Heterogeneous semantic graph embedding assisted edge sensitive learning for cross-domain recommendation

IAES International Journal of Artificial Intelligence (IJ-AI)
Vol. 14, No. 3, June 2025, pp. 1839~1852
ISSN: 2252-8938, DOI: 10.11591/ijai.v14.i3.pp1839-1852  1839
Journal homepage: http://ijai.iaescore.com
Heterogeneous semantic graph embedding assisted edge-
sensitive learning for cross-domain recommendation
Pravin Ramchandra Patil, Pramod Halebidu Basavaraju
Department of Information Science and Engineering, Adichunchanagiri Institute of Technology affiliated with Visvesvaraya
Technological University, Karnataka, India
Article Info ABSTRACT
Article history:
Received Oct 7, 2024
Revised Feb 15, 2025
Accepted Mar 15, 2025
In the digital age, recommendation systems navigate vast alternatives.
Content-based, collaborative filtering, deep-driven, and cross-domain
recommendation (CDR) have been studied significantly but face cold-start
and data sparsity. Though CDR methods outperform others, they struggle to
optimize user-item matrices. Recent graph-based CDR methods improve
efficiency by leveraging additional user-item interactions; however,
optimizing graph features remains an open research area. Moreover, current
techniques do not consider the impact of noise items (unrelated) on
recommendation accuracy. To address this gap, this paper develops a
heterogeneous semantic graph-embedding (HSGE) edge-pruning model that
leverages user ratings and item metadata in the source and target domains to
recommend items to target domain users. To achieve it, at first Word2Vec
method is applied to explicit and implicit details, followed by Node2Vec-
driven graph embedding matrix generation. Our HSGE method obtains user-
user, user-item, and item-item connections to achieve more semantic
features. To improve accuracy, our model prunes edges that drop source
domain items and allied edges unrelated to the target domain users.
Subsequently, the retained HSGE matrices from both domains are processed
for element-wise attention. A multi-layer perceptron with cosine similarity
processed combined features matrices to generate top-N recommendations
with superior hit-rate (HR) and normalized discounted cumulative gain
(NDCG).
Keywords:
Attention method
Cross-domain recommendation
Edge pruning
Heterogeneous semantic graph
embedding
Node2Vec
Word2Vec
This is an open access article under the CC BY-SA license.
Corresponding Author:
Pravin Ramchandra Patil
Department of Information Science and Engineering
Adichunchanagiri Institute of Technology affiliated with Visvesvaraya Technological University
Belagavi 590018, Karnataka, India
Email: mr.patil1234@gmail.com
1. INTRODUCTION
Over the past few years, there has been high growth in digital data and allied applications serving
various purposes, including business communication, and recommendation systems. As rising digital data
volume requires robust analytics to help users navigate numerous options, as the challenge of choosing the
right one can often lead to unexpected decisions [1], [2]. Over the rising population and allied digital data,
recommender systems have gained widespread attention to alleviate challenges as mentioned earlier [3] in
online digital ecosystems [4]. Recommender systems often exploit users’ previous digital behavior,
purchases, and preferences to recommend certain suitable items like products or services. Recommender
systems are segmented into three primary categories [4]. These exist: i) collaborative filtering, ii) content-
based recommendation, and iii) hybrid recommendation. The collaborative filter methods [5], [6] derive

 ISSN: 2252-8938
Int J Artif Intell, Vol. 14, No. 3, June 2025: 1839-1852
1840
user’s preferences by exploiting their interaction with users and items without exploiting the user’s or item’s
characteristics. On the contrary, content-based methods identify user preferences by assessing the similarity
or shared features of the items that the user has brought or availed of [7]. Hybrid recommendation methods
combine collaborative filtering and content-based recommenders, but achieving optimal solutions is
challenging due to data sparsity issues [8], [9]. Presently, most at-hand recommender solutions rely on
collaborative filtering methods [10] that hypothesize that users with similar interests or behaviors would have
similar item choices. Though a history-based approach is easier to realize; yet, the allied data sparsity
problem remains a bottleneck. Among the most employed item-based collaborative filtering methods [5]
measure item similarity through item-user interactions. These methods rely on inner product
(i.e., linear combination) between the user and item latent vector in a one-hot coding approach, which can
limit performance and fail to capture complex interactions. Accuracy for new users is limited due to
insufficient information about new entrants [7], a problem known as data sparsity. This issue can be
addressed by amalgamating these methods, which employ both user reviews and item metadata for
recommendation decisions.
Recently, different methods like topic modeling [11] and deep networks [12] have been designed to
exploit content information to make recommendation decisions. Unlike these methods, cross-domain
recommendation (CDR) methods [13] exploit user preferences and item characteristics from other domains to
improve data in the receiving domain for recommendation. For instance, CDR can recommend movies based
on a user's book purchases or suggest hotels based on travel history. The cross-domain recommender can
exploit supplementary details from the varied domains to alleviate the problem of cold start and data sparsity
[14]. A few approaches applied transfer learning methods [15], which were later applied to derive cross-
domain collaborative filtering (CDCF) [16], [17] by transferring information from the originating domain to
the receiving domain for an accurate recommendation based on matrix factorization (MF) [18], [19].
However, these methods primarily learn over the shallow and linear characteristics of the user and allied
items. A few approaches apply deep learning methods for the latent feature extraction from users and items
[1], [20]; yet, such methods employ one-hot vector as input and hence often fail in employing collaborative
features between users and items. Such approaches are found suitable only for single-domain
recommendation tasks [21]. Unlike single-domain recommendation models, CDRs are found more efficient
in addressing data sparsity problems, where it can use various transfer learning methods [22], but they face
challenges like costly, complex computation, and underutilization of in-domain data, affecting the
recommender's accuracy.
New developments like CDR have recently addressed these problems by using insights from related
areas, like recommending hotels based on travel history or movie recommendations based on book purchases.
Deep learning approaches provide more complex feature extraction, whereas transfer learning and MF aid in
information transfer across domains. Nevertheless, the efficacy and precision of these suggestions are still
affected by obstacles such as computational complexity and the underutilization of domain data.
A literature review will examine previous work and key figures to highlight developments and
issues in the field. The methodology section will describe the strategy and methods used, while the results
section evaluates findings and potential improvements. Together, these sections clarify the present state and
future of advanced recommendation systems.
2. LITERATURE SURVEY
The CDR methods are classified as semantic, factorial, tag, and graph-based [23]–[25]. However,
among these approaches, the graph-based methods have shown better efficacy. In this method, users and
items represent the nodes, and their inter-element (say, user-item) relationship is referred to as edges, where
the inter-element relation and allied interactions can have a decisive impact on the recommendation accuracy.
Some recent literature discusses the CDR systems exploiting reviews (text-data), transfer learning, deep
learning, and graph-embedding approaches.
2.1. Text-data (review) based recommendation model
To address the data-sparsity problems, utilizing text data is crucial in recommendation systems.
Srifi et al. [26] have used topic modeling to uncover hidden topics in user review text. A MF model called
TopicMF was proposed, which combined user ratings and reviews to obtain latent topics. Social MF [27] and
topic MF were employed to assess data fusion effectiveness for better recommendations. A probabilistic
method [28] combined an integrated latent topic model with a random walk and restart technique for
recommendation. These methods often ignore text order and use a bag-of-words (BOW) method to derive
latent topic variables. Instead, embedding methods created word vectors, learned via convolutional neural
networks (CNN) for prediction.

Int J Artif Intell ISSN: 2252-8938 
Heterogeneous semantic graph embedding assisted edge-sensitive learning … (Pravin Ramchandra Patil)
1841
2.2. Deep-learning-based recommendation model
In the past, numerous efforts have been made by applying deep-learning methods for
recommendation. Zhao [29] applied denoising auto-encoders to learn over the user-item matrix for user and
item profiling. Xue et al. [30] developed an improved neural network-driven MF model, where at first a user-
item matrix was generated over the explicit ratings and non-preference implicit review text. Once obtaining the
user-item matrix, a deep neural network (DNN) was applied to learn the low-dimensional feature space (user-
item matrix). Applying a binary loss function, their proposed deep network performed the recommendation.
Research by Zhang et al. [31], a long short-term memory (LSTM)-based approach was applied to generate text
features for further recommendation decisions. Research by Chen et al. [32], a neural attentional regression
was applied in conjunction with review-level explanations for recommendation decisions, focusing on review
text but ignoring item metadata, and primarily addressed single-domain recommendations.
2.3. Cross-domain recommendation
CDRs exploit the information from one domain to make recommendations in other domains.
Cheema et al. [33] designed a latent user profiling approach that generates a user profile by gathering diverse
user behaviors across multiple data domains. This domain-independent profile and contextual data enable
relevant CDRs. The transfer learning method was applied to make serendipitous item recommendations [34].
A generalized cross domain-multi-dimension tensor factorization model to balance impact amongst domains
effectively [35]. Hu et al. [36] applied a transfer meeting hybrid model for CDR by applying item reviews
and article titles. They used attentively mined semantic features from review text to transfer vital information
from an originating domain to perform learning and recommendation. Yet, these methods could not exploit
user reviews and item metadata together. To alleviate the aforesaid limitation, the authors suggested
transferring knowledge amongst the varied domains to make CDRs [3], [19]. An adaptive codebook transfer
learning (ACTL) was proposed that identifies the suitable codebook scale to reduce computational cost and
improve CDR [19]. Yu et al. [37] developed CDCF, for addressing data sparsity by utilizing rating
information from two auxiliary domains: user-side and item-side, both characterized by dense rating data.
The recommendation task is then reframed as a classification problem based on derived intrinsic features,
which is solved using an SVM. The algorithm [38] addresses data sparsity in the receiving domain by
augmenting user and item features with latent factors from both auxiliary domains, using Funk-singular value
decomposition (SVD) to extract additional features. Two-sided CDCF algorithms [39], like two-stage stacked
ensemble autoencoder (TSSEAE), enhance performance by combining user and item auxiliary domains,
framing the CDCF problem as an ensemble of classifiers. The authors designed the framework in which one
neural network was used to learn the reduced dimensional illustration from the one-hot coding of users and
items, while another network leveraged auxiliary information in a different latent space [40].
Recently, a graph-based CDR method uses user-item interactions from the originating domain to
recommend items in receiving. However, state-of-the-art graph-based recommenders may include irrelevant
items from the originating domain, adding noise and reducing effectiveness. None of the existing methods
could address the problem of noise items. To alleviate this, we propose a heterogeneous semantic graph-
embedding (HSGE)-assisted model that prunes irrelevant edges in the originating domain. Our proposed
method uses the Word2Vec semantic embedding method for the user’s review, ratings, and item details.
Then, the Node2Vec captures user-user, user-item, and item-item interactions from both domains. Here, the
amalgamation of the different interactions constitutes the HSGE method. The model improves semantic
feature embedding vectors by removing unrelated items from the originating domain for better learning and
prediction. Using a multilayer feedforward neural network-multilayer perceptron (MFNN-MLP) with cosine
similarity, it predicts top-N recommendations in the receiving domain, achieving better hit-rate (HR) and
normalized discounted cumulative gain (NDCG) scores than other methods.
3. METHOD
The proposed method encompasses a heterogenous semantic graph embedding matrix, inter-domain
relativeness-driven edge pruning, attention method, and MFNN-MLP for CDR. It processes user reviews and
item details into input layers for each domain, creating semantic graph embeddings of user-user, user-item,
and item-item interactions using Node2Vec. These embeddings are refined through edge pruning to remove
unrelated items, followed by element-wise attention to form a combined feature model. This model is then
used for learning and prediction via an MFNN-MLP with SoftMax for top-N recommendations at the output
layer. The proposed model is depicted in Figure 1.
The cross-domain prediction model encompasses two domains, 𝑝 (origin) and 𝑞 (receiving), each
with users and items containing explicit (user’s rating and review) and implicit (item(s) details) details. By,
the amalgamation of the review and the item’s metadata creates a profile for a user 𝑢 aimed at improving
recommendation accuracy in the receiving domain. In real-time data models, there can be certain users that

 ISSN: 2252-8938
1842
overlap, which means certain users can exist in both the origin (say, p) and the receiving domain (say, q),
serving as a bridge connecting both domains. This forms the basis for the cross-domain recommender design.
The implementation schematic of the planned CDR model with five main components is shown in Figure 1.
These are: i) input layer, ii) HSGE layer, iii) feature amalgamation layer, iv) MFNN-MLP layer, and
v) prediction layer or the output layer.
User (Review) Item (Content)
Word2Vec Word2Vec
Node2Vec Node2Vec
User-User User-Item Item-Item
Edge Pruning
Heterogeneous Semantic Graph Embedding
Metrics
Element-Wise Attention
User (Review) Item (Content)
Word2Vec Word2Vec
Node2Vec Node2Vec
User-User User-Item Item-Item
Edge Pruning
Heterogeneous Semantic Graph Embedding
Metrics
MFNN MLP Layer
Output Layer (Predicted top-N Recommended Items in the Target Domain)
Source Domain Target Domain
1
2
3
4
5
Figure 1. Proposed CDR model
3.1. Input layer
This work uses both explicit user details, like reviews and ratings, and implicit item information,
such as metadata and feedback, as input. The explicit information constructed a profile for each user in both
source as well as receiving domains. The user-item details were then fed as input to the semantic graph
embedding layer.
3.2. Heterogenous semantic graph embedding layer
The user’s review and item details from the originating (𝑝) and the receiving domain (𝑞) are fed as
input to the semantic graph embedding layer. Unlike traditional one-hot encoding, which suffers from
sparsity and noise issues, this novel graph-embedding method leverages user-user, item-user, and item-item
interactions for improved predictions. It can be vital for CDR in both domains (i.e., 𝑝 and 𝑞). Our proposed
method uses heterogeneous graph embedding matrices for user-user, user-item, and item-item interactions in
both source and receiving domains. Here, the user’s rating and the item’s content information related to each
domain (i.e., 𝑝 and 𝑞) are leveraged to constitute heterogenous graph embedding metrics. Unlike one-hot
coding, Word2Vec offers richer semantic, and low-dimensional embedding metrics. We used Gensim's
Word2Vec with skip-gram and negative sampling (SGNS) to transform content-review information of both
𝑝 and 𝑞 domains into corresponding semantic values, distinctly and also to transform each review or item
metadata like titles, sub-titles, and descriptions into a semantic latent vector.
The Node2Vec method extracted HSGE metrics from latent embedding vectors by leveraging
heterogeneous interactions, improving CDRs while addressing data sparsity and cold-start issues. The
Node2Vec method extracted HSGE metrics from latent embedding vectors by leveraging heterogeneous
interactions, improving CDRs while addressing data sparsity and cold-start issues. HSGE was obtained for
both source and receiving domains, resulting in two heterogeneous graph embedding matrices. Noticeably,
the HSGE graphs obtained contained nodes signifying users and items in the originating domain, while
weighted edges represented the user’s review and metadata content similarities obtained for the receiving
domain, respectively. It improved the quality of user and item embedding matrices for learning and
prediction. The homeland security and emergency management (HSEM) model consist of three functional
elements: i) inter-domain relatedness-based edge pruning, ii) Doc2Vec (document) embedding, iii) semantic
graph embedding and HSGE generation. The detail of these sub-components is given as follows.

1843
3.2.1. Inter-domain relatedness-based edge pruning
This work introduces an inter-domain correlation-driven edge pruning model to improve
recommendation accuracy by removing unrelated user-item edges in the receiving domain, preventing false
positive results. To project aforesaid features to the latent embedding matrix or vector, the proposed model
estimates projection vectors in both the originating domain as well as the receiving domain (their
corresponding documents, i.e., 𝑑𝑠,𝑑𝑡). Let, the projection matrices for originating and the receiving domain
be 𝛷𝑠 ∈ ℝ𝑑𝑠×𝑑𝐶𝐶𝐴 and 𝛷𝑡 ∈ ℝ𝑑𝑡×𝑑𝐶𝐶𝐴. These projection vectors are obtained by performing canonical
correlation analysis (CCA) over user’s semantic embedding vector output (i.e., Word2Vec embedding vector
for user) in both domains. In other words, let the user’s embedding vector in originating and receiving
domains be (1) and (2), respectively.
𝐸𝑢𝑠
= [𝑒𝑠
1
, 𝑒𝑠
2
,… , 𝑒𝑠
𝑈
] ∈ ℝ𝑑𝑠×𝑈
(1)
𝐸𝑢𝑡
= [𝑒𝑡
1
, 𝑒𝑡
2
, … , 𝑒𝑡
𝑈
] ∈ ℝ𝑑𝑡×𝑈
(2)
To achieve it, the proposed method performs maximization of the cost function defined in (3).
(𝛷
̂𝑠,𝛷
̂𝑡) = arg max
𝛷𝑠,𝛷𝑡
𝛷𝑠
𝑇𝐶𝐸𝑈𝑠,𝐸𝑈𝑡
𝛷𝑡
√𝛷𝑠
𝑇𝐶𝐸𝑈𝑠,𝐸𝑈𝑠
𝛷𝑠 √𝛷𝑡
𝑇𝐶𝐸𝑈𝑡,𝐸𝑈𝑡
𝛷𝑡
(3)
In (3), the parameters 𝛷𝑠 and 𝛷𝑡 represents the projection embedding vectors, which can mathematically be
derived as per (4) and (5), correspondingly.
𝛷𝑠 = ℝ𝑑𝑠×𝑑𝐶𝐶𝐴 (4)
𝛷𝑡 = ℝ𝑑𝑡×𝑑𝐶𝐶𝐴 (5)
The projection embedding metrics for source and receiving domain documents were obtained by
maximizing the objective function defined in (3). The objective function (3) solved eigenvalue problem that
eventually yielded (4) and (5) for the original and the receiving domains, respectively. The embedding matrix
per domain was obtained as per the (6) and (7), conditioned with the (8) and (9), correspondingly.
𝑒̂𝑠
𝑖
∈ ℝ𝑑𝐶𝐶𝐴 (6)
𝑒̂𝑡
𝑖
∈ ℝ𝑑𝐶𝐶𝐴 (7)
𝑒̂𝑠
𝑖
= 𝛷𝑠
𝑇
𝑒𝑠
𝑖
(8)
𝑒̂𝑡
𝑖
= 𝛷𝑡
𝑇
𝑒𝑡
𝑖
(9)
The proposed model compared the items in originating and receiving domain in respective
embedding latent space. Noticeably, only items with high correlation and similarity were retained. The noise
proneness of the items in the originating domain applies the local density function over the semantic
embedding metrics in source document 𝑑𝑠. In this manner, the difference between the local density of item 𝑖𝑠
and the local density of the neighboring item 𝑜. The noise density for originating domain items is estimated
as per (10).
𝐿𝑅𝐷(𝑖𝑠) =
1
(
∑ 𝑅𝑒𝑎𝑐ℎ−𝑑𝑖𝑠𝑡𝑣(𝑖𝑠,𝑜)
𝑜∈𝑁𝑣(𝑖𝑠)
|𝑁𝑣(𝑖𝑠)|
)
(10)
In (10), 𝑁𝑣(𝑖𝑠) represents a set of 𝑣 adjoining items (in the vicinity of the source item 𝑖𝑠), while 𝐿𝑅𝐷(𝑖𝑠) be
the inverse of the mean of the reachability distance via 𝑣 nearest neighbors. Additionally, 𝑅𝑒𝑎𝑐ℎ −
𝑑𝑖𝑠𝑡𝑣(𝑖𝑠,𝑜) signifying the reachability distance between 𝑖𝑠 and 𝑜 is measured as per (11).
𝑅𝑒𝑎𝑐ℎ − 𝑑𝑖𝑠𝑡𝑣(𝑖𝑠,𝑜) = 𝑚𝑎𝑥(𝑣 − 𝑑𝑖𝑠𝑡(𝑖𝑠),𝑑(𝑖𝑠,𝑜)) (11)
In (11), 𝑣 − 𝑑𝑖𝑠𝑡(𝑖𝑠) represents the distance in between item 𝑖𝑠 and the 𝑣-th neighboring items. Similarly, the
distance between 𝑖𝑠 and 𝑜 items is given by 𝑑(𝑖𝑠,𝑜). Assuming aforesaid outlier rank (10) as the (inter-item)

 ISSN: 2252-8938
1844
distance indicator (i.e., the distance between 𝑖𝑠 and the item set, then the local reachability density of the item
𝑖𝑠 is measured as per (12).
𝐼𝑂𝐹(𝑖𝑠) =
∑
𝐿𝑅𝐷(𝑜)
𝐿𝑅𝐷(𝑖𝑠)
𝑜∈𝑁𝑣(𝑖𝑠)
|𝑁𝑣(𝑖𝑠)|
(12)
In (12), the parameter 𝐼𝑂𝐹(𝑖𝑠) retrieves the extent to which the item 𝑖𝑠 can be called as the noise or the
outlier item having no significant relationship with the receiving domain user. Eventually, the semantic
embedding feature for originating domain 𝑒̂𝑠
𝑖
∈ ℝ𝑑𝐶𝐶𝐴 is obtained after pruning the noisy or irrelevant edges
in originating domain. Now, once obtaining the pruned item semantic vector in the originating domain, the
graph embedding vectors or matrices were obtained for the user-user, user-item, and item-item interactions.
3.2.2. Doc2Vec embedding
The proposed model combines user-user, user-item, and item-item interaction relationships in
heterogeneous semantic embedding matrices and is preserved as documents, creating unit graphs for both
source and receiving domains. Content similarity between users and items is estimated using multi-source
content information. The retained embedding metrics are converted into unit documents for each domain,
collecting a semantic embedding review for each user. Being a multi-source document embedding method,
the semantic embedding review 𝐶𝑖∗ for a user 𝑢𝑖 is collected. For an 𝑖 −th user (say, 𝑢𝑖), the allied user
profile 𝑢𝑝𝑖 is obtained and saved in a single document 𝑑𝑖. For an item 𝑣𝑗, the reviews 𝐶∗𝑗 on the item and
(item) detail 𝑖𝑑𝑗 are collected in the same document 𝑑𝑚+𝑗. Subsequently, the embedding vectors in aforesaid
documents 𝐷 = {𝑑1, 𝑑2, …, 𝑑𝑚+𝑛 } are segmented by using 𝑆𝑡𝑎𝑛𝑓𝑜𝑟𝑑𝐶𝑜𝑟𝑒𝑁𝐿𝑃 tool [41], which extracts
user and item vectors from the input documents in both source and receiving domains. The Doc2Vec method
was applied to map 𝐷 into the text vectors 𝑈𝐶 and 𝑉𝐶 for users and items semantic embedded vectors.
3.3. Semantic graph construction
In this work, the users and items were mapped by exploiting their interaction relationships. The
relative weight of each interaction was measured in sync with the local reachability density (12), which is
derived as the normalized ratings (13).
𝑁𝑜𝑟𝑚𝑎𝑙𝑖𝑧𝑒𝑑 𝑟𝑎𝑡𝑖𝑛𝑔 𝑓𝑜𝑟 𝑒𝑑𝑔𝑒𝑠 =
𝑅
𝑚𝑎𝑥(𝑅)
(13)
In reference to the targeted heterogenous (semantic) graph, the synthetic edges were obtained in between the
two users or the two items based or corresponding normalized edge weights (13). In this manner, the
likelihood 𝑃(𝑖,𝑙) of the edge in between the two users 𝑢𝑖and 𝑢𝑗 is estimated by using (14).
𝑃(𝑖, 𝑙) = 𝛼. 𝑠𝑖𝑚(𝑈𝐶𝑖,𝑈𝐶𝑙) (14)
In (14), 𝛼 refers to a hyper-parameter controlling the sampling likelihood, while 𝑠𝑖𝑚(𝑈𝐶𝑖,𝑈𝐶𝑙) represents
the normalized cosine similarity (NCS) in between 𝑈𝐶𝑖 and 𝑈𝐶𝑙. Be noted, here 𝑈𝐶𝑖 and 𝑈𝐶𝑙 represent the
user profile documents for the 𝑖 − 𝑡ℎ and 𝑙 − 𝑡ℎ user. Similarly, the edges (probability) for item-item are also
measured. Thus, based on the user-item, item-item and user-user interaction relationship, semantic
heterogeneous graphs are obtained for both originating domain 𝐺𝑝
and the receiving domain 𝐺𝑞
.
3.4. Feature amalgamation layer
The semantic embedding matrix is estimated and combined using an element-wise attention
mechanism. This creates an optimal weight set for common users learned across domains, with combined
embeddings for the common users U
̃ from each domain represents the features learned over the different
fractions. In attention methods, a specific fraction of the feature representation is selected, which are
subsequently given higher weights to generate the combined feature [42]. For a user ui, it gives more
attention to the elements possessing higher information from each pair of items in Ui
p
and Ui
q
. The proposed
element-wise attention method generates two embedding matrices U
̃
i
p
and U
̃
i
q
for the common user ui for the
original and the receiving domain, correspondingly. The combined HSEM embedding feature U
̃ i
a
after
attention mechanism for a user ui in originating domain is derived using (15).
𝑈
̃
𝑖
𝑝
= 𝑊𝑝
⊙ 𝑈
̃
𝑖
𝑝
+ (1 − 𝑊𝑝) ⊙ 𝑈𝑖
𝑞
(15)

1845
In (15), ⊙ represents the element-wise attention function, while 𝑊𝑝
∈ ℝ𝑚𝑝×𝑘
refers to the weight matrix
retrieved at the attention layer for domain 𝑝. Thus, we obtained the cumulative embedding matrix for a user
𝑢𝑖 as 𝑈
̃
𝑖
𝑞
for the receiving domain 𝑞. The combined embedding matrices from both domains are fed as input
to the MLP layers (Figure 1) for learning and prediction. An MLP layer snippet is shown as follow.
3.5. Multi-layer perceptron layer
The study uses a fully-connected MLP to capture non-linear user-item interactions in each domain.
This multi-layer feed-forward neural network learns complex interactions, unlike traditional methods that
rely on linear element-wise attention. To learn over the latent features or HSEM embedding metrics of the
users and items, the heterogeneous graph embedding matrix (features) from each domain is fed to the MLP
that obtained the latent embedding matrix of the 𝑝 − 𝑡ℎ domain using (16).
𝑋0
𝑝
= 𝑈𝑒
𝑝
𝑋1
𝑝
= 𝑓(𝑊1
𝑋
.𝑋0
𝑝
+ 𝐵1
𝑋
)
𝑋𝑙
𝑝
= 𝑓(𝑊𝑙
𝑋
.𝑋𝑙−1
𝑝
+ 𝐵𝑙
𝑋
)
𝑋𝑠
= 𝑓(𝑊𝐿𝑚
𝑋
. 𝑋𝐿𝑚−1
𝑝
+ 𝐵𝐿𝑚
𝑋
) (16)
In (16), 𝑋1
𝑝
,… , 𝑋𝐿𝑚
𝑝
and 𝐵1
𝑋
, … , 𝐵𝐿𝑚
𝑋
represent the weight and the bias matrices for the MLP. Here, 𝐿𝑚 states
the total number of MLP layers, while 𝑓(∗) be the rectified linear unit (ReLU) activation function. In (16),
𝑋𝑝
∈ ℝ𝑚×𝑐
be the user’s latent feature matrix for the p-th domain learnt by the MLP method. Thus, the latent
(semantic) embedding matrix for items pertaining to the 𝑝 − 𝑡ℎ domain is obtained as per (17).
𝑌0
𝑝
= 𝑉
𝑒
𝑝
𝑌
1
𝑝
= 𝑓(𝑊1
𝑌
.𝑌0
𝑝
+ 𝐵1
𝑝
)
𝑌𝑙
𝑝
= 𝑓(𝑊𝑙
𝑌
.𝑌𝑙−1
𝑝
+ 𝐵𝑙
𝑝
)
𝑌𝑠
= 𝑓(𝑊𝐿𝑚
𝑌
. 𝑌𝐿𝑚−1
𝑝
+ 𝐵𝐿𝑚
𝑝
) (17)
In (17), 𝑊1
𝑌
,… , 𝑊𝐿𝑚
𝑌
and 𝐵1
𝑌
, …, 𝐵𝐿𝑚
𝑌
represent the weights and the bias matrices for the MLP network.
𝑌𝑠
∈ ℝ𝑛𝑝×𝑐
be the item latent embedding metrics for the 𝑝 − 𝑡ℎ domain.
3.6. Output layer
This layer handles the user-item prediction for the receiving domains. The proposed model executes
MLP training over the input combined user-item latent (say, HSGE) metrics, where the training model is
designed based on the loss in between the predicted and the measured user-item interaction relationship.
Thus, the proposed model is trained 𝑝 by using the following objective function to make predictions for the
receiving domain 𝑞. In (18), 𝑙(𝑦, 𝑦
̂) refers a loss function between the observed user-item interactions
(say, 𝑦) and the predicted user-item interaction 𝑦
̂. In (18), 𝑌𝑝+
and 𝑌𝑝−
represents the observed and the
predicted user-item interactions over the originating domain 𝑝, correspondingly. ‖𝑃𝑝
‖𝐹
2
+ ‖𝑄‖𝐹
2
refers the
regularize, while 𝜆 represents the hyper-parameter controlling the level of significance of the regularize.
min
𝑝𝑝,𝑄𝑝,⊝𝑝
∑ 𝑙(𝑦,𝑦
̂)
𝑦∈𝑌𝑝+⋃𝑌𝑝− + 𝜆(‖𝑃𝑝
‖𝐹
2
+ ‖𝑄‖𝐹
2
) (18)
To alleviate over-fitting towards 𝑌+
, a definite amount of the unobserved user-item interaction
relationships is selected arbitrarily as the negative instances, given by 𝑌𝑠𝑎𝑚𝑝𝑙𝑒𝑑
−
. 𝑌𝑠𝑎𝑚𝑝𝑙𝑒𝑑
−
was used to
substitute 𝑌−
, which helps achieving swift and accurate learning over non-linear input features. Thus, using
user’s review, for a user 𝑢𝑖 and an item 𝑣𝑖, the corresponding user-item interactions 𝑦𝑖𝑗 was obtained (19).
𝑦𝑖𝑗 = {
𝑟𝑖𝑗 𝑖𝑓 𝑦𝑖𝑗 ∈ 𝑌+
0 𝑖𝑓 𝑦𝑖𝑗 ∈ 𝑌𝑠𝑎𝑚𝑝𝑙𝑒𝑑
−
𝑛𝑢𝑙𝑙 𝑂𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
(19)
In this work, a normalized cross-entropy loss function was applied to perform learning, which is
derived as (20).
𝑙(𝑦,𝑦
̂) =
𝑦
𝑚𝑎𝑥(𝑅)
log 𝑦
̂ + (1 −
𝑦
𝑚𝑎𝑥(𝑅)
) log(1 − 𝑦
̂) (20)

 ISSN: 2252-8938
1846
In (20), 𝑚𝑎𝑥(𝑅) represents the maximum rating on the originating domain 𝑎. In this work, MLP is used to
represent a non-linear association amongst the users and items. Let, for the MLP network the input
embedding matrices for the users and items over the originating domain be 𝑀𝑖𝑛
𝑝
= [𝑈
̃𝑝
; 𝑈𝑝𝑑
] and 𝑄𝑖𝑛
𝑝
= 𝑉𝑝
,
correspondingly. Here, 𝑈
̃𝑝
states the combined HSGE embedding matrix pertaining to the common users for
the domain 𝑝, while the embedding matrix of the distinct users in domain 𝑝 be 𝑈𝑝𝑑
. Thus, for user 𝑢𝑖 and
item 𝑣𝑗, the corresponding embedding matrices in the output layer of the MLP are obtained as (21) and (22).
𝑀𝑖
𝑝
= 𝑀𝑜𝑢𝑡𝑖
𝑝
= 𝑓 (…𝑓 (𝑓 (𝑀𝑖𝑛𝑖
𝑝
.𝑊
𝑀1
𝑝
) .𝑊
𝑀2
𝑝
)) (21)
𝑁𝑗
𝑝
= 𝑁𝑜𝑢𝑡𝑗
𝑝
= 𝑓 (…𝑓 (𝑓 (𝑁𝑖𝑛𝑗
𝑝
.𝑊
𝑁1
𝑝
) . 𝑊
𝑁2
𝑝
)) (22)
In (21) and (22), ReLU activation function 𝑓(∗) was applied, while 𝑊
𝑀1
𝑝
and 𝑊
𝑀2
𝑝
…, and 𝑊
𝑀1
𝑝
, 𝑊
𝑁2
𝑝
… be the
weights of the multi-layer networks in varied layers in 𝑝 domain for 𝑀𝑖𝑛𝑖
𝑝
and 𝑁𝑖𝑛𝑗
𝑝
, correspondingly. Finally,
in output layer, the predicted user-item interaction 𝑦
̂𝑖,𝑗 between the user 𝑢𝑖 and the item 𝑣𝑗 on domain 𝑝 is
given by the (23).
𝑦
̂𝑖,𝑗
𝑝
= 𝑐𝑜𝑠𝑖𝑛𝑒(𝑀𝑖
𝑝
,𝑁𝑗
𝑝
) =
𝑀𝑖
𝑝
.𝑁𝑗
𝑝
‖𝑀𝑖
𝑝
‖‖𝑁𝑗
𝑝
‖
(23)
Similarly, the predicted interaction 𝑦
̂𝑖,𝑗
𝑝
on the receiving domain 𝑞 were obtained to perform top-N
recommendation. Thus, the cross-domain recommender model was developed using the aforementioned
methods, as shown in Figure 2, with simulation results and insights discussed in the subsequent section.
Explicit Implicit Explicit Implicit
Layer 1 Layer 1 Layer 1 Layer 1
Layer X Layer X Layer X Layer X
ReLU ReLU ReLU ReLU
ReLU ReLU ReLU ReLU
ReLU ReLU ReLU ReLU
Source Domain p Target Domain q
Training Training Top-N
Prediction
Top-N
Prediction
Element-wise Attention
1
2
3
4
5
1 Input Layer 2 HSGM Layer 3 Attention Layer 4 MLP Layer 5 Output Layer
Figure 2. Proposed CDR model realization
4. RESULTS AND DISCUSSION
4.1. Experimental datasets
In alignment with the CDR task, this research considered two distinct benchmarks datasets.
featuring real-time transaction details that varied significantly in data nature and item types. Specifically, the
study focused on Douban datasets, which include book and movie subsets. We obtained the user and items
with at least four interactions to exploit more behavioral aspects, reduce noise probability, and enhancing
prediction accuracy in the originating domain by using the MFNN-MLP layer. A snippet of the dataset used

1847
in this work is given in Table 1. As depicted in Table 1, the movie dataset encompassed a set of interactions
with 2,718 users and 9,565 items that cumulatively gave rise to the data size of 1,133,420. Similarly, the
book dataset possessed 2,718 users with 6,777 items with a total data size of 96,041. Our considered datasets
encompassed both explicit data and implicit details, where the first embodies the reviews and ratings, while
the latter contains item metadata. The proposed algorithm was developed in Python language, where the
simulation environment encompassed Python 3.4, with the CPU armored with 8 GB memory running over
Intel i5 processor. It also applied different Python libraries including NumPy, maths, SciPy, Gensim
Word2Vec, and TensorFlow. The simulations were made on the Google Colab platform with the processor
configured with the native graphical processing unit.
Table 1. Data specifications
Data User Items Data size
Movie 2,718 9,565 1,133,420
Books 2,718 6,777 96,041
4.2. Simulation parameters
To ensure optimal performance, the proposed model employed numerous value additions in terms of
the design parameters or the hyperparameter tuning. In the deployed recommender model, especially for
HSEM, the sampling probability 𝛼 was assigned as 0.05. In Gensim Word2Vec embedding of the ratings and
items, we assigned window size as 3. The Node2Vec and Doc2Vec models' hyper-parameters were adjusted,
and an MFNN-driven MLP was designed and tuned for different layer architectures. Initialized parameters
were Gaussian distribution, and training was done with 8 negative instances per positive instance,
100 epochs, fixed learning rate, regularization parameter, and 64 batch sizes. Here, the key motive behind
lower batch size was to improve (intrinsic) feature learning over the heterogenous embedding features which
can have high non-linearity over millions of the user’s profiles. Furthermore, we performed feature
embedding concerning the different dimensions including 4, 8, 16, 32, and 64.
4.3. Evaluation parameters
The study focuses on a top-N recommendation task for users in a receiving domain, ranking target
items based on interactions with uninteracted items. A leave-one-out approach is adopted, selecting recent
interactions with a test item for each user. The performance parameters are HR and NDCG, which measure
the recall rate and the quality of the ranking. The aim is to evaluate the effectiveness of this ranking-based
method. The proposed cross-domain recommender model generates a (top-N) ranked list of items (in the
receiving domain) for the target user. If a target item is in this list, it's a hit. We define hit rate (HR) as in
(24). In (24), the denominator |𝑇| represents the total number of interactions in the test set. In this work,
NDCG was applied to provide (or assess) the hit position by assigning higher scores to the hits (especially at
the top K ranks).
𝐻𝑅@𝐾 =
𝑁𝑜.𝑜𝑓 𝐻𝑖𝑡𝑠@𝐾
|𝑇|
(24)
Mathematically, we derived NDCG@K as per the (25). In (25), the parameter 𝑟𝑖 states the ranked
relevance of the target item at the 𝑖-th position. Therefore, in this context, 𝑟𝑖 = 1.
𝑁𝐷𝐶𝐺@𝐾 = ∑
2𝑟𝑖−1
log2(𝑖+1)
𝐾
𝑖=1 (25)
In otherwise case, we label 𝑟𝑖 = 0. Additionally, we also estimated the root mean square error
(RMSE) performance, which was derived as per the (26). In (26), 𝑇 represents the total number of test
ratings, while 𝑅𝑠
(𝑖,𝑗) states the real rating and the measured or predicted rating be 𝑅
̂𝑠
(𝑖, 𝑗).
𝑅𝑀𝑆𝐸 = √
1
𝑇
∑ (𝑅𝑠
(𝑖, 𝑗) − 𝑅
̂𝑠
(𝑖, 𝑗))
2
𝑠,𝑖,𝑗 (26)
This work characterizes overall performance and involves both intra- and inter-model assessments.
Here, the intra-model assessment states the performance characterization with the different embedding
dimensional (say, latent dimensions). These assessments are conducted across different input datasets to
analyze performance trends. These assessments are conducted across different input datasets.

 ISSN: 2252-8938
1848
4.3.1. Intra-model assessment
To evaluate the effectiveness of the inter-domain relatedness-driven edge pruning method, we
simulated our model with and without edge pruning (by feeding direct HSGE metrics to the attention layer). In
addition, we evaluated various latent dimensions (4, 8, 16, 32, and 64), with simulation results for the Douban
book and movie datasets presented in Table 2. It is often hypothesized that the high dimensional features often
yield superior results but increased computation. we simulated our CDR model using five embedding
dimensions (k=4, 8, 16, 32, 64), where 𝑘 is the embedding dimensions. The simulation results in Table 2 show
that as 𝑘 increases, the HR@K also rises. Training on various input data subsets, including Book and Movie,
shows that increasing latent and intrinsic feature learning improves both HR and NDCG. As depicted in Table
2, the average HR@4 was 0.5675, while NDCG@4 was observed as 0.3470. Similarly, in another test case with
HSGE embedding dimension of 8, we measured average HR@8 as 0.5791, while NDCG@8 was found as
0.3805. For k=16, we found HR@16 as 0.5799 and NDCG@16 as 0.3990. On the contrary, for 𝑘 = 32, the
model yielded HR@32 as 0.5998 and NDCG@32 as 0.4171. Similarly, over simulation, it depicted HR@64 of
0.6070 and NDCG@64 as 0.3796. Table 3 shows the average HR and NDCG across various embedding
dimensions, with higher sizes leading to decreased RMSE, demonstrating the model's robustness and efficiency
in providing accurate top-N recommendations. The results indicate that higher embedding dimensions lead to
better HR and NDCG for reliable CDRs. A consistent efficiency pattern across datasets (books and movies)
confirms the model's suitability for real-time recommendation tasks.
To evaluate the impact of inter-domain edge pruning on top-N recommendations, we simulated our
model with and without pruning. Results are shown in Table 4. The result in Table 4 can easily be inferred
that the use of the proposed edge pruning method with HSGE mechanism enables superior HR and
corresponding NDCG performance. The average HR performance increased from 0.5410 without edge-
pruning to 0.5866 with edge-pruning. It reveals that the use of our proposed edge pruning model with HSGE
enables almost 9.2% better performance (i.e., HR). The results confirm that the edge pruning method with the
HSGE model outperforms methods relying solely on user-item interactions, as it reduces noise, significantly
improving both HR and RMSE shown in Tables 3 and 4. The HSGE-CDR model's performance was
evaluated against other leading methods. The results are detailed in the subsequent section.
Table 2. Performance over the different latent embedding dimensions
Latent dimensions or embedding size (K) Data HR@K NDCG@K RMSE
4 Book 0.4531 0.2831 1.409
Movie 0.6820 0.4109 1.389
8 Book 0.4861 0.2994 1.091
Movie 0.6721 0.4617 1.077
16 Book 0.4824 0.2993 0.987
Movie 0.6775 0.4987 0.988
32 Book 0.5015 0.3740 0.938
Movie 0.6982 0.4603 0.874
64 Book 0.5126 0.3779 0.880
Movie 0.7014 0.4813 0.871
Table 3. Average HR and NDCG performance over the different latent dimensions
Latent dimensions or embedding size (K) Average
HR NDCG RMSE
4 0.5675 0.3470 1.399
8 0.5791 0.3805 1.084
16 0.5799 0.3990 0.987
32 0.5998 0.4171 0.906
64 0.6070 0.4296 0.875
Table 4. Efficacy of edge pruning on HSGE driven CDR
Latent dimensions or embedding size (K) Data Without edge pruning With edge-pruning
HR@K NDCG@K RMSE HR@K NDCG@K RMSE
4 Book 0.3831 0.2189 1.998 0.4531 0.2831 1.409
Movie 0.5993 0.3776 1.849 0.6820 0.4109 1.389
8 Book 0.4473 0.2851 1.248 0.4861 0.2994 1.091
Movie 0.6110 0.4294 1.216 0.6721 0.4617 1.077
16 Book 0.4361 0.2840 1.198 0.4824 0.2993 0.987
Movie 0.6485 0.4401 1.233 0.6775 0.4987 0.988
32 Book 0.4791 0.3392 1.194 0.5015 0.3740 0.938
Movie 0.6620 0.4261 1.130 0.6982 0.4603 0.874
64 Book 0.4899 0.3482 1.078 0.5126 0.3779 0.880
Movie 0.6538 0.3399 1.004 0.7014 0.4813 0.871

1849
4.3.2. Inter-model assessment
To assess the performance of the proposed CDR approach, we analyzed results from various state-
of-the-art methods at embedding sizes of 8, 16, 32, and 64. The considered reference methods employ cosine
similarity to perform top-N recommendations in cross-domain applications. Typically, the cosine similarity
method the final component in interaction association analysis for a recommendation, where it provides 𝐾
estimates of most similar items, whereas the cosine similarity method is applied to measure similarity
amongst the items. We assessed performance using HR and NDCG metrics. Before discussing the empirical
results and their implications, we briefly outline these methods:
‒ GMF [20]: it represents a generalized form of the MF method, utilizing a sophisticated neural network
model that learns user-item interaction relationships by employing certain activation functions over the
linear combination of the element-wise multiplication of the input vectors. This approach receives user
identification and corresponding items as one-hot vectors (that often impose computational exhaustion).
According to Ma et al. [2], the retrieved one-hot vectors of the users and items are projected as input to
the dense vectors, which are used as input to the neural network to obtain the prediction scores for the
user-item pairs.
‒ Deep matrix factorization (DMF) [30]: the method generates an implicit user-item embedding matrix,
which is input to a DNN for high-dimensional user and item vectors. The vectors are projected to a low-
dimensional latent space, and the cosine similarity method is applied for recommendation results.
Xue et al. [30] employed a dual-layer DMF method to generate recommendation results.
‒ 𝑁𝑒𝑢𝑀𝐹𝐶: it is an extended approach of 𝑁𝑒𝑢𝑀𝐹 model that employs clustering as a component to make
recommendation results.
‒ 𝑁𝑒𝑢𝑀𝐹
𝑚𝑝: this approach applies Word2Vec on user-item details to generate a semantic vector, which is
then applied as input to the neural network to perform recommendation results. Unlike 𝑁𝑒𝑢𝑀𝐹, this
method applies Word2Vec embedded feature vector as input to the 𝑁𝑒𝑢𝑀𝐹 to perform
recommendation.
Table 5 presents the HR@K performance over the different embedding dimensions, indicating that
larger embedding sizes improve HR performance, with HR@64 showing superior over the other embedding
sizes. The GMF-based CDR model performs the maximum HR for HR@64 as 0.2901 and 2997 for book and
movies data subsets, respectively. On the other hand, the DMF-based model exhibits maximum performance
for HR@64 and was found as 0.3004 and 0.2899 with Book and Movie datasets, respectively. The 𝑁𝑒𝑢𝑀𝐹𝐶
method too exhibits superior with K=64, where it achieves 0.2832 and 0.3421, correspondingly with book
and movie datasets. 𝑁𝑒𝑢𝑀𝐹
𝑚𝑝 as an improved solution than the 𝑁𝑒𝑢𝑀𝐹𝐶 method. The simulation reveals
that the 𝑁𝑒𝑢𝑀𝐹
𝑚𝑝 method exhibits HR@64 of 0.5126 and 0.7014 for book and movie datasets, respectively.
In comparison to the state-of-art methods (i.e., GMF, DMF, 𝑁𝑒𝑢𝑀𝐹𝐶, and 𝑁𝑒𝑢𝑀𝐹
𝑚𝑝), our proposed method
shows superior results of HR@64 as 0.3779 and 0.4813 for book and movie. Similar performances were
obtained with other embedding sizes as well. The NDCG performance in Table 6 reveals a similar
performance as that of HT@K. As depicted in the following result in Table 6, the NDCG@K performance
reveals that the proposed CDR model performance is superior over other approaches (NDCG@64 of 0.3779
and 0.4813, respectively over book and movie datasets). Interestingly, the depth assessment also reveals that,
unlike K=64, our proposed method exhibits superior outputs even with lower K (i.e., K=4, 8, 16). The results
in Table 5 indicate that the average HR@K for K=4, 8, and 16 exhibits 0.5675, 0.5791 and 0.5799,
respectively. Considering K=8 as a moderate embedding size (assumption), our depth analysis reveals that
the GMF and DMF methods show.
Table 5. Relative HR@K efficacy analysis
Method Data HR@K
K=4 K=8 K=16 K=32 K=64
GMF Book 0.0985 0.1094 0.1375 0.2017 0.2901
Movie 0.1246 0.1898 0.2311 0.2458 0.2997
DMF Book 0.1304 0.1969 0.2119 0.2499 0.3004
Movie 0.1332 0.2331 0.2398 0.2503 0.2899
𝑁𝑒𝑢𝑀𝐹𝐶 Book 0.1731 0.2316 0.2390 0.2570 0.2832
Movie 0.1996 0.2402 0.2441 0.2898 0.3421
𝑁𝑒𝑢𝑀𝐹𝑚𝑝 Book 0.2742 0.4362 0.2973 0.3117 0.3568
Movie 0.2995 0.4962 0.3991 0.3996 0.5310
Proposed Method Book 0.4531 0.4861 0.4824 0.5015 0.5126
Movie 0.6820 0.6721 0.6775 0.6982 0.7014
Average NDCG@8 as 0.2776, and 0.2604, respectively. While 𝑁𝑒𝑢𝑀𝐹𝐶 and 𝑁𝑒𝑢𝑀𝐹
𝑚𝑝 methods
output NDCG@8 as 0.2947 and 0.301, respectively. In comparison to these results, our proposed model

 ISSN: 2252-8938
1850
exhibits an average NDCG of 0.3805, which is higher than any other existing approaches and hence confirms
the superiority of the suggested solution over other state-of-the-art approaches About the HR@8 as well,
GMF, DMF, 𝑁𝑒𝑢𝑀𝐹𝐶, and 𝑁𝑒𝑢𝑀𝐹
𝑚𝑝 show outputs as 0.1496, 0.215, 0.2359, and 0.4662. On the contrary,
our proposed CDR model exhibits an average HT@8 of 0.5791, which outperforms any other method. Thus,
it confirms the superiority of the proposed model over other state-of-the-art models.
Table 6. Relative NDCG efficacy analysis
Method Data NDCG@K
K=4 K=8 K=16 K=32 K=64
GMF Book 0.2300 0.2413 0.3318 0.3384 0.3460
Movie 0.3167 0.3139 0.3660 0.3341 0.3952
DMF Book 0.2416 0.2508 0.2440 0.3312 0.3229
Movie 0.2667 0.2701 0.2510 0.2351 0.3550
𝑁𝑒𝑢𝑀𝐹𝐶 Book 0.2412 0.2811 0.2615 0.3691 0.3212
Movie 0.3046 0.3084 0.2801 0.3416 0.3666
𝑁𝑒𝑢𝑀𝐹𝑚𝑝 Book 0.2615 0.2893 0.2729 0.3711 0.3991
Movie 0.3247 0.3127 0.2842 0.3415 0.3779
Proposed method Book 0.2831 0.2994 0.2993 0.3740 0.3779
Movie 0.4109 0.4617 0.4987 0.4603 0.4813
5. CONCLUSION
This paper presents a novel HSGE-assisted edge-sensitive learning approach for cross-domain
recommender design. The model uses user ratings and item metadata in both original and receiver domains to
recommend items for target users. The model employs the Word2Vec method for explicit and implicit
details, followed by Node2Vec-driven graph embedding matrix generation. The HSGE method obtains user-
user, user-item, and item-item interactions for additional representative semantic features. The model also
employs an inter-domain relatedness-sensitive edge pruning method to improve accuracy. The model's
superior HR and NDCG results make it ideal for real-time cross-domain recommender tasks without cold-
start and sparsity problems. The model can alleviate cold-start and sparsity problems and maintain high CDR
efficiency with low computational cost or complexity. Further, the enhanced CDR can be optimized through
advancements in transfer learning techniques, emphasizing the development of more sophisticated models
that effectively capture inter-domain relationships. Exploring domain adaptation techniques and employing
graph neural networks to model complex cross-domain interactions may improve recommendation accuracy.
FUNDING INFORMATION
Authors state no funding involved.
AUTHOR CONTRIBUTIONS STATEMENT
This journal uses the Contributor Roles Taxonomy (CRediT) to recognize individual author
contributions, reduce authorship disputes, and facilitate collaboration. According to the CRediT, the
following table presents the individual author contributions.
Name of Author C M So Va Fo I R D O E Vi Su P Fu
Pravin Ramchandra Patil ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
Pramod Halebidu Basavaraju ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
C : Conceptualization
M : Methodology
So : Software
Va : Validation
Fo : Formal analysis
I : Investigation
R : Resources
D : Data Curation
O : Writing - Original Draft
E : Writing - Review & Editing
Vi : Visualization
Su : Supervision
P : Project administration
Fu : Funding acquisition
CONFLICT OF INTEREST STATEMENT
Authors state no conflict of interest.

1851
DATA AVAILABILITY
Data availability does not apply to this paper, as no new data were created. The dataset used in this
study is publicly available at Kaggle.
REFERENCES
[1] C. D. Wang, Z. H. Deng, J. H. Lai, and P. S. Yu, “Serendipitous recommendation in e-commerce using innovator-based
collaborative filtering,” IEEE Transactions on Cybernetics, vol. 49, no. 7, pp. 2678–2692, 2019, doi:
10.1109/TCYB.2018.2841924.
[2] J. Ma, J. Wen, M. Zhong, W. Chen, and X. Li, “MMM: multi-source multi-net micro-video recommendation with clustered hidden
item representation learning,” Data Science and Engineering, vol. 4, no. 3, pp. 240–253, 2019, doi: 10.1007/s41019-019-00101-4.
[3] L. Huang, Z. L. Zhao, C. D. Wang, D. Huang, and H. Y. Chao, “LSCD: low-rank and sparse cross-domain recommendation,”
Neurocomputing, vol. 366, pp. 86–96, 2019, doi: 10.1016/j.neucom.2019.07.091.
[4] S. Pyo, E. Kim, and M. Kim, “LDA-based unified topic modeling for similar TV user grouping and TV program
recommendation,” IEEE Transactions on Cybernetics, vol. 45, no. 8, pp. 1476–1490, 2015, doi: 10.1109/TCYB.2014.2353577.
[5] V. Kumar, A. K. Pujari, S. K. Sahu, V. R. Kagita, and V. Padmanabhan, “Collaborative filtering using multiple binary maximum
margin matrix factorizations,” Information Sciences, vol. 380, pp. 1–11, 2017, doi: 10.1016/j.ins.2016.11.003.
[6] C. Shi et al., “Deep collaborative filtering with multi-aspect information in heterogeneous networks,” IEEE Transactions on
Knowledge and Data Engineering, vol. 33, no. 4, pp. 1413–1425, 2021, doi: 10.1109/TKDE.2019.2941938.
[7] D. Wang, Y. Liang, D. Xu, X. Feng, and R. Guan, “A content-based recommender system for computer science publications,”
Knowledge-Based Systems, vol. 157, pp. 1–9, 2018, doi: 10.1016/j.knosys.2018.05.001.
[8] A. K. Sahu and P. Dwivedi, “User profile as a bridge in cross-domain recommender systems for sparsity reduction,” Applied
Intelligence, vol. 49, no. 7, pp. 2461–2481, 2019, doi: 10.1007/s10489-018-01402-3.
[9] L. Hu, S. Jian, L. Cao, Z. Gu, Q. Chen, and A. Amirbekyan, “HERS: modeling influential contexts with heterogeneous relations
for sparse and cold-start recommendation,” The Thirty-Third AAAI Conference on Artificial Intelligence (AAAI-19),
pp. 3830–3837, 2019, doi: 10.1609/aaai.v33i01.33013830.
[10] Y. Zhang, C. Yin, Q. Wu, Q. He, and H. Zhu, “Location-aware deep collaborative filtering for service recommendation,” IEEE
Transactions on Systems, Man, and Cybernetics: Systems, vol. 51, no. 6, pp. 3796–3807, 2021, doi: 10.1109/TSMC.2019.2931723.
[11] H. Guo, R. Tang, Y. Ye, Z. Li, and X. He, “DeepFM: a factorization-machine based neural network for CTR prediction,”
Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, pp. 1725–1731, 2017, doi:
10.24963/ijcai.2017/239.
[12] O. Barkan, A. Caciularu, O. Katz, and N. Koenigstein, “Attentive Item2vec: neural attentive user representations,” ICASSP 2020 -
2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Barcelona, Spain, 2020, pp. 3377-
3381, doi: 10.1109/ICASSP40776.2020.9053071.
[13] H. Zhao, Q. Yao, J. Li, Y. Song, and D. L. Lee, “Meta-graph based recommendation fusion over heterogeneous information
networks,” Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining,
pp. 635–644, 2017, doi: 10.1145/3097983.3098063.
[14] L. Hu, L. Cao, J. Cao, Z. Gu, G. Xu, and J. Wang, “Improving the quality of recommendations for users and items in the tail of
distribution,” ACM Transactions on Information Systems, vol. 35, no. 3, 2017, doi: 10.1145/3052769.
[15] L. Nie, X. Song, and T.-S. Chua, Learning from multiple social networks. Cham: Springer International Publishing, 2016, doi:
10.1007/978-3-031-02300-2.
[16] H. Kanagawa, H. Kobayashi, N. Shimizu, Y. Tagami, and T. Suzuki, “Cross-domain recommendation via deep domain
adaptation,” Advances in Information Retrieval, pp. 20–29, 2019, doi: 10.1007/978-3-030-15719-7_3.
[17] A. K. Sahu and P. Dwivedi, “Knowledge transfer by domain-independent user latent factor for cross-domain recommender
systems,” Future Generation Computer Systems, vol. 108, pp. 320–333, 2020, doi: 10.1016/j.future.2020.02.024.
[18] A. K. Sahu and P. Dwivedi, “Matrix factorization in cross-domain recommendations framework by shared users latent factors,”
Procedia Computer Science, vol. 143, pp. 387–394, 2018, doi: 10.1016/j.procs.2018.10.410.
[19] M. He, J. Zhang, and S. Zhang, “ACTL: adaptive codebook transfer learning for cross-domain recommendation,” IEEE Access,
vol. 7, pp. 19539–19549, 2019, doi: 10.1109/ACCESS.2019.2896881.
[20] X. He, L. Liao, H. Zhang, L. Nie, X. Hu, and T. S. Chua, “Neural collaborative filtering,” 26th International World Wide Web
Conference, pp. 173–182, 2017, doi: 10.1145/3038912.3052569.
[21] Z. H. Deng, L. Huang, C. D. Wang, J. H. Lai, and P. S. Yu, “DeepCF: a unified framework of representation learning and
matching function learning in recommender system,” The Thirty-Third AAAI Conference on Artificial Intelligence (AAAI-19),
pp. 61–68, 2019, doi: 10.1609/aaai.v33i01.330161.
[22] W. Pan, “A survey of transfer learning for collaborative recommendation with auxiliary data,” Neurocomputing, vol. 177,
pp. 447–453, 2016, doi: 10.1016/j.neucom.2015.11.059.
[23] C. Zhao, C. Li, and C. Fu, “Cross-domain recommendation via preference propagation graphnet,” in CIKM '19: Proceedings of
the 28th ACM International Conference on Information and Knowledge Management, 2019, pp. 2165–2168, doi:
10.1145/3357384.3358166.
[24] C. Gao et al., “Cross-domain recommendation without sharing user-relevant data,” in WWW '19: The World Wide Web
Conference, 2019, pp. 491–502, doi: 10.1145/3308558.3313538.
[25] D. Yang, J. He, H. Qin, Y. Xiao, and W. Wang, “A graph-based recommendation across heterogeneous domains,” in CIKM '15:
Proceedings of the 24th ACM International on Conference on Information and Knowledge Management, 2015, pp. 463–472, doi:
10.1145/2806416.2806523.
[26] M. Srifi, A. Oussous, A. A. Lahcen, and S. Mouline, “Recommender systems based on collaborative filtering using review texts-
A survey,” Information, vol. 11, no. 6, 2020, doi: 10.3390/INFO11060317.
[27] G. N. Hu, X. Y. Dai, Y. Song, S. J. Huang, and J. J. Chen, “A synthetic approach for recommendation: combining ratings, social
relations, and reviews,” Proceedings of the Twenty-Fourth International Joint Conference on Artificial Intelligence (IJCAI 2015),
pp. 1756–1762, 2015.
[28] S. Feng, J. Cao, J. Wang, and S. Qian, “Recommendations based on comprehensively exploiting the latent factors hidden in items’
ratings and content,” ACM Transactions on Knowledge Discovery from Data, vol. 11, no. 3, 2017, doi: 10.1145/3003728.
[29] J. Zhao, L. Wang, D. Xiang, and B. Johanson, “Collaborative denoising auto-encoders for top-N recommender systems,” SIGIR |
Special Interest Group on Information Retrieval, Paris, France. Amazon Science, pp. 1-6, 2016.

 ISSN: 2252-8938
1852
[30] H. J. Xue, X. Y. Dai, J. Zhang, S. Huang, and J. Chen, “Deep matrix factorization models for recommender systems,” in
Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence (IJCAI-17), pp. 3203–3209, 2017, doi:
10.24963/ijcai.2017/447.
[31] Q. Zhang, J. Wang, H. Huang, X. Huang, and Y. Gong, “Hashtag recommendation for multimodal microblog using co-attention
network,” Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence (IJCAI-17), pp. 3420–3426,
2017, doi: 10.24963/ijcai.2017/478.
[32] C. Chen, M. Zhang, Y. Liu, and S. Ma, “Neural attentional rating regression with review-level explanations,” in WWW '18:
Proceedings of the 2018 World Wide Web Conference, 2018, pp. 1583–1592, doi: 10.1145/3178876.3186070.
[33] A. A. Cheema, M. S. Sarfraz, M. Usman, Q. Uz Zaman, U. Habib, and E. Boonchieng, “KT-CDULF: knowledge transfer in
context-aware cross-domain recommender systems via latent user profiling,” IEEE Access, vol. 12, pp. 102111–102125, 2024,
doi: 10.1109/ACCESS.2024.3430193.
[34] G. Pandey, D. Kotkov, and A. Semenov, “Recommending serendipitous items using transfer learning,” in International
Conference on Information and Knowledge Management, Proceedings, 2018, pp. 1771–1774, doi: 10.1145/3269206.3269268.
[35] A. Taneja and A. Arora, “Cross domain recommendation using multidimensional tensor factorization,” Expert Systems with
Applications, vol. 92, pp. 304–316, 2018, doi: 10.1016/j.eswa.2017.09.042.
[36] G. Hu, Y. Zhang, and Q. Yang, “Transfer meets hybrid: a synthetic approach for cross-domain collaborative filtering with text,”
in WWW '19: The World Wide Web Conference, 2019, pp. 2822–2829, doi: 10.1145/3308558.3313543.
[37] X. Yu, Y. Chu, F. Jiang, Y. Guo, and D. Gong, “SVMs classification based two-side cross domain collaborative filtering by
inferring intrinsic user and item features,” Knowledge-Based Systems, vol. 141, pp. 80–91, 2018, doi:
10.1016/j.knosys.2017.11.010.
[38] X. Yu, F. Jiang, J. Du, and D. Gong, “A cross-domain collaborative filtering algorithm with expanding user and item features via
the latent factor space of auxiliary domains,” Pattern Recognition, vol. 94, pp. 96–109, 2019, doi: 10.1016/j.patcog.2019.05.030.
[39] X. Yu, Q. Peng, L. Xu, F. Jiang, J. Du, and D. Gong, “A selective ensemble learning based two-sided cross-domain collaborative
filtering algorithm,” Information Processing and Management, vol. 58, no. 6, 2021, doi: 10.1016/j.ipm.2021.102691.
[40] J. He, R. Liu, F. Zhuang, F. Lin, C. Niu, and Q. He, “A general cross-domain recommendation framework via bayesian neural
network,” in 2018 IEEE International Conference on Data Mining (ICDM), Singapore, 2018, pp. 1001–1006, doi:
10.1109/ICDM.2018.00125.
[41] C. D. Manning, M. Surdeanu, J. Bauer, J. Finkel, S. J. Bethard, and D. McClosky, “The stanford CoreNLP natural language
processing toolkit,” in Proceedings of 52nd Annual Meeting of the Association for Computational Linguistics: System
Demonstrations, Baltimore, Maryland, 2014, pp. 55–60, doi: 10.3115/v1/p14-5010.
[42] D. Bahdanau, K. H. Cho, and Y. Bengio, “Neural machine translation by jointly learning to align and translate,” 3rd International
Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings, CoRR, 2015.
BIOGRAPHIES OF AUTHORS
Pravin Ramchandra Patil received a degree in Bachelor of Information Science
and Engineering from Shivaji University, Kolhapur, Maharashtra, India, and M.E. in
computer science and engineering from Shivaji University, Kolhapur, Maharashtra, India. He
is pursuing a Ph.D. in information science and engineering at Adichunchanagiri Institute of
Technology, Chikkamagaluru affiliated with Visvesvaraya Technological University,
Belgaum Karnataka, India. He is currently working as an assistant professor in the Department
of Information Technology at Vasantdada Patil College of Engineering and Visual Arts, Sion,
Mumbai, Maharashtra, India. He has 14 years of teaching experience. His research interest
includes recommendation system, artificial intelligence, deep learning, and machine learning.
She has published papers in conferences and international journals. He can be contacted at
email: mr.patil1234@gmail.com or pravin_patil@pvppcoe.ac.in.
Dr. Pramod Halebidu Basavaraju has an experience of 12 and above years as
an academician, currently working as an associate professor in the Department of Information
Science and Engineering, Adichunchanagiri Institute of Technology, Chikkamagaluru
affiliated with Visvesvaraya Technological University, Karnataka, India. In his credit there are
19 research papers were published in reputed journals, 9 research papers, and 10 papers have
been presented at international and national conferences respectively. He received a Bachelor
of Engineering degree in computer science and engineering from the Visvesvaraya
Technological University in 2007, and a Master of Technology degree in computer science
from the University of Mysore in 2012. He received a doctorate, Ph.D. in the field of wireless
sensor networks from the Department of Computer Science and Engineering, Shri
Venkateshwara University, Uttar Pradesh in 2019. His research area includes wireless sensor
networks, network security, and data analytics. He can be contacted at email:
hbpramod@aitckm.in.

Heterogeneous semantic graph embedding assisted edge sensitive learning for cross-domain recommendation

More Related Content

More from IAESIJAI (20)

Recently uploaded (20)

Heterogeneous semantic graph embedding assisted edge sensitive learning for cross-domain recommendation