Tabular and Graph Automated Deep Learning (TGADL): An Open-Source Automated Deep Learning Package for Tabular and Graph Data

Tabular and Graph Automated Deep Learning (TGADL): An Open-Source Automated Deep Learning Python Package For Tabular and Graph Data 1st Nick Cliffel Computer Science and Engineering The Ohio State University Columbus Ohio, United States clif[email protected] 2nd Rajiv Ramnath Computer Science and Engineering The Ohio State University Columbus Ohio, United States [email protected] Abstract—As non-computer science researchers look to leverage deep learning to predict outcomes in a variety of domains, obtaining the specialized knowledge and experience required to design and train neural networks can be a daunting task. Automated Machine Learning (AutoML) approaches can allow non-computer science researchers to leverage machine learning without advanced statistical or machine learning knowledge. While automated deep learning frameworks already exist, many focus on vision-related tasks, where clear benchmarks have been established. Motivated by the plethora of real-world tabular and graph structured data that could be used to predict interesting phenomena, we designed and developed Tabular and Graph Automated Deep Learning (TGADL), a Python package with an easy-to-use interface focused on automated deep learning for tabular and graph datasets on regression and classification tasks. To automate the deep learning process, TGADL combines an evolutionary approach to neural architecture and hyperparameter search with feature engineering. TGADL works by generating an initial random model and then employing a hillclimbing procedure to optimize a given function. Changes in model architecture, hyperparameters, and feature engineering are created as neighbors and the effect on performance is explored in the evaluation stage of the hill-climbing procedure. TGADL has 4 major architecture components: Optimizer (HillClimbOptimizer), Model (MLPModel, SkipConnectionMLPModel, and GraphModel), Dataset (TabularDataset, GraphDataset), and Track, which handles interactions with the file system. This modular design allows for high cohesion and low coupling, making the TGADL framework easy to use and extend. To utilize TGADL a user defines a split function or picks from a list of defaults, creates a dataset instance, creates an optimizer instance, and calls the optimizer’s run method, requiring approximately 3 lines of code. A Model object and a dictionary of error metrics are returned. Using the Track instance automatically created by the optimizer, with 1 line of code, users can write models, standard scalers, and pca objects to file for later use or sharing. TGADL also supports optimizer options (HillClimbOptions) which more advanced users can use to control values such as evaluation accuracy, neighbor generation and runtime. As well as being easy to use, TGADL is highly extendable, for instance, to optimize an SVM, a user would just have to write an SVMModel class that implements the Model interface. To use another kind of optimization, such as a more traditional evolutionary approach, a user would just have to write a new Optimizer based on the functions defined in the Model interface. TGADL has been applied to a variety of tabular datasets including team and player sports predictions, generating stateof-the-art results, reducing mean absolute error for nfl margin of victory predictions from 11.522 points per game to 10.08 points per game, representing a 12.44% reduction in error. Index Terms—AutoML, Deep Learning, Neural Networks, Neural Architecture Search REFERENCES [1] M. Feurer, A. Klein, K. Eggensperger, J. Springenberg, M. Blum, F. Hutter, ”Efficient and Robust Automated Machine Learning”, Conference on Neural Information Processing Systems, 2015. [2] T. Elksen, J. H. Metzen, and F. Hutter, ”Simple and Efficient Architecture Search For Convolutional Neural Networks”, International Conference on Learning Representations, November 2017. [3] X. He, K. Zhao, and X. Chu, ”AutoML: A survey of the state-of-the-art”, Knowledge-Based Systems, vol. 212, January 2021.