
Forward Neural Network for Time Series
Anomaly Detection
ZHANG Rong
Tencent Company
zr9558@gmail.com
NIE Xin
Tencent Company
michaelnie@tencent.com
DONG Shandong
National University of Singapore
E0009088@u.nus.edu
XIAO Shiguang
Tencent Company
philipxiao@tencent.com
December 21, 2018
Abstract
Time series anomaly detection is usually formulated as finding outlier data points relative to some usual data, which is also an
important problem in industry and academia. To ensure systems working stably, internet companies, banks and other companies
need to monitor time series, which is called KPI (Key Performance Indicators), such as CPU used, number of orders, number of
online users and so on. However, millions of time series have several shapes (e.g. seasonal KPIs, KPIs of timed tasks and KPIs of
CPU used), so that it is very difficult to use a simple statistical model to detect anomaly for all kinds of time series. Although some
anomaly detectors have developed many years and some supervised models are also available in this field, we find many methods
have their own disadvantages. In this paper, we present our system, which is based on deep forward neural network and detect
anomaly points of time series. The main difference between our system and other systems based on supervised models is that we do
not need feature engineering of time series to train deep forward neural network in our system, which is essentially an end-to-end
system.
I. Introduction
To ensure systems working stably and efficiently, internet
companies need to monitor huge time series every minute,
whose names are KPIs (Key Performance Indicators). For
example, in the industry and academia, KPIs contain
several kinds of time series, including CPU, online page
views, online users of some application, the number of
failures and successes of logging some website. In our
opinion, different KPIs have different shapes and trends,
so it is difficult to use a simple statistical model to detect
all anomaly points of KPIs. Before using machine learning
models, we wrote rules and used 3-sigma method to
detect anomaly points of time series. However, rules
become more and more complicated and it is impossible
for us to check all rules at regular time intervals in order
to guarantee precision and recall of the whole system.
Therefore, we try our best to build a new system which is
based on human experience and machine learning theory,
in order to increase the precision and recall of the our
system for anomaly detection of time series.
Anomaly detection has been an active research area in
the fields of machine learning and statistics. Statistical
methods, control chart theory [
1
], ARIMA and seasonal
ARIMA models [
2
],[
3
],[
4
], Holt-Winters model [
5
] are pro-
posed for time series anomaly detection. Beside statistical
models, in machine learning theory, there are also a lot
of methods to detect anomaly points of time series, such
as supervised and unsupervised models. Most existing
anomaly detection approaches, including classification-
based methods [
6
], isolation forest [
7
], one-class SVM [
8
],
clustering-based methods, construct normal pattern from
samples, then identify anomaly points as those which do
not satisfy the normal pattern. In the field of time series
anomaly detection, some scholars provided supervised
models bases on feature engineering [
9
] and unsuper-
vised models [10] to detect anomaly points of KPIs.
As discussed before, to ensure high precision and recall
of the system, we must provide a stable and efficient
algorithm. This paper proposes a different approach that
detects anomalies by neural networks, without relying
on any feature engineering or time series detectors. The
contributions of the paper can be summarized as follows:
•
Deep forward neural network is an end-to-end
model, which can be trained from normalized raw
datas to corresponding labels. The main technique
in our system is that there is no need of feature engi-
neering of time series.
•
The offline precision and recall of deep forward neu-
ral network are higher than XGBoost with a lot of
feature engineering from human experience.
•
The output of hidden layers of trained forward neu-
ral networks can be taken as the features of time
1
arXiv:1812.08389v1 [cs.LG] 20 Dec 2018