BIG DATA MINING AND ANALYTICS

ISSN 2096-0695 03/06 pp128–136

Volume 1, Number 2, June 2018
DOI: 10.26599/BDMA.2018.9020012

An Improved Hybrid Collaborative Filtering Algorithm Based on Tags

and Time Factor

Chunxia Zhang, Ming Yang , Jing Lv, and Wanqi Yang

Abstract: The Collaborative Filtering (CF) recommendation algorithm, one of the most popular algorithms in
Recommendation Systems (RS), mainly includes memory-based and model-based methods. When performing
rating prediction using a memory-based method, the approach used to measure the similarity between users or
items can significantly influence the recommendation performance. Traditional CFs suffer from data sparsity when
making recommendations based on a rating matrix, and cannot effectively capture changes in user interest. In
this paper, we propose an improved hybrid collaborative filtering algorithm based on tags and a time factor (TT-
HybridCF), which fully utilizes tag information that characterizes users and items. This algorithm utilizes both tag
and rating information to calculate the similarity between users or items. In addition, we introduce a time weighting
factor to measure user interest, which changes over time. Our experimental results show that our method alleviates
the sparsity problem and demonstrates promising prediction accuracy.

Key words: recommendation system; similarity; tag; time factor

recommendation techniques that have been successfully

1 Introduction applied to various fields, e.g., recommendations
With the rapid development of information technology regarding Amazon shopping[3] , music[4] , and the
and networks, the volume of data is increasing and news[5] .
network information transmission is rapid, which brings Existing recommendation algorithms mainly
convenience but also information overload to daily life. include content-based recommendations[6, 7] ,
In this era of information explosion, the development of collaborative filtering[8, 9] , and knowledge-based
methods for extracting information from massive data recommendations[10, 11] . Of these, collaborative filtering
in a timely manner is becoming the focus of increasing has attracted the attention of researchers in both
attention. academia and industry due to its high accuracy and
Recommendation Systems (RS)[1, 2] , a hot research wide range of recommendations. Collaborative filtering
topic, use a kind of information filtering technology utilizes memory-based methods[12–14] , model-based
that mines historical user behavior to excavate methods[15–17] , and hybrid approaches. Memory-
information and identify the individual needs of based methods first calculate the similarity between
users in today’s environment of inforation overload. users/items, obtain the top k-neighbor users/items that
Thus far, researchers have proposed a variety of are most similar to the target user/item, then generate
prediction results based on these neighbors. In contrast,
 Chunxia Zhang, Ming Yang, Jing Lv, and Wanqi Yang
model-based methods learn a model from training data
are with the School of Computer Science and Technology,
Nanjing Normal University, Nanjing 210023, China. E-mail: using machine learning or other techniques based on
1608195294@qq.com; m.Yang@njnu.edu.cn; jinglv njnu@ a user-item rating matrix, and then make predictions.
163.com; yangwq@njnu.edu.cn. The most commonly used model-based methods[18]
 To whom correspondence should be addressed. utilize linear regression models, Bayesian models, and
Manuscript received: 2017-12-27; accepted: 2018-01-03 graph models.
 Chunxia Zhang et al.: An Improved Hybrid Collaborative Filtering Algorithm Based On Tags and Time Factor 129

Rating prediction is an important issue in RS. When 2 Related Work

making predictions based on memory-based methods,
the key step is measuring the similarity between users The recommendation algorithm is the core
(or items), which directly affects the performance component of a recommendation system. With
of the recommendation algorithm. The traditional respect to the different recommendation approaches,
collaborative filtering method makes recommendations recommendation algorithms can be divided into three
based on a rating matrix. However, with the continuing main categories: content-based, collaborative filtering,
expansion of network scales and the explosive growth in and knowledge-based recommendation algorithms, of
the numbers of items and users, traditional collaborative which collaborative filtering is the most popular. The
filtering algorithms inevitably suffer from problems collaborative filtering algorithm, first proposed by
with data sparsity and cold starts due to the extreme Goldberg et al.[22] in 1992, has made significant
sparsity of the rating matrix. Researchers have made progress in the past two decades. This algorithm
many efforts to solve this problem. The authors in constructs a user’s interest model by analyzing
Ref. [19] used clustering and association rules to a user’s historical behavior and then generates
solve the sparsity problem, whereas those in Ref. [20] recommendations for target users. Collaborative-
introduced tag information preset by experts to the filtering recommendation algorithms comprise a large
process of calculating similarity to compensate for the category containing many different types of algorithms,
deficiencies of sparse rating matrixes. The authors in which can generally be divided into memory-based
Ref. [21] obtained the optimal neighbor set by subspace and model-based methods. Memory-based methods
clustering to alleviate data sparsity. To complicate the first calculate the similarity between users (or items),
problem, user interest evolves over time and users obtain the top k-neighbor users (items) that are
do not maintain their inerest in items they previously most similar to the target users (items), then make
liked. Traditional collaborative filtering methods cannot predictions based on these neighbors. Memory-based
capture changing user interest. In this paper, we methods include User-based Collaborative Filtering
propose an improved collaborative filtering algorithm (UserCF)[23, 24] and Item-based Collaborative Filtering
(TT-Hybrid CF) that utilizes tag and rating information (ItemCF)[25, 26] . Model-based methods first construct a
to calculate the similarity between users (or items). In prediction model based on a user-item rating matrix and
addition, it adds a hot-item penalty into the process of then make predictions. In this paper, we mainly focus
calculating users’ similarity to offset the influence of on the memory-based collaborative filtering algorithm.
the hot item in their co-rated items. In addition, the TT- Table 1 lists the main mathematical symbols we use in
Hybrid CF also takes the time factor into account when this article.
measuring user interest, since it changes over time. Our 2.1 User-based collaborative filtering
contributions in this paper are sumarized as follows:
Resnick et al.[27] first proposed UserCF and verified
 We use tag information in the calculation of items’
its validity with the MovieLens dataset in 1994. The
(or users) similarity, and introduce a hot-item penalty
UserCF algorithm is based on the assumption that
into the user-similarity calculation process.
users are likely to purchase the same items if they
 We take into account the time factor when
share similar interests, and it identifies users who share
measuring user interest, which changes with time.
 We propose a new hybrid collaborative filtering
model (TT-HybridCF) that combines improved user- Table 1 Definitions of mathematical symbols.
based collaborative filtering (TT-UserCF) and item- Symbol Definition
based collaborative filtering (TT-ItemCF). U The set of all users. 8u; v 2 U
I The set of all items. 8i; j 2 I
 Our extensive experimental results show that our I uv The set of co-rated items of user u and user v.
proposed method is effective. I u ; Iv The set of items which are rated by user u, user v.
The rest of this paper is arranged as follows. In U ij The set of co-rated users on item i and item j.
The rating of user u on item j,
Section 2, we introduce related work. In Section 3, we Ruj ,Rvj
the rating of user v on item, respectively.
present our method. We report our experimental results The average rating of user u
RN u ,RN v
and analysis in Section 4, and we draw our conclusion and user v respectively.
in Section 5. N(u), N(i) The neighbor set of user u and item i respectively.
 130 Big Data Mining and Analytics, June 2018, 1(2): 128-136

similar interests with a target user as neighbor users. Step three: Predict the unknown rating of the target
Therefore, to make a recommendation to a user, UserCF user u for item i.
first identifies neighbor users by mining the user’s According to the known score of a neighbor set, we
historical behavior. Then, it makes recommendations to can predict the unknown rating of target user u for item
the target user based on the behavior of these neighbor i. Using UserCF, we can predict the rating of user u for
users. The main steps in UserCF are as follows: item i, which can be expressed as shown in Eq. (2).
si m.u; v/.Rvi RN v /
P
Step one: Calculate the similarity between users.
In the UserCF, the measurement of user similarity v2N.u/
PUserCF .Rui / D RN u C P
is a critical step that directly impacts the accuracy si m.u; v/
v2N.u/
of recommendations. There are several commonly
(2)
used similarity measures in RS, including the Pearson
The above equation indicates that we can determine
Correlation Coefficient (PCC), Euclidean distance,
the unknown rating of the target user u for item i based
cosine similarity, and modified cosine similarity
on the weighted sum of the known rating of the user’s
methods, the specific formulas of which are listed in
neighbor set for item i. We calculate sim(u, v) using Eq.
Table 2. Of these, the PCC is the most frequently used
(1) .
and most accurate similarity measure strategy. PCC
calculates the similarity between users based on an item 2.2 Item-based collaborative filtering
set that is co-rated by users. The PCC formula is shown Linden et al.[3] first proposed ItemCF, which assumes
in Eq. (1) below. that a user may like items that are similar to items they
sim.u; v/ D have previously liked. The main steps of ItemCF are as
follows:
RN u /.Rvj RNj /
P
.Ruj Step one: Calculate the similarity between items.
j 2Iuv
r P r P (1) The calculation of the similarity between items i and
.Ruj RN u /2 .Rvj RN v /2 j is based on a user set in which users rate items i and j
j 2Iuv j 2Iuv
together. We again use PCC to calculate the similarity
We use the average score for users u and v to of items, as shown in Eq. (3).
eliminate any difference in different users’ scoring
scales and to ensure similarity accuracy. For example, si m.i; j / D
RN i /.Rvj RNj /
P
some users tend to give high ratings for items, whereas .Rui
u2Uij
other users are more demanding and give lower scores, r P r P (3)
despite the fact that these users may have the same .Rui RN i /2 .Rvj RNj /2
u2Uij u2Uij
interest.
Step two: Find the target user’s neighborhood set. Step two: Find the target item’s neighborhood set.
After Step one, we obtain the user similarity matrix We then sort the obtained similarity matrix based on
and sort this matrix according to the degree of the degree of similarity. Then, we can obtain the top k-
similarity. Then, we can determine the top k-neighbor neighbor items that are most similar to the target item,
users who are most similar to the target user, which we which we label as N.i /.
label as N(u). Step three: Predict the unknown rating of the target
item.
Equation (4) shows the use of ItemCF to predict the
Table 2 Traditional methods for measuring similarities. rating of user u for item i.
si m.i; j /.Ruj RNj /
n
P P
Ruj Rvj
j D1
Cosine sim(u,v)= s j 2N.i /
PI t e mCF .Rui / D RN i C
s
n n
similarity P 2
Ruj
P 2
Rvj
P
j D1 j D1
si m.i; j /
N u /.Rvj N v/ j 2N.i /
P
.Ruj R R
Modified cosine sim(u,v)= r P
j 2Iuv
r P (4)
similarity .Ruj RN u /2 .Rvj R N v /2
j 2Iu j 2Iv This equation indicates that the unknown rating of
N u /.Rvj N v/
P
.Ruj R R
j 2Iuv user u for target item i can be represented by the known
Pearson sim(u,v)= r P r P
correlation .Ruj N u /2
R .Rvj N v /2
R rating of the target item i’s neighbor set. sim(i, j)
j 2Iuv j 2Iuv
coefficient
represents the similarity of items i and j, which is
 Chunxia Zhang et al.: An Improved Hybrid Collaborative Filtering Algorithm Based On Tags and Time Factor 131

calculated using Eq. (3). matrix is very sparse, so they experience sparsity
From the above, we can see that UserCF recommends problems. The authors in Ref. [19] proposed an
hotspots in a group in which members have the improved collaborative filtering algorithm based on
same interest as those of a target user, i.e., the combination of tags and ratings, known as UTR-
UserCF emphasizes socialization. In contrast, ItemCF CF. In this paper, we use this UTR-CF method to
emphasizes individualization, in that it recommends calculate the tag similarity between users or items.
similar items based on a user’s historical behavior. In the MovieLens dataset, the movie tag information
However, both methods suffer from problems of is the genre, e.g., action, adventure, animation, and
sparsity and cold start. the user tag information user comprises demographic
2.3 Hybrid CF characteristics, e.g., fMan, 28, ‘educator’g. Before
calculating the similarity of a tag set, we first transform
When using just one recommended algorithm, the tag set and other text information into digital form to
the resulting recommendation accuracy is not facilitate the modeling process. Assume that two users
very high, since UserCF and ItemCF both have (or two movies) are converted to digital information and
advantages and disadvantages. Therefore, to make are represented as two vectors in m-dimensional space:
up for the shortcomings of individual algorithms, t= (t1 , t2 , ..., tm ), s = (s1 , s2 , ..., sm ). Then, we use the
some scholars have proposed the integration of cosine similarity to calculate the similarity of the tag
different recommendation algorithms when making vectors, as follows:
recommendations. The authors of Ref. [24] integrated Pm
tk sk
UserCF and ItemCF in making recommendations, as si m.t; s/ D qP kD1 P (6)
m 2 m 2
shown in Eq. (5) below. . kD1 k t s
kD1 k /
si m.u; v/.Rvi RN v /
P
3.2 Hot item punishment
0 1
v2N.u/
Rui D  @RN u C AC In UserCF, the traditional method for calculating user
B C
P
si m.u; v/
v2N.u/ similarity is to consider item ratings co-rated by two
0 P
si m.i; j /.Ruj RNj /
1 users without considering the influence of the hot items.
j 2N.i / For example, if two users buy a Xinhua Dictionary,
.1 /  @RN i C A (5)
B C
this does not mean that they have the same interest
P
si m.i; j /
j 2N.i / because most Chinese people have bought this book.
The first part of Eq. (5) comprises the prediction However, if both users buy a book with the title
based on UserCF and the second part on ItemCF. The Machine Learning, we can consider that they have the
parameter is an adjustment parameter, which controls same interest because only those who study this field of
the degree to which this method relies on UserCF and research would buy this book. In summary, if two users
ItemCF. buy hot items, this does not indicate that they have the
same interests. As such, we introduce a weight wi to
3 Our Proposed Model reduce the influence of hot items on user similarity, as
In this paper, we propose an improved collaborative shown in Eq. (7).
filtering method based on tags and a time factor (TT- 1
wi D (7)
Hybrid CF) for RS. In calculating similarity, the TT- lg.1 C Ni /
Hybrid CF algorithm utilizes tag and rating information In this equation, i2I uv , N i represents the number of
to calculate the similarity of users (or items). In users who have rated the item i.
addition, it employs a hot-item penalty when calculating 3.3 Temporal weight
users’ similarity to penalize the influence of a hot item
in their co-rated items. In the prediction phase, TT- Actually, user interest in items fluctuates, but traditional
Hybrid CF takes the time factor into account to measure collaborate filtering algorithms do not take this into
user interest, which changes over time. consideration. The recent behavior of users is more
influential than their earlier behavior. If a user liked
3.1 Calculating similarity with tags
an item last month, this does not mean that he (or she)
Existing collaborative filtering algorithms only use a still likes that item this month. Recent behavior is
rating matrix to calculate similarity, but this rating more likely to indicate a user’s current interest. So,
 132 Big Data Mining and Analytics, June 2018, 1(2): 128-136

we introduce the f (tui ) to represent a time weight, (2) TT-ItemCF

which penalizes earlier behavior and highlights recent In contrast to traditional ItemCF, TT-ItemCF utilizes
behavior, as shown in Eq. (8). tag and rating information to calculate the similarity
f.tui / D 1 exp. tui / (8) between items. We refer to the rating similarity
calculated using Eq. (3) as simrat i ng (i, j) and the tag
where tui is the rating time of user u on item i. We
similarity calculated using Eq. (6) as sim t ag (i, j). Thus,
use the exponential function to indicate user interest,
the integrated similarity is as follows.
which decays over time. The value of f (tui ) increases
with time. A larger tui value indicates that the time of simunif y .i; j/ D ˛simrat i ng .i; j/C.1 ˛/sim t ag .i; j/
the rating behavior is more recent, so the time weight is (12)
greater. In Eq. (12) above, parameter ˛ is an adjustment
Considering the above, we propose three algorithms: parameter, which controls the proportion of rating and
the TT-UserCF, TT-ItemCF, and TT-Hybrid CF tag information in the process of calculating similarity.
algorithms. The TT-ItemCF formula for predicting the rating of
(1) TT-UserCF user u for item iP is as follows:
si munif y .i; j /.Ruj RNj /f .tuj /
In contrast to the traditional UserCF algorithm, TT- j 2N.i /
UserCF adds the hot-item penalty weight wi introduced Rui D RN i C
si munif y .i; j /f .tuj /
P
above to the calculation of the rating similarity between j 2N.i /
users, which reduces the influence of hot items on user (13)
similarity. This improved method for calculating the (3) TT-Hybrid CF
rating similarity of users is shown in Eq. (9) below. The TT-Hybrid CF algorithm employs the following
formula to predict the rating of user u for item i.
simrati ng .u; v/ D 0 P
simunif y .u; v/.Rvi RN v /f .tvi /
1
RN u /.Rvj RNj /wi
P v2N.u/
.Ruj Rui D@RuC
B
AC
C
simunif y .u; v/f .tvi /
P
j 2Iuv
r P r P (9) v2N.u/
.Ruj RN u /2 wi .Rvj RN v /2 wi
si munif y .i; j /.Ruj RNj /f .tuj /
0 P 1
j 2Iuv j 2Iuv

We indicate the similarity calculated using Eq. (6) as j 2N.i /

.1 /@RN i C
B C
si munif y .i; j /f .tuj /
P A
sim tag (u, v) and, using our proposed method, we utilize
j 2N.i /
tag and rating information to compute the similarity (14)
between users or items. This integrated similarity is Equation (14) is used to predict the behavior of user
calculated as shown as Eq. (10). u with respect to item i. The first part of this formula
simunif y .u; v/ D ˇsimrat i ng .u; v/C.1 ˇ/sim t ag .u; v/ is the improved TT-UserCF and the second part is the
(10) improved TT-ItemCF. Parameter  is an adjustment
simunif y (u, v) is the integrated similarity between user parameter, which controls the degree to which the
u and user v. We calculate simt ag (u, v) using Eq. (9) method relies on TT-UserCF and TT-ItemCF.
to obtain the rating similarity between users u and v.
The parameter ˇ is an adjustment parameter, which 4 Experimental
controls the proportions of rating and tag information 4.1 Dataset
when calculating similarity.
In this study, we used the MovieLens1 dataset collected
The calculation process of the TT-UserCF algorithm
by the GroupLens research team at the Uniersity of
in predicting the rating of user u for item i is as follows:
Minnesota. This dataset includes 100 000 rating data
si munif y .u; v/.Rvi RN v /f .tvi /
P
v2N.u/ from 943 users with respect to 1682 films. Each user
Rui D RN u C P  was asked to rate at least 20 films, with a rating
si munif y .u; v/f .tvi /
v2N.u/ range from 1 to 5, whereby the higher the score, the
(11) more interested the user was in the movie. We chose
f (tvi ) penalizes the influence of the past interest of this dataset because it conains the label information
neighbor user v and can thus make more accurate used in our method and because most researchers have
recommendations. conducted experiments based on this dataset.
 Chunxia Zhang et al.: An Improved Hybrid Collaborative Filtering Algorithm Based On Tags and Time Factor 133

4.2 Evaluation metric recommendation. On this basis, we found the optimal

The Mean Absolute Error (MAE), a commonly used value of parameter ˇ to be 0.7, which means that the
evaluation metric in collaborative filtering RS, is used to proportion of rating similarity is 0.3 and the proportion
evaluate the accuracy of a recommendation system. The of tag similarity is 0.7.
MAE is the average absolute error between predicted Experiment 2: Impact of Parameter ˛
and real ratings. The lower the MAE value, the higher In the TT-ItemCF algorithm, we use Eq. (12) to
is the accuracy of system’s recommendation. Assume calculate the similarity between items, in which factor ˛
that the predicted ratings are (Pu1 , Pu2 , Pu3 , ..., PuN ) is very important. This parameter balances the influence
and the real ratings are .ru1 , ru2 , ru3 , ..., ruN ), then the of the tag and rating information. In this experiment, we
MAE is as follows: set the size of the neighbor set to 20 and changed the
P
jPui rui j value of ˛ from 0 to 1 with steps of 0.1. Figure 2 shows
MAE D
i 2N
(15) our experimental testing results of the MAE values for
N the TT-ItemCF with different ˛ values.
4.3 Experimental results Figure 2 shows the impact of parameter ˛ on the
To confirm the superiority of our approach, we MAE results, in which we can see that as the value of ˛
compared the results of our methods with those of increases, the value of MAE decreases until it reaches a
UserCF[27] , ItemCF[3] , HybridCF[24] , and UTR-CF[20] , minimum, after which it increases. We determined the
and we present our detailed analysis in this section. optimal value of parameter ˛ to be 0.4, which means
Experiment 1: Impact of Parameter ˇ that the optimal proportion of the rating similarity is 0.4
In the TT-UserCF algorithm, parameter ˇ in Eq. (10) and that of tag similarity is 0.6.
plays an important role in calculating the similarity Experiment 3: Impact of Parameter 
between users. It controls the proportions of tag and In the TT-HybridCF algorithm,  is an important
rating information in the similarity calculation. In this factor that balances the influence of the TT-UserCF and
experiment, we set the size of the user’s neighbor set to TT-ItemCF algorithms. In this experiment, we set the
20 and changed the value of ˇ from 0 to 1 with steps size of the neighbor set to 10 and changed the value of
of 0.1. Figure 1 shows our experimental results for the  from 0 to 1 with steps of 0.1. Figure 3 shows our
testing set, which shows the MAE values of TT-UserCF experimental testing results of the MAE values for the
for different ˇ values. TT-HybridCF for different  values.
Figure 1 shows the impact on the MAE in the TT- Figure 3 shows the impact of  on the MAE results,
UserCF algorithm, from which we can observe that the in which we can see that as the  value increases, the
MAE value decreases and reaches a minimum, then MAE value decreases until it reaches a minimum, after
increases as the value of ˇ increases. Obviously, we can which it increases. We determined the optimal value
conclude that the value of ˇ affects the accuracy of the of parameter  to be 0.4, which means that the optimal

Fig. 1 Effect of parameter ˇ on MAE of TT-UserCF Fig. 2 Effect of parameter ˛ on MAE values of TT-ItemCF
algorithm. algorithm.
 134 Big Data Mining and Analytics, June 2018, 1(2): 128-136

two other contrastive algorithms

In this experiment, we compared the TT-ItemCF
algoithm with ItemCF and UTR-ItemCF algorithms.
We set the number of neighbor K to range from 10
to 50 with steps of 5. Table 4 lists the MAE values
of all the algorithms. The MAE value varies with the
number of K. We can clearly see that our proposed
TT-ItemCF approach outperforms the other approaches
(ItemCF and UTR-ItemCF).
Experiment 6: Comparison of TT-Hybrid CF with
two other contrastive algorithms
In this experiment, we compared the TT-Hybrid CF
with the UserCF, ItemCF, HybridCF, and UTR-ItemCF
algorithms. We set the numer of neighbor Ks to range
Fig. 3 Effect of parameter  on MAE by the TT-HybridCF from 10 to 50 with steps of 5. Table 5 shows the MAE
algorithm. values of all the algorithms. A downward trend in the
MAE values as the size of the neighbor set increases, so
proportion of TT-UserCF is 0.4 and that of TT-ItemCF we can clearly see that the ac uracy of our method TT-
is 0.6. Hybrid CF outperforms the other three methods. That
Experiment 4: Comparison of TT-UserCF is, taking into consideration both tag information and
algorithm with two other contrastive algorithms the time factor is very important.
In this experiment, we compared the TT-UserCF
algorithm with the UserCF and UTR-UserCF 5 Conclusion
algorithms. We set the number of neighbor K to
In this paper, we proposed an improved hybrid
range from 10 to 50 with steps of 5. Table 3 shows the
MAE values of all the algorithms.
Table 4 MAE values of the ItemCF, UTR-ItemCF, and TT-
Table 3 shows that as the value of K increases, ItemCF algorithms
the MAE value has a downward trend. Obviously,
K ItemCF UTR-ItemCF TT-ItemCF
the number of neighbor users is very important.
10 0.7725 0.7469 0.7091
By taking more neighbors into considerations, the 15 0.7595 0.7404 0.7013
recommendation becomes more accurate. In addition, 20 0.7528 0.7372 0.6973
the performance of our method is better than 25 0.7489 0.7359 0.6972
those of the traditional UserCF and UTR-UserCF 30 0.7466 0.7354 0.6964
algorithms, which means that it is effective to take into 35 0.7450 0.7352 0.6962
consideration the tag information and changing users’ 40 0.7440 0.7351 0.6968
interest. 45 0.7436 0.7352 0.6968
Experiment 5: Comparison of TT-ItemCF with 50 0.7433 0.7354 0.6972

Table 3 MAE values of UserCF, UTR-UserCF, and TT- Table 5 MAE values of the UserCF, ItemCF, HybridCF,
UserCF algorithms UTR-Hybrid, and TT-HybridCF algorithms

K UserCF UTR-UserCF TT-UserCF K UserCF ItemCF HybridCF UTR-Hybrid TT-HybridCF

10 0.7736 0.7666 0.7496 10 0.7740 0.7725 0.7517 0.7308 0.7151
15 0.7617 0.7565 0.7305 15 0.7617 0.7595 0.7445 0.7281 0.7066
20 0.7555 0.7514 0.7227 20 0.7555 0.7528 0.7404 0.7269 0.7027
25 0.7523 0.7488 0.7197 25 0.7523 0.7489 0.7382 0.7267 0.7009
30 0.7505 0.7470 0.7168 30 0.7505 0.7466 0.7370 0.7267 0.6991
35 0.7489 0.7459 0.7145 35 0.7489 0.7450 0.7361 0.7269 0.6990
40 0.7482 0.7453 0.7131 40 0.7482 0.7440 0.7355 0.7270 0.6982
45 0.7475 0.7449 0.7113 45 0.7475 0.7436 0.7352 0.7272 0.6980
50 0.7471 0.7447 0.7119 50 0.7471 0.7433 0.7350 0.7274 0.6980
 Chunxia Zhang et al.: An Improved Hybrid Collaborative Filtering Algorithm Based On Tags and Time Factor 135

collaborative filtering method based on tags and [10] A. Felfernig, Koba4MS: Selling complex products
the time factor (TT-Hybrid CF). In the process of and services using knowledge-based recommender
calculating similarity, we used both tag and rating technologies, in Proc. 7th IEEE Int. Conf. E-Commerce
Technology, Munich, Germany, 2005, pp. 92–100.
information. In addition, the TT-Hybrid CF introduces [11] A. Felfernig and K. Shchekotykhin, Debugging user
a hot-item penlty to the calculation of users’similarity interface descriptions of knowledge-based recommender
to penalize the influence of a hot item among co- applications in Proc. 11st Int. Conf. Intelligent User
rated items. In the process of rating prediction, TT- Interfaces, Sydney, Australia, 2006, pp. 234–241.
[12] J. Wang, A. P. De Vries, and M. J. T. Reinders,
Hybrid CF takes into consideration the users’interest by
Unifying user-based and item-based collaborative filtering
introducing a temporal weight to measure the changing approaches by similarity fusion, in Proc. 29th Annual
user interest over time. Compared with four other Int. ACM SIGIR Conf. Research and Development in
collaborative filtering algorithms (UserCF, HybridCF, Information Retrieval, Seattle, WA, USA, 2006, pp. 501–
ItemCF, and UTR-CF), our proposed TT-Hybrid CF 508.
[13] B. L. Wang, J. H. Huang, L. B. Ou, and R. Wang, A
realizes a great improvement in recommendation
collaborative filtering algorithm fusing user-based, item-
performance. In future work, we will continue to based and social networks, in Proc. 2015 IEEE Int. Conf.
research the problems of sparsity and cold start in Big Data, Santa Clara, CA, USA, 2015, pp. 2337–2343.
traditional collaboraive filtering. [14] H. Koohi and K. Kiani, User based collaborative filtering
using fuzzy c-means, Measurement, vol. 91, pp. 134–139,
Acknowledgment 2016.
[15] X. Y. Liu, C. Aggarwal, Y. F. Li, X. N. Kong, X. Y.
This work was supported by the National Natural Science
Sun, and S. Sathe, Kernelized matrix factorization for
Foundation of China (Nos. 61432008 and 61272222).
collaborative filtering, in Proc. 2016 Siam Int. Conf. Data
References Mining, Miami, FL, USA, 2016, pp. 378–386.
[16] T. Hofmann, Latent semantic models for collaborative
[1] G. Adomavicius and A. Tuzhilin, Toward the next filtering, ACM Trans. Inf. Syst., vol. 22, no. 1, pp. 89–115,
generation of recommender systems: A survey of the state- 2004.
[17] R. Salakhutdinov, A. Mnih, and G. Hinton, Restricted
of-the-art and possible extensions, IEEE Trans. Knowl.
Boltzmann machines for collaborative filtering, in Proc.
Data Eng., vol. 17, no. 6, pp. 734–749, 2005.
[2] L. Y. Lü, M. Medo, C. H. Yeung, Y. C. Zhang, Z. K. Zhang, 24th Int. Conf. Machine Learning, Corvalis, OR, USA,
and T. Zhou, Recommender systems, Phys. Rep., vol. 519, 2007, pp. 791–798.
[18] H. L. Xu, X. Wu, X. D. Li, and B. P. Yan, Comparison
no. 1, pp. 1–49, 2012.
[3] G. Linden, B. Smith, and J. York, Amazon.Com Study of internet recommendation system, (in Chinese), J.
recommendations: Item-to-Item collaborative filtering, Softw., vol. 20, no. 2, pp. 350–362, 2009.
IEEE Int. Comput., vol. 7, no. 1, pp. 76–80, 2003. [19] M. K. Najafabadi, M. N. Mahrin, S. Chuprat, and H. M.
[4] O. Celma and X. Serra, FOAFing the music: Bridging the Sarkan, Improving the accuracy of collaborative filtering
semantic gap in music recommendation, Web Semant., vol. recommendations using clustering and association rules
6, no. 4, pp. 250–256, 2008. mining on implicit data, Comput. Hum. Behav., vol. 67,
[5] F. Hopfgartner, T. Brodt, J. Seiler, B. Kille, A. pp. 113–128, 2017.
Lommatzsch, M. Larson, R. Turrin, and A. Serény, [20] N. Gao and M. Yang, An improved unifying tags and
Benchmarking news recommendations: The CLEF ratings collaborative filtering for recommendation system,
NewsREEL use case, ACM SIGIR Forum, vol. 49, no. 2, (in Chinese), J. Nanjing Nor. Univ. (Nat. Sci. Ed.), vol. 38,
pp. 129–136, 2015. no. 1, pp. 98–103, 2015.
[6] M. Balabanović and Y. Shoham, Combining content-based [21] H. Koohi and K. Kiani, A new method to find neighbor
and collaborative recommendation, Commun. ACM, vol. users that improves the performance of collaborative
40, no. 3, pp. 66–72, 1997. filtering, Expert Syst. Appl., vol. 83, no. C, pp. 30–39,
[7] J. B. Shu, X. X. Shen, H. Liu, B. L. Yi, and Z. L. 2017.
Zhang, A content-based recommendation algorithm for [22] D. Goldberg, D. Nichols, B. M. Oki, and D. Terry, Using
learning resources, Multimed. Syst., doi: 10.1007/s00530- collaborative filtering to weave an information tapestry,
017-0539-8. Commun. ACM, vol. 35, no. 12, pp. 61–70, 1992.
[8] Y. Koren, Factorization meets the neighborhood: A [23] Z. D. Zhao and M. S. Shang, User-based collaborative-
multifaceted collaborative filtering model, in Proc. 14th filtering recommendation algorithms on hadoop, in Proc.
ACM SIGKDD Int. Conf. Knowledge Discovery and Data 3rd Int. Conf. Knowledge Discovery and Data Mining,
Mining, Las Vegas, NV, USA, 2008, pp. 426–434. Phuket, Thailand, 2010, 478–481.
[9] B. Sarwar, G. Karypis, J. Konstan, and J. Riedl, Item- [24] H. Ji, J. F. Li, C. R. Ren, and M. He, Hybrid collaborative
based collaborative filtering recommendation algorithms, filtering model for improved recommendation, in Proc.
in Proc. 10th Int. Conf. World Wide Web, Hong Kong, 2013 IEEE Int. Conf. Service Operations and Logistics,
China, 2001, pp. 285–295. and Informatics, Dongguan, China, 2013.
 136 Big Data Mining and Analytics, June 2018, 1(2): 128-136

[25] S. Y. Wei, N. Ye, S. Zhang, X. Huang, and J. Zhu, Item- Information and Knowledge Management, Atlanta,
based collaborative filtering recommendation algorithm GA, USA, 2001, pp. 247–254.
combining item category with interestingness measure, [27] P. Resnick, N. Iacovou, M. Suchak, P. Bergstrom, and J.
in Proc. 2012 Int. Conf. Computer Science and Service Riedl, GroupLens: An open architecture for collaborative
System, Nanjing, China, 2012, pp. 2038–2041. filtering of Netnews, in Proc. 1994 ACM Conf. Computer
[26] G. Karypis, Evaluation of item-based top-N Supported Cooperative Work, Chapel Hill, NC, USA,
recommendation algorithms, in Proc. 10th Int. Conf. 1994, pp. 175–186.

Chunxia Zhang is currently a master Jing Lv is currently a PhD candidate

student in the School of Computer in the School of Mathematical Sciences,
Science and Technology, Nanjing Normal Nanjing Normal University. Her research
University, and received the BS degree interests include machine learning and
from Jiangsu University of Science and pattern recognition.
Technology in 2015. Her research interests
include machine learning and recommend
system.

Ming Yang received the PhD degree Wanqi Yang received the PhD degree from
from Southeast University at Nanjing in Nanjing University in 2015. Her research
2004. He received the MS degree from interests include multiview learning,
University of Science & Technology of feature selection, multimodal fusion,
China, and BS degree from Anhui Normal abnormal event detection and activity
University, in 1990 and 1987, respectively. recognition. She has published several
He is currently a professor in the school papers in top conferences and journals,
of computer science and technology at e.g., IEEE TNNLS, CVIU. Her work
Nanjing Normal University. His research interests include data currently focuses on multiview dimension reduction, crossview
mining and knowledge discovery, machine learning, pattern correlation analysis and their applications to real-world problems
recognition, and their applications. in image/video analysis.