Life-Stage Modeling by Customer-Manifold Embedding
∗
Jing-Wen Yang
†
, Yang Yu
†
, Xiao-Peng Zhang
‡
†
National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China
†
Collaborative Innovation Center of Novel Software Technology and Industrialization, Nanjing, China
‡
Tencent Inc., China
yangjw@lamda.nju.edu.cn, yuy@lamda.nju.edu.cn, xpzhang@tencent.com
Abstract
A person experiences different stages throughout
the life, causing dramatically varying behavior pat-
terns. In applications such as online-shopping,
it has been observed that customer behaviors are
largely affected by their stages and are evolving
over time. Although this phenomena has been rec-
ognized previously, very few studies tried to model
the life-stage and make use of it. In this paper, we
propose to discover a latent space, called customer-
manifold, on which a position corresponds to a
customer stage. The customer-manifold allows us
to train a static prediction model that captures dy-
namic customer behavior patterns. We further em-
bed the learned customer-manifold into a neural
network model as a hidden layer output, resulting
in an efficient and accurate customer behavior pre-
diction system. We apply this system to online-
shopping recommendation. Experiments in real
world data show that taking customer-manifold into
account can improve the performance of the rec-
ommender system. Moreover, visualization of the
customer-manifold space may also be helpful to un-
derstand the evolutionary customer behaviors.
1 Introduction
There are various stages through everyone’s life. The focuses,
concepts and consumption ability of a person can be largely
influenced by his/her stages. Some marketing researchers and
sociologists have recognized that life stages would have a
huge impact on customers’ purchasing behaviors
[
Wells and
Gubar, 1966
][
Bojanic, 2011
]
. For example, a woman would
buy vitamins during her pregnancy stage, baby food and dia-
pers after baby’s birth, and then toys and fairy tale books as
baby grows. Obviously, considering life-stage would help us
understand customers better, and can help build better service
systems such as recommender systems.
Recommender systems aim at helping customers discover
most useful information from a massive amount of data. They
∗
This research was supported by the NSFC (61375061), Jiang-
suSF (BK20160066), Foundation for the Author of National Excel-
lent Doctoral Dissertation of China (201451).
have been widely adopted in areas such as social networks, en-
tertainment, and e-commerce recent years. Researchers and
engineers are always pursuing a better personalized system
for great value of both research and business. Most existing
recommender systems are based on, e.g., collaborative filter-
ing techniques, but do not take the evolutionary stage infor-
mation of customers into account.
A few recent studies have tried to improve recommender
systems by utilizing the information from the evolutionary
custom behaviors. Extended collaborative filtering model
[
Ding and Li, 2005
][
Liu et al., 2010
]
, matrix factorization
[
Koren, 2010
][
Xiong et al., 2010
]
, hidden Markov model
[
Li et al., 2011
]
, and Gaussian processes
[
Liu, 2015
]
were
studied for incorporating temporal effects in the models. The
time-based functions incorporated have been shown to be
able to capture some temporal patterns, but they mostly op-
erate in the customer behavior space where it would be too
many different temporal patterns to be well captured.
As an important matter of fact, people in different stages
may behave similarly in a current time slice, but have dramat-
ically different future behaviors. For example, one is moving
to a new house and another is planing to redecorate his old
apartment. They both buy paints, brushes and wallpapers. At
this time point, they behave almost the same. But for the few
next weeks, the first one purchases furnitures, while the latter
one buys a potted plants. Therefore, an approach processing
in the customer behavior space would get confused in this
situation, meanwhile, if an approach can understand the cus-
tomer stage, it would distinguish their behaviors easily and
make correct recommendations. From the view of this aspect,
modeling customer life-stage can be helpful to eliminate the
complexity of modeling evolution dynamics. With the infor-
mation of life-stage, we could have a deeper understanding
of customer behavior, thus helping us capture various modes
and make the evolution trend predictable.
In this paper, we propose to model customer stages by
learning the customer-manifold space. The intuition is that
the stage of a customer can be reflected by his behavior his-
tory. Therefore, we collect customers’ daily record to build
a novel similarity matrix and employ manifold learning to
embed those sequences into a metric space, which we call
as the customer-manifold space. Within this space, similar
stages are kept aligned, a stage can now be represented by a
point and the stage evolution of a customer can be captured