Building an AutoML Application

Machine learning (ML) has been developed and growing fast. It also depends on an expert in ML to develop the ML model, i.e. pre-processing, building model, feature engineering, tuning of hyper-parameters, etc. Data scientists have many demanding responsibilities, and challenges. These challenges include the selection of the best algorithm, modification of this algorithm over many iterations, and identification and tuning of the hyper-parameters many times. Despite going through difficult processes, it is worth noting to watch the ML model beat expected accuracy. There is the myth that the developers love to overdo the coding. Many steps that are performed by data scientists are repeatable and time-consuming which makes these steps ideal candidates for automation [1].

Figure 1

AutoML allows users with little knowledge to generate multiple ML models with a few clicks. By using AutoML, data scientist experts can do away with the pain points of developing the model and get to the core of problem-solving. For these reasons, AutoML has become popular recently.

Machine Learning 

ML is used for solving different tasks. ML algorithms are used for spam filtering, network intruders deduction, and computer vision.

There are two types of ML: classical and AutoML. In classical ML, the data scientists do some feature engineering, i.e. generating the new aspects from the data, selecting vital aspects, and then the ML algorithm is trained on the train set and is evaluated on the test sets. In classical ML, results are not accurate. The classical ML is demonstrated below in Figure 2 [3].

Figure 2

AutoML is quite different from classical ML. In this automatic approach, there is no need for a human. A program makes the feature engineering and the model is automated. AutoML is the open domain that the same algorithm can implement on different types of data. AutoML saves money and human resources as the solution is scalable to many other tasks. A single algorithm can be used for any task. Although humans can get more accurate results, it takes more time and specialist analysis. AutoML is demonstrated below in Figure 3 [3].

Figure 3

AutoML

AutoML makes the developers have less expertise in ML to train the high-quality models that are specific to the business type. It builds custom ML models fast. AutoML reduces the barriers to entry for ML model building by the process of streamlining, thereby allowing the non-technical user to harness the ML power. The availability of the AutoML also helps the data scientists by sparing their time to explore the other data areas. In short, the users can provide input to the AutoML system. This system is used for model building, i.e. feature transformations, feature selections, hyper-parameters optimizations, etc., and then it returns predictions as output. It would be a great task if one built their own AutoML app that they can custom tailor to the content of the heart. This AutoML app development is twofold, with model deployment that helps to complete the data life cycle and AutoML that helps to make the ML accessible to non-technical users [2].

Building of AutoML

The process of building ML models and then choosing the best model is a long process. It usually takes a number of lines of code and it takes time to complete. Data science and ML are generally associated with the process of automation. Below, Figure 4 shows the ML workflow. AutoML covers only the pre-processing of data, choosing a model, and hyper-parameter-tuning. The users still have to understand the main goals and explore the dataset to prepare the data.

Figure 4

Which ML Processes to Automate?

AutoML aims to automate the steps of the ML process that include [4]:

• Data pre-processing: This process improves the quality of data and converts the raw data to a structured and refined format by cleaning, transforming, integrating, and reducing data.
• Features engineering: AutoML creates a feature that is more compatible with the ML algorithm by the analysis of the input data.
• Features extraction: In this process, there is a combination of the different features or the datasets for generating the new aspects that give better results and reduces the data size that is being processed.
• Features selection: AutoML automates the selection of only features for processing that have benefits.
• Algorithm selections and hyper-parameter optimizations: AutoML tools choose the optimal hyper-parameters and algorithms.

Automated systems can refine and tune the data, algorithms, features, and hyper-parameters of the algorithms for generating accurate models relying on the established ML knowledge and trial-and-error. 

Benefits of AutoML

Some benefits of AutoML are given below [5]:

• The cost is reduced.
• The productivity of data scientists is increased.
• The demand for data scientists is reduced by the democratization of ML.
• It increases revenues and the satisfaction of customers.
• More models roll out with accuracy can improve other and less tangible business results. 

Limitations and Conclusions

AutoML is still being researched. The recent approaches of AutoML include the solving of tasks, i.e. classifications and regressions, and it can configure neural networks. However, AutoML has its own limitations:

• Some problems include the fact that semi-supervised/unsupervised learning and reinforcement learning are not been tackled by the AutoML community yet.
• AutoML algorithms rely on the data that is relevant. Data cleaning and features engineering are not well supported by AutoML.
• AutoML process takes time to give the best solution.

ML and artificial intelligence (AI) have become more accessible with time. Meanwhile, some high-level libraries like Keras hide the underlying complexity of the deep learning model. The AutoML approach takes one more step and provides more feasible ML models from the inputs. This smooth pathway is provided by the ML for the users that have little knowledge. AutoML provides a production-ready model for small startups.

This doesn’t mean that AutoML is only for non-experts. AutoML also doesn’t mean that there will be no need for an ML expert for collecting, pre-processing, ingesting, cleaning, monitoring, and evaluating data are the parts of the ML pipelines. In short, for ML to be available to non-experts the advancements in Auto ML are done with time [6].   

References

1. Build machine learning models with and without AutoML (Samaya Madhavan, Austin Eovito, 2020)
2. How to Build an AutoML App in Python (Nantasenamat, 2021)
3. How to build AutoML from scratch (Mamaev, 2019)
4. Automated Machine Learning for Supervised Learning (Part 1) (Rendyk, 2021)
5. AutoML: In-depth Guide to Automated Machine Learning [2022] (Dilmegani, 2022)
6. Introduction to Automated Machine Learning (AutoML) (Simsek, 2019)