Source-linked AI summary
NeuTM: A Neural Network-based Framework for Traffic Matrix Prediction in SDN
Abdelhadi Azzouni, Guy Pujolle
TL;DR
Traffic-matrix prediction supports network planning, management, and security, but predicting full matrices is more challenging than predicting aggregate traffic. NeuTM addresses this gap with a large-scale LSTM framework that learns from historical traffic data, and evaluations report fast convergence and superior prediction performance over traditional and feed-forward methods.
Problem
Traffic-matrix prediction matters for network planning, management, and security, yet prior neural-network work mostly predicts aggregate traffic rather than full traffic matrices.
Method
NeuTM uses a large-scale LSTM recurrent-neural-network framework, transforming traffic matrices into traffic vectors and training on historical data to predict future matrices.
Results
LSTM models converge quickly and outperform traditional linear methods and feed-forward neural networks in traffic-matrix prediction.
Takeaways & Limitations
The proposed preprocessing and recurrent-network feeding technique achieves high traffic-matrix prediction accuracy with short training time.
Abstract
from arXiv · showhide
This paper presents NeuTM, a framework for network Traffic Matrix (TM) prediction based on Long Short-Term Memory Recurrent Neural Networks (LSTM RNNs). TM prediction is defined as the problem of estimating future network traffic matrix from the previous and achieved network traffic data. It is widely used in network planning, resource management and network security. Long Short-Term Memory (LSTM) is a specific recurrent neural network (RNN) architecture that is well-suited to learn from data and classify or predict time series with time lags of unknown size. LSTMs have been shown to model long-range dependencies more accurately than conventional RNNs. NeuTM is a LSTM RNN-based framework for predicting TM in large networks. By validating our framework on real-world data from GEEANT network, we show that our model converges quickly and gives state of the art TM prediction performance.
I. INTRODUCTION
Accurate, timely traffic-matrix prediction supports network management, planning, and security, but predicting full traffic matrices is more challenging than aggregate traffic. The paper introduces a large-scale LSTM framework, implements it on an SDN using GÉANT data, and reports fast convergence and state-of-the-art performance.
- Motivation: Accurate and timely traffic matrices support accounting, scheduling, rerouting, capacity planning, network design, and anomaly detection.The paper connects early prediction with faster high-volume traffic detection and DDoS prevention.
- Motivation: Network traffic exhibits self-similarity, multiscalarity, long-range dependence, and strong nonlinearity, which affect its predictability.
- Research gap: Neural-network traffic predictors outperform linear models, but prior work mostly predicts aggregate traffic rather than the more challenging traffic matrix.
- Contributions: The paper presents a LSTM-based framework for large-scale traffic-matrix prediction.
- Contributions: NeuTM is deployed on a Software Defined Network, trained with real-world GÉANT data, and evaluated across model configurations against traditional methods.
- Scope: The study assumes historical traffic-matrix data is already accurately obtained and does not address traffic-matrix estimation.
II. TIME SERIES PREDICTION
The paper reviews linear time-series predictors, including ARMA, ARAR, and Holt-Winters, which use recursive statistical formulations to forecast future values. These methods differ in how they model dependence, memory, trend, and smoothing.
- Linear prediction: Traditional linear prediction methods reviewed include ARMA, ARAR, and Holt-Winters algorithms.
- ARMA: ARMA models stationary time series using autoregressive and moving-average terms, with predictions generated recursively.
- ARAR: ARAR shortens long or moderate memory before fitting an autoregressive model to the transformed series.
- ARAR: The ARAR formulation uses recursive relationships to determine linear predictors from the transformed series.
- Holt-Winters: Holt-Winters is an exponential-smoothing method that recursively forecasts series containing a trend.
- Holt-Winters: Holt-Winters estimates the trend-function level and slope through recursive equations, with smoothing parameters selected randomly or by minimizing one-step errors.
2) Neural Networks for Time Series Prediction:
Neural networks model time series from observed data rather than an analytical model, using layered interconnected neurons. Their flexible nonlinear modeling makes them suitable for complex time-series relationships.
- Neural networks: Neural networks are widely used for time-series modeling because they can learn complex nonlinear patterns from data.
- Neural networks: Neural networks estimate linear or nonlinear functions, with architecture and parameters determined from the dataset.
- Architecture: A neural network contains weighted interconnected neurons organized into input, hidden, and output layers.
- Architecture: In feed-forward networks, information moves only forward from the input layer toward the output layer.
III. LONG SHORT TERM MEMORY NEURAL NETWORKS
Recurrent neural networks preserve temporal context through cyclic connections, making them better suited than feed-forward networks for sequence modeling. LSTMs address conventional RNN training problems that limit long-range dependency modeling.
- Temporal modeling: FFNNs use fixed-size windows and cannot adequately model historical dependencies, whereas RNNs feed prior activations into current predictions.RNN internal states retain temporal contextual information across time steps.
- Temporal modeling: RNNs store activations from each time step in internal states, supporting sequence modeling tasks such as time-series prediction.
- RNN limitations: Conventional RNN training with BPTT suffers from vanishing and exploding gradients, limiting useful long-range dependencies to roughly 5–10 time steps.The influence of inputs can decay or grow exponentially around recurrent connections.
- LSTM motivation: LSTMs were designed to address vanishing and exploding gradients and have performed better than conventional RNNs on context-free and context-sensitive language learning.
A. LSTM Architecture
An LSTM is organized into memory blocks whose cells retain temporal state while gates regulate information entering the cell and leaving it for the network.
- Memory blocks: Each LSTM memory block contains memory cells with self-connections that store the network’s temporal state.
- Gates: Input, output, and forget gates control input flow, output flow, and adaptive resetting or forgetting of cell memory.The forget gate scales the internal state before feeding it back through the cell’s recurrent connection.
- Peephole connections: Modern LSTM architectures use peephole connections from internal cells to gates to learn precise output timing.
B. LSTM Equations
The LSTM cell-state equation combines retained prior state with gated candidate input, while the accompanying notation defines gates, activations, weights, biases, and element-wise operations.
- Cell-state update: The cell state c_t combines the forget-gated previous state c_t−1 with the input-gated candidate activation.The equation is ct = ft ⊙ct−1 + it ⊙g(Wcxxt + Wcmmt−1 + bc).
- Notation: The notation identifies i, f, o, and c as the input gate, forget gate, output gate, and cell activation vectors.
- Notation: m denotes the output activation vector, ⊙ denotes element-wise multiplication, and g and h denote cell input and cell output activation functions.
- Notation: Biases are represented by b terms, weights by W terms, and σ by the logistic sigmoid function.
IV. TRAFFIC MATRIX PREDICTION USING LSTM RNN
NeuTM trains a deep LSTM with backpropagation through time on historical traffic data to predict future traffic matrices. The task represents time-indexed node-to-node traffic and estimates the next matrix from a sequence of prior matrices.
- Prediction framework: NeuTM trains a deep LSTM architecture with backpropagation through time to learn traffic characteristics and predict future traffic matrices.
- Traffic representation: A traffic matrix Y contains node-to-node volumes, and adding time produces an N-by-N-by-T tensor of traffic observations.Each tensor entry records traffic from node i to node j at time t.
- Problem definition: The prediction target is the traffic matrix at time t, estimated from the historical sequence Y_t−1, Y_t−2, Y_t−3, ..., Y_t−T.
- Problem definition: The central modeling challenge is capturing inherent relationships among historical traffic data accurately enough to predict the next matrix.
B. Feeding The LSTM RNN
NeuTM converts each traffic matrix into a traffic vector for LSTM input and uses historical vectors within a fixed learning window to predict current traffic.
- Each N×N traffic matrix Y_t is flattened row-wise into a traffic vector X_t of size N×N.The mapping preserves entries through n = i × N + j.
- The predictor uses a sequence of historical traffic vectors to estimate the current vector.
- Predicting individual origin-destination components independently is presented as insufficient because prior traffic across all ODs is needed for a more accurate traffic-vector prediction.
- A fixed learning window W limits the historical time-slots used for prediction, reducing the computational burden of continuous learning.The window forms a W-by-N^2 traffic-over-time matrix from X_t−1 through X_t−W.
C. Performance Metric
NeuTM evaluates prediction accuracy with mean squared error and reports that deeper networks improve accuracy while a six-layer model trains in under five minutes for 20 epochs.
- Mean squared error (MSE) quantifies prediction accuracy by averaging squared differences between forecasted and actual values.
- MSE is scale dependent and measures the difference between forecasted and actual values of the predicted quantity.
- Deeper networks provide better prediction accuracy according to the MSE results across different hidden-layer depths.
- Less than 5 minutes are required to train a six-layer network for 20 epochs.
- The evaluation compares prediction errors across methods and reports superiority of LSTM.
VI. RELATED WORK
Prior work includes linear, nonlinear, and traffic-matrix prediction methods; NeuTM addresses the more challenging task of predicting full traffic matrices with LSTM-based modeling.
- Existing traffic-matrix prediction studies compare methods including linear predictors, FARIMA, independent node prediction, total matrix prediction, and principal-component prediction.
- Earlier results report that neural networks outperform traditional linear prediction methods for network traffic forecasting.
- NeuTM presents LSTM architectures as well suited for traffic-matrix prediction and introduces data preprocessing and RNN feeding techniques.
- The evaluations report that LSTMs outperform traditional linear methods and feed-forward neural networks by many orders of magnitude.