=Paper=
{{Paper
|id=Vol-2068/wii5
|storemode=property
|title=Improved Listwise Collaborative Filtering with High-Rating-Based Similarity and Temporary Ratings
|pdfUrl=https://ceur-ws.org/Vol-2068/wii5.pdf
|volume=Vol-2068
|authors=Yoshiki Tsuchiya,Hajime Nobuhara
|dblpUrl=https://dblp.org/rec/conf/iui/TsuchiyaN18
}}
==Improved Listwise Collaborative Filtering with High-Rating-Based Similarity and Temporary Ratings==
https://ceur-ws.org/Vol-2068/wii5.pdf
Improved Listwise Collaborative Filtering with High-
Rating-Based Similarity and Temporary Ratings
Yoshiki Tsuchiya Hajime Nobuhara
University of Tsukuba University of Tsukuba
Tsukuba, Japan Tsukuba, Japan
tsuchiya@cmu.iit.tsukuba.ac.jp nobuhara@iit.tsukuba.ac.jp
ABSTRACT [7,10] and listwise ranking-oriented [2]. Pairwise ranking-
In this paper, we make two proposals to speed up listwise oriented CF predicts the order of pairs of items but requires
collaborative filtering and improve its accuracy. The first is large computation time.
to speed up computation by only using a subset of the rating
information (the high ratings). The second is to improve
accuracy using temporary ratings that estimate the rating
scores that neighboring users are not rating. Experiments
using MovieLens datasets (1M and 10M) demonstrate that
these proposals effectively reduce computation time about
1/50 and improve accuracy 2.22% compared with ListCF, a
well-known listwise collaborative filtering algorithm.
Author Keywords
Recommender system; Ranking-oriented collaborative
filtering;
Figure 1. Collaborative filtering classification
ACM Classification Keywords
• Information systems, Collaborative filtering In contrast, listwise ranking-oriented CF predicts the order
INTRODUCTION of the complete list of items. Although this produces better
In recent years, due to the development of the Web, accuracy than a typical pairwise CF algorithm, calculating
recommender systems have become increasingly important the required similarities is time-consuming and there is
in various situations; many researchers are now focusing on room to improve the ranking accuracy. In this paper, we
recommendation technologies and systems [2,6,7,9,10]. propose an efficient listwise ranking-oriented CF algorithm
Collaborative filtering (CF) is a widely used that is both faster and has higher ranking accuracy.
recommendation algorithm that is based on the similarity The proposed method implements two improvements. First,
between users or items, as calculated from a user and rating when calculating the similarity between users, it only
matrix. Various CF algorithms have been proposed, and considers the highest-rated items, greatly speeding up the
they can be divided into two types: rating-oriented [6,9] and calculation. Second, it introduces temporary ratings when
ranking-oriented [2,7,10], as shown in Fig. 1. Rating- making ranking predictions. Experimental comparisons
oriented CF algorithms, such as item-based CF [9], predict using MovieLens 1M (6,040 users, 3,952 movies,
the ratings of items that have not been evaluated by users 1,000,209 ratings) and 10M (71,567 users, 10,681 movies,
and make recommendations by calculating the similarity 10,000,054 ratings) confirm that the proposed method
between users or items. On the contrary, ranking-oriented reduces about 1/50 computation time of similarity and
CF uses user similarity to predict the item ranking and improves 2.22% ranking accuracy than a conventional CF
recommends items based on this. We will focus on this algorithm.
method due to its performance. Ranking-oriented CF can be
further divided into two types: pairwise ranking-oriented RELATED WORK
Overview of ListCF
© 2018. Copyright for the individual papers remains with the authors.
Copying permitted for private and academic purposes.
In this section, we give an overview of ListCF [2], a well-
WII'18, March 11, Tokyo, Japan. known ranking-oriented listwise CF algorithm. ListCF
proceeds in two phases, first calculating the similarities
between users and then predicting ranks for the target user’s
unrated items. The first phase is based on a probability
distribution of item permutations, calculated by combining
the Plackett–Luce [8] and top-k probability [1] models and
finding each user’s neighboring users.
Similarity calculation
In ListCF, the similarity of a pair of users 𝑢 and 𝑣 is
calculated based on a probability distribution of item
permutations for each user, calculated using the Plackett–
Luce model [8], which is a representative permutation
probability model. The flow of similarity calculation is
shown on the left of Fig. 2 and Fig. 3. Let the set 𝐼 =
{𝑖1 , 𝑖2 , ⋯ , 𝑖𝑛 } of items 𝜋 𝑖 = (𝜋1𝑖 , 𝜋2𝑖 , ⋯ , 𝜋𝑛𝑖 ) (∈ 𝐼 𝑛 ) be an
ordered list where 𝜋𝑗𝑖 ∈ 𝐼 and 𝜋𝑗𝑖 ≠ 𝜋𝑘𝑖 if 𝑗 ≠ 𝑘 , and let
Ω𝐼 (⊂ 𝐼 𝑛 ) be the set of all possible permutations of 𝐼. Given
the item ratings (𝑟𝜋𝑖 , 𝑟𝜋𝑖 , ⋯ , 𝑟𝜋𝑛𝑖 ) (e.g., on a real interval
1 2
in the case of MovieLens [1,5]), where 𝑟𝜋𝑖 is the rating Figure 3. Procedure for similarity calculation of conventional
𝑗 ListCF.
score of 𝜋𝑗𝑖 , the probability of 𝜋 𝑖 , 𝑃(𝜋 𝑖 ), is defined using an
by users 𝑢 and 𝑣, and 𝑃𝑢 and 𝑃𝑣 (∈ [0,1]) as the probability
increasing and strictly positive function 𝛷(∙) ≥ 0 as 𝐼
follows: distributions over 𝑔𝑘𝑢,𝑣 (⊂ Ω𝐼𝑢,𝑣 ) calculated by Eq. (3) using
the users’ rating scores. The similarity score is now
𝑛 𝛷 (𝑟𝜋𝑖 ) obtained from the Kullback–Leibler (KL) divergence [5]
𝑖) 𝑗
𝑃(𝜋 =∏ (∈ [0,1]), (1) calculated from 𝑃𝑢 and 𝑃𝑣 . Given a pair of users 𝑢 and 𝑣,
𝑛
𝑗=1 ∑𝑘=𝑗 𝛷 (𝑟𝜋𝑘
𝑖) the KL divergence of 𝑃𝑢 and 𝑃𝑣 is defined as
where the function is defined as 𝛷(𝑟) = 𝑒 𝑟 . However, this 𝑃𝑢 (𝑔)
𝐷𝐾𝐿 (𝑃𝑢 ∥ 𝑃𝑣 ) = ∑ 𝑃𝑢 (𝑔) 𝑙𝑜𝑔2 ( ). (4)
requires us to consider 𝑛! different permutations of the 𝑛 𝑃𝑣 (𝑔)
𝐼
items, which would take a long time to compute. To speed 𝑔∈𝑔𝑘𝑢,𝑣
up the computation, the top-k probability model 𝑔𝑘 [1] is
The KL divergence is asymmetric, but ListCF defines the
introduced as follows:
similarity 𝑠(𝑢, 𝑣)(∈ (−∞, 1 ]) of users 𝑢 and 𝑣 to be
𝑔𝑘 (𝑖1 , 𝑖2 , ⋯ , 𝑖𝑘 ) = symmetric as follows:
𝑛! 1
{𝜋 𝑙 |𝜋 𝑙 ∈ Ω𝐼 , 𝜋𝑗𝑙 = 𝑖𝑗 , 𝑗 = 1, ⋯ , 𝑘, 𝑙 = 1, ⋯ , } 𝑠(𝑢, 𝑣) = 1 − [𝐷𝐾𝐿 (𝑃𝑢 ∥ 𝑃𝑣 ) + 𝐷𝐾𝐿 (𝑃𝑣 ∥ 𝑃𝑢 )]. (5)
(𝑛 − 𝑘)! 2
(⊂ Ω𝐼 ), (2) If the set 𝐼𝑢,𝑣 only includes a few items, the similarity will
be high, so this is relaxed by multiplying by the similarity
and the probability of the top-k permutation is calculated as
function by min{𝐼𝑢,𝑣 ⁄𝑐 , 1}, where 𝑐 is a threshold. Each
𝑘
𝛷 (𝑟𝜋𝑗 ) user’s neighboring users can then be found from the
𝑃(𝑔𝑘 (𝑖1 , 𝑖2 , ⋯ 𝑖𝑘 )) = ∏ (∈ [0,1]), similarities calculated using Eqs. (3)–(5).
∑𝑛 𝛷(𝑟𝜋𝑙 )
𝑗=1 𝑙=𝑗
Ranking prediction
∀𝑗 = 1, ⋯ , 𝑘 ∶ 𝜋𝑗 = 𝑖𝑗 . (3) The flow of ranking prediction is shown on the left of Fig. 4.
Let U be the set of users, 𝑁𝑢 (⊂ 𝑈) be the set of the user 𝑢’s
Let 𝑔𝑘𝐼 (⊂ Ω𝐼 ) be the set of top-k permutations of 𝐼, and let neighbors and 𝑇𝑢 (⊂ 𝐼 ∖ 𝐼𝑢 ) be the set of items whose ranks
the probabilities of these permutations form the probability are to be predicted. Let 𝑃̂𝑢 (∈ [0,1]) be the probability
distribution. Then, define 𝐼𝑢,𝑣 (⊂ 𝐼) as the set of items rated 𝑇
distribution of the top-k permutations 𝑔𝑘𝑢 (⊂ Ω𝑇𝑢 ) of 𝑇𝑢 ,
written as
𝜑𝑢,𝑔
𝑃̂𝑢 (𝑔) = , (6)
∑𝑔′ ∈𝑔𝑇𝑢 𝜑𝑢,𝑔′
𝑘
𝑇
where {𝜑𝑢,𝑔 |∀𝑔 ∈ 𝑔𝑘𝑢 } are unknown variables assigned to
the top-k permutations. In ListCF, the cross entropy is used
as a loss function for prediction. Consider a target user 𝑢,
for whom you want to rank a set of items 𝑇𝑢 , and a
neighboring user 𝑣 ∈ 𝑁𝑢 . Let the set of items rated by 𝑣 be
𝑇
𝐼𝑣 , with 𝑇𝑢,𝑣 = 𝑇𝑢 ∩ 𝐼𝑣 , and 𝑔𝑘𝑢,𝑣 (⊂ Ω𝑇𝑢,𝑣 ) be the set of
Figure 2. Similarity calculation: conventional ListCF (left), top-k permutations of 𝑇𝑢,𝑣 . The cross entropy is calculated
proposed method (right).
Figure 5. Procedure of similarity calculation of proposed
method.
propose a method of calculating user similarity by focusing
Figure 4. Ranking prediction: conventional ListCF (left), only on high ratings. The flow of similarity calculation is
proposed method (right). shown on the right of Fig. 2 and Fig.5. Let 𝐻𝑢 (⊂ 𝐼𝑢 ) be the
𝑇 set of items that were rated highly by the target user 𝑢, and
using probability distributions 𝑃̂ ′𝑢 and 𝑃′𝑣 over 𝑔𝑘𝑢,𝑣 as 𝐻𝑢,𝑣 (⊂ 𝐼𝑢,𝑣 ) be the set of items that both 𝑢 and 𝑣 rated
follows: highly. Given the user and rating matrix, we define the
similarity between 𝑢 and 𝑣 as follows:
𝐸(𝑃̂ ′𝑢 , 𝑃′𝑣 ) = − ∑ 𝑃′ 𝑣 (𝑔) log 2 𝑃̂ ′𝑢 (𝑔) . (7)
𝑇
𝑔∈𝑔𝑘 𝑢,𝑣 |𝐻𝑢,𝑣 |
𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦(𝑢, 𝑣) = (∈ [0,1]), (12)
|𝐻𝑢 |
ListCF makes predictions by minimizing the following
cross entropy weighted sum: where 𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦(𝑢, 𝑣) is asymmetric. We define
𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦(𝑢, 𝑣) = 0 if 𝑢 = 𝑣 not to add him/herself to
arg min ∑ 𝑠(𝑢, 𝑣) ∙ 𝐸(𝑃̂ ′𝑢 , 𝑃′ 𝑣 ), (8) neighborhood users when predicting each user’s ranking. In
𝜑𝑢
𝑣∈𝑁𝑢 conventional ListCF, based on the top-k probability model,
𝑇 the time complexity is 𝑛!⁄(𝑛 − 𝑘)! , which is smallest
s. t ∀𝑔 ∈ 𝑔𝑘𝑢 ∶ 𝜑𝑢,𝑔 ≥ 0. (i.e., 𝑛 ) when 𝑘 = 1 . As the proposed method’s time
The objective function 𝐹(𝜑𝑢 ) in Eq. (8) is then transformed complexity is 𝑛, it is never worse than that of conventional
as follows, using Eqs. (5) and (7): ListCF, and often better. Changing how the high ratings are
determined changes the ranking accuracy, as explained in
𝐹(𝜑𝑢 ) = ∑ 𝑠(𝑢, 𝑣) ∙ the experimental section.
𝑣∈𝑁𝑢 Improving ranking accuracy using temporary ratings
We also propose to improve the ListCF’s ranking accuracy
by introducing temporary ratings. In ListCF, the cross
∑ 𝑃′ 𝑣 (𝑔) [log 2 ( ∑ 𝜑𝑢,𝑔′ ) − log 2 (𝜑𝑢,𝑔 )] (9) entropy in Eq. (7) is calculated from the probability
𝑇 𝑇 𝑇
𝑔∈𝑔𝑘 𝑢,𝑣 𝑔′ ∈𝑔𝑘 𝑢,𝑣 distribution of 𝑔𝑘𝑢,𝑣 , the set of permutations of 𝑇𝑢,𝑣 .
However, optimizing the objective function may fail if 𝑇𝑢,𝑣
Equation (9) is optimized by the gradient descent method.
Partially differentiating F with respect to 𝜑𝑢,𝑔 , we obtain has too few elements, e.g., if it only includes one element
and that item received a low rating from the neighboring
𝜕𝐹 user 𝑣 . In this case, despite the item’s low rating,
= optimization generates a high item rank, which is unhelpful
𝜕𝜑𝑢,𝑔
for user 𝑢. Fig. 6 shows how the probability distribution for
𝑠(𝑢, 𝑣) ∑𝑣∈𝑁𝑢 𝑠(𝑢, 𝑣)𝑃′ 𝑣 (𝑔) target user 𝑢’s unrated items before optimization (upper)
∑ − . (10)
ln 2 ∙ ∑ 𝑇𝑢,𝑣 𝜑𝑢,𝑔′
𝑔′∈𝑔𝑘
ln 2 ∙ 𝜑𝑢,𝑔 and neighboring user 𝑣’s distribution (middle) give a graph
𝑣∈𝑁𝑢
like the one on the lower. Item 1 was given a low rating by
For a given learning rate 𝜂, 𝜑𝑢,𝑔 is updated as follows: 𝑣 but, since no other items were rated, it is guaranteed to
top the rankings. As a result, 𝑢’s probability distribution is
𝜕𝐹
𝜑𝑢,𝑔 ← 𝜑𝑢,𝑔 − 𝜂 . (11) updated to place item 1 at a higher position. The proposed
𝜕𝜑𝑢,𝑔 method therefore makes ranking predictions after giving
PROPOSED METHOD temporary estimated ratings to items that were not rated by
the neighboring user 𝑣. The flow of ranking prediction is
Rapid similarity calculation using only high ratings shown on the right of Fig 4 and Fig.7. Let 𝑟𝑢,𝑖 be the rating
Conventional ListCF is slow because it has to calculate given by user 𝑢 to item 𝑖, and let unrated items have a score
similarities using all rating scores. To speed up computation, of zero.
we now
𝑠(𝑢, 𝑣) in Eq. (8) with 𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦(𝑢, 𝑣) (Eq. (12)), ranking
predictions can then be made in the same way as for ListCF
by using Eqs. (8)–(11). Calculating temporary ratings
allows the neighbors’ probability distributions in situations
such as Fig. 6 to be more reasonable. As a result, since the
parameters are less frequently updated in an undesirable
way, this should improve the ranking accuracy.
EXPERIMENTAL EVALUATION
Overview of the experiment
To confirm the effectiveness of the two proposed changes,
we experimentally compared the methods in terms of
computation time and ranking accuracy using the
MovieLens 1M and 10M datasets, details of which are
shown in Table 1. We selected 10 sets of ratings from each
user’s rating information to create the test dataset and used
the rest as training data. In this experiment, since it is
shown that the ranking accuracy does not improve greatly
by increasing the value of k [2], we compared the proposed
method with conventional ListCF based on the top-1
Figure 6. Examples of unsuitable parameter updates that probability model, as in [3,4]. Also, the threshold 𝑐 of the
occur when users have only one low-rated item in common.
The horizontal axis represents the item number.
similarity function is 300. The normalized discounted
cumulative gain (NDCG) metric was used to assess ranking
prediction accuracy. The NDCG value, considering the top
𝑛 predicted rankings for user 𝑢, is defined as follows:
𝑛 𝑝
2𝑟𝑢 − 1
𝑁𝐷𝐶𝐺𝑢 @𝑛 = 𝑍𝑢 ∑ , (14)
log 2 (1 + 𝑝)
𝑝=1
where 𝑍𝑢 is a normalization term that ensures the NDCG
𝑝
value of the correct ranking is 1 and 𝑟𝑢 is the rating of the
pth-ranked item for user 𝑢. For a set of users 𝑈, the overall
NDCG@n score is calculated as follows:
|𝑈|
1
𝑁𝐷𝐶𝐺@𝑛 = ∑ 𝑁𝐷𝐶𝐺𝑢 @𝑛 . (15)
|𝑈|
𝑢=1
Comparison of similarity computation time
Figure 7. Procedure for calculating temporary ratings.
In this experiment, we measured the similarity computation
For a given neighborhood user 𝑣 ∈ 𝑁𝑢 and set 𝑇𝑢 = time. We compared the time taken to calculate the user
{𝑡1, 𝑡2 , ⋯ , 𝑡𝑝 } (⊂ 𝐼 ∖ 𝐼𝑢 ) of items not rated by 𝑢 , 𝑣 ’s similarities using both the proposed and ListCF methods,
temporary ratings are defined as follows: and then compared the ranking accuracy of ListCF
predictions made using the calculated similarities. The
𝑡𝑒𝑚𝑝𝑜𝑟𝑎𝑟𝑦 𝑟𝑎𝑡𝑖𝑛𝑔(𝑣, 𝑡𝑗 ) = criterion used to determine high ratings was changed from 1
to 5 in increments of 1, and similarities were obtained for
𝑟𝑣,𝑡𝑗 (if 𝑟𝑣,𝑡𝑗 is nonzero) each criterion. A criterion of 1 means all ratings are used,
∑𝑣′ ∈𝑁𝑣 𝑟𝑣′ ,𝑡𝑗 (13) while a criterion of 5 means that only items rated as 5 are
(if 𝑟𝑣,𝑡𝑗 is zero) , (𝑗 = 1, ⋯ , 𝑝), used.
|𝑁 ′ 𝑣,𝑡𝑗 |
{ MovieLens 1M MovieLens 10M
where 𝑁′𝑣,𝑡𝑗 is the set of 𝑣’s neighbors 𝑣′ ∈ 𝑁𝑣 who have
Users 6,040 71,567
rated item 𝑡𝑗 . If none of 𝑣’s neighbors have rated 𝑡𝑗 , the
temporary rating will still be zero. In that case, we can Items 3,952 10,681
obtain a nonzero temporary rating by calculating the Ratings 1,000,209 10,000,054
temporary ratings 𝑡𝑒𝑚𝑝𝑜𝑟𝑎𝑟𝑦 𝑟𝑎𝑡𝑖𝑛𝑔(𝑣′, 𝑡𝑗 ) for each
neighbor 𝑣′ of 𝑣. By calculating the cross entropy of Eq. (7) Table 1. Detailed Explanation of Dataset
using these temporary ratings (Eq. (13)) and replacing
12 80 80 80
10
8
6 75 75 75
4
2
0 70 70 70
(a) (b) (c) (d)
Figure 8. (a) is the comparison of calculation times of similarity and (b)-(d) are comparison of ranking accuracy by using
MovieLens 1M. All horizontal axes represent criteria of high rating and each vertical axis represents minutes (a), NDCG@1 (b),
NDCG@3 (c) and NDCG@5 (d) respectively.
70 80 80 80
60
50
40
75 75 75
30
20
10
0 70 70 70
(a) (b) (c) (d)
Figure 9. (a) is the comparison of calculation times of similarity and (b)-(d) are comparison of ranking accuracy by using
MovieLens 10M. All horizontal axes represent criteria of high rating and each vertical axis represents hours (a), NDCG@1 (b),
NDCG@3 (c) and NDCG@5 (d) respectively.
Figs. 8(a) and 9(a) show the computation time results for Comparison of ranking accuracy
MovieLens 1M and 10M, respectively, while Figs. 8(b)-(d) Next, we conducted experiments to examine the effect of
and 9(b)-(d) show the corresponding ranking accuracy using temporary ratings on ranking prediction accuracy. We
results. The horizontal axes show the high rating criterion in compared the accuracy of ranking predictions made using
all graphs, while the vertical axes show calculation time in both the proposed and ListCF methods using the similarities
Figs. 8(a) and 9(a), NDCG@1 in Figs. 8(b) and 9(b), calculated in the previous subsection. We used an initial
NDCG@3 in Figs. 8(c) and 9(c) and NDCG@5 in Figs. 𝜑𝑢,𝑔 value of 10 for both datasets with learning rates of
8(d) and 9(d) respectively. As can be seen from Figs. 8(a) 0.025 and 0.01 for MovieLens 1M and 10M, respectively.
and 9(a), similarity computation is considerably more rapid The gradient descent method was repeated 50 times. The
for the proposed method than for conventional ListCF. In computation time results are shown in Fig. 10, while
addition, Figs. 8(b)-(d) and 9(b)-(d) show that when scores ranking accuracy results are shown in Figs. 11 and 12.
of 4 and 5 are considered to be high ratings, the rankings
are as accurate as those of the conventional method.
ListCF prediction proposed prediction ListCF prediction proposed prediction
160 60
50
120
40
minutes
seconds
80 30
20
40
10
0 0
ListCF 1 2 3 4 5 ListCF 1 2 3 4 5
(a) (b)
Figure 10. Ranking prediction times for MovieLens 1M (a) and MovieLens 10M (b).
ListCF prediction proposed prediction ListCF prediction proposed prediction ListCF prediction proposed prediction
85 85 85
80 80 80
75 75 75
70 70 70
ListCF 1 2 3 4 5 ListCF 1 2 3 4 5 ListCF 1 2 3 4 5
(a) (b) (c)
Figure 11. Ranking accuracy for MovieLens 1M. Each vertical axis represents NDCG@1 (a), NDCG@3 (b) and NDCG@5 (c).
ListCF prediction proposed prediction ListCF prediction proposed prediction ListCF prediction proposed prediction
85 85 85
80 80 80
75 75 75
70 70 70
ListCF 1 2 3 4 5 ListCF 1 2 3 4 5 ListCF 1 2 3 4 5
(a) (b) (c)
Figure 12. Ranking accuracy for MovieLens 10M. Each vertical axis represents NDCG@1 (a), NDCG@3 (b) and NDCG@5 (c).
Fig. 10 shows that the computation times for the proposed REFERENCES
ranking prediction method were shorter in all cases. In 1. Zhe Cao, Tao Qin, Tie-Yan Liu, Ming-Feng Tsai, and
addition, Figs. 11 and 12 show that its NDCG@1, Hang Li. 2007. Learning to rank: from pairwise
NDCG@3 and NDCG@5 values were better for all cases. approach to listwise approach. In Proceedings of the
Since MovieLens 1M has less data than MovieLens 10M, 24th international conference on Machine learning
the number of temporary ratings increases, and the (ICML '07), 129-136.
difference between the proposed method and the http://dx.doi.org/10.1145/1273496.1273513
conventional ListCF becomes larger than MovieLens 10.
2. Shanshan Huang, Shuaiqiang Wang, Tie-Yan Liu, Jun
CONCLUSION
Ma, Zhumin Chen, and Jari Veijalainen. 2015. Listwise
In this paper, we have made two proposals. The first is to
Collaborative Filtering. In Proceedings of the 38th
calculate user similarity scores using only high ratings,
International ACM SIGIR Conference on Research and
instead of all ratings, to speed up computation. The second
Development in Information Retrieval (SIGIR '15),
is to introduce temporary estimated ratings for items that
343-352.
have not been rated by neighboring users to improve
http://dx.doi.org/10.1145/2766462.2767693
ranking accuracy.
3. Kalervo Järvelin and Jaana Kekäläinen. 2000. IR
In experiments using the MovieLens 1M and 10M datasets,
evaluation methods for retrieving highly relevant
we have compared the computation time and accuracy of
documents. In Proceedings of the 23rd annual
both proposals with those of conventional ListCF. The
international ACM SIGIR conference on Research and
results demonstrated that the first proposal provided a
development in information retrieval (SIGIR '00), 41-
considerable reduction in computation time, compared with
48.
conventional ListCF, while maintaining equal or greater
http://dx.doi.org/10.1145/345508.345545
ranking accuracy. In addition, the second proposal
shortened the prediction time in most cases while always 4. Kalervo Järvelin and Jaana Kekäläinen. 2002.
improving the ranking accuracy. Cumulated gain-based evaluation of IR techniques.
ACM Trans. Inf. Syst. 20, 4: 422-446.
In the future, we plan on improving the way neighboring
users are selected and search for a better objective function. 5. S.Kullback. 1997. Information Theory and Statistics.
Courier Corporation.
6. Linden, G., Smith, B., & York, J. 2003. Amazon. com
recommendations: Item-to-item collaborative filtering.
IEEE Internet computing, 7, 1: 76-80.
7. Nathan N. Liu and Qiang Yang. 2008. EigenRank: a
ranking-oriented approach to collaborative filtering. In
Proceedings of the 31st annual international ACM
SIGIR conference on Research and development in
information retrieval (SIGIR '08), 83-90.
http://dx.doi.org/10.1145/1390334.1390351
8. J. I. Marden. 1996. Analyzing and modeling rank data.
CRC Press.
9. Badrul Sarwar, George Karypis, Joseph Konstan, and
John Riedl. 2001. Item-based collaborative filtering
recommendation algorithms. In Proceedings of the
10th international conference on World Wide Web
(WWW '01), 285-295.
http://dx.doi.org/10.1145/371920.372071
10. Shuaiqiang Wang, Jiankai Sun, Byron J. Gao, and Jun
Ma. 2014. VSRank: A Novel Framework for Ranking-
Based Collaborative Filtering. ACM Trans. Intell. Syst.
Technol. 5, 3: 24.