Abstract
Modeling time-series data with asynchronous, multi-cardinal, and uneven patterns presents several unique challenges that may impede convergence of supervised machine learning algorithms, or significantly increase resource requirements, thus rendering modeling efforts infeasible in resource-constrained environments. The authors propose two approaches to multi-class classification of asynchronous time-series data. In the first approach, they create a baseline by reducing the time-series data using a statistical approach and training a model based on gradient boosted trees. In the second approach, they implement a fully convolutional network (FCN) and train it on asynchronous data without any special feature engineering. Evaluation of results shows that FCN performs as well as the gradient boosting based on mean F1-score without computationally complex time-series feature engineering. This work has been applied in the prediction of customer attrition at a large retail automotive finance company.
TopBackground
Time-series sequences have been modeled with Hidden Markov Models (HMMs) (Rabiner, 1989; Ephraim & Merhav, 2002) and Bayesian Networks (BNs) (Heckerman et al., 1995; Nielsen et al., 2009). However, HMMs and BNs are not designed for asynchronous sequences because they require specification of a constant time interval between consecutive events (Wu et al., 2018). Asynchronous data needs to be reshaped and synchronized to fit HMMs and BNs. TS can be reshaped to synchronicity at the data preprocessing step. Reshaping and synchronization methods often obfuscate original data or create artificial data points that are not initially present in the dataset. This results in information loss and requires imputation techniques that incur significant computational and memory costs when datasets are large, and the degree of synchronicity is low.
Key Terms in this Chapter
Gradient Tree Boosting: A supervised machine learning algorithm that consists of an ensemble of decision trees.
Convolutional Neural Network: A type of neural network with an architecture that consists of kernels that learn to perform the matrix convolution operation on inputs to find patterns that have spatial proximity such as images or time-series.
Long-short Term Memory Network: A type of recurrent neural network where regular neurons are replaced by long-short term memory units designed to allow deep networks to “remember” inputs over larger number of steps and mitigate the exploding/vanishing gradient problem.
Neural Network: A supervised machine learning algorithm that searches for a function to fit existing data via an iterative training process. Neural Networks are characterized by multiple hidden layers that consist of neurons with activation functions that adjust weights using the backpropagation algorithm and a loss function.
Time-Series Data: Are data points spread across the time dimension. Time-series data is sequential and ordered based on timestamps. Examples include series of event occurrences and temperature measurements over time.
Gated Recurrent Unit: A modification of the long-short term memory unit with fewer parameters and no output gate.
Fully Convolutional Network: A type of convolutional network that does not contain any dense layers in its architecture. This type of network only contains layers that are specific to convolutional neural networks, such as convolution, pooling, and batch normalization.
Recurrent Neural Network: A neural network with an added dimension to represent the sequence or time component of sequential or temporal data.
Deep Learning: A subfield of machine learning that specializes in neural network based algorithms that have more than one hidden layer.