Page 382 - Special Topic Session (STS) - Volume 2

STS506 L. Leticia R. et al.
Recurrent Neural Networks (RNNs) have been successfully used in
prediction of time ordered data (see da Silva et al., 2017). Cho et al. 2014
defines a Recurrent Neural Network (RNN) as a neural network with a hidden
state that operates on a variable length sequence. Here, the hidden state is
updated using its current value and a new input sequence. If we segment the
sequence in fixed length subsequences, and feed them to the network, the
hidden state will dependent on the past observed sequence values. This
structure then allows training the stimuli on previous sequence values, which
is why these networks are referred as having memory.
We use neural networks to find an efficient forecaster for an outbreak
given the network of contacts and the infectious agent parameters. To train
the RNN, we generate random networks using the configuration model on
degree sequences sampled from probability distributions (1) Poisson, (2)
Power Law and (3) Polylogarithmic (Newman, 2002); and graph generation
models: (4) Watts- Strogatz and (5) Barabasi-Albert. These were selected with
parameters such that we originate diverse topological network features. On
each network, we simulate a SIR epidemic outbreak with parameters varying
in a range of values reported in diverse literature for influenza (Cori et al., 2012
and Biggerstaff et al., 2014). From the observed outbreak we recreate the
surveillance information, where the individuals entering state are aggregated
at regular time intervals.
We employ Gated Recurrent Unit (GRU) layers Cho, et al. (2014) in order to
prevent the vanishing/exploding gradient problem that affects standard
neural networks. We use the Mean Squared Error (MSE) as error metric for
training and evaluation of our models. The model is fed with information on
consecutive time series points of the number of individuals in each
compartment, and some network topological features (Mean, Variance and
Asymmetry of its degree sequence; order and size of the network, edge
density, clustering coefficient, assortativity and variance of its eigencentrality
coefficients).

Figure 1a: Initial RNN architecture. Figure 1a: Initial RNN architecture.

The recurrent layer transforms the input sequence into a fixed length

vector ( −1 , … , − ) that encodes relevant features of the time series and the
371 | I S I W S C 2 0 1 9