Sakib Abrar

Manar Samad, Sakib Abrar

Fully connected deep neural networks (DNN) often include redundant weights leading to overfitting and high memory requirements. Additionally, the performance of DNN is often challenged by traditional machine learning models in tabular data classification. In this paper, we propose periodical perturbations (prune and regrow) of DNN weights, especially at the self-supervised pre-training stage of deep autoencoders. The proposed weight perturbation strategy outperforms dropout learning in four out of six tabular data sets in downstream classification tasks. The L1 or L2 regularization of weights at the same pretraining stage results in inferior classification performance compared to dropout or our weight perturbation routine. Unlike dropout learning, the proposed weight perturbation routine additionally achieves 15% to 40% sparsity across six tabular data sets for the compression of deep pretrained models. Our experiments reveal that a pretrained deep autoencoder with weight perturbation or dropout can outperform traditional machine learning in tabular data classification when fully connected DNN fails miserably. However, traditional machine learning models appear superior to any deep models when a tabular data set contains uncorrelated variables. Therefore, the success of deep models can be attributed to the inevitable presence of correlated variables in real-world data sets.

Sakib Abrar, Ali Sekmen, Manar D. Samad

Deep learning methods in the literature are commonly benchmarked on image data sets, which may not be suitable or effective baselines for non-image tabular data. In this paper, we take a data-centric view to perform one of the first studies on deep embedding clustering of tabular data. Eight clustering and state-of-the-art embedding clustering methods proposed for image data sets are tested on seven tabular data sets. Our results reveal that a traditional clustering method ranks second out of eight methods and is superior to most deep embedding clustering baselines. Our observation aligns with the literature that conventional machine learning of tabular data is still a robust approach against deep learning. Therefore, state-of-the-art embedding clustering methods should consider data-centric customization of learning architectures to become competitive baselines for tabular data.

Manar D Samad, Sakib Abrar, Norou Diawara

Missing values in tabular data restrict the use and performance of machine learning, requiring the imputation of missing values. The most popular imputation algorithm is arguably multiple imputations using chains of equations (MICE), which estimates missing values from linear conditioning on observed values. This paper proposes methods to improve both the imputation accuracy of MICE and the classification accuracy of imputed data by replacing MICE's linear conditioning with ensemble learning and deep neural networks (DNN). The imputation accuracy is further improved by characterizing individual samples with cluster labels (CISCL) obtained from the training data. Our extensive analyses involving six tabular data sets, up to 80% missingness, and three missingness types (missing completely at random, missing at random, missing not at random) reveal that ensemble or deep learning within MICE is superior to the baseline MICE (b-MICE), both of which are consistently outperformed by CISCL. Results show that CISCL plus b-MICE outperforms b-MICE for all percentages and types of missingness. Our proposed DNN based MICE and gradient boosting MICE plus CISCL (GB-MICE-CISCL) outperform seven other baseline imputation algorithms in most experimental cases. The classification accuracy on the data imputed by GB-MICE is improved by proposed GB-MICE-CISCL imputed data across all missingness percentages. Results also reveal a shortcoming of the MICE framework at high missingness (>50%) and when the missing type is not random. This paper provides a generalized approach to identifying the best imputation model for a data set with a missingness percentage and type.