A PREFIXED-ITEMSET-BASED IMPROVEMENT FOR APRIORI ALGORITHM

Jan Zizka et al. (Eds) : CCSIT, SIPP, AISC, CMCA, SEAS, CSITEC, DaKM, PDCTA, NeCoM - 2016
pp. 287–296, 2016. © CS & IT-CSCP 2016 DOI : 10.5121/csit.2016.60124
A PREFIXED-ITEMSET-BASED
IMPROVEMENT FOR APRIORI
ALGORITHM
Yu Shoujian1
, Zhou Yiyang2
College of computer science and technology,
Donghua University, Shanghai, 201600, China
1
jackyysj@dhu.edu.cn
2
yiyang0203@foxmail.com
ABSTRACT
Association rules is a very important part of data mining. It is used to find the interesting
patterns from transaction databases. Apriori algorithm is one of the most classical algorithms
of association rules, but it has the bottleneck in efficiency. In this article, we proposed a
prefixed-itemset-based data structure for candidate itemset generation, with the help of the
structure we managed to improve the efficiency of the classical Apriori algorithm.
KEYWORDS
Data mining, association rules, Apriori algorithm, prefixed-itemset, hash map
1. INTRODUCTION
With the rapid development of computer technology in various sectors, the data generated by
different industries are becoming more and more, but how to get valuable information from the
big data has become a new problem. Data mining, that is data knowledge discovery, came into
being in this backdrop. Data mining is to excavate the implied, unknown, interesting knowledge
and rules from a large number of data [1]
. Association rules is an important part of data mining, it
was first put forward by R.Agrawal, mainly to solve the customer transaction association rules
between sets of items in the transaction library [2]
. In the following year, R.Agrawal proposed the
most classical algorithm to calculate association rules, that is Apriori algorithm [3]
, which is to
infer the (k+1) – itemsets by the k- itemsets.
However, due to the computing bottleneck of Apriori algorithm when calculating the candidate
set, in recent years there have been many improved algorithms of the traditional Apriori
algorithm from different aspects. Chun-Sheng Z proposed an improved Apriori algorithm based
on classification [4]
. Jia Y improves the algorithm from the aspect of transaction database
partitioning and dynamic itemset planning [5]
. Shuangyue L proposed an improved algorithm
based on the matrix of database to enhance the efficiency of calculating [6]
. Wang P proposed an
optimization method to reduce the search times of the transaction library to improve the
efficiency [7]
. Vaithiyanathan V uses the method of compressing the transactions of the similar
interests in the database to improve the efficiency of the algorithm [8]
. Lin X implements Apriori

288 Computer Science & Information Technology (CS & IT)
algorithm based on Map Reduce to improve the candidate sets of large amounts of data
generation efficiency [9]
. Zhang first analyze the characteristic of the data, that is medical data,
and then combine the characteristics of the data to improved Apriori algorithm [10]
. Wu Huan
proposed an improved algorithm IAA, which adopts a new count-based method to prune
candidate itemsets and uses generation record to reduce total data scan amount [11]
. Wang Yuan
proposes an improved item constrain association rules mining algorithm, which improves
traditional algorithm in two aspects: trimming frequent itemsets and calculating candidate
itemsets [12]
. Lin Ming-Yen proposes three algorithms, named SPC, FPC, and DPC, to investigate
effective implementations of the Apriori algorithm in the MapReduce framework [13]
. Chai Sheng
proposes a novel algorithm so called Reduced Apriori Algorithm with Tag (RAAT), which
reduces one redundant pruning operations of C2 [14]
.
This article will be focus on the two concrete steps of classical Apriori algorithm, namely
connecting step and the pruning step, using a new prefix-itemset-based storage, combining the
fast lookup feature of hash tables to improve the efficiency. This paper will first describe the
classical Apriori algorithm and its shortcomings, then specifically describe the improvements,
and finally introduce the comparisons of efficiency of classical Apriori algorithm and improve
Apriori algorithm on specific data sets.
2. APRIORI ALGORITHM
2.1. Apriori algorithm introduction
Apriori algorithm is a classical algorithm for frequent itemset mining association rules, the basic
idea of the algorithm is to use an iterative approach layer by layer to find the frequent. The
algorithm will first obtain k-itemsets, and then use the k- itemsets to explore (k+1)-itemsets. First,
let’s introduce the priori knowledge of frequent itemsets, which is, any subset of a frequent
itemset is also a frequent itemset. Apriori algorithm uses the prior knowledge of frequent
itemsets, first to find the collection of frequent 1-itemsets, denoted L1. Then use the 2-itemsets of
L1 to get L2, and then L3, and so on, until you cannot find the frequent k-itemsets. Apriori
algorithm mainly consists of the following three steps:
(1) Connecting step: connecting k- frequent itemsets to generate (k+1)-candidate sets, denoted by
Ck+1. The connect condition of the connecting step is that the two k-itemsets have the same
first (k-1) items and different k-th items. Denote li[j] is j-th item of li, the condition is:
lଵሾ1ሿ = lଶሾ1ሿ ∧ lଵሾ2ሿ = lଶሾ2ሿ ∧ … … ∧ lଵሾk − 1ሿ = lଶሾk − 1ሿ ∧ lଵሾkሿ ≠ lଶሾkሿ
In which l1 and l2 are k-item subset of the set collection Lk, l1[k] ≠ l2[k] is to ensure not to
generate duplicate k- itemsets. Itemsets generated by the l1 and l2 connection as follows:
{lଵሾ1ሿ, lଵሾ2ሿ, lଵሾ3ሿ, … … , lଵሾkሿ, lଶሾkሿ}
(2) Pruning step: To pick out the true frequent itemsets Lk+1 from the candidate set Ck+1. Because
the candidate set Ck+1 is the superset of the true frequent itemsets Lk+1. According to the
nature of Apriori: any subsets of frequent set must also be frequent, that is any (k-1)- items
subsets of k-items must also be frequent. With this property we can find out if the k- items
subsets of Ck+1 are in Lk, if not, then remove the candidate (k + 1) - itemset is removed from
the Ck+1.
(3) Counting step: scanning the database, accumulate the number of candidates appearing in the
database. If the appear times of a candidate is less than the given minimum support threshold,
the candidate itemset will be removed.

Computer Science & Information Technology (CS & IT) 289
2.2. Shortage of Apriori algorithm
Apriori is one of the most classical algorithms for mining association rules, but it also has the
shortage of low efficiency. The time Apriori algorithm consumes lies mainly in the following
three aspects:
(1) In connection step, when connects k-itemsets to generate (k+1)-itemsets, it compares too
many times to determine if the itemsets meets the connection conditions. When Lk has m k-
itemsets, the time complexity of the connection step is Ο(k*m^2).
(2) In the pruning step, when determine if a subset of candidate set Ck+1 is in the frequent set Lk,
the best situation is to simply scan once to get the result, while the worst-case is that is needs
to scan k times to find that the k-th subset of Ck+1 is not in the Lk. So the average times need
to scan and compare the Lk is |Ck+1| * |Lk| * k / 2.
(3) In counting step, when accumulate the support times of itemsets in Ck+1, we need to scan the
database for |Ck+1|times.
Taking into account these three aspects of time-consuming steps of classical Apriori algorithm,
this article presents an improved Apriori algorithm based on prefix-itemset.
3. IMPROVED APRIORI ALGORITHM
3.1. Improved Apriori algorithm
In 1.2 we have analyzed the shortcomings of classical Apriori algorithm, so its improvements
also focus on the three steps mentioned in 1.2. Since the records are already sorted by the
dictionary, therefore the candidate set generated by Apriori algorithm is ordered.
(1）Prefixed-itemset-based storage
In the improved algorithm we proposed a new method to store the itemsets. For each itemset in
Lk, we use a structure similar to Map <key, value> to store them, in which we save the forward
(k-1)- item content as the key while the last item content as the value. After having all the
itemsets saved in the new format, we group all the itemsets with the same key and store the union
of their values as the new value.
For example: the database is shown in Table 1and the minimum support is 2.
Table 1 database
TID Itemset
T1 A,B,E
T2 B,D
T3 B,C
T4 A,B,D
T5 A,C
T6 B,C
T7 A,C
T8 A,B,C
T9 A,B,C,E

The traditional Apriori algorithm will scan the database to obtain the times each item appears in
the database, to form the 1- itemsets, and then to generate the 2-itemsets that meets the minimum
support, that is 2. Here is the content generated by the classical Apriori algorithm.
Table 2 classical Apriori algorithm
1-itemset 2- itemset
Item Count Item Count
A 6 AB 4
B 7 AC 4
C 6 AE 2
D 2 BC 4
E 2 BD 2
BE 2
While Table 3 shows how we store the itemsets with the prefix-itemset-based storage.
Table 3 prefix-itemset-based storage
Prefixed-key Value
1-itemset NULL {A, B, C, D, E}
2-itemset A {B, C, E}
B {C, D, E}
As shown in Table 3, 1- itemset has only one item, so the key of 1- itemset is NULL. Besides, we
can infer the length of the itemset from the length of the key because the length of the value of
the key stores all the items in the itemset but the last item.
(2) Prefixed-itemset-based connecting step
After the establishing of prefix-itemset-based storage, when we have to generate (k+1)- itemset
by connecting the two k- itemsets, we can simply combine two different items in the value, and
then generate new itemset with the key. For example, when connecting the 2-itemset with the
prefix-key of A in Table 3, we can generate the 3- itemset by combine the value and get the result
as {{B, C},{B, D},{C, D}}.
(3) Prefixed-itemset-based pruning step
In chapter 1.1 we know that (k+1)- itemsets are generate from two k- itemsets, and if any k-
itemset subset of the (k+1)-itemset does not exist in Lk, then we have to remove the (k+1)-
itemset from Ck+1.
Theorem: If we generate a (k+1)-itemset by connecting two k-itemsets, l1 and l2, and one k-
itemset of all the k-itemset subset does not exist in Lk, then the k-itemset subset must contains
both l1[k] and l2[k].

Prove: Assume l1 and l2 are both k-itemset, and the (k+1)-itemset generated by connecting l1 and
l2 is {lଵሾ1ሿ, lଵሾ2ሿ, lଵሾ3ሿ, … … , lଵሾkሿ, lଶሾkሿ}. If the k- itemset dose not contains both l1[k] and l2[k],
then the possible options are {lଵሾ1ሿ, lଵሾ2ሿ, lଵሾ3ሿ, … … , lଵሾkሿ}and {lଵሾ1ሿ, lଵሾ2ሿ, lଵሾ3ሿ, … … , lଶሾkሿ}, that
is l1 and l2, and both l1 and l2 come from Lk, so if the k-itemset does not belong to Lk, then it must
contains both lଵሾkሿ and lଶሾkሿ.
So in prefixed-itemset-based pruning step, we can simply consider the subset of (k+1)- itemset
which contains both the last two items. With the example from Table 3, we can get the result as
follow.
Table 4 pruning step
Subset of 3-itemset If belong to L2
B,C yes
B,E yes
C,E no
C,D no
C,E no
D,E no
As shown in Table 4, only {{B,C},{B,E}} are possible 2-itemset subsets, plus the corresponding
prefix-key, that is {{A,B,C},{A,B,E}}, namely the candidate set Cଷ.
After the pruning step, we have to scan the database to accumulate the times the itemset appears.
After accumulating the times of itemsets after pruning step, we can find that both {A,B,C} and
{A,B,E} meet the minimum support, and then we add them to the prefix-itemset-based storage as
follows.
Table 5 3-itemset storage
prefix-key Value
3-itemset A,B {C,E}
3.2. Algorithm
The algorithm is described as follow:
Input：Database D，minimum support min_sup
Output：frequent itemsetsL
1) Lଵ=1-itemset of D
2) Map<String[],String[]> map;
3) Import L1 to map, set the key as null，value as the union of items in L1
4) for(k=2;L୩ିଵ ≠ ϕ;k++){
5) C୩=pre_apriori_gen(map,k-2);
6) count the appear times of every itemset of C୩, L୩={cϵC୩|c.count>min_sup}
7) }
8) Return L୩;
procedurepre_apriori_gen(map:Map<String[],String[]>;k:int)
1) for each key in map{

2) if(key.length()==k){
3) c:=key plus two items from value
4) if(map.containsKey(c[0:k])){
5)
6)
7)
8) }
9） else continue；
10) }
11) Return C୩
4. EXPERIMENT AND RESULTS
The data of the experimental is a total of 120000 patient
in Ruijin Hospital, the data record
experimental machine is configured to Core i5
memory.
This experiment compares the classical Apriori
algorithm from two aspects, one is to compare the operation efficiency with fixed total tests and
variable minimum support, the other is to compare the operation efficiency with fixed minimum
support and variable total tests.
The result of the first experiment
Figure 1. Time
The picture above shows the time consuming
and variable min_sup. We can infer that the less min_sup is, the more operation efficiency the
improved algorithm improves. And when the min_sup increases to a certain point, the classical
Apriori algorithm and the improved algorithm are of the same efficiency.
Computer Science & Information Technology (CS & IT)
if(key.length()==k){
c:=key plus two items from value
if(map.containsKey(c[0:k])){
If( any (k+1)- itemset subset belong to (key,value)){
put c into Ck
}
ESULTS
The data of the experimental is a total of 120000 patients with diabetes clinical prescription data
the data records the prescription drug number per user per visit. This
experimental machine is configured to Core i5 2.7GHz 8GB processor, 1866MHz LPDDR3 Intel
compares the classical Apriori algorithm and the prefixed-itemset
from two aspects, one is to compare the operation efficiency with fixed total tests and
, the other is to compare the operation efficiency with fixed minimum
experiment is as follows.
Figure 1. Time consuming with variable min_sup
The picture above shows the time consuming of the two algorithms when given fixed total test
and the improved algorithm are of the same efficiency.
itemset subset belong to (key,value)){
with diabetes clinical prescription data
per visit. This
2.7GHz 8GB processor, 1866MHz LPDDR3 Intel
itemset-based
from two aspects, one is to compare the operation efficiency with fixed total tests and
, the other is to compare the operation efficiency with fixed minimum
of the two algorithms when given fixed total test

Table 6 improvements under variable min_sup
Min_sup
(total 12w)
Classical Apriori
Time(ms)
600 210696
1200 53822
1800 26317
2400 19359
3000 15508
3600 12393
4200 9017
4800 5705
5400 4868
Table 6 shows the specific operation time of the
Apriori algorithm and the comparison
The results of the second experiment are as follows.
Figure 2. Time
The picture above shows the time consuming of the two algorithm
and variable total test. And we can tell that when the total test becomes larger, the
become more obvious.
Table 6 improvements under variable min_sup
Classical Apriori
Time(ms)
Improved Apriori
Time(ms)
Improvement (%)
210696 25192 81.44%
53822 11648 68.63%
26317 7614 51.88%
19359 7127 38.99%
15508 6753 56.45%
12393 5842 52.86%
9017 5424 39.85%
5705 5175 9.29%
4868 5161 -6.02%
Table 6 shows the specific operation time of the classical Apriori algorithm and the
comparison between them.
The results of the second experiment are as follows.
Figure 2. Time consuming with variable total test
above shows the time consuming of the two algorithms when given fixed min_sup
and variable total test. And we can tell that when the total test becomes larger, the improvements
293
and the improved
when given fixed min_sup
improvements

Table 7 improvements under variable total test
Total tests
(min_sup%
=2%)
Classical
Apriori
Time(ms)
Improved
Apriori
Time(ms)
Improvement
(%)
10k 1686 390 76.87%
20k 2839 695 75.52%
30k 4088 1229 69.94%
40k 5729 1846 67.78%
50k 8141 2409 70.41%
60k 9833 3197 67.49%
70k 12848 3630 71.75%
80k 13004 4339 66.63%
90k 16007 5442 66.00%
100k 17438 5588 67.96%
Table 7 shows the specific operation time of the two algorithm and we can learn from the table
that when the min_sup is fixed to 2% of the total test, the improvement rate is about 70%.
Experiments have shown that the prefix-itemset-based Apriori algorithm is effective and feasible.
4. SUMMARY
In this paper, we described the Apriori algorithm specifically, and pointed out some limitations of
the classical Apriori algorithm during the two steps of the algorithm, namely the connection and
the paper cutting steps, and proposed the method of prefixed-itemset-based data storage and the
improvements based on it. With the help of prefixed-itemset-based data storage, we managed to
finish the connecting step and the pruning step of the Apriori algorithm much faster, besides we
can store the candidate itemsets with smaller storage space. Finally, we compare the efficiency of
classical Apriori algorithm and improve Apriori algorithm on the aspect of support degree and
the total number, and the experimental results on both aspects proved the feasibility of the
prefixed-itemset-based algorithm.

REFERENCES
[1] Han J, Kamber M, Pei J, et al. Data mining.Concepts and techniques. 3rd ed[J]. San Francisco, 2001,
29(S1):S103–S109.
[2] RakeshAgrawal T. Imielinski,andArun Swami. Miningassociationrules between setsof items inlarge
databases[J]. Inproceedings Ofthe AcmSigmod Conference, 1993:207--216.
[3] Agrawal, Rakesh, and RamakrishnanSrikant. "Fast algorithms for mining association rules." Proc.
20th int. conf. very large data bases, VLDB. Vol. 1215. 1994.
[4] Chun-Sheng Z, Yan L. Extension of local association rules mining algorithm based on apriori
algorithm[C]//Software Engineering and Service Science (ICSESS), 2014 5th IEEE International
Conference on. IEEE, 2014: 340-343.
[5] Jia Y, Xia G, Fan H, et al. An Improved Apriori Algorithm Based on Association Analysis[C]//2012
Third International Conference on Networking and Distributed Computing. 2012.
[6] Shuangyue L, Li P. Analysis of Coal Mine Hidden Danger Correlation Based on Improved A Priori
Algorithm[C]//Intelligent Systems Design and Engineering Applications, 2013 Fourth International
[7] Wang P, An C, Wang L. An improved algorithm for Mining Association Rule in relational
database[C]//Machine Learning and Cybernetics (ICMLC), 2014 International Conference on. IEEE,
2014, 1: 247-252.
[8] Vaithiyanathan V, Rajeswari K, Phalnikar R, et al. Improved apriori algorithm based on selection
criterion[C]//Computational Intelligence & Computing Research (ICCIC), 2012 IEEE International
[9] Lin X. MR-Apriori: Association Rules algorithm based on MapReduce[C]//Software Engineering and
Service Science (ICSESS), 2014 5th IEEE International Conference on. IEEE, 2014: 141-144.
[10] Zhang, Wenjing, Donglai Ma, and Wei Yao. "Medical Diagnosis Data Mining Based on Improved
Apriori Algorithm." Journal of Networks 9.5 (2014): 1339-1345.
[11] Wu, Huan, et al. "An improved apriori-based algorithm for association rules mining." Fuzzy Systems
and Knowledge Discovery, 2009. FSKD'09. Sixth International Conference on. Vol. 2. IEEE, 2009.
[12] Wang, Yuan, and Lan Zheng. "Endocrine Hormones Association Rules Mining Based on Improved
Apriori Algorithm." Journal of Convergence Information Technology 7.7 (2012).
[13] Lin, Ming-Yen, Pei-Yu Lee, and Sue-Chen Hsueh. "Apriori-based frequent itemset mining algorithms
on MapReduce." Proceedings of the 6th international conference on ubiquitous information
management and communication. ACM, 2012.
[14] Chai, Sheng, Jia Yang, and Yang Cheng. "The research of improved Apriori algorithm for mining
association rules." Service Systems and Service Management, 2007 International Conference on.
IEEE, 2007.
[15] Han J, Pei J, Yin Y. Mining Frequent Patterns without Candidate Generation[J]. Proceeding of
AcmSigmod International Conference Management of Data, 1999, 29(2):1-12.

AUTHORS
Yu Shoujian, the vice professor, main research direction: Web services, enterprise application integration,
database and data warehouse;
Zhou Yiyang, master, the main research direction: data mining, machine learning

A PREFIXED-ITEMSET-BASED IMPROVEMENT FOR APRIORI ALGORITHM

More Related Content

What's hot (20)

Viewers also liked (10)

Similar to A PREFIXED-ITEMSET-BASED IMPROVEMENT FOR APRIORI ALGORITHM (20)

Recently uploaded (20)

A PREFIXED-ITEMSET-BASED IMPROVEMENT FOR APRIORI ALGORITHM