INT J COMPUT COMMUN, ISSN 1841-9836
Vol.7 (2012), No. 3 (September), pp. 494-508

Mining Temporal Sequential Patterns Based on Multi-granularities

N. Li, X. Yao, D. Tian

Naiqian Li, Xinhui Yao, Dongpin Tian
Department of Computer Science
Baoji University of Arts and Sciences
Baoji 721016, Shaanxi, China
E-mail: xalnq@hotmail.com, baojiyxh@163.com,
tdp211@163.com

Abstract:
Sequential pattern mining is an important data mining problem that can extract fre-
quent subsequences from sequences. However, the times between successive items
in a sequence is typically used as user-specified constraints to pre-process the in-
put data or to prune the pattern search space. In either cases, the times cannot be
used to identify item intervals of sequential patterns. In this paper, we introduce
a form of multi-granularity sequence patterns, which is a sequential pattern where
each transition time is annotated with multi-granularity boundary interval and aver-
age time derived from the source data rather than the user-predetermined time interval
or only a typical time. Then we present a novel algorithm, MG-PrefixSpan, of mul-
tiple granularity sequential patterns based on PrefixSpan[, which discovers all such
patterns. Empirical evaluation shows that MG-PrefixSpan scales up linearly as the
size of database, and has a good scalability with respect to the length of sequence and
the size of transaction.
Keywords: Data Mining Algorithm, Sequential Pattern Mining, Sequential Data,
Time Granularity, Temporal Patterns.

1 Introduction

Among various types of data mining applications [1, 2], the sequential pattern mining, which discov-
eries interesting sequential patterns hidden in sequence of events, is an important data mining problem
with broad applications, including market analysis, decision support, the prediction of occurrences of
recurrent illnesses, system performance analysis and telecommunication network analysis etc.

The problem of mining sequential patterns was first proposed by Agrawal and Strikant [3]: Given a
data set of sequences, each sequence is a list of transactions, where each transaction is a set of items. The
sequential pattern mining is to find all subsequences that is more frequent than a user-specified minimum
support threshold while maintaining their item occurrence order.

For example, in the database of a book-club, a sequential pattern might be “5 percent of customers
bought ‘Foundation’, then ‘Foundation and Empire’, and then ‘Second Foundation”’ [4]. Although the
discovered sequential patterns reveal what items are frequently bought together and in what order, they
cannot reveal how long time the items will be bought after the preceding items. Unfortunately, not
knowing the time means that we cannot exactly predict when the next purchase will happen. In addition,
some sequential patterns could occur in different periods with different time granularities. For example,
“HP stock could rise within 5 days after IBM stock rose" and “HP stock could fall within 6 month after
IBM stock rose", which reveal HP stock rise or fall with respect to IBM stock rise at different time
granularities (day and month), are two useful different patterns. However, these traditional sequence
patterns can only tell us “HP stock could rise after IBM stock rose" and “HP stock could fall after IBM
stock rose". This situation means that the two patterns are useless. Another situation may be “HP stock

Copyright c⃝ 2006-2012 by CCC Publications



Mining Temporal Sequential Patterns Based on Multi-granularities 495

could rise 5 days later after IBM stock rose" and “HP stock could rise 6 months later after IBM stock
rose". Although these two patterns are completely different with regard to different time granularities,
these traditional sequence patterns could treat the two patterns as the same pattern “HP stock could rise
after IBM stock rose", which make the extracted pattern less precise and some useful information lost.

Given the above reasons, in this paper we generalize the problem definition given in [1, 8, 9, 13] to in-
corporate the maximum, minimum and average time between successive transactions, which are derived
from the source data, and different time granularities in traditional sequential patterns. We present a novel
algorithm of multiple granularity sequence patterns based on PrefixSpan [6] , called MG-PrefixSpan,
which discovers all such patterns. Empirical evaluation shows that MG-PrefixSpan scales up linearly as
the size of database, the length of sequence and the size of transaction.

The rest of this paper is organized as follows. Related work is discussed in section 2. We give a
formal description of the problem of mining temporal sequence patterns based multiple granularities in
section 3. In section 4, we describe MG-PrefixSpan, an algorithm for finding such patterns, and then
Section 5 provides empirical evaluation of the performance of MG-PrefixSpan. Finally, we conclude the
paper in section 6.

2 Related Work

Sequential pattern mining, in general, can be grouped into two categories. One category, called
un-temporal sequence pattern mining or traditional sequence pattern mining, considers only the item
occurrence order in a sequential pattern, but does not deal with time-related data [1, 3–6]. The other
category, called temporal sequence pattern mining, consider not only the item occurrence order in a
sequence pattern, but also the time between successive items in a sequential pattern, such as [8, 9, 11–13].

2.1 Un-temporal Sequence Pattern Mining

Agrawal and Strikant [3] introduced the notion of sequential pattern mining, and based on the prop-
erty that any sub-pattern of a frequent pattern must be frequent, three Apriori-based algorithms were
proposed: AprioriSome, DynamicSome and AprioriAll. Two of these algorithms were designed only to
find maximal sequential patterns. The third algorithm, AprioriAll, finds all patterns. Briefly, AprioriAll
is decomposed into two phases: (1) generating candidate sequences; (2) scanning the sequential database
to check the support of each candidate to determine frequent sequence patterns according to minimal sup-
port threshold. Although AprioriAll is not efficient, it is the basis of many efficient algorithms developed
later. SPADS [5] is an algorithm proposed to find frequent patterns using efficient lattice search technol-
ogy and simple joins. It decomposes the original search space (lattice) into smaller pieces (sub-lattices),
which can be processed independently in the main memory. Due to adopting a vertical id-list database
format to count the number of frequent patterns, all the sequential patterns are discovered with only a
few passes over the database. SPADS outperforms ApriorAll [2,3]. PrefixSpan[6] is another more effi-
cient algorithm for mining sequential patterns comparing with the aprior-based algorithm ApriorAll and
SPADS, especially in dealing with very large databases. It mainly adopts a projection-based, sequential
pattern-growth method to make the database for next pass much smaller and consequently make the algo-
rithm more speedy. Also in PrefixSpan there is no need to generate candidates, only recursive projection
of database according to their prefix. Our method is based on the PrefixSpan algorithm.

Some have tried to exert constraints on the mining of sequential patterns so that only those sequential
patterns interesting to users are discovered rather than the whole possible sequential patterns[2,5,6,7].
Strikant and Agrawal [4] generalized their definition of sequential patterns in [3] to integrate with time
constraints, sliding time window, and user-defined taxonomy, and proposed algorithm GSP. Mannila et
al. [7] proposed a method of mining frequent episodes in a sequence of events, in which episodes are
essentially constraints on events in form of acyclic graphs. Garofalakis et al. [8] proposed a family of



496 N. Li, X. Yao, D. Tian

SPIRIT algorithms of mining user-specified sequential patterns by using regular expression constraints.
Pei et al. [9] developed an extended framework based on a sequential pattern growth for constraint-based
sequential pattern mining. Although time between successive items is typically used as a user-specified
constraint to shrink the pattern search space to make the computation more efficient, it is not used in the
output frequent sequence patterns.

2.2 Temporal Sequence Pattern Mining

Yoshida et al.[10] proposed a notation of delta pattern, which is a temporal sequence pattern with

temporal constraints in the form A
[0,7]
→ B

[3,5]
→ C of bounding intervals. An example of delta pattern has

the form denoting a sequential pattern A → B → C that frequently appears in the database with transition
times from A to B and from B to C that contained in [0, 7] and in [3, 5] respectively. However, Yoshida
et al. only provided a heuristics for finding some frequent delta patterns, and did not investigate the
problem of finding all of them. Along the same direction, Chen et al. [11] introduced a form of temporal
sequence pattern by inserting pseudo items into the original sequential pattern. Pseudo items are user-
defined time interval segmentations in advance. When counting the support of a sequential pattern,
only the sequence, in which the pseudo item between successive items is same, supports the sequential
pattern. Two algorithms, I-Apriori and I-PrefixSpan, were proposed. I-Apriori is based on Apriori
algorithm [12], and I-PrefixSpan is based on PrefixSpan[6]. Hirate et al. [13] proposed generalized
sequential pattern mining with item intervals. They extended sequences which are defined by inserting
pseudo items based on the interval itemization function and exerting four interval constraints on items.
However, as they adopted some user-predefined pseudo items or constraints on the time intervals between
successive items, it is difficult for a user to specify optimal constraints related to item interval and cannot
reveal the time interval between successive items in the patterns precisely, it may result in some useful
sequential patterns not being found. Giannotti et al. [14] introduced another form of sequential pattern,
called Temporally-Annotated Sequence, TAS in short, where each transition is annotated with temporal

information representing a typical time derived from the source data. For instance, TAS A
t1→ B

t2→C
denotes the fact that a sequential pattern A → B → C frequently appears in the database and that the
time for getting from event A to B is close to t1 and from event B to C is close to t2. Although this method
using typical time to annotate transition between two successive events can reveal how much time the
event will occur after the preceding event, they cannot distinguish completely patterns with regard to
different time granularities. For example, from event A to B is close to t1 days, other from event A to
B may be close to t2 months, these two different patterns are treated as the same. It is useful to be able
to distinguish the patterns to understand not only what event will follow, but also when these events will
occur.

Finally, we mention the work in [15], where event structures that have temporal constraints with
multiple granularities are introduced. Event structures essentially are used as a flexible user-defined
constraint specification to define the pattern discovery problem with these structures that enable users to
focus on their interested sequential patterns. This method can only find the patterns that satisfy the event
structures. Furthermore, an event structure consists of a number of variables representing events and
temporal constraints among these variables, an efficient event structure is difficult to define beforehand
even if you are a domain expert.

In our work, we introduce a new form of sequential pattern with multi-granularities, which is a se-
quential pattern where each transition is annotated with multi-granularity boundary interval and average
time derived from the source data rather than user-predetermined time intervals [8,9] or only a typical
time [14]. We also define the pattern discovery problem involving multi-granularities for these pattern
and study efficient algorithm based on PrefixSpan to solve it.



Mining Temporal Sequential Patterns Based on Multi-granularities 497

3 Problem Formulation

3.1 Time granularity

In order to formally define temporal sequence pattern that involves time granularities, we first review
the notion of a time granularity [15].

Definition 1. A granularity is a mapping µ from the set of the positive integers ( the time ticks) R to
R2 (the set of absolute time sets) such that for all positive integer i and j with i ≤ j, the following two
condition are satisfied:

(1) µ(i) , ∅∧µ( j) , ∅ implies that each number in µ(i) is less than all the numbers in µ( j), and
(2) µ(i) , ∅ implies µ( j) , ∅.
Each set µ(i), if non-empty, is called a granule of the µ. Property (1) says that granules do not overlap

and the order on time ticks follow the order on the corresponding granules. Property (2) says that the
subset of the time ticks corresponding to the granules forms a set of contiguous integers. The set µ(i) of
reals is said to be the ith tick of µ, or tick i of µ . For example, hour, day, week, month, and year, satisfy
the above definition. We can also define more complex granularities like business-week, weekend and
so on.

When dealing with temporal types, it is needed to determine the tick (if any) of a temporal type µ that
covers a given tick z of anther temporal type v. Formally, for each positive integer z and temporal types
µ and v, if ∃ z’ (necessarily unique) such that v(z) ⊆µ(z’) then ⌈z⌉µv = z′, otherwise ⌈z⌉

µ
v is undefined[15].

In this paper, all timestamps in a sequence are assumed to be in terms of a fixed granularity g0, and
abbreviate ⌈z⌉µv as ⌈z⌉µ if v= g0.
Definition 2. A multi-granularity schema is a tuple of the form Gm = (gm,gm−1, · · · ,g2,g1), where each
gi (1 ≤ i ≤ m) is a granularity.

For simplicity, we require that in Gm and g0, each unit of gi is contained in a unit gi+1 (0 ≤ i ≤ m−1).

3.2 Temporal sequence and multi-granularity sequence pattern

Let I = {i1, i2, · · · , in} be a set of items. An itemset si is a non-empty subset of items (without loss of
generality, we assume that items of an itemset are sorted alphabetically) denoted as i1, · · · , ik, where i j
(1 ≤ j ≤ k) is an item. A traditional data sequence is an ordered list of itemsets, which is sorted by the
order of priority of the transaction time and denoted as (s1 → s2 →···→ sn), where si (1 ≤ i ≤ n) is an
itemset [1,3]. In practice, a data sequence of a customer is also formed as an ordered list of itemsets and
time stamps [11], we call it a temporal sequence.

Definition 3. A temporal sequence s is represented as ⟨(s1, t1) → (s2, t2) → ··· → (sn, tn)⟩, where si is
an itemset and ti stands for the time at which si occurs, ti < t j for 1 ≤ i < j ≤ n.

When adding the itemset time information in the sequence, the time interval value between any two
elements in the sequence can be computed as follows:

Ti j = T j − Ti,where 1 ≤ i < j ≤ n

We now formally define multi-granularity schema, multi-granularity sequence patterns and its’ sup-
port, then formalize our novel mining problem as the discovery of all frequent multi-granularity sequence
patterns in the database D.

Definition 4. Given a multi-granularity schema Gm = (gm,gm−1, · · · ,g2,g1), a multi-granularity se-
quence pattern is represented as follows:

α=α1
[L1,M1,U1]µ1→ α2

[L2,M2,U2]µ2→ ···
[Ln−1,Mn−1,Un−1]µn−1→ αn



498 N. Li, X. Yao, D. Tian

where (1) αi (1 ≤ i ≤ n) is an itemset; (2) µi ∈ {g j| j = 1, · · · ,m} (1 ≤ i ≤ n − 1) a granularity; (3)
Li, Mi, and Ui (1 ≤ i ≤ n − 1) are respectively the lower bound, average time and upper bound of the
time that αi+1 occurs after α1, α2, · · · and αi with the granularity µi. [Li, Mi,Ui]µi is called temporal
annotation of αi+1, and the tuple ([Li, Mi,Ui]µi,αi+1) also called a temporal item (1 ≤ i ≤ n − 1).

Example α = α1
[1,25,5]day
→ α2

[2,33,4]week
→ α3 is a multi-granularity sequential pattern, where α1,α2,α3

are itemsets or events, day and week are time granularities. The pattern indicates that α2 occurs in 1 day
to 5 days or average 2.5 days after α1, then 2 weeks to 4 weeks or average 2.5 weeks later, α3 occurs.

The total number of items in a multi-granularity sequence pattern is referred as the length of the
pattern. A multi-granularity sequence pattern whose length is k is also represented as k-multi-granularity
sequence pattern.

Definition 5. Given a temporal sequence s = (s1, t1) → (s2, t2) →···→ (sn, tn) and a multi-granularity
schema Gm = (gm,gm−1, · · · ,g2,g1). Let α be a multi-granularity sequence pattern
α = α1

[L1,M1,U1]µ1→ α2
[L2,M2,U2]µ2→ ···

[Lk−1,Mk−1,Uk−1]µk−1→ αk, where µi ∈ g j| j = 1, · · · ,m (1 ≤ i ≤ k − 1). s is
said to support( or contain ) α if and only if there exists integers 1 ≤ i1 < i2 < · · ·< ik ≤ n such that

(1) ∀ j1 ≤ j ≤ k, αi ⊆ si j
(2) ∀ j1 ≤ j ≤ k−1, if µ j = gh (1 ≤ h ≤ m−1) then gh(1) ≤ Ti ji j+1 <gh+1(1), otherwise Ti ji j+1 ≥ gm(1).

The condition (2), in fact, is a constraint which divides the time between successive items into non-
overlap partitions according to the sizes of different time granules. Without this constraint, the too long
or too short time interval between successive items can be treated as the same in a pattern, and may result
in some useful patterns not being found.

Definition 6. Let α = α1
[L1,M1,U1]µ1→ α2

[L2,M2,U2]µ2→ ···
[Lk−1,Mk−1,Uk−1]µk−1→ αk be a multi-granularity sequen-

tial pattern, Gm be a multi-granularity schema, and S l be a set of temporal sequence that support

αl+1 = α1
[L1,M1,U1]µ1→ α2

[L2,M2,U2]µ2→ ···
[Ll,Ml,Ul]µl→ αl+1 (1 ≤ l ≤ k − 1) in database D. The annotation

[L1, M1,U1]µl of αl+1 is defined as follows:

Ul = max
s∈S l

max
i1,··· ,il+1

⌈
Tilil+1

⌉µl
Ll = min

s∈S l
min

i1,··· ,il+1

⌈
Tilil+1

⌉µl
Ml =

1
|S l|

∑
i1,··· ,il+1

⌈
Tilil+1

⌉µl
where i1, i2, · · · , il, il+1 are the indexes of the elements in s ∈ S l that match the elements of αl+1, µl ∈ Gm.

Definition 7. Let D be the database of temporal sequence, Gm be a multi-granularity schema. For a
multi-granularity sequence pattern α, its support in database D is defined by:

S upportD(α) =
|{s|s ∈ D ∧ s supports α}|

|D|

α is said to be frequent if its support is no less than the user-specified minimum support threshold min-sup
i.e. S upportD(α) ≥ min − sup.

Given a temporal sequence database D, multi-granularity schema Gm and user-specified minimum
support threshold min-sup, the task of mining multi-granularity sequence pattern is to find all frequent
multi-granularity sequence patterns in the database D.



Mining Temporal Sequential Patterns Based on Multi-granularities 499

4 Algorithm for Mining Multi-Granularity Sequence Patterns

In this section, we propose a novel multi-granularity sequence pattern mining algorithm based on
the well-known PrefixSpan[6], called as MG-PrefixSpan. The PrefixSpan algorithm solves the pattern
mining problem by a divide-and-conquer, pattern-growth principle as follows: for each frequent item a,
a projection of the initial sequence database D is created, denoted by D|a, i.e., which (1) contains only
the suffix of sequences in D with respect to a; (2) contains only the frequent items; (3) in general is much
smaller than D. Then, for each frequent item b in D|a, appends b to a to form a sequential pattern a′
which starts with a. Finally a new, smaller projection D|a′ can be constructed recursively and used for
finding longer patterns starting with a′.

However, although the PrefixSpan algorithm is very efficient for mining sequential pattern, it cannot
be used straightforwardly to find multi-granularity sequence patterns. We extend the sequence database
projection operation in PrefixSpan to be able to deal with the temporal relationship between succes-
sive items in the sequential pattern based on multi-granularities. Before introduction of the algorithm
MG-PrefixSpan, we first extend definitions of prefix, suffix, projection and projected database based on
PrefixSpan[6].

Definition 8. Given a temporal sequence s = ⟨(s1, t1) → (s2, t2) →···→ (sn, tn)⟩ and a multi-granularity
schema Gm = (gm,gm−1, · · · ,g2,g1). A multi-granularity sequence pattern α=α1

[L1,M1,U1]µ1→ α2
[L2,M2,U2]µ2→

···
[Lk−1,Mk−1,Uk−1]µk−1→ αk (k ≤ n) is a multi-granularity sequence pattern prefix of s if and only if
(1) αi = si for 1 ≤ i ≤ k − 1;
(2) αk ⊆ sk and all the items in (sk −αk) are alphabetically after those in αk;
(3) ∀i1 ≤ i ≤ k − 1, if µi = gh (1 ≤ h < m) then gh(Li) ≤ Tii+1 ≤ gh(Mi).

For example, let multi-granularity schema be G2 = (hour, minute), g0 = second. If there is a temporal
sequence α= ⟨(c,5) → (abc,80) → (ab,320) → (d,5400) → (c f,6400)⟩, β1 = (c) is a multi-granularity
sequence pattern prefix of and so is β2

[1,2.4,4]minute
→ (ab).

Definition 9. Given a temporal sequence s and a multi-granularity sequence pattern α such that s supports
α. A subsequence β of s is called a projection of s with respect to α if and only if

(1) β has a multi-granularity sequence pattern prefix α, and
(2) there exists no proper super-sequence β′ of β such that β′ is a subsequence of s and also has a

multi-granularity sequence pattern prefix α.

If a temporal sequence α = ⟨(c,5) → (abc,80) → (ab,320) → (d,5400) → (c f,6400)⟩ is pro-
jected with respect to β2

[1,2.4,4]minute
→ (ab), then two projection are ⟨(c,5) → (abc,80) → (ab,320) →

(d,5400) → (c f,6400)⟩ and ⟨(c,5) → (abc,80) → (ab,320) → (d,5400) → (c f,6400)⟩. If α is pro-
jected with respect to β1 = (c), then three different projections are ⟨(c,5) → (abc,80) → (ab,320) →
(d,5400) → (c f,6400)⟩, ⟨(c,5) → (abc,80) → (ab,320) → (d,5400) → (c f,6400)⟩ and ⟨(c f,6400)⟩
respectively.

In the above, a temporal sequence α that is projected with respect to a prefix produces more than
one projection. To differentiate these projections from the same temporal sequence, the tag [Sid, item,
t] is attached to each projected sequences, where Sid is the identifier of the temporal sequence, item is
last item in the multi-granularity sequence pattern and t is the time of element in temporal sequence that
matches the last element of the multi-granularity sequence pattern.

Consequently, if sequence α = ⟨(c,5) → (abc,80) → (ab,320) → (d,5400) → (c f,6400)⟩ is
projected with respect to the β2, then different projections are presented as follows:

[Sid, b, 80]: ⟨(c,5) → (abc,80) → (ab,320) → (d,5400) → (c f,6400)⟩;
[Sid, b, 320]: ⟨(c,5) → (abc,80) → (ab,320) → (d,5400) → (c f,6400)⟩.



500 N. Li, X. Yao, D. Tian

Definition 10. Given a temporal sequence s and a multi-granularity schema Gm = (gm,gm−1, · · · ,g2,g1).
Let α = ⟨(s1, t1) → (s2, t2) → ··· → (sn, tn)⟩ be the projection of s with respect to a multi-granularity
sequence pattern prefix

α=α1
[L1,M1,U1]µ1→ α2

[L2,M2,U2]µ2→ ···
[Lk−1,Mk−1,Uk−1]µk−1→ αk,(k ≤ n)

Temporal sequence γ = ⟨(sk′′, tk) → (sk+1, tk+1) → ··· → (sn, tn)⟩ is called the suffix of s with
respect to α, denoted as γ= s/α, where sk′′= sk − sk′.

Note, if α is not a multi-granularity sequence pattern prefix of s, the suffix of s with respect to α is
empty.

This definition indicates that the suffix of a temporal sequence can be obtained by directly remov-
ing the prefix from its projection. In the example above, if the sequence α = ⟨(c,5) → (abc,80) →
(ab,320) → (d,5400) → (c f,6400)⟩ is projected with respect to the β2, then three different suffix are
obtained as follows:

[Sid, b, 80]: ⟨(c,80) → (ab,320) → (d,5400) → (c f,6400)⟩;
[Sid, b, 320]: ⟨(d,5400) → (c f,6400)⟩.

Definition 11. Let α be a multi-granularity sequence pattern in a temporal sequence database D, The α-
projected databases, denoted as D|α, is the collection of suffixes of temporal sequences in D with respect
to α.

Based on the above description, the newly proposed algorithm, MG-PrefixSpan, is shown as follow-
ing:
Input: D: temporal sequence database; Gm: multi-granularity schema; min-sup: minimal support thresh-
old.
Output: The complete set of frequent multi-granularity sequence patterns.
Method:

(01) scan D once, to find all frequent items, which are denoted by L1.
(02) for each a L1 do begin
(03) construct candidate temporal items α= (0,a) and add α in L1;
(04) end
(05) output L1;
(06) for each s ∈ D do begin
(07) for each α ∈ L1 do begin
(08) for each β suffix of s with respect to α do begin
(09) if β, null null then add β and its tag to D|α;
(10) end
(11) if D|α, null null then call MG-PrefixSpan (α,1, D|α);
(12) end
(13) end
Subroutine MG-PrefixSpan (α,1, D|α)
Parameter: α a multi-granularity sequence pattern; k the length of α; D|α the α-projected database.
(20) if Lk = null then return;
(21) for all a ∈ Lk and all µ ∈ Gm do begin
(22) construct candidate temporal item (A, a), where A=[L, M, U]µ or 0, add it in Ck;
(23) end
(24) for each s s ∈ D|α with tag [Sid, b, t] do begin
(25) for each (A,a) ∈ Ck do begin
(26) for each (si, ti) in s where a a ∈ si do begin
(27) if ti == t, then add 1to the counting number of (0, a) for different sid;



Mining Temporal Sequential Patterns Based on Multi-granularities 501

(28) if g j(1) ≤ ti − t <g j+1(1) or gm(1) ≤ ti − t for 1 ≤ j ≤ m, then do begin
(29) increase the counting number of (A, a) for different sid;
(30) let L = min{L,⌈ti − t⌉g j}, U = max{U,⌈ti − t⌉g j}, M = M + ⌈ti − t⌉g j ;
(31) end
(32) end
(33) end
(34) end
(35) for each (A,a) ∈ Ck do begin
(36) if counting number of (A, a) ≥ min-sup |D|, then do begin
(37) let M=M/the total number of pulsing M (A,0);
(38) append (A, a) to α to form k+1-multi-granularity sequential pattern γ, add a in Lk+1;
(39) let γ. support=min α. support, counting number of (A, a)/|D|, add γ in Lk+1;
(40) end
(41) end
(42) output Lk+1

(43) for each γ ∈ Lk+1 do begin
(44) for each s ∈ D|α do begin
(45) for each suffix of s with respect to γ: β do begin
(46) if β, null then add β and its tag to D|γ;
(47) end
(48) end
(49) if D|γ= null then return;
(50) call MG-PrefixSpan (γ,k + 1, D|γ);
(51) end
(52) return
The overall MG-PrefixSpan algorithm is summarized in the above. The most importance difference

lies in that MG-PrefixSpan algorithm is able to deal with the temporal relationship between successive
items in the patterns based on multi-granularities. Steps 21-34 handle the frequent items in D|α and
correspond annotations with multi-granularities, where steps 21-23 construct candidate temporal items,
steps 24-30 derive the low bound, upper bound and average time of each candidate temporal item from
the projected database D|γ. If a candidate temporal item is frequent, then it can be appended to the prefix
to form k+1-multi-granularity sequence pattern γ in the steps 35-41, where when a annotation of a item
is 0, it only insert the item into the last element of the prefix, otherwise appends the temporal item to the
prefix as an new element. Steps 43-48 construct a new, smaller projection D|γ, recursively find temporal
items in D|γ and yield longer multi-granularity sequence patterns until all the multi-granularity sequence
patterns are found.

5 Experimental Results and Performance Analysis

In this section, we provide an experimental assessment of the proposed algorithm on synthetic data
sets. The purpose is to test the performance of our algorithm MG-PrefixSpan. We will analyze the
effects of input parameters on execution times and compare the performances with those of the most
efficient algorithm PrefixSpan [6] and I-PrefixSpan [11]. The first algorithm being the fastest algorithm
for sequential pattern mining without time intervals [6] and the second algorithm the most effective
algorithm in finding sequential patterns with user-defined time intervals, It is natural to compare our
algorithm with its.



502 N. Li, X. Yao, D. Tian

Table 1: PARAMETERS USED IN THE GENERATION OF DATASET
|D| Number of customers
|C| Average number of transactions of per customer
|T| Average number of items per transaction
|S| Average length of maximal potentially large sequences
|I| Average size of itemsets in maximal potentially large sequences
|Ns| Number of maximal potentially large sequences
|NI| Number of maximal potentially large itemsets
|N| Number if items
|TT| Average length of time intervals

Table 2: PARAMETER SETTINGS
Name |C| |T| |S| |I| |D| Size (Mb)

C10T2S4I1 10 2 4 1 10K 1.733

C10T2S4I2 10 2 4 2 10K 1.743

C10T2S8I1 10 2 8 1 10K 1.746

C10T4S4I1 10 4 4 1 10K 2.509

C20T2S4I1 20 2 4 1 10K 3.317

5.1 Evaluation Environment

The two algorithms are implemented by Sun Java language and experiments were conducted on a
1.7MHz Intel Pentium IBM laptop with 512MB main memory, running Microsoft Windows XP and Bor-
land JBiulder 9.0 as the Java execution environment. Detailed algorithm implementation of PrefixSpan
and I-PrefixSpan is according to the algorithms described in [4,9], but with the pseudo projection turned
off; MG-PrefixSpan is implemented as described in this paper.

5.2 Synthetic dataset generation

In this work we extended the IBM synthetic generator described in [1,10] to generate synthetic data
sets. Basically, each data-sequence is a list of transactions, where each transaction is a set of items, called
itemset. However, the transaction data is extended so that the items in different itemsets are assigned
different time values and that those in the same itemsets are assigned the same time values. The interval
time value between successive itemsets for each customer obeys a Poisson distribution with mean w. the
value w is drawn repetitively from a Poisson distribution with mean TT for this particular customer [11].
After that, we determine the time t for each transaction of this customer by summing the interval times
before the transaction.

Table 1 lists the parameters used in the generation of the simulating dataset. The parameters except
for the last one are the classical ones used in the previous research. The parameter TT is a new one used
to generate the time for each transaction.

We generate datasets by setting NS = 500, NI = 2500, N = 1000 and TT = 60. Table 2 summarizes
the dataset parameter settings.



Mining Temporal Sequential Patterns Based on Multi-granularities 503

Table 3: PART OF THE EXACTED PATTE
Extracted patterns Support(%)

< (80) > 22.89

< (80)
[1,3,6]day
→ (80) > 1.71

< (80)
[1,2,5]day
→ (80) > 1.97

< (80)
[1,3,6]day
→ (80)

[1,2,5]day
→ (80) > 0.54

< (80)
[2,3,6]day
→ (569) > 0.62

< (80)
[1,2,4]day
→ (569) > 0.84

< (80,432) > 1.19

< (80,432)
[1,2,3]day
→ (944) > 0.60

Table 4: PART OF THE EXACTED PATTERN
Extracted patterns Support(%)

< (80) > 22.89

< (80) → (80) > 3.03
< (80) → (80) → (80) > 0.71
< (80) → (569) > 1.39
< (80,432) > 1.19

< (80,432) → (944) > 0.62

5.3 Time Granularity Selection

In the experiment, we use multi-granularity schema: G3 = (Week, Day, Hour), and all the timestamps
in the synthetic data sets are Hour, i.e. g0 = Hour.

In order to compare the performance of algorithm I-PrefixSpan to that of MG-PrefixSpan, all the
intervals between successive items are partitioned into five intervals: {l0, l1, l2, l3, l4}, where l0 : t = 0 :
0 < t < 1, l2 : 1 ≤ t < 24, l3 : 24 ≤ t < 168, and l4 : 168 ≤ t <∞.

5.4 Comparison of Extracted Sequential Pattern Quality

The first test is a comparison of the quality of extracted patterns by algorithm PrefixSpan[6], I-
PrefixSpan[11] and MG-PrefixSpan. The test is designed using the data set C10T4S5I2 and all minimum
support thresholds is set to 0.005(0.5%). Table 3 shows part of the extracted frequent multi-granularity
sequence patterns using MG-PrefixSpan algorithm, and Tables 4 and 5 show part of the extracted frequent
sequential patterns using the algorithm PrefixSpan and I-PrefixSpan, respectively.

As shown in Table 3, we can made the following observation:
(1) once item 80 occurs, then item 80 will occur again within 1 day to 6 days, or with average time 3

days, with probability (1.71/22.89) × 100% = 7.4%; within 1 week to 5 weeks, or with average time 2
weeks, with probability (1.97/22.89) × 100% = 8.6%.

(2) once item 80 occurs and item 80 occurs within 1day to 6 days, or with average time 3 days,
then item 80 will occur again within 1 week to 5 weeks, or with average time 2 weeks, with probability
(0.54/1.19)100% = 45.4%.



504 N. Li, X. Yao, D. Tian

0

500

1000

1500

2000

2500

3000

0.005 0.01 0.015 0.02

Su pport thre shol d
R

u
n

ti
m

e 
(i

n
 s

ec
o

n
d

s)

MG-PrefixSpan
PrefixSpan
I-PrefixSpan

 

Figure 1: Performance of the of the algorithms on data set C10T2S4I1.

0

500

1000

1500

2000

2500

3000

0.005 0.01 0.015 0.02

Su pport thre s hol d

R
u

n
ti

m
e 

(i
n

 s
ec

o
n

d
s)

MG-PrefixSpan
PrefixSpan
I-P refixSpan

 

 

Figure 2: Performance of the algorithms on data set C10T2S4I2.

(3) once item 80 occurs, then item 569 will occur within 2 days to 6 days, or with average time 3
days, with probability (0.62/22.89) × 100% = 2.7%; within 1 week to 4 weeks, or with average time 2
weeks, with probability (0.84/22.89) × 100% = 3.7%.

(4) once item 80 and 432 occur, then item 94 will occur within 1 days to 3 days, or with average time
2 days, with probability (0.60/1.19) × 100% = 50.4%;

As shown in Table 4, although the patterns are similar to those in Table 3, the time intervals between
successive itemsets are not tighter than those in table 3. Thus users cannot precisely predict when items
569 and 944 occur and 80 occurs again. On the other hand, as shown in Table 5, there is no item
annotation information based on multi-granularities in extracted sequences. Thus, users are not able to
predict how long time item 569, 944 will occur and item 80 will occur again. Furthermore, the pattern
< (80) → (569) > also makes users unable to distinguish periods of item 80 and 569 occurrence.

These results indicate that the multi-granularity sequence patterns mined by algorithm, MG-PrefixSpan,
are more useful than those mined by algorithms, I-PrefixSpan, or PrefixSpan.

5.5 Comparison of the Execution Time

The second test of the three algorithms would compare the run times for different minimum supports.
The comparison is on the five data sets shown in the table 2, where the minimum support threshold
is varied from 0.5% to 2%. Fig.1 to Fig. 5 summarizes the results. It is clearly show that how the
performance of the three algorithms changes as varying of the parameters |C|, |T|, |S| and |I|, and the
differences among the three algorithms.



Mining Temporal Sequential Patterns Based on Multi-granularities 505

 

 0500
1000

1500

2000

2500

3000

0.005 0.01 0.015 0.02

S u pport th re sh ol d
R

u
n

ti
m

e 
(i

n
 s

ec
o

n
d

s)

MG-PrefixSpan
PrefixSpan
I-PrefixSpan

 

Figure 3: Performance of the algorithms on data set C10T2S8I1.

0

1000

2000

3000

4000

5000

6000

7000

8000

9000

10000

0.005 0.01 0.015 0.02

S upport th re sh ol d

R
u

n
ti

m
e 

(i
n

 s
ec

o
n

d
s)

MG-P refixSpan
PrefixSpan
I-P refixSpan

 

Figure 4: Performance of the algorithms on data set C10T4S4I1.

The results in the Fig. 1 Fig. 3 indicate that the processing time of algorithm MG-PrefixSpan is
at different support threshold is no significant difference. The efficiency of MG-PrefixSpan is little less
than that of the I-PrefixSpan. This result matches our expectation, because the MG-PrefixSpan algorithm
does some more complicated computes for the bounds and average interval time in different granularities
than that of the I-PrefixSpan.

On the other hand, the results in the Fig. 1, Fig. 2 and Fig. 3 show that the run time of three
algorithms at different support threshold is also no significant difference. The speed of MG-PrefixSpan
and I-PrefixSpan are little less than that of PrefixSpan. It is correct, because the algorithm PrefixSpan
does not some complicated computes for interval time.

When we see the Fig. 3, Fig. 4 and Fig. 5 we can find that the run time of the three algorithms is
increase rapidly as the minimum support threshold values vary from small to large. It is correct, because
larger number of frequent sequential patterns could be found as the minimum support threshold value
became smaller. However, it is worth note that the algorithm MG-PrefixSpan and I-PrefixSpan take less
time than the PrefixSpan algorithm and the time difference of processing become larger as the minimum
support threshold declines although the MG-PrefixSpan and I-PrefixSpan are more complicated than the
PrefixSpan. In order to find the reason, Let us note the data sets. Fig. 3 test is performed on the data set
C10T2S8I1, where the parameter average length of maximal potentially large sequences S increased from
4(Fig.1) to 8. On average, a potentially frequent sequential pattern consists of 8 transactions, which mean
that although the average number of transactions in a potentially frequent sequential pattern is set to 8, the
number of transactions in a sequence is still set to 10. This situation can not cause the number and length
of frequent sequential pattern increase considerably, because the data set C10T2S8I1 (1.746MB) is as
sparse as C10T2S4I1 (1.733MB); Fig.4 on the data set C10T4S4I1, where the parameter average number



506 N. Li, X. Yao, D. Tian

0
1000
2000
3000
4000
5000
6000
7000
8000
9000

10000
11000
12000
13000
14000
15000

0.005 0.01 0.015 0.02

Support threshold
R

u
n

ti
m

e 
(i

n
 s

ec
on

d
s)

MG-PrefixSpan
PrefixSpan
I-PrefixSpan

 

Figure 5: Performance of the algorithms on data set C20T2S4I1.
 

 

0

1000

2000

3000

4000

5000

6000

10 15 20 25 30

Database  si z e  (i n  K se qu e n ce s)

R
u

n
ti

m
e 

(i
n

 s
ec

o
n

d
s)

Threshold=0.01
Threshold=0.015

 

Figure 6: Scalability test of the algorithm MG-PrefixSpan on data set T2S4I1.

of items per transaction T increased from 2 (Fig.1) to 4, which mean, on average, a sequence consists of
10 transactions, each transaction composed of 4 items. It may result in the number of frequent sequential
pattern increasing significantly; Fig.5 on the data set C20T2S4I1, where the parameter average number
of transactions of per customer C increased from 10 (Fig.1) to 20, which mean, on average, a sequence
consists of 20 transactions, each transaction composed of 2 items. It may result in the length of frequent
sequential pattern increasing greatly. This both situations may cause the number and length of frequent
sequential pattern increase considerably, because the data sets C10T4S4I1 and C20T2S4I1 are denser
than the C10T2S8I1 and C10T2S4I1. When we use PrefixSpan to find all frequent sequential patterns,
much more time would be spent in dealing with these more frequent sequential patterns although many
of frequent sequential patterns may be useless. On the contrary, to the MG-PrefixSpan and I-PrefixSpan,
although adding multiple time granularities or pseudo items to the traditional sequential pattern may
make each of them to produce the several multiple granularities or time-interval patterns and need spend
some time to does some complicated computation, many of them may not be frequent and some patterns
found frequent by PrefixSpan may be infrequent by MG-PrefixSpan and I-PrefixSpan too. This reason
reducing the numbers of frequent sequential patterns is why the MG-PrefixSpan and I-PrefixSpan are
faster than the PrefixSpan.

5.6 Related global performances

The final test is the scalabilities of the MG-PrefixSpan algorithm. Four tests are designed using the
data set C10T2S4I1 and all minimum support thresholds are set to 0.01 and 0.015. In each test, test how
the runtime of the MG-PrefixSpan algorithm scales as one parameter is increased. Fig. 6 shows the result



Mining Temporal Sequential Patterns Based on Multi-granularities 507

 

 010002000
3000

4000

5000

6000

5 10 15 20
Average number of transaction per

customer

R
u

n
ti

m
e(

in
 s

ec
on

d
s)

Threshold=0.01
Threshold=0.015

 

Figure 7: Performance of the algorithms on data set T2S10I1.
 

 01000
2000

3000

4000

5000

6000

2 3 4 5
Ave rage  n u mbe r of tran sacti on

pe r cu stome r

R
u
n
ti

m
e(

in
 s

ec
o
n
d
s)

Threshold=0.01
Threshold=0.015

 

Figure 8: Performance of the algorithms on data set C10S10I1.

of scalability of the algorithm as the data set size grows from 10000 to 30000; Fig.7 shows the result as
the average number of transactions of per customer varies from 5 to 20; Fig. 8 shows the result as the
average number of items per transaction varies from 2 to 5. The results indicate that the MG-PrefixSpan
algorithm has good scalabilities for its runtime increases linearly as each parameter varies from small to
large respectively.

6 Conclusions

In this paper, we have proposed a novel approach for mining multi-granularities sequence patterns
based on PrefixSpan [6], called MG-PrefixSpan. The multi-granularities sequence pattern not only re-
veals what items occur frequently together and in what order, but also the time interval that the next items
occur after the preceding items. We use boundary interval and average time with multi-granularities,
which are derived from the source data, rather than user-predetermined the time interval or only a typical
time to annotate time interval between successive items in the patterns. The performance analysis shows
that MG-PrefixSpan scales up linearly as the size of database, and has a good scalability with respect to
length of sequence and the size of transaction.

The future work along this line of research includes several aspects as follows: (1) validation on
large, real databases; (2) low-level optimizing mechanism of the algorithm.

Acknowledgment

This work was supported by the Department of Education of Shaanxi Province of China under Grant
05JK137 and the Natural Science Foundation of Shaanxi Province of China under Grant 2005F11.



508 N. Li, X. Yao, D. Tian

Bibliography

[1] Constantinescu, Z, Marinoiu, C, Vladoiu, M, “Driving Style Analysis Using Data Mining Tech-
niques,” International Journal Of Computers Communications and Control, Vol.5, No.5, pp. 654-
663, Dec 2010.

[2] Andonie, R, “Extreme Data Mining: Inference from Small Datasets,” International Journal Of
Computers Communications and Control, Vol.5, No.3, pp. 280-291, Sep 2010.

[3] R. Agrawal and R. Srikant, “Mining sequential patterns,” Proc. of the 7th International Conference
on Data Engineering (ICDE’95), pp. 3-14, March, 1995.

[4] R. Srikant and R. Agrawal, “Mining sequential patterns: Generalizations and performance im-
provements,” Proc. of the 5th International Conference on Extending Database Technology, pp.
3-17, March, 1996.

[5] M. Zaki, “An Efficient Algorithm for Mining Frequent Sequences,” Machine Learning, Vol. 40, pp.
31-60, 2000.

[6] J. Pei, J. Han and H. Pinto et al, “Mining Sequential Pattern-Growth: The PrefixSpan Approach,”
IEEE Transactions on Knowledge and Engineering, Vol.16, No.11, pp. 1424-1440, 2004.

[7] H. Manila, H. Toivonen, and A.I. Verkamo, “Discovery of frequent episodes in event sequences,”
Data Mining and Knowledge Discovery, Vol. 1, No. 3, pp. 256-289, 1997.

[8] M.N. Garofalakis, R. Rastogi and K. Shim, “SPIRIT: Sequential Pattern Mining with Regular
Expression Constraints,” Proc. of the 25th International Conference on Very Large Data Bases
(VLDB’99), pp. 223-234, September, 1999.

[9] J. Pei, J. Han and W. Wang, “Constraint-based Sequential Pattern Mining: The Pattern-growth
Methods,” Journal of Intelligent Information Systems, Volume 28, Issue 2, pp. 133-160, April,
2007.

[10] M. Yoshida et al. “Mining sequential patterns including time intervals”, Proc. of SPIE Conf.-
DMKD, pp. 213-220, April, 2000.

[11] Y.-L. Chen, M.-C. Chiang and M.-T. Ko, “Discovering time-interval sequential patterns in sequence
databases,” Expert System with Applications, Volume 25, Issue3, Pp. 343-354, October, 2003.

[12] R. Algawal and R. Srikant, “Fast algorithm for mining association rules in Large Databases.,” Proc.
of the 20th International Conference on Very Large Data bases (VLDB’94), pp. 487-499, September
1994.

[13] Y. Hirate, H. Yamana, “Generalized Pattern Mining with Item Intervals,” Journal of Computers,
Vol.1, No3, pp. 51-60, June, 2006.

[14] F. Giannotti, M. Nanni, and D. Pedreschi. “Efficient Mining of Temporally Annotated Sequences,”
Proc. of the 6th SIAM International Conference on Data Mining, pp. 346-357, April, 2006.

[15] C. Bettini, X.S. Wang and S. Jajodia et al, “Discovering Temporal Relationships with Multiple
Granularities in Time Sequences,” IEEE Transactions on Knowledge and Data Engineering, Vol.
10 (2), pp. 222-237, 1998.