IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 13, No. 1, March 2024, pp. 786~797

ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i1.pp786-797  786

A recommender system-using novel deep network

collaborative filtering

Shruthi Nagaraj, Blessed Prince Palayyan

Department of Computer Science and Engineering, Presidency University, Bangalore, India

Article Info ABSTRACT

Article history: The recommendation model aims to predict the user’s preferred items
among million through analyzing the user-item relations; furthermore,
Received May 4, 2023 Collaborative Filtering has been utilized as one of the successful
Revised Oct 25, 2023 recommendation approaches in last few years; however, it has the issue of
Accepted Nov 7, 2023 sparsity. This research work develops a deep network collaborative filtering
(DeepNCF), which incorporates graph neural network (GNN), and novel
network collaborative filtering (NCF) for performance enhancement. At first
Keywords: user-item dual network is constructed, thereafter-custom weighted dual
mode modularity is developed for edge clustering. Furthermore, GNN is
Deep network collaborative utilized for capturing the complex relation between user and item. DeepNCF
filtering is evaluated considering the two distinctive. The experimental analysis is
Dual network carried out on two datasets for Amazon and movielens dataset for recall@20
Graph neural network and recall@50 and the normalized discounted cumulative gain (NDCG)
Network collaborative filtering metric is evaluated for Amazon Dataset for NDCG@20 and NDCG@50.
Recommender system The proposed method outperforms the most relevant research and is accurate
enough to give personalized recommendations and diversity.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Shruthi Nagaraj
Department of Computer Science and Engineering, Presidency University
Bangalore, India
Email: [email protected]

1. INTRODUCTION
Nowadays, it is customary to make purchases online. Consumers sometimes find it difficult to select
the most interesting product from among the various ones offered by E-platforms. A substantial amount of
internet information may now be available rapidly with advances in recent technological developments and
the popularity of online services. For a diversity of online services and goods, users can leave reviews,
comments, and ratings. Because of recent advancements in global computing, the problem of online data
overload has arisen. Finding meaningful and helpful information online is becoming more and more
challenging because of the data flood. Yet, several contemporary approaches with reduced computing
demands may now more effectively and efficiently direct users to the relevant material. As a result,
generating recommender systems has gained a lot of traction lately. Generally considering, recommender
systems serve as information filtration tools, providing customers with pertinent and practical information
[1]–[3]. For each page of the website, we are on, a recommendation algorithm is used. For instance, e-
commerce companies typically provide "guess your favorite" on the front page. There are many more
websites, including those for movies, videos, news, books, restaurants, maps, and other content, that have
built-in recommendation engines, amongst which a few of the ones are already in use.
Collaborative, content-based, and hybrid systems are the three primary categories of recommender
systems (RSs). Without taking into account any information about other users, content-based filtering (CBF)

Journal homepage: http://ijai.iaescore.com

Int J Artif Intell ISSN: 2252-8938  787

predicts a user's preferences based on his or her information (gender, age and activities on social media).
Given that it employs several strategies to provide the user with the information they need, CBF might be
seen as an information-filtering task. Instead of looking for particular data within an incoming stream,
filtering is frequently considered of as the removal of undesirable data (viewed as noise) from that stream.
The most often used strategy is based on the semantic content of an item. In this scenario, a number of the
guiding principles from the information retrieval discipline, upon which it is built, are applied: products are
suggested based on comparisons between their content and the user profile. This profile is shown as a table
with the user-specified key elements and weights. This method of information retrieval is straightforward,
quick, and effective [4]–[7]. The majority of current research has been on recommendation systems based on
deep learning. Deep learning model topologies may be easily changed to accommodate different
recommendation situations and specific application conditions. By mining feature combinations and fitting
data patterns, deep learning recommendation models outperform traditional recommendation algorithms. Yet,
the majority of these current recommendation algorithms ignore the potential worth of supplementary data,
user ratings, and user reviews. Auxiliary information, a type of implicit feedback, contains user
characteristics and item characteristics. Hence, it is possible to mine auxiliary data for implicit attributes of
objects. Yet, it is important to remember that different users may have different rating habits. Ratings from
users may seem to indicate a user's preference for a certain item. As an illustration, some people constantly
offer great ratings, whilst others may have high expectations and frequently offer negative ones. In this
situation, it is challenging to evaluate an item's quality just based on customer reviews. In addition, just
because two people give a product a high rating does not always mean that they have the same opinion on it.
Because of his rating habits, one individual could just give everything that is not invasive, a good review
while the other might truly appreciate the product. We will incorporate user feedback into the
recommendation system to address these issues efficiently. Both user ratings and reviews are explicit
observations, but since reviews are more customized than ratings, they are more likely to represent people's
emotional tendencies [8]–[11].
The internet, books, e-learning, travel, movies, music, e-commerce, news, specialized research
resources and television shows. are just a few of the applications in a variety of recommender systems. To
offer its customers personalized recommendations across a range of applications, it is critical to create better
and more efficient recommender systems. For a wider range of applications, the current generation of
recommender systems must be improved to offer better useful and suitable recommendations. It is necessary
to conduct further research on recent studies on recommender systems that focus on different applications.
Further, research contribution is given as follows,
− Deep network collaborative filtering (DeepNCF) comprises a novel network collaborative filtering and
graph neural network (GNN) for performance enhancement; user-item based dual mode network is
constructed through rating matrix; later custom weighted dual mode modularity is developed for edge
clustering.
− GNN is adopted for optimization of edge clustering and understand the more complex relationship
between user as well as item. Integrated approach of network collaborative filtering (NCF) and GNN.
Furthermore, Edge reduction approach is used for selection of non-redundant edges.
− DeepNCF is evaluated considering the two distinctive dataset of Amazon Books and MovieLens
considering the metrics normalized discounted cumulative gain (NDCG) and recall with different
variants.
This Particular research is organized as follows: First section of the research work starts with the
background of recommendation system, importance of collaborative filtering and integration of deep learning
towards recommendation system. Second section discusses the existing work for recommendation system.
DeepNCF is designed and develop in third section along with mathematical modelling and architecture;
DeepNCF is evaluated considering the different dataset along with comparative analysis.

2. RELATED WORK
Graph neural network (GNN) has been one of the key research areas in recommendation system
since development of earlier one; this section discuss little recent existing mechanism that adopted GNN for
exploiting the optimal relationship among user and items. To develop precise session embedding and train
item representations from session graphs, gated graph neural networks were employed [12]. Identified the
local and long-range interdependence of session-based recommendations using GNN and self-attention
approaches [13]. To locate and create item and user choice embedding, several weighted graph attention
networks were used [14]. Removed nonlinear activations and feature transformations from the network
architecture to address graph convolutional networks [15]. They uses GNN to handle a variety of
characteristics and created a component to look at the relationship between potential neighbor nodes [16].

A recommender system-using novel deep network collaborative filtering (Shruthi Nagaraj)

788  ISSN: 2252-8938

Object interactions were captured in using a self-attentive graph neural network and a soft attention
method [17]. Yet, statistics on user-item interaction are frequently scarce. By including item attribute data in
interaction characteristics, some researchers have started to combine multimodal data to alleviate the data
shortage. To find user and object embedding for the goal of making recommendations, consider persons and
objects as knowledge networks, including their traits and interactions; Proposes linking user-item interaction
graphs with user-user social networks to improve social recommendation [18], [19]. A technique for
capturing the intricate relationship between explicit user decisions and edge data was developed by Inx [20].
To get over these limitations and provide more precise item recommendations, the review text and
node attributes for graph neural network recommendation (RTN-GNNR) model combines the review text and
node attributes for graph neural network recommendation. There are four parts to the RTN-GNNR. A
bidirectional, gated network is suggested by the module for obtaining review text attributes. Bidirectional
encoder representation from transformers (BERT) and an attention mechanism are used in the recurrent unit
(Bi-GRU) text analysis method to assist the model choose the most important reviews [21], [22]. Construct a
context-aware recency-based attention network (CARAN) that uses the attention mechanism to give recent
visitation spots the highest priority based on the temporal and spatial context and the weather. Using linear
interpolation and spatiotemporal matrices to characterize spatial distance smoothly, enables interaction
between non-adjacent check-ins.

3. PROPOSED METHODOLOGY
This research develops DeepNCF, which integrates the graph neural network (GNN) and network
collaborative filtering for performance enhancement in terms of diversity and recommendation accuracy.
Figure 1 shows the DeepNCF workflow; at first novel dual mode graph is designed, later novel NCF with
edge clustering and edge reduction is carried out; further we adopt GNN architecture and design Deep NCF
train DeepNCF to make the recommendation. Figure 2 shows the proposed workflow. The first step here is to
develop a bipartite graph, which performs link segmentation that generates a dual mode graph community.
On this basis, a link segmentation algorithm divides the so that it has a robust structure. The second step is
the execution of a deep reinforcement learning (DRL) into a specific single community such as the
neighborhood chosen among the group that generates a varied type of recommendations. The in-depth
discussion of the proposed model is discussed here.

Figure 1. DeepNCF workflow Figure 2. Proposed workflow

3.1. Problem statement

The main goal of this work is to draw attention to the problem of recommendation overfitting while
ensuring recommendation accuracy and maintaining variation without the need for additional data. The

Int J Artif Intell, Vol. 13, No. 1, March 2024: 786-797

Int J Artif Intell ISSN: 2252-8938  789

traditional user-based collaborative filtering method wherein a neighborhood consists of the user’s relevant
ratings that focus on specific items. This will allow the users that are present in the neighborhood to retain the
similarity with the target user, which consists of various choices. Because the links here are connected to
various items that do not efficiently reflect a user’s choice. Upon considering an example of users rating to
meet the personalized requirements, henceforth ensuring that the candidate neighbor ensures the same rating
features irrespective of the correlated dense items with the target user. To enhance the recommendation
system, the users recommend a specific category of items. The existing techniques are based on the
mechanism of user-based collaborative filtering which enhances the performance of the recommendation
system recently. Solving the overfitting in user-based collaborative filtering is further applicable to various
techniques used in recent Recommendation systems. The proposed system (PS) approach is shown below.

3.2. Edge clustering

Simultaneously, the dual mode network has gained widespread use for various events such as
customer-item relationships. Here there exist two categories of the non-overlapping category of nodes for
uppermost and bottommost nodes. A specific single link here collaborates with a set of nodes that belongs to
various categories. The two-mode graph is an essential measure to enhance the community structure of the
two-mode graph. To attain an effective community identification model that proposes a dual mode graph
which states that there exists top 𝑘 top nodes and 𝑒 below nodes in the dual mode graph. Henceforth the two-
mode graph here is considered as 𝑘 ∗ 𝑒 incidence matrix 𝐵 that focuses on the communication among the top
and bottom nodes whereas 𝐵𝑚,𝑛 = 1 if there exists a link between uppermost node 𝑚 and bottommost node
𝑛 and 𝐵𝑚,𝑛 = 0. Hence 𝐶𝑚,𝑛 shows the probabilities associated with any random model through a link that is
present between the uppermost node 𝑚 and bottommost node 𝑛. Accordingly, to the connection matrix 𝐵
and 𝐶, for the two-mode graph, 𝐷 is shown as,
𝐵𝑚,𝑛
𝐷 = ∑𝑘𝑚=1 ∑𝑒𝑛=1 ( − 𝐶𝑚,𝑛 )𝜑(𝐻(𝑚), 𝐻(𝑛)) (1)
𝐾

1 𝐵𝑚,𝑛 𝑃(𝑚)𝑃(𝑛)
= ∑𝑘𝑚=1 ∑𝑒𝑛=1 ( - ) 𝜑(𝐻(𝑚), 𝐻(𝑛)) (2)
𝐾 𝐾

Here K denotes the number of links in 𝐵. 𝐻(𝑚) In addition, 𝐻(𝑛) denotes the degree of nodes 𝑚 and 𝑛
irrespectively. The 𝜑 function here shows the 𝜑(𝐻(𝑚), 𝐻(𝑛)) to equalize the nodes 𝑚 and 𝑛 are segmented
into a similar community else 0. The RS consists of customer-item relation; it shows the dual mode graph.
An RS is transformed into 𝐼𝐵𝐺 = 𝐻(𝑇, 𝐽, 𝑆), which shows the |𝑇| denotes the user nodes, |𝐽| consists of item
nodes and |𝑆| denotes the number of links. 𝑆𝑡,𝑗 shows that a customer 𝑡 has rated an item𝑗. Henceforth 𝐼𝐵𝐺 is
denoted as |𝑇|*|𝐽| adjacent matrix 𝐵. The adjacency matrix 𝐵 in (3) shows the correlation degree between the
node 𝑡 and item𝑗, which consists of two links that accommodate the transformation. The links consisting of
degree 𝜶≤𝑙𝑡,𝑗 ≤1 are denoted as strong links and 0 ≤𝑔𝑡,𝑗 ≤ 𝜶 are denoted as weak links.

𝑣 𝑐𝑡,𝑗
𝐺𝑡,𝑗 = 𝑠(𝑡, 𝑗) = 0 𝑒𝑙𝑠𝑒 0 < 𝑐𝑡,𝑗 < 𝜇 1.0 𝜇 ≤ 𝑐𝑡,𝑗 ≤ 1 (3)
𝑀𝑎𝑥𝑗∈𝐽{

Here each link denotes the parameter for each community, to analyze a co-efficient 𝑃𝑟𝑜𝑏(𝑙𝑓 |𝑆𝑡,𝑗 )
that denotes the link 𝑆𝑡,𝑗 that denotes the community 𝑙𝑓 . In this step, each link is associated with many
communities at a specific time and the main advantage levied on the dual mode graph is weighted by the co-
efficient. The dual mode graph is reframed upon substitution by 𝜑(𝑆(𝑡), 𝑆(𝑗)) with different parameters as
𝜓𝑡,𝑗,𝑓 and 𝜔𝑡,𝑗,𝑓 irrespectively.

1 𝑃(𝑡)𝑃(𝑗)
𝑋𝑣 = ∑𝐹𝑓=1 ∑|𝑇| |𝐽| 𝑣
𝑡=1 ∑𝑗=1 (𝜓𝑡,𝑗,𝑓 , 𝐺𝑡,𝑗 − 𝜔𝑡,𝑗,𝑓 ) (4)
|𝑆| |𝑆|

𝜓𝑡,𝑗,𝑓 = 𝑝𝑟𝑒𝑑(𝑆𝑓 |𝑆𝑡,𝑗 ) (5)

The two co-efficient 𝜓𝑡,𝑗,𝑓 and 𝜔𝑡,𝑗,𝑓 are constructed. For 𝜓𝑡,𝑗,𝑓 the probability that is associated
with 𝑆𝑡,𝑗 to the model belonging to a community𝑆𝑓 . 𝜓𝑡,𝑗,𝑓 Is equivalent to 1 if the customer node 𝑡 and item
node 𝑗 are segmented into similar community 𝑆𝑓 if 𝑆(𝑇) = 𝑆(𝐽) = 𝑆𝑓 this is equivalent to 0. Henceforth the
probability of two nodes associated with a similar community is high in comparison if the nodes considered
are neighbours. We can state that,

𝜔𝑡,𝑗,𝑓 = 𝑝𝑟𝑒𝑑 𝐽 (𝑆𝑓 |𝑆𝑡,𝑗 )𝑝𝑟𝑒𝑑 𝑇 (𝑆𝑓 |𝑆𝑡,𝑗 ) (6)

A recommender system-using novel deep network collaborative filtering (Shruthi Nagaraj)
790  ISSN: 2252-8938

∑𝑗∈𝐽 𝑝𝑟𝑒𝑑(𝑆𝑓 |𝑆𝑡,𝑗 )

𝑝𝑟𝑒𝑑 𝐽 (𝑆𝑡,𝑗 ) = (7)
|𝐽|

∑𝑡∈𝑇 𝑝𝑟𝑒𝑑(𝑆𝑓 |𝑆𝑡,𝑗 )

𝑝𝑟𝑒𝑑 𝑇 (𝑆𝑡,𝑗 ) = (8)
|𝑇|

Accordingly, to (4) -(8) a dual mode weighted graph is,

1 𝑃(𝑡)𝑃(𝑗)
𝑋𝑣 = ∑𝐹𝑓=1 ∑|𝑇| |𝐽| 𝑣 𝐽 𝑇
𝑡=1 ∑𝑗=1 ( 𝑝𝑟𝑒𝑑(𝑆𝑓 |𝑆𝑡,𝑗 )𝐺𝑡,𝑗 − 𝑝𝑟𝑒𝑑 (𝑆𝑡,𝑗 )𝑝𝑟𝑒𝑑 (𝑆𝑡,𝑗 ) ) (9)
|𝑆| |𝑆|

A high-end value 𝑋 𝑣 denotes a robust community structure for , the 𝑋 𝑣 shows the difference amidst a dual
mode graph after the community is identified and an unstructured network doesn’t contain a community
structure. The wider the change the clearer the community structure is Assumption 𝑋 𝑣 = 0 then all the links
belong to a similar community, if so then 𝐹 = 1 we get, ∀𝑗 ∈ 𝐽, 𝑡 ∈ 𝑇 𝑝𝑟𝑒𝑑 (𝑆𝑡,𝑗 ) = 1. Simultaneous to (7)
and (8) we can find the equation,

∑𝑗∈𝐽 𝑝𝑟𝑒𝑑(𝑆1 |𝑆𝑡,𝑗 ) |𝐽|

𝑝𝑟𝑒𝑑 𝐽 (𝑆𝑡,𝑗 )= = =1 (10)
|𝐽| |𝐽|

∑𝑡∈𝑇 𝑝𝑟𝑒𝑑(𝑆𝑡,𝑗 ) |𝑇|

𝑝𝑟𝑒𝑑 𝑇 (𝑆𝑡,𝑗 ) = = =1 (11)
|𝑇| |𝑇|

The 9 − 𝑡ℎ equation is simply reduced to,

1 |𝑇| |𝐽| 𝑃(𝑡)𝑃(𝑗)

𝑋 𝑣 = |𝑆| ∑𝐹𝑓=1 ∑𝑡=1 ∑𝑗=1 ( 𝑝𝑟𝑒𝑑 (𝑆𝑡,𝑗 )𝐺𝑡,𝑗
𝑣
− 𝑝𝑟𝑒𝑑 𝐽 (𝑆𝑡,𝑗 )𝑝𝑟𝑒𝑑 𝑇 (𝑆𝑡,𝑗 ) |𝑆|
) (12)

1 𝑃(𝑡)𝑃(𝑗)
= ∑|𝑇| |𝐽| 𝑣
𝑡=1 ∑𝑗=1 (𝐺𝑡,𝑗 − )
|𝑆| |𝑆|

Henceforth if all inks are focused on the same community, then we have𝑋 𝑣 = 0. The 𝑋 𝑣 uses local data the
details of the link associated with each community. Community identification comes under the category of
minima or maxima. A global value is introduced here as 𝑋 𝑣 , the 𝑀𝑎𝑥𝑑𝑒𝑔 is related to global data, and
integrate 𝑀𝑎𝑥𝑑𝑒𝑔 into 𝑋 𝑣 . The 𝑀𝑎𝑥𝑑𝑒𝑔 is denoted as,

1 1
𝑀𝑎𝑥𝑑𝑒𝑔 = ∑|𝑇| |𝐽|
𝑡=1 ∑𝑗=1 1 ≤ 𝑓 ≤ 𝐹𝑚𝑎𝑥𝑝𝑟𝑒𝑑(𝑆𝑓 |𝑆𝑡,𝑗 ) (13)
|𝑈| |𝐽|

Here the dual mode graph is denoted by,

𝑣 2𝑀𝑎𝑥𝑑𝑒𝑔 +𝑋 𝑣
𝑋𝐺𝑙𝑜𝑏𝑙𝑜𝑐 = (14)
3

𝑣
Consequently, by (14) we can show that 𝑋𝐺𝑙𝑜𝑏𝑙𝑜𝑐 that focuses on local as well as global information whereas
𝑣
the variation exists in the [0, 1] range. Increasing the value 𝑋𝐺𝑙𝑜𝑏𝑙𝑜𝑐 , the effective result is generated in terms
of community detection. Parallely, a novel algorithm is developed for a dual mode weighted graph for each
link here 𝑆𝑡,𝑗 the initialization is done as 𝑝𝑟𝑒𝑑(𝑆𝑓 |𝑆𝑡,𝑗 ) for ∑𝐹𝑓 =1 𝑝𝑟𝑒𝑑(𝑆𝑓 |𝑆𝑡,𝑗 )=1.

[𝑝𝑟𝑒𝑑(𝑡)+𝜏]∗[𝑝𝑟𝑒𝑑(𝑡)+𝜗))
𝑝𝑟𝑒𝑑(𝑆𝑡,𝑗 ) = (15)
∑𝐹′ [𝑝𝑟𝑒𝑑(𝑡)+𝜏]∗∑𝐹′ [𝑝𝑟𝑒𝑑(𝑡)+𝜗))
𝑓 =1 𝑓 =1

Here 𝜏 and 𝜗 are considered as the hyper-parameter so that the denominator is not 0. Based on the
values of𝑝𝑟𝑒𝑑(𝑆𝑓 |𝑆𝑡,𝑗 ) shown by (16) we try to compute the equation 𝑝𝑟𝑒𝑑(𝑆𝑓 |𝑡 ) and 𝑝𝑟𝑒𝑑(𝑆𝑓 |𝑗 ).

∑𝑗∈𝐽(𝑡) 𝑝𝑟𝑒𝑑(𝑆𝑓 |𝑆𝑡,𝑗 )

𝑝𝑟𝑒𝑑(𝑡) = (16)
∑𝐹′ ∑𝑗∈𝐽(𝑡) 𝑝𝑟𝑒𝑑(𝑆𝑓′ |𝑆𝑡,𝑗 )
𝑓 =1

∑𝑡∈𝑇(𝑗) 𝑝𝑟𝑒𝑑(𝑆𝑓 |𝑆𝑡,𝑗 )

𝑝𝑟𝑒𝑑(𝑡) = ∑𝐹 ∑
(17)
𝑝𝑟𝑒𝑑(𝑆𝑓′ |𝑆𝑡,𝑗 )
𝑓′ =1 𝑡∈𝑇(𝑗)

Int J Artif Intell, Vol. 13, No. 1, March 2024: 786-797

Int J Artif Intell ISSN: 2252-8938  791

Based on the results considered 𝑝𝑟𝑒𝑑(𝑡) and 𝑝𝑟𝑒𝑑(𝑡) the value of 𝑝𝑟𝑒𝑑(𝑆𝑓 |𝑆𝑡,𝑗 ) is re-evaluated by
(15). The evaluation is carried out above until the 𝑝𝑟𝑒𝑑(𝑆𝑓 |𝑆𝑡,𝑗 ) that reaches the community to achieve the
large value for 𝑝𝑟𝑒𝑑(𝑆𝑓 |𝑆𝑡,𝑗 ) as the final community that links the 𝑆𝑡,𝑗 . The custom dual mode weighted
𝑣 𝑣
graph 𝑋𝐺𝑙𝑜𝑏𝑙𝑜𝑐 is evaluated at each level from 𝐹𝑚𝑖𝑛 to 𝐹𝑚𝑎𝑥 . The higher the value for 𝑋𝐺𝑙𝑜𝑏𝑙𝑜𝑐 , the robust the
𝑣
community is, the maximum value obtained by 𝑋𝐺𝑙𝑜𝑏𝑙𝑜𝑐 by considering an optimal number of 𝐹
communities.the value 𝐹𝑚𝑖𝑛 is estimated to be 2, whereas the 𝐹𝑚𝑎𝑥 value is given by 𝐹𝑚𝑎𝑥 = (|𝑇| ∗ |𝐽|)0.5 .
Edge clustering algorithm as shown in Algorithm 1.

Algorithm 1. Edge clustering algorithm

Input Dual Mode Network 𝐼𝐵𝐺 = 𝐻(𝑇, 𝐽, 𝑆), the number of communities F.
Step 1 Partition the 𝐻(𝑇, 𝐽, 𝐿) into F-number of communities {𝑙1 , 𝑙2 , … … , 𝑙𝑓 } .
Step 2 for each link 𝑆𝑡,𝑗 do
Step 3 Initialize probability 𝑃𝑟𝑜𝑏(𝑙𝑓 |𝑆𝑡,𝑗 ), for ∑𝐹𝑓=1 𝑃𝑟𝑜𝑏(𝑆𝑡,𝑗 ) = 1
Step 4 end for
Step 5 for each link 𝑆𝑡,𝑗 do
Step 6 repeat
Step 7 Evaluate 𝑃𝑟𝑜𝑏(𝑙𝑓 |𝑡) and 𝑃𝑟𝑜𝑏(𝑙𝑓 |𝑗)
Step 8 Use the values of 𝑃𝑟𝑜𝑏(𝑙𝑓 |𝑡) and 𝑃𝑟𝑜𝑏(𝑙𝑓 |𝑗) to evaluate 𝑃𝑟𝑜𝑏(𝑆𝑡,𝑗 ) accordingly.
Step 9 Till it converges 𝑃𝑟𝑜𝑏(𝑆𝑡,𝑗 )
Step 10 Choose a community that has the largest 𝑃𝑟𝑜𝑏(𝑆𝑡,𝑗 ) that is the final community that connects 𝑆𝑡,𝑗
Step 11 Update the value in {𝑙1 , 𝑙2 , … … , 𝑙𝑓 }
Step 12 end for
Step 13 Return 𝑘 number of communities {𝑙1 , 𝑙2 , … … , 𝑙𝑓 }
output F number of communities {𝑙1 , 𝑙2 , … … , 𝑙𝑓 }

3.3. Edge reduction

Once the edge clustering is done there exist many redundant links that may reduce the performance
of the system with the 𝐼𝐵𝐺 denotes as redundant links. Consider customers 𝑐1 , 𝑐2 𝑎𝑛𝑑 𝑐3 by considering the
items 𝐼1 , 𝐼2 , 𝐼3, 𝐼4, 𝐼5 and 𝐼6 is reduced into a similar community. Additionally, 𝑐1 , 𝑐2 𝑎𝑛𝑑 𝑐3 generate
comparisons with a target customer 𝑡𝑐 and the comparisons denoted as ranks are depicted in the table. It is
essential to ensure that the community includes this customer, hence 𝑐1 𝑎𝑛𝑑 𝑐2 is selected. In the next phase,
the information of customers𝑐1 , 𝑐2 𝑎𝑛𝑑 𝑐3 . Via Figure 3 to evaluate a 𝑆(𝑐1 )= {𝑠1,1 , 𝑠1,2 , 𝑠1,3 } , 𝑆(𝑐2 )=
{𝑠2,1 , 𝑠2,2 , 𝑠2,3 , 𝑠2,4 , 𝑠2,5 } and 𝑆(𝑐3 )= {𝑠3,1 , 𝑠3,2 , 𝑠3,3 , 𝑠3,6 } . the links are connected to a similar item that is
equal. 𝑠1,2 is equivalent to 𝑠2,2 upon computation with 𝑐2 that has additional links {𝑠2,4 , 𝑠2,5 } in comparison
with 𝑐1 , indicates 𝑐2 that recommends items 𝐼𝑀4 and 𝐼𝑀5 additionally along with 𝑐1 . Edge reduction as
shown in Algorithm 2.
The suggestions from the united 𝑐1 and 𝑐2 communities will be identical to those from 𝑐2 alone.
Thus, establishing connections with 𝑐1 is irrational. Used to foster community The connections to 𝐼𝑀1 , 𝐼𝑀2 ,
and 𝐼𝑀3 are the same for 𝑐2 and 𝑐3 but 𝑐3 also has a connection to 𝐼𝑀6 . This suggests that while 𝑐2 cannot
approve 𝐼𝑀6 , 𝑐3 can. Therefore, to assess similarity links leading to 𝑐3 must be preserved. So, 𝑐2 , 𝑐3 , rather
than 𝑐3 , 𝑐2 , will be a superior decision if we chose the top two users as neighbours. As 𝑐3 's link set includes
𝑐1 , it is a stronger contender for acceptance into the community than 𝑐1 . Duplicate 𝐼𝐵𝐺 connections are
covered in this section. Let𝑐𝑎 stand in for a single user node, and let 𝐶 represent the group of user nodes in
the 𝐼𝐵𝐺. Connections in S(𝑐𝑎 )are judged redundant if another user node already exists such that S(𝑐𝑎 ) S(𝑐𝑏 ).
The remaining linkages in a 𝐼𝐵𝐺 are known as nonredundant links once all redundant connections have been
eliminated, at which point the 𝐼𝐵𝐺 ceases to be reducible. Duplicate links in the user's link list will be
removed. Duplicate links prevent the algorithm from determining how similar a user is to a target user. As a
result, this user is removed from the target user's list of prospective neighbors. In this study, we postulate that
removing redundant connections produced by user nodes like 𝑐1 may broaden the variety of ideas.
Comprehensive edge reduction, data are provided by the second technique. It is essential to stress that we
conduct edge reduction is in each local community rather than the entire 𝐼𝐵𝐺 for two reasons.
a) It is possible to lessen the computational difficulty of edge reduction.
b) Any number of unnecessary links may be deleted.
First, following edge clustering, each community's size is significantly reduced to the size of the
𝐼𝐵𝐺 as a whole, representing the edge reduction that is performed in a lesser region. In addition, edge
reduction is performed concurrently in many communities, hence decreasing computing time by a factor of 𝑇
compared to processing the whole 𝐼𝐵𝐺. The second argument is because it is simple to derive that redundant
A recommender system-using novel deep network collaborative filtering (Shruthi Nagaraj)
792  ISSN: 2252-8938

links in the 𝐼𝐵𝐺 are also considered redundant links in a single community; yet, a redundant link in a single
community, redundant connections may be found throughout𝐼𝐵𝐺. The whole 𝐼𝐵𝐺 features an enormous
amount of user nodes and item nodes. Despite the number of connections that exist between each user node
and the link set, due to its small size, this user node, which only appears within other user nodes, is
challenging to locate. Finding out if a user node's link set is present in the link sets of other user nodes will be
easier after the domain has been limited from the whole IBG to each distinct community. Novel NCF
algorithm as shown in Algorithm 3. Moreover, these three-algorithm lead development of novel NCF;
further, it is integrated with GNN for enhancement of performance which is carried out in the next section of
the research.

Algorithm 2. Edge reduction

Input G community 𝑙𝑓
Step 1 Fix the value 𝑙𝑓𝑐 = 𝑙𝑓
Step 2 for 𝑎 = 1 𝑡𝑜 |𝑇(𝑙𝑓 )|𝑑𝑜
Step 3 for b =1 to |𝑇(𝑙𝑓 )|𝑑𝑜
Step 4 if S(𝑙𝑎 ) ⊆ S(𝑙𝑏 ) then
Step 5 if S(𝑙𝑎 ) ⊆𝑙𝑓𝑐 then
Step 6 𝑙𝑓𝑐 = 𝑙𝑓𝑐 − S(𝑙𝑎 )
Step 7 end if
Step 8 end if
Step 9 end for
Step 10 end for
Step 11 return G community to 𝑙𝑓𝑐
output G community to 𝑙𝑓𝑐

Algorithm 3. Novel NCF algorithm

Input Item matrix 𝐼𝑀 and a target customer 𝑡𝑐, 𝐾 recommendations for the target customer 𝑡𝑐
𝐾: the total items recommended to the target customer 𝑡𝑐
𝑣
𝑋𝐺𝑙𝑜𝑏𝑙𝑜𝑐 :the custom dual mode model
𝐽𝑡𝑐 : The item set which the target customer has not rated yet.
𝑝𝑟𝑒𝑑𝑡𝑐,𝑗 : predicted target score for the target customer 𝑡𝑐 for item 𝑗
Step 1 Map a IM to an item Dual Mode graph 𝐼𝐵𝐺 = 𝐻(𝑇, 𝐽, 𝑆)
𝑣
Step 2 𝑋𝐺𝑙𝑜𝑏𝑙𝑜𝑐 = 0; 𝐹 = 0
′
Step 3 for 𝐹 = 𝑓𝑚𝑖𝑛 to 𝑓𝑚𝑎𝑥 do
Step 4 Use edge clustering algorithm (alg 1) to get the communities {𝑙1 , 𝑙2 , … … , 𝑙𝑓 } =
𝑝𝑠(𝐻(𝑇, 𝐽, 𝑆), 𝐹 ′ ).
𝑣
Step 5 Evaluate the 𝑋𝐺𝑙𝑜𝑏𝑙𝑜𝑐 value for {𝑙1 , 𝑙2 , … … , 𝑙𝑓 }
𝑣 𝑣
Step 6 If 𝑋𝐺𝑙𝑜𝑏𝑙𝑜𝑐 ≤ 𝑋𝐺𝑙𝑜𝑏𝑙𝑜𝑐 , then
𝑣 𝑣
Step 7 𝑋𝐺𝑙𝑜𝑏𝑙𝑜𝑐 ≤ 𝑋𝐺𝑙𝑜𝑏𝑙𝑜𝑐 , 𝐹 = 𝐹′,
Step 8 𝐻(𝑇, 𝐽, 𝑆) is clustering into {𝑙1 , 𝑙2 , … … , 𝑙𝑓 }
Step 9 end if
Step 10 end for
𝑣
Step 11 Compute 𝐻(𝑇, 𝐽, 𝑆) with community 𝐹 and peak 𝑋𝐺𝑙𝑜𝑏𝑙𝑜𝑐
Step 12 for 𝑓 = 1 to F do
Step 13 Utilize ( alg 2) to remove redundant links in 𝑙𝑓 , such that 𝑙𝑓𝑐 = 𝑝𝑠(𝑙𝑓 )
Step 14 end for
Step 15 Fetch a set of communities {𝑙1𝑐 , 𝑙2𝑐 , … . , 𝑙𝑓𝑐 }
Step 16 for each item 𝑗 ∈ 𝐽𝑡𝑐 do
Step 17 Extract the community 𝑙𝑓𝑐 that consists of the target customer 𝑡𝑐 and item 𝑗
Step 18 Evaluate the similarity between the target customer 𝑡𝑐 for each customer 𝑐 ∈𝑙𝑓𝑐
Step 19 Choose customers in 𝑙𝑓𝑐 having maximum similarity for the neighbourhood 𝑁𝑖𝑡𝑐 of the target
customer 𝑡𝑐
Step 20 Predicting the score 𝑝𝑟𝑒𝑑𝑡𝑐,𝑗 for item 𝑗 for 𝑁𝑖𝑡𝑐
Step 21 end for
Step 22 The 𝐾 items with 𝑝𝑟𝑒𝑑𝑡𝑐,𝑗 recommended to the target customer 𝑡𝑐
𝑣
output 𝐾 recommendations for the target customer 𝑡𝑐, 𝑋𝐺𝑙𝑜𝑏𝑙𝑜𝑐 , 𝐽𝑡𝑐 , 𝑝𝑟𝑒𝑑𝑡𝑐,𝑗

Int J Artif Intell, Vol. 13, No. 1, March 2024: 786-797

Int J Artif Intell ISSN: 2252-8938  793

4. RESULT SECTION
In this section, we compare the effectiveness and robustness of the proposed PS approach
comparison with the existing state-of-art methods on two datasets i.e. Amazon Dataset and MovieLens
Dataset, which are the benchmark datasets. We evaluate recall for Amazon dataset and MovieLens dataset
for recall @20 and recall@ 50, whereas the NDCG metric is evaluated for the amazon database for metrics
NDCG@20 and NDCG@50. The various state-of-art techniques used for comparison are MostPopular,
YouTube DNN, GRU4Rec, MIND, ComiRec-DR, ComiRec-SA and Fat technique.

4.1. Methods considered for evaluation

The methods considered for evaluations are: i) MostPopular is a traditional recommendation
method, which will recommend the most popular items to users, ii) YouTube DNN [23] is one of the most
successful deep learning models of industry recommender system, iii) GRU4Rec [24] is the first time to use a
gated recurrent neural network to model the sequence relationship between items in the sequence,
iv) MIND [25] is a recommendation model of multi-interest extraction using a capsule network, v) ComiRec-
DR [26] is the capsule network model based on the original dynamic routing method that has better
performance than MIND, vi) ComiRec-SA [26] is a multi-interest extraction model based on self-attention,
vii) Future aware diverse trends (FAT) [27] is a recent state-of-the-art model, which takes into consideration
of users’ feature sequence. FAT uses ComiRec-DR methods to get the future sequence multi-interest, and it
computes correlations among different users by using a collaborative filtering method.
Moreover, the above all method has been compared in the existing model [28], thus proposed
DeepNCF is compared with existing model.

4.1.1. Dataset details

The dataset details are: i) Amazon Dataset: This consists of product reviews and metadata from
Amazon. In our experimental analysis, we use the book category of the Amazon dataset. The length of each
training sample is truncated to 20, and ii) MovieLens Dataset: The benchmark data set in the field of
recommendations is the MovieLens dataset. These facts consist of user demographics (age, gender), movie
ratings, and movie information (style kind, age and employment). In this case, we use the Movielens-1M
dataset, which contains 1 million score records, to examine how the PS framework works.

4.2. Recall
We use the better interpretable average per user rather than the global average. The recall is
evaluated here for two databases, one is amazon Dataset and the other is the MovieLens dataset. For both of
these datasets, the recall is evaluated at metrics@20 and metrics@50.

𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒
𝑅𝑒𝑐𝑎𝑙𝑙 =
𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒+𝐹𝑎𝑙𝑠𝑒 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒

− Recall@20 for Amazon dataset

In Figure 3 and Table 1 we can see that the recall@20 is evaluated for Amazon Dataset, the method
MostPopular method gives a value of 1.368 which is the least value, whereas we can see that ComiRec –SA
denotes an average value of 5.103 and FAT technique shows a value of 5.177. However, the existing
approach EA attains a value of 6.761, means–ends analysis (MEA) technique denotes a higher value than the
existing system as 7.084 and in comparison, with the existing techniques, and our proposed model ensures
better performance and gives a value of 13.89.
− Recall@50 for Amazon dataset
In Figure 4 and Table 2 the recall@20 is evaluated for Amazon Dataset, the method MostPopular
method gives a 1.368 recall value, which is the least value; whereas the method self efficacy academic (SEA)
method generates an average value of 8.172 and future aware diverse trends (FAT) technique gives a value of
8.161. However, the existing approach existing approach (EA) attains a value of 10.625, gives a higher value
than the existing system as 11.054, in comparison with the existing techniques our proposed model ensures
better performance and gives a value of 21.84.
− Recall@20 for MovieLens Dataset
In Figure 5 and Table 3 we can see that the recall@20 is evaluated for the MovieLens dataset, the
method MostPopular method gives a value of 4.636, which is the least value, whereas the ComiRec -SA
denotes an average value of 11.205 and the MIND technique shows a value of 11.155. However, the existing
approach (EA) attains a value of 10.958, YouTube-DNN technique denotes a higher value than the existing
system as 13.558, in comparison with the existing techniques our proposed model ensures better performance
and gives a value of 20.41.

A recommender system-using novel deep network collaborative filtering (Shruthi Nagaraj)

794  ISSN: 2252-8938

In Figure 6 and Table 4 we can see that the recall@50 is evaluated for MovieLens Dataset, the
method MostPopular method gives a value of 10.829 which is the least value, whereas we can see that MIND
denotes an average value of 23.889 and the SEA technique shows a value of 24.325. However, the existing
approach EA attains a value of 23.478, YouTube-DNN technique denotes a higher value than the existing
system as 26.596 and in comparison, with the existing techniques, and our proposed model ensures better
performance and gives a value of 47.05.

Figure 3. Comparison recall@20 for Figure 4. Comparison of recall@50 for various

Amazon dataset techniques

Figure 5. Comparison of recall@20 for Figure 6. Comparison of recall@50 for

MovieLens dataset MovieLens dataset

Table 1. Comparison recalls@20 for Table 2. Comparison of recall@50 for

various technique various techniques
Method Recall@20 Method Recall@20 Method Recall@50 Method Recall@50
MostPopular 1.368 FAT 5.177 MostPopular 2.4 FAT 8.161
GRU4Rec 4.057 SEA 5.576 GRU4Rec 6.501 SEA 8.172
MIND 4.035 EA 6.761 YouTube DNN 6.544 EA 10.625
ComiRec-DR 4.299 MEA 7.084 ComiRec-DR 6.957 MEA 11.054
YouTube DNN 4.573 PS 13.89 ComiRec-SA 7.458 PS 21.48
ComiRec-SA 5.103 MIND 7.638

Table 3. Comparison of recall@20 for Table 4. Comparison of recall@50 for

MovieLens dataset MovieLens dataset
Method Recall@20 Method Recall@20 Method Recall@50 Method Recall@50
MostPopular 4.636 ComiRec-DR 11.205 MostPopular 10.829 MEA 23.962
ComiRec-SA 10.834 MEA 11.366 ComiRec-DR 22.391 SEA 24.325
FAT 10.892 SEA 11.645 ComiRec-SA 23.137 GRU4Rec 29.467
EA 10.985 YouTube DNN 13.558 FAT 23.364 YouTube DNN 26.596
MIND 11.115 PS 20.41 EA 23.478 PS 47.05
MIND 23.889

4.3. NDCG
Normalized Discounted Cumulative Gain (NDCG) considers the location of the correct project. In
Figure 7 and Table 5 the NDCG@20 is evaluated for Amazon Dataset, the method MostPopular method

Int J Artif Intell, Vol. 13, No. 1, March 2024: 786-797

Int J Artif Intell ISSN: 2252-8938  795

gives a value of 2.259, which is the least value, whereas we can see that ComiRec-DR gives an average value
of 4.508. The FAT technique shows a value of 4.07, however, the existing approach EA attains a value of
5.613, MIND technique denotes a higher value than the existing system as 7.933, in comparison with the
existing techniques our proposed model ensures better performance and gives a value of 7.95.
In Figure 8 and Table 6 the NDCG@50 is evaluated for Amazon dataset, the method MostPopular
method gives a value of 3.936, which is the least value, whereas YouTube DNN denotes an average value of
5.039 and ComiRec-DR technique shows a value of 5.591. However, the existing approach EA attains a
value of 7.184, MEA technique denotes a higher value than the existing system as 7.238, in comparison with
the existing techniques our proposed model ensures better performance and gives a value of 10.02.

Figure 7. Comparison of NDCG@20 for Figure 8. Comparison of NDCG@50 for

MovieLens dataset MovieLens dataset

Table 5. Comparison of NDCG@20 for MovieLens Table 6. Comparison of NDCG@50 for MovieLens
dataset dataset
Method NDCG@20 Method NDCG@20 Method NDCG@50 Method NDCG@50
MostPopular 2.259 YouTube DNN 4.673 MostPopular 3.936 ComiRec-DR 5.591
ComiRec-SA 3.807 EA 5.613 ComiRec-SA 4.41 GRU4Rec 5.678
SEA 3.99 MEA 5.816 FAT 4.824 EA 7.184
FAT 4.07 MIND 7.933 MIND 4.907 MEA 7.238
GRU4Rec 4.163 PS 7.95 YouTube DNN 5.039 PS 10.02
ComiRec-DR 4.508 SEA 5.231

4.4. Comparative analysis

The comparative analysis carried out here by the proposed system with the existing system, for
Amazon Dataset the metric Recall@20 the existing system gives a value of 6.761. Table 7 shows the
comparative analysis. Our proposed model denotes a value of 13.89 the improvisation done is 69.0427%, for
Recall@50 the existing system gives a value of 10.625 and our proposed model denotes a value of 21.48 the
improvisation done is 67.62%. For MovieLens Dataset the metric Recall@20 the existing system gives a
value of 10.895 and our proposed model gives a value of 20.41 the improvisation done is 60.0414%,
MovieLens Dataset the metric Recall@50 the existing system gives a value of 23.478 and our proposed
model denotes a value of 47.05 the improvisation done is 66.8444%. , for Amazon Dataset the metric
NDCG@20 the existing system gives a value of 5.613 and our proposed model denotes a value of 7.95 the
improvisation done is 34.4614%. , for Amazon Dataset the metric NDCG@50 the existing system gives a
value of 7.184 and our proposed model denotes a value of 10.02 the improvisation done is 32.9691%, upon
the comparative analysis of the proposed system with the existing system it can be concluded that our
proposed model outperforms the existing model.

Table 7. Comparative analysis

Dataset Details ES DeepNCF Improvisation
Amazon Dataset Recall@20 6.761 13.89 69.0427%
Amazon Dataset Recall@50 10.625 21.48 67.62%
MovieLens Dataset Recall@20 10.985 20.41 60.0414%
MovieLens Dataset Recall@50 23.478 47.05 66.8444%
Amazon Dataset NDCG@ 20 5.613 7.95 34.4614%
Amazon Dataset NDCG@ 50 7.184 10.02 32.9691%

A recommender system-using novel deep network collaborative filtering (Shruthi Nagaraj)

796  ISSN: 2252-8938

5. CONCLUSION
In this paper, a DeepNCF model is developed that addresses the several issues such as overfitting,
sparsity and performance metrics; DeepNCF incorporated the dual mode network into the collaborative
filtering; DeepNCF designs user-item dual mode network from rating matrix; further considering the edges in
network, it carries out edge clustering through custom weighted dual model modularity for higher densing.
Furthermore, GNN is utilized for learning the complex relation among the user and items. The performance
is better of the DeepNCF model in comparison with other existing approaches on two datasets. The edge-
clustering algorithm is applicable here for developing a dual mode graph and implementing a edge reduction
algorithm for every single community. The DeepNCF attains higher metrics in comparison with the existing
methods in aspects of recommendation accuracy and variation in attaining the item from the customer for
rating data and discarding irrelevant information.

REFERENCES
[1] A. Gatzioura, J. Vinagre, A. M. Jorge, and M. Sanchez-Marre, “A Hybrid Recommender System for Improving Automatic
Playlist Continuation,” IEEE Transactions on Knowledge and Data Engineering, vol. 33, no. 5, pp. 1819–1830, 2021, doi:
10.1109/TKDE.2019.2952099.
[2] K. Al Fararni, F. Nafis, B. Aghoutane, A. Yahyaouy, J. Riffi, and A. Sabri, “Hybrid recommender system for tourism based on
big data and AI: A conceptual framework,” Big Data Mining and Analytics, vol. 4, no. 1, pp. 47–55, 2021, doi:
10.26599/BDMA.2020.9020015.
[3] Z. Yang and M. Zhang, “TextOG: A Recommendation Model for Rating Prediction Based on Heterogeneous Fusion of Review
Data,” IEEE Access, vol. 8, pp. 159566–159573, 2020, doi: 10.1109/ACCESS.2020.3020942.
[4] X. Yang, Y. Guo, Y. Liu, and H. Steck, “A survey of collaborative filtering based social recommender systems,” Computer
Communications, vol. 41, pp. 1–10, 2014, doi: 10.1016/j.comcom.2013.06.009.
[5] X. He, L. Liao, H. Zhang, L. Nie, X. Hu, and T. S. Chua, “Neural collaborative filtering,” in 26th International World Wide Web
Conference, WWW 2017, in WWW ’17. International World Wide Web Conferences Steering Committee, Apr. 2017, pp. 173–
182. doi: 10.1145/3038912.3052569.
[6] M. Iqbal et al., “Kernel Context Recommender System (KCR): A Scalable Context-Aware Recommender System Algorithm,”
IEEE Access, vol. 7, pp. 24719–24737, 2019, doi: 10.1109/ACCESS.2019.2897003.
[7] L. Wu, L. Chen, R. Hong, Y. Fu, X. Xie, and M. Wang, “A Hierarchical Attention Model for Social Contextual Image
Recommendation,” IEEE Transactions on Knowledge and Data Engineering, vol. 32, no. 10, pp. 1854–1867, 2020, doi:
10.1109/TKDE.2019.2913394.
[8] J. Zhang, X. Shi, S. Zhao, and I. King, “STAR-GCN: Stacked and reconstructed graph convolutional networks for recommender
systems,” in IJCAI International Joint Conference on Artificial Intelligence, in IJCAI-2019, vol. 2019- August. International Joint
Conferences on Artificial Intelligence Organization, 2019, pp. 4264–4270. doi: 10.24963/ijcai.2019/592.
[9] W. Wang, G. Zhang, and J. Lu, “Hierarchy Visualization for Group Recommender Systems,” IEEE Transactions on Systems,
Man, and Cybernetics: Systems, vol. 49, no. 6, pp. 1152–1163, 2019, doi: 10.1109/TSMC.2017.2699690.
[10] B. Xiao, X. Xie, C. Yang, and Y. Wang, “RTN-GNNR: Fusing Review Text Features and Node Features for Graph Neural
Network Recommendation,” IEEE Access, vol. 10, pp. 114165–114177, 2022, doi: 10.1109/ACCESS.2022.3218882.
[11] A. Breitfuss, K. Errou, A. Kurteva, and A. Fensel, “Representing emotions with knowledge graphs for movie recommendations,”
Future Generation Computer Systems, vol. 125, pp. 715–725, 2021, doi: 10.1016/j.future.2021.06.001.
[12] S. Wu, Y. Tang, Y. Zhu, L. Wang, X. Xie, and T. Tan, “Session-based recommendation with graph neural networks,” 33rd AAAI
Conference on Artificial Intelligence, AAAI 2019, 31st Innovative Applications of Artificial Intelligence Conference, IAAI 2019
and the 9th AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, vol. 33, no. 01, pp. 346–353, 2019,
doi: 10.1609/aaai.v33i01.3301346.
[13] C. Xu et al., “Graph contextualized self-attention network for session-based recommendation,” in IJCAI International Joint
Conference on Artificial Intelligence, in IJCAI-2019, vol. 2019- August. International Joint Conferences on Artificial Intelligence
Organization, 2019, pp. 3940–3946. doi: 10.24963/ijcai.2019/547.
[14] R. Qiu, Z. Huang, J. Li, and H. Yin, “Rethinking the item order in session-based recommendation with graph neural networks,” in
International Conference on Information and Knowledge Management, Proceedings, in CIKM ’19. ACM, Nov. 2019, pp. 579–
588. doi: 10.1145/3357384.3358010.
[15] X. He, K. Deng, X. Wang, Y. Li, Y. D. Zhang, and M. Wang, “LightGCN: Simplifying and Powering Graph Convolution
Network for Recommendation,” in SIGIR 2020 - Proceedings of the 43rd International ACM SIGIR Conference on Research and
Development in Information Retrieval, in SIGIR ’20. ACM, 2020, pp. 639–648. doi: 10.1145/3397271.3401063.
[16] Z. Duan, H. Xu, Y. Huang, J. Feng, and Y. Wang, “Multivariate Time Series Forecasting with Transfer Entropy Graph,” Tsinghua
Science and Technology, vol. 28, no. 1, pp. 141–149, 2023, doi: 10.26599/TST.2021.9010081.
[17] S. M. Al-Ghuribi and S. A. Mohd Noah, “Multi-Criteria Review-Based Recommender System-The State of the Art,” IEEE
Access, vol. 7, pp. 169446–169468, 2019, doi: 10.1109/ACCESS.2019.2954861.
[18] S. Fan et al., “Metapath-guided heterogeneous graph neural network for intent recommendation,” in Proceedings of the ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining, in KDD ’19. ACM, 2019, pp. 2478–2486. doi:
10.1145/3292500.3330673.
[19] W. Fan et al., “Graph neural networks for social recommendation,” in The Web Conference 2019 - Proceedings of the World
Wide Web Conference, WWW 2019, in WWW ’19. ACM, 2019, pp. 417–426. doi: 10.1145/3308558.3313488.
[20] J. Zhao et al., “IntentGC: A scalable graph convolution framework fusing heterogeneous information for recommendation,” in
Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, in KDD ’19. ACM,
2019, pp. 2347–2357. doi: 10.1145/3292500.3330686.
[21] F. Luo, G. Ranzi, X. Wang, and Z. Y. Dong, “Social Information Filtering-Based Electricity Retail Plan Recommender System
for Smart Grid End Users,” IEEE Transactions on Smart Grid, vol. 10, no. 1, pp. 95–104, 2019, doi: 10.1109/TSG.2017.2732346.
[22] M. B. Hossain, M. S. Arefin, I. H. Sarker, M. Kowsher, P. K. Dhar, and T. Koshiba, “CARAN: A Context-Aware Recency-Based
Attention Network for Point-of-Interest Recommendation,” IEEE Access, vol. 10, pp. 36299–36310, 2022, doi:

Int J Artif Intell, Vol. 13, No. 1, March 2024: 786-797

Int J Artif Intell ISSN: 2252-8938  797

10.1109/ACCESS.2022.3161941.
[23] P. Covington, J. Adams, and E. Sargin, “Deep neural networks for youtube recommendations,” in RecSys 2016 - Proceedings of
the 10th ACM Conference on Recommender Systems, in RecSys ’16. ACM, Sep. 2016, pp. 191–198. doi:
10.1145/2959100.2959190.
[24] B. Hidasi, A. Karatzoglou, L. Baltrunas, and D. Tikk, “Session-based recommendations with recurrent neural networks,” 4th
International Conference on Learning Representations, ICLR 2016 - Conference Track Proceedings, 2016.
[25] G. Zhou et al., “Deep interest network for click-through rate prediction,” in Proceedings of the ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, in KDD ’18. ACM, 2018, pp. 1059–1068. doi:
10.1145/3219819.3219823.
[26] Y. Cen, J. Zhang, X. Zou, C. Zhou, H. Yang, and J. Tang, “Controllable Multi-Interest Framework for Recommendation,” in
Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, in KDD ’20. ACM,
2020, pp. 2942–2951. doi: 10.1145/3394486.3403344.
[27] Y. Lu et al., “Future-aware diverse trends framework for recommendation,” in The Web Conference 2021 - Proceedings of the
World Wide Web Conference, WWW 2021, in WWW ’21. ACM, Apr. 2021, pp. 2992–3001. doi: 10.1145/3442381.3449791.
[28] D. Yin and S. Feng, “Enhanced Attention Framework for Multi-Interest Sequential Recommendation,” IEEE Access, vol. 10, pp.
67703–67712, 2022, doi: 10.1109/ACCESS.2022.3185063.

BIOGRAPHIES OF AUTHOR

Shruthi Nagaraj is a Research Scholar in the school of Computer Science and

engineering at Presidency University, Bengaluru. She has a Bachelor's degree in Computer
Science Engineering from VTU, Master's degree in Software Engineering from Ramaiah
Institute of Technology, VTU. She has two years of experience working in the IT industry in
Communication and Network Security. Her area of interests in research are Recommendation
Systems, Deep Learning and Machine Learning Technologies, Artificial Intelligence and
Intelligent Information Retrieval. She can be contacted at: [email protected].

Blessed Prince Palayyan is an Associate Professor with the school of computer

science and engineering at Presidency University. He has served in the education sector for
more than 17 years at college level and among the top university in India and has served with
university abroad. He has received his bachelor’s degree from Manonmaniam Sundaranar
University, Master’s degree from Annamalai University and Doctorate from Anna University,
Chennai. His research interest includes major areas of computer science, particularly security
in computing, Recommendation System, Image Processing, Artificial Intelligence, Robotics
and IoT. He can be contacted at this email: [email protected].

A recommender system-using novel deep network collaborative filtering (Shruthi Nagaraj)