Computational Intelligence - August 2017 - 49

shopping data set3 from Tianchi Data Lab
for evaluating the effectiveness of the proposed method. This dataset includes two
parts: customer's historical behavior dataset containing 1,103,702 users' historical
records in one year and a human-labeled
goods match ground truth dataset containing 23,105 records. We extracted all
users' history records as the basic transaction database (denoted as T ). Note that
some of the items in labeled ground truth
sets do not appear in T. Thus we
removed the items from the initial ground
truth data that do not appear in T and
finally obtained 4,009 records as the
goods match ground truth dataset. Each
customer's purchased goods can be
viewed as a transaction. For a customer
who wants to buy a good 'G', we first
collect all records in T containing 'G' as a
transaction database. Based on this transaction database, we aim to recommend a set
of interesting goods to the customer.
Print area recommendation in
SmartPrint. Assume that we have the
log transaction database recording how
previous users clipped the Web pages
from a Web site, where each transaction
refers to a set of content areas (or clips)
selected on a Web page. Given a Web
page from the same Web site, we aim to
recommend the informative print areas
in SmartPrint for this Web page. Let I
be the complete set of distinct clips in
the database. For a given Web page in
SmartPrint, we can get a set S 3 I of
clips which are included in this Web
page. Thus, the task is to select a subset
of S for the print area recommendation.
We adopted the ground truth data of
selected print areas on the 2,000 Web
pages from the 100 major print-worthy
Web sites used in [4] for evaluating the
effectiveness of our proposed method.
In addition, in order to show the
efficiency of the proposed algorithm, we
adopt the IBM synthetic data generator
for generating large data sets with different characteristics [35].
2) Comparison Methods
In the following, we show that the
-p roposed algorithm MOPM is both
3

https://tianchi.aliyun.com/datalab/dataSet.htm?id=13

effective and efficient for solving the
problem of task-oriented pattern mining. To this end, we compare MOPM
with several state-of-the-art algorithms
and its variants.
1) MFI: This baseline method is based
on maximal frequent itemset mining [21] (MFI for short) over the
transaction database. To be specific,
we first obtain all the maximal frequent itemsets, and then among
them select the one with the largest
number of items for pattern recommendation. This baseline needs users
to specify a prior parameter min_sup.
2) DOFIA: DOFIA [4] is currently the
best algorithm for the problem of
task-oriented pattern mining over
the transaction database shown in
Equation (1). This baseline is needed
to specify the prior parameters min_
sup, min_occ and m for users.
3) SGA: We extended a single genetic
algorithm [36] (named SGA) to solve
the problem shown in Equation (1).
The SGA was initially developed for
solving continuous optimization
benchmark test problems. Note that
SGA still needs to specify the prior
parameter m for users.
4) MOPM (Random): In order to
show the effectiveness of the proposed initialization strategy used in
MOPM, we also compare MOPM
with a variant of MOPM, MOPM
(Random), where all components
keep the same as MOPM except for
the population initialization by random strategy.
5) MOPM (-Area): In order to show
the effectiveness of the proposed
measure Area, we also compare
MOPM with a variant of MOPM,
MOPM (-Area). MOPM (-Area)
means that MOPM without the metric area, i.e., using only two objectives
support and occupancy.
Note that MOPM does not need to
specify the prior parameters min_sup,
min_occ, m for users. All algorithms were
implemented with C++ language and
the experiments were performed using
Red Hat on a server with 8 ) Intel(R)
E5620 @2.40GHz and 24GB of main
memory. For MOPM, we adopt the rec-

ommended parameter values used in
NSGA-II [6] as follows. The crossover
probability p c is set to 1.0 and the
mutation probability p m is set to 1 I
( I denotes the numbers of items). The
population size popSize is set to
^ I /50 + 1 h # 50. The algor ithm
stops when the current population
h a s more than 90% individuals
which are unchanged in the successive
5 -generations.
3) Evaluation Metrics
For the application of goods match recommendation, we adopt leave-one-out
cross validation to evaluate the recommendation performance. In other words,
we iteratively select one record from the
goods match ground truth dataset as the
query and select the first item f in this
query as the preference of the user, and
the records containing f in historical
behavior dataset of customers are used
to generate a transaction database for
recommendation. Then, we can average
these performance values over the 4,009
records to get the average performance.
Similarly, for the application of print
area recommendation, we iteratively
select one page as the query and the log
data on the left Web pages from the
same Web site are used to generate the
transaction database for recommendation. The recommended pattern refers
to a set of print areas on the given Web
page. Then, we can get the average
-performance by averaging these performance values over 2,000 Webpages.
For the evaluation metrics, based on
the recommended pattern and the
ground truth query, we can calculate the
precision Pre, recall Rec and F1 score:
M (R, G)
,
M (R)
M (G, R)
,
Rec =

M (G)
F1 = 2 # Pre # Rec ,
Pre + Rec
Pre =

(6)

where M ^R, G h is the number of
c-ommon items between recommended pattern R and ground truth pattern G and M (R) is the number of
items in R.

AUGUST 2017 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE

Table of Contents for the Digital Edition of Computational Intelligence - August 2017