conducted for performance comparison of the proposed model with existing recommendation models,

demonstrate that it can indeed provide better recommendations and can be used as a strong baseline for top-N

recommendations.

Key-Words: - ranking-based recommendation, implicit feedback, Bayesian Personalized Ranking (BPR),

Received: March 11, 2023. Revised: October 15, 2023. Accepted: December 19, 2023. Published: February 14, 2024.

browsing to sophisticated Internet of Things (IoT)

services, by assisting users to discover what they

s/he has not yet ranked. In contexts where explicit

item ratings of other users may not be available,

ranking prediction can be more important than

rating prediction. Most of the existing ranking-based

prediction approaches consider items as having

equal weights, which is not always the case.

Different weights of items could be regarded as a

reflection of items’ importance, or desirability, to

users.

the case, and (ii) the similarity matrix is optimized

by minimizing the squared error, which is not

adaptable for recommendations from implicit

feedback. In recommendation scenarios, it is

common knowledge that the importance of items

varies from one to another due to different

popularity or other features of items. Thus, in SLIM,

the predicted score on an unrated item i of user u is

calculated by the aggregation of items that have

been rated by this user. Among these rated items,

the model using Bayesian Personalized Ranking

(BPR), [3]. Extensive experiments, conducted on

three benchmark datasets, confirm that the proposed

model outperforms the existing ranking-based

models (which do not use item weights) on several

different evaluation metrics.

follows. Section 2 briefly presents the related work

in this area. Section 3 explains the background.

Section 4 describes the proposed model. Section 5

A common task of recommender systems is to

improve user experience via personalized

recommendations based on different sorts of user

behavior, [4]. The objective of the recommendation

can be treated either as a rating estimation problem

besides the user–item ratings, [15], [16], [17]. In

real-world recommendation situations, however,

additional data sources may not always be available

but there is still intrinsic information within the

user–item rating matrix to utilize for improving the

recommendation performance. Concerning this

problem, the current paper proposes to exploit item

weights from the item correlation graph constituted

by the user–item interactions and utilize these

weights within the SLIM model, [1], with

PageRank is a very successful technique for

ranking nodes in a graph, e.g., for hyperlink analysis

in web graphs or user interaction analysis in social

networks, [18]. It computes the importance score for

each node in the graph according to the graph

connectivity. PageRank has also been used as a

weight-learning method in recommender systems,

based on different degrees of correlation between

users or items, [19], [20]. Given a graph

  󰇛 󰇜, where  is the set of nodes and  is

where  denotes the normalized connectivity matrix

for graph ,  denotes the decay factor,  denotes

the vector representing the ranking score of each

node, and  is a  long vector of ones.

calculate the ranking score of items. ItemRank is a

scoring algorithm based on PageRank, which can

rank the items according to users’ preferences in

recommender systems. One advantage of ItemRank

is that it works with both explicit and implicit

item may have a higher contribution. For example,

suppose that item  and  are two items that user 

has seen in the past, but item  is more popular

(important) than item . In both models described

above, similarities of item  to items  and 

optimize the parameters of recommendation models,

[3]. The assumption behind BPR is that the user

prefers a consumed/seen item to an

unconsumed/unseen item, aiming to maximize the

following posterior probability:

and a very large set of negative/unrated items from

the users’ histories. BPR estimates parameters by

minimizing the loss function defined in, [3], as

follows:

regularization term.

Most of the current ranking-based CF models treat

common sense that some items are more important

than others, which can be represented by assigning

bigger weights to them. This idea is incorporated

into the proposed ranking-based CF model,

WeightedSLIM, which takes into account the item

where PR(l) denotes the weight of item l.

proposed model, is the following one. Suppose that,

in the past, user  has rated two items,  and .

For determining the rank of two unrated items  and

Suppose that 󰇛 󰇜  , 󰇛 󰇜  ; and

The performance of the proposed WeightedSLIM

the user–item matrix.

user–movie pair having a rating is regarded as an

observed interaction. However, for ML-100k and

kept as part of the observed positive feedback. The

as MAE and RMSE, are not suitable. Instead, the

In the conducted experiments, the proposed

WeightedSLIM model was compared to the

following existing ranking-based recommendation

models:

items that user  has rated. 󰇛󰇜   triples

󰇛  󰇜 :   

for each user  for all datasets, where  is an

observed item by user , and  is an unobserved

item. The number of iterations was set to 30. For

regularization parameter , several values were tried

best AUC result was chosen, namely  for all

recommendation performance comparison results of

item correlation matrix computation and item

computation is 󰇛󰇜 using Sparse Basic Linear

weights learning it is 󰇛󰇜, where  is the number

of iterations and  is the number of items. In the

model training stage, the total time complexity of

learning the model parameters is 󰇛󰇜, where

users, and  is the average number of items that

user likes. Thus, the worst-case total time

complexity of WeightedSLIM is the same as for

[4] Y. Hu, Y. Koren, and C. Volinsky,

[6] G. B. Guo, J. Zhang, and N. Yorke-Smith, "A

[7] Y. Koren, R. Bell, and C. Volinsky, "Matrix

[9] M. Weimer, A. Karatzoglou, Q. V. Le, and A.

[10] W. Pan, H. Zhong, C. Xu, and Z. Ming,