect scenario. The world is not as ideal as we think: how we draw a mathematical or theoretical formula is not exactly as it seems. With encryption, this uncertainty is actually much higher — that's why you need to tweak your mathematical formula or completely customize it. For this reason, my team has a development platform where they can customize code when it fails.
The data preparation capabilities need to be improved. Using this product, I can not prepare the data very much and this is a bottleneck in machine learning. There are some features that are not supported, so I have to use either Python or R to accomplish these tasks.
The AutoML feature is very basic and they should improve it by using a more robust algorithm. It lacks deep learning type algorithms but works great for the basic classification and regression models.
I used Azure Machine Learning in a free trial and I had a complete preview of the service. A problem that I encountered was that I had a model that I wanted to deploy and use on Azure Machine Learning, but there wasn't any option that that model can be used in the designer. I didn't find any option to upload my model, so that I can create my own block and use it in Azure Machine Learning designer. I believe this is a problem because sometimes you have your model created on some other device and you just have a file that you think can be uploaded to Azure Machine Learning and can be tested through a simple drag and drop tool.
The solution should be more customizable. There should be more algorithms. The solution needs more functionality.
On the customer side, the solution should do more to push companion marketing. When you use different Microsoft tools, there are different pricing metrics. It doesn't make sense. The pricing metrics are quire difficult to understand and should be either clarified or simplified. It would help us sell the solution to customers. The solution should simplify switching between platforms in the studio.
The data cleaning functionality is something that could be better and needs to be improved. There should be special pricing for developers so that they can learn this solution without paying full price.
Some of the terminologies, or the way that the questions are asked, could be stronger. When people use local colloquialisms, it would be better if it understood rather than forwarding it to an agent. If the frontline efficiencies were improved then we could pass this on to our clients. Integration with social media would be a valuable enhancement.
If you want to be able to deploy your tools outside of Microsoft Azure, this is not the best choice. One of the problems that we had was that you could only execute the model inside the machine learning environment. Comparing this to Databricks, if you create a pipeline, it could be in a notebook and you have all the code and then you can export your notebook to some other tool directly, for example in Jupyter and Spark. If you change tools then you won't lose your assets. I would like to see improvements to make this solution more user-friendly. They need to have some tools, like Apache Airflow, for helping to build workflows. Better tools are needed to bring the data from existing storage into the environment where they can play with it and start to analyze what they already have, on-site. This is what the majority of people would like to do. A feature that would be useful is to have some standard data transportation functions. They have ADF, Azure Data Factory, but it's a little bit heavy to manipulate. If they could have something more user-friendly, like Apache Airflow, it would be very nice.
Operability with R could be improved.
We all know it's really hard to get good pricing and cost information.
Please share what you can so you can help your peers.
Let the community know what you think. Share your opinions now!