Cognitive-Based Enhanced Accuracy and Relevance in Cross-Domain Recommendations

1  Introduction

The rise of online platforms has significantly transformed the way users interact with information, products, and services across various domains. Whether in entertainment, e-commerce, education, or healthcare, recommendation systems (RS) have become essential to enhancing user experience by helping users navigate the overwhelming amount of available content [1]. These systems typically analyze user preferences within a specific domain (e.g., recommending movies on a streaming platform or books on an e-commerce site). However, users often engage with multiple domains simultaneously, introducing a new challenge for traditional RS, which are limited to domain-specific user interactions [2,3]. To bridge this gap, cross-domain RSs (CDRS) have emerged as a promising approach that leverages user data from one domain (the source domain) to enhance recommendations in another (the target domain).

The motivation for cross-domain recommendations arises from several key observations. First, user preferences are often interrelated across domains. For instance, a user who enjoys action movies might also prefer high-adrenaline books, or someone who frequently purchases fitness equipment may be interested in health-related products or services. Traditional RSs overlook this opportunity by focusing solely on single-domain user behavior, thereby limiting their ability to offer more comprehensive, personalized recommendations. Second, in many cases, the target domain may suffer from data sparsity, meaning there is insufficient user interaction data to accurately predict preferences [4]. By incorporating data from the source domain, CDRS can mitigate this issue, allowing the system to provide more accurate and diverse recommendations [5]. Finally, cross-domain recommendations enhance user engagement and satisfaction by providing a seamless experience across platforms and domains, thereby increasing retention and conversion rates for businesses.

Despite these advantages, several challenges hinder the implementation of effective CDRS [6–8]. The primary challenge is modeling and transferring knowledge between domains, especially when the domains are highly heterogeneous (e.g., movies and books). Traditional collaborative filtering (CF) approaches, successful in single-domain settings, struggle in cross-domain scenarios due to the complexity of aligning user-item interactions across different domains [9]. Additionally, ensuring that the transferred knowledge remains relevant and accurate in the target domain remains a key challenge, as not all user preferences in one domain are applicable in another.

This research addresses these challenges by introducing a novel cognitive similarity-based cross-domain (Cognitive-CDR) recommendation approach. The proposed method leverages cognitive similarity between users in different domains to transfer knowledge from a well-populated source domain to a sparsely populated target domain. Cognitive similarity captures semantic relationships among user preferences across domains, enabling the system to understand and predict user preferences even when direct interactions with items in the target domain are limited. Our approach is designed to overcome the limitations of traditional cross-domain methods by explicitly accounting for the cognitive connections between user behaviors across domains. This is particularly beneficial in cases where user preferences exhibit cross-domain coherence (e.g., users who prefer particular genres of movies and books). Compared with existing cross-domain recommendation methods based on latent-factor alignment, adversarial adaptation, or graph propagation, the proposed Cognitive-CDR introduces a user-centric cognitive similarity framework. Instead of relying solely on latent embedding correspondence, Cognitive-CDR integrates behavioral, semantic, and textual preference signals into a unified cognitive representation. This design provides both effective cross-domain transfer and enhanced interpretability through explicit similarity computation and visualization. The contributions of this study are described as follows.

•   Proposes a Cognitive-CDR model that bridges user preferences across domains by leveraging shared cognitive features.

•   Develops an algorithm to predict user preferences in a target domain using user interactions from a source domain.

•   Uses MDS and heatmaps to visually represent cognitive similarity among users, thereby enhancing interpretability.

The remainder of this manuscript is organized as follows. The next Section 2 details related work on cross-domain methods. The proposed method is described in Section 3. In Section 4, the datasets, settings, and experimental results are presented, including a comparison of the proposed methods and baselines on recommendation accuracy and performance. This section also describes the results of our MDS-based visualization and provides insight into the visualizations. Finally, we conclude the study and outline directions for future work in Section 5.

2  Related Work

CDRS has emerged as a prominent research area, offering solutions for transferring knowledge across domains to enhance the performance of recommendations. This section presents related works, categorized into the following areas: factorization-based models, domain adaptation approaches, adversarial learning methods, graph-based models, and knowledge transfer techniques. Recent advances have increasingly explored foundation-model and contrastive-learning paradigms for cross-domain recommendation [10]. Several studies employ graph contrastive learning to align user representations across domains [11,12], while others leverage large language models and semantic embeddings to improve cross-domain transferability and cold-start performance [13]. Despite their effectiveness, these approaches often rely on complex neural architectures and provide limited interpretability. In contrast, the proposed Cognitive-CDR focuses on explicit cognitive preference modeling and interpretable similarity computation while maintaining competitive recommendation performance.

2.1 Factorization-Based Models

Factorization methods are widely used in cross-domain recommendation systems (CDRS) because they learn compact latent representations that can be shared across domains. A typical multi-domain objective jointly factorizes domain-specific interaction matrices while regularizing shared user representations:

min{U,V(d)}∑d∈DL(d)(R(d),UV(d)⊤)+λ(‖U‖2+∑d‖V(d)‖2)+ΩCD,(1)

where U represents shared user factors, V(d) denotes domain-specific item factors, and ΩCD enforces cross-domain consistency.

Several approaches extend this framework to enhance transfer. Loni et al. [14] apply Factorization Machines (FM) to incorporate side features (e.g., user IDs, item attributes, or domain tags), enabling soft information sharing across domains. Sahu and Dwivedi [15] propose domain-invariant user representations, where a shared latent user space improves generalization for sparse target domains. Study [16] introduce cluster-level latent factorization, transferring knowledge through user/item clusters to improve robustness under sparsity. Zhang et al. [17] further enhance collective matrix factorization by integrating knowledge graphs, which provide semantic relations across domains and strengthen cross-domain transfer.

2.2 Domain Adaptation Approaches

Domain adaptation (DA) reduces the distribution gap between source and target domains, enabling knowledge transfer with minimal negative shift. A general DA objective combines task losses with a distribution-alignment term:

minθs,θtLsrc(θs)+Ltgt(θt)+λD(fs(Xs;θs),ft(Xt;θt)),(2)

where fs and ft are representation functions and D(⋅,⋅) measures cross-domain discrepancy (e.g., adversarial alignment, MMD, or CORAL).

Several studies extend this principle for cross-domain recommendation. Kanagawa et al. [18] first apply deep domain adaptation to align latent representations across domains while preserving user privacy. Following on this idea, Tiroshi and Kuflik [19] introduce domain selection, transferring knowledge only from the most relevant source domains to reduce negative transfer. Focusing further on user-level consistency, Zhao et al. [5] align user interest distributions across domains so that semantically similar users remain close in the shared representation space. Other, Gao et al. [20] address practical deployment issues by proposing privacy-preserving cross-domain recommendation that exchanges representations instead of raw user data. Overall, DA-based CDR methods align latent spaces, select useful source domains, and preserve user privacy. However, they may struggle when user preferences differ significantly across domains or when alignment overly constrains target learning. Our Cognitive-CDR complements DA with cognitive-similarity modeling, domain alignment, and a cross-domain adapter, balancing transferable knowledge with domain-specific signals, which is particularly effective under cold-start and sparse data conditions.

2.3 Adversarial Learning Models

Adversarial objectives help disentangle domain-invariant from domain-specific signals by training encoders to fool a domain discriminator while still optimizing recommendation loss. A typical setup solves

minE,RmaxDLrec(R∘E)+λLadv(D(E(⋅))),(3)

where E encodes users/items, R scores user-item pairs, and D predicts domain labels; the adversarial term encourages domain-invariant embeddings. Hao et al. [21] propose adversarial feature translation between domains (with consistency constraints), yielding substantial gains by aligning representational geometry across domains. Similarly, Liu et al. [22] couple adversarial alignment with attention to emphasize transferable user signals, effectively capturing cross-domain preferences while preserving domain-specific nuances.

2.4 Graph-Based Models

Graphs naturally capture higher-order relations among users and items within and across domains. Let each domain d∈{A,B} be represented as a bipartite graph G(d)=(U∪I(d),ℰ(d)) with normalized adjacency A~(d). A L-layer GCN encoder propagates information as

H(l+1,d)=σ(A~(d)H(l,d)W(l)),Z(d)=H(L,d),(4)

where shared weights W(l) encourage transferable structures across domains. Liu et al. [23] extend this framework with a bi-directional transfer graph model that propagates signals between domains (source ↔ target), enabling cross-domain diffusion of user–item signals. To further align representations, cross-domain losses can be applied on shared users:

Lalign=∑u∈𝒰∩‖zu(𝒜)−zu(ℬ)‖22orLInfoNCE(u)=−log⁡exp⁡(⟨zu(𝒜),zu(ℬ)⟩/τ)∑vexp⁡(⟨zu(𝒜),zv(ℬ)⟩/τ).(5)

Beyond user–item graphs, Liu et al. [24] incorporate cross-domain knowledge graphs (KGs) to model semantic relations among items. Items and entities are learned using KG objectives (e.g., DistMult or TransE) and combined with collaborative signals:

max∑(h,r,t)∈𝒯ϕ(eh,r,et)−∑(h′,r,t′)∈𝒯−ϕ(eh′,r,et′),y^ui(d)=g(zu(d),zi(d)⊕ei).(6)

These graph-based approaches are effective under sparsity and cold-start conditions by exploiting high-order neighborhoods and semantic relations. However, they may incur higher computational cost and require regularization to avoid over-smoothing across domains.

2.5 Knowledge Transfer Method

A central theme in cross-domain recommendation is to transfer user preferences or latent factors learned in a source domain to a target domain while avoiding negative transfer. Zhang et al. [25] formulate transfer in a probabilistic manner by imposing a prior on target latent factors conditioned on source factors, e.g.,

p(U(t)∣U(s))∝exp⁡(−12σ2‖U(t)−U(s)M‖F2),(7)

which yields a MAP objective with a quadratic coupling term that regularizes the target space toward a mapped source space via M (learned). In a related vein, Man et al. [26] bridge domains by embedding & mapping: both domains learn their factors, and an explicit linear (or nonlinear) mapping aligns source to target, typically through

minΘLsrc+Ltgt+λ‖U(t)−ϕ(U(s),Θ)‖F2,(8)

where ϕ captures cross-domain correspondences beyond a simple rotation. To mitigate user heterogeneity, Zhu et al. [27] advocate personalized transfer: each user u, has a transfer weight αu∈[0,1] controlling how strongly the source guides the target, e.g.,

minLtgt+∑uαu‖uu(t)−ϕ(uu(s))‖22withαu=gate(sparsityu,similarityu)(9)

so that highly sparse or well-aligned users benefit more from transfer, while others rely primarily on target evidence. Complementarily, Tiroshi and Kuflik [19] introduce domain ranking to prioritize which sources to transfer from, selecting a subset 𝒮⋆ that maximizes estimated relevance/utility to the target:

𝒮⋆=arg⁡max𝒮⊆Dsrc,|𝒮|=k∑d∈𝒮ω(d,tgt)⋅Δ(d→tgt),(10)

thereby reducing negative transfer from weakly related domains.

2.6 Content and Tag-Based Models

A complementary line of work transfers knowledge through descriptive signals such as item content, tags, and reviews, allowing semantically similar users or items to align across domains even when interactions are sparse. Let xu and xi(d) denote user and item content features in domain d. A simple content-aware scoring function is:

y^ui(d)=⟨W(d)xu,Q(d)xi(d)⟩,(11)

where W(d) and Q(d) are learnable projections. Cross-domain coupling can be achieved by sharing or aligning content projections across domains.

Several approaches extend this idea. Sahebi and Walker [28] propose segmented content transfer, where users or items are grouped by semantic facets (e.g., genres or topics), and segment-specific models are jointly learned across domains:

min{Θs}∑s(Ls(𝒜)+Ls(ℬ))+λ∑s‖Θs(𝒜)−Θs(ℬ)‖22.(12)

This restricts transfer to semantically coherent segments, thereby reducing negative transfer. Tags can also serve as cross-domain bridges. Shi et al. [29] learn tag embeddings zt in a user–tag–item graph and incorporate aggregated tag representations into the scoring function:

y^ui(d)=g(pu,qi(d)⊕1|Ti(d)|∑t∈Ti(d)zt),(13)

encouraging items with shared tags to align across domains. Finally, review texts provide richer semantic signals. Xin et al. [30] encode reviews into item embeddings ei(d) and apply contrastive alignment to match the same items across domains:

LCL=−∑i∈ℐ∩log⁡exp⁡(sim(ei(𝒜),ei(ℬ))/τ)∑jexp⁡(sim(ei(𝒜),ej(ℬ))/τ).(14)

Such semantic signals help stabilize cross-domain transfer when interaction data is limited.

3  Proposed Method

This section outlines the process of visualizing cross-domain recommendations using cognitive similarity, which involves several steps. These steps include defining cognitive similarity, developing a recommendation algorithm, and visualizing the results.

3.1 User-Based Feature Extraction

In CDRS, the first step is extracting features from user–item interactions to represent user preferences and behaviors across domains (e.g., books and movies). The goal is to construct a feature vector for each user that summarizes their interactions within a domain and enables cross-domain comparison. Let U={u1,…,uN} denote the set of users and Id the set of items in domain d. For each user ui, we build a feature vector uid that captures interaction characteristics such as ratings, genres, and review keywords. These vectors form the basis for computing cognitive similarity across domains.

•   Average Rating Feature: For a user ui in domain d, the average rating feature is calculated as the mean of the user’s ratings for all items Id they have interacted with.

avg_ratingid=1|Iid|∑j∈Iidrij(15)

where Iid is the set of items in domain d rated by user ui. rij is the rating given by user ui to item j∈Iid.

•   Genre Preferences Feature: Each item j∈Id belongs to one or more genres from a set of G distinct genres. We define the genre preference vector gid for user ui in domain d, where each element represents the fraction of items interacted with in a specific genre gk.

gid=[|Iid(g1)||Iid|,|Iid(g2)||Iid|,…,|Iid(gG)||Iid|](16)

where |Iid(gk)| is the number of items that user ui has interacted with in genre gk. G is the total number of distinct genres in the domain d.

•   Keywords from Reviews Feature: User reviews are processed to extract keywords or topics reflecting user interests. The keyword vector kid for user ui in domain d is constructed from the set of keywords W derived from the user’s reviews.

kid=[w1d,w2d,…,wTd](17)

where wjd represents the importance (e.g., TF-IDF weight) of keyword j for user ui in domain d. T is the total number of keywords considered in domain d.

The complete feature vector uid for user ui in domain d combines the extracted features [avg_ratingid,gid,kid]. This vector includes the average rating feature, the genre preference vector, and the keywords vector extracted from reviews.

3.2 Cognitive Similarity

Cognitive similarity refers to the degree of similarity between two users’ cognitive patterns, preferences, or behaviors across different domains. It is computed by comparing users’ feature vectors across domains and quantifying the alignment or overlap between them. Feature alignment does not require learning an additional neural mapping function. Instead, heterogeneous information from different domains is transformed into a common cognitive feature space. Specifically, rating statistics are normalized into comparable scales, genre preferences are represented as probability distributions, and textual reviews are converted into TF-IDF vectors. This unified representation enables direct similarity computation between users originating from different domains while preserving semantic interpretability. This study defines cognitive similarity as follows.

Definition 1: Let uid1 represent the feature vector of user ui in domain d1, and ujd2 represent the feature vector of user uj in domain d2. The cognitive similarity between users ui and uj across domains d1 and d2, denoted as Sijd1,d2, can be defined as:

Sijd1,d2=uid1⋅ujd2‖uid1‖‖ujd2‖,(18)

where uid1⋅ujd2 is the dot product of the feature vectors of users ui and uj from domains d1 and d2, respectively. ‖uid1‖ and ‖ujd2‖ are the Euclidean norms (magnitudes) of the feature vectors uid1 and ujd2, respectively. This formula calculates the cosine similarity between the two users’ feature vectors, yielding a value between −1 and 1, where Sijd1,d2=1 indicates perfect cognitive similarity, 0 indicates no cognitive similarity, and −1 indicates perfect cognitive dissimilarity.

The cognitive similarity between users across domains is estimated in two steps.

•   Feature alignment across domains: Let uid1 and ujd2 denote the feature vectors of user ui in domain d1 (e.g., Movies) and user uj in domain d2 (e.g., Books). Since features may differ across domains, we align them into a shared feature space F (e.g., average rating, genre preferences, and review keywords):

uid1=[f1d1(ui),…,fFd1(ui)],ujd2=[f1d2(uj),…,fFd2(uj)].

Before computing cognitive similarity, domain-specific user features are projected into a shared cognitive representation space. Let ϕd:𝒳d→ℱ denote a feature harmonization function that transforms domain-specific attributes into a common cognitive space ℱ. The transformed user representations are:

zid1=ϕd1(uid1),zjd2=ϕd2(ujd2).

The harmonized representations contain comparable dimensions corresponding to normalized rating behavior, preference distributions, and textual-interest features. Cognitive similarity is then computed between zid1 and zjd2.

•   Cognitive similarity computation: The similarity between users ui and uj across domains is computed using cosine similarity from Eq. (18):

Sijd1,d2=zid1⋅zjd2‖zid1‖‖zjd2‖.

The score ranges from −1 (completely dissimilar) to 1 (perfectly similar). Cosine similarity was selected because it is scale-invariant and robust to differences in feature magnitude across domains. Since cognitive profiles combine ratings, genre distributions, and textual features with heterogeneous ranges, cosine similarity focuses on preference orientation rather than absolute values. Alternative measures such as Euclidean distance, Pearson correlation, and Jaccard similarity were considered; however, cosine similarity offers a favorable balance between computational efficiency, interpretability, and robustness for high-dimensional sparse representations.

Example 1: Assume the feature vectors for user ui in the Movies domain and user uj in the Books domain are uiMovies=[4.2,giMovies,kiMovies],andujBooks=[4.5,gjBooks,kjBooks], where g represents genre preferences and k represents review-keyword features. Following Eq. (18), the cosine similarity SijMovies,Books=uiMovies⋅ujBooks‖uiMovies‖‖ujBooks‖ measures how similar the two users’ preferences are across domains. A higher value indicates stronger cognitive similarity.

3.3 Recommendations Generation

The proposed cross-domain recommendation algorithm exploits cognitive similarity between users to transfer knowledge from a source domain to a target domain, improving recommendation accuracy. It combines user-based collaborative filtering with cross-domain similarity. Although experiments in this study focus on the common two-domain setting (Books and Movies), the proposed formulation is directly applicable to multi-source recommendation scenarios.

•   Input: user set U, item sets ℐd1 (source) and ℐd2 (target), rating matrices Rd1 and Rd2, and cognitive similarity scores Sijd1,d2 between users across domains.

•   Output: a ranked list of recommended items from ℐd2 for user ui.

•   Procedure: First, cognitive similarity between the target user ui and source-domain users is computed using cosine similarity of their feature vectors. Next, ratings for unseen target items are predicted using a weighted sum of ratings from the top-N similar users. Items are then ranked by predicted scores, and the top-N items are returned as recommendations (Algorithm 1).

The proposed framework can be naturally extended to multiple source domains. Assume a target domain dt and a set of source domains 𝒟s={d1,d2,…,dm}. For each source domain dk, cognitive similarity scores are computed independently, Sij(dk,dt). The final prediction can be obtained through weighted aggregation:

r^u,i=∑k=1mωkr^u,i(dk),

where ∑k=1mωk=1 and ωk represents the relevance of source domain dk to the target domain.

3.4 Multidimensional Scaling Visualization

To visualize cognitive similarity between users across domains, we use Multidimensional Scaling (MDS), a dimensionality-reduction technique that maps high-dimensional user profiles into a two-dimensional space while preserving pairwise distances. Each point in the resulting plot represents a user, and the distance between points reflects their cognitive similarity. First, the cognitive similarity matrix Sijd1,d2 is converted into a dissimilarity matrix using Dijd1,d2=1−Sijd1,d2, where larger values indicate lower similarity. MDS then projects the data into a 2D space by finding coordinates xi∈R2 that best preserve these dissimilarities by minimizing the stress function:

Stress(X)=∑i,j(Dijd1,d2−‖xi−xj‖)2.(19)

The resulting visualization allows us to observe clusters of cognitively similar users, detect outliers, and better understand relationships between users across domains.

4  Experiments

4.1 Datasets

To evaluate the performance of our proposed Cognitive-CDR, we use two subsets of the Amazon Product Reviews dataset: “Books” and “Movies and TV” (the latter is referred to as “Movies”). This dataset is widely used in RS research due to its diverse product offerings, large user base, and high volume of ratings and reviews.

The Books dataset contains 8,898,041 reviews and 22,507,155 ratings, while the Movies dataset includes 1,697,533 reviews and 4,607,047 ratings. Both datasets provide reviews, ratings, and metadata that support feature extraction and cross-domain analysis. Dataset statistics are summarized in Table 1. We retain users with at least five interactions in either domain and binarize interactions (rating ≥4 as positive). Following a leave-one-out protocol, the latest interaction is used for testing, the second latest for validation, and the rest for training. For each test instance, 99 negative items are sampled, yielding candidate lists of 100 items. Results are averaged over five random seeds.

4.2 Settings

Regarding implementation, we set the embedding size to 64, train with Adam (lr = 10−3), batch size 1024, and ℓ2 regularization tuned over {10−5,10−4,10−3}. We stop on validation NDCG@10 after 10 epochs of patience to avoid overfitting while preserving strong early-rank performance. For each user, we sample 99 negatives to form a size-100 candidate set at test time; the same random seeds are used for negative sampling and parameter initialization. All results are averaged over five seeds, and we report means. We compare the proposed Cognitive-CDR method with the following baselines.

•   PopRec: Rank by item popularity in the target domain.

•   ItemCF: Item-based CF with cosine similarity.

•   BPR-MF [31]: Matrix factorization trained with BPR loss.

•   NeuMF [32]: GMF+MLP fusion for implicit feedback.

•   NGCF [33]: Graph-based CF over user-item graph.

•   LightGCN [34]: Simplified GCN for recommendation.

•   CMF [35]: Collective Matrix Factorization for cross-domain.

•   CoNet [36]: Cross-connection networks for cross-domain transfer.

For each user u, we sample 99 negative items and combine them with the held-out positive item y^u to form a candidate set of 100 items. The model ranks all items, and the position of the true item is denoted by ru=ranku(y^). Let U be the set of users and I{⋅} the indicator function. We evaluate top-K recommendation performance (K∈{5,10}) using standard metrics: HR@K=1|U|∑u∈UI{ru≤K}, NDCG@K=1|U|∑u∈UI{ru≤K}log2⁡(1+ru), MRR=1|U|∑u∈U1ru, MAP@K=1|U|∑u∈UI{ru≤K}ru, and P@K=1|U|∑u∈U1KI{ru≤K}. The experiment scenarios are as follows:

•   S1: In-Domain (ID) Train and evaluate within the same domain (Books-only or Movies-only) to measure pure in-domain performance without transfer.

•   S2: Cross-Domain (CD) Train on a source domain and evaluate on a target domain for shared users (Books → Movies, Movies → Books) to test cross-domain knowledge transfer.

•   S3: Cold-Start (CS) Evaluate users with no target-domain interactions but available source-domain history.

•   S4: Sparsity (SP) Randomly remove {20%,40%,60%} of training interactions to test robustness under limited data.

•   S5: Ablation (AB) Remove key components (e.g., cognitive similarity, adapters, alignment losses) to measure their impact.

•   S6: Sensitivity (SN) Vary the cognitive weight α∈{0.0,0.1,0.3,0.5,0.7} to evaluate model stability.

4.3 Experimental Results

To assess statistical significance, we conducted paired t-tests across five independent random seeds. Cognitive-CDR consistently outperformed the strongest baseline methods, and the observed improvements in HR@10 and NDCG@10 were statistically significant (p<0.05). Table 2 summarizes performance across of the S1 experiment scenarios. Across both Books and Movies & TV, the ranking quality improves monotonically as we move from heuristic popularity (PopRec) to memory-based CF (ItemCF), then to latent-factor methods (BPR-MF, NeuMF), and finally to graph-based recommenders (NGCF, LightGCN). This pattern is consistent with the intuition that models capturing higher-order connectivity and more expressive user-item interactions concentrate probability mass at the top of the list, thereby improving HR@10 and NDCG@10. Among the baselines, LightGCN performs best in both domains, reaffirming the effectiveness of lightweight propagation on the user-item graph without heavy transformations.

On Books, Cognitive-CDR improves over LightGCN by +0.027 HR@10 (+4.9%), +0.021 NDCG@10 (+5.9%), and +0.017 MRR (+6.4%), with similar gains in MAP@10 and P@10. On Movies & TV, improvements remain consistent: +0.027 HR@10 (+4.8%), +0.020 NDCG@10 (+5.5%), and +0.019 MRR (+7.0%). These gains indicate that Cognitive-CDR promotes the true item to higher positions within the top-10, improving early-rank accuracy. By incorporating cross-domain cognitive similarity, the model regularizes user representations so that users with similar preferences across domains are placed closer together in the embedding space, benefiting sparse users. Compared with cross-domain baselines such as CMF and CoNet, Cognitive-CDR better balances graph structure and cross-domain alignment, leading to stronger early-rank performance.

Absolute metrics are slightly higher on Movies & TV than on Books for most methods (e.g., HR@10: PopRec 0.448 vs. 0.426; LightGCN 0.564 vs. 0.552; Cognitive-CDR 0.591 vs. 0.579), likely due to differences in interaction density or item popularity. Notably, the largest relative gains appear in MAP@10 and MRR, suggesting that improvements mainly occur at very early ranking positions. Regarding the other experiment scenarios, the following is a detailed on the results:

•   Cross-domain Transfer Scenarios: Across both directions, Cognitive-CDR outperforms the strongest baseline (CoNet) on all metrics, indicating more effective knowledge transfer between domains (refer Table 3). From Books → Movies, gains over CoNet are consistent at early ranks: HR@10 increases from 0.559 to 0.596 (+0.037, ≈+6.6% rel.), NDCG@10 from 0.364 to 0.390 (+0.026, ≈+7.1%), and MRR from 0.270 to 0.292 (+0.022, ≈+8.2%). In the reverse direction Movies → Books, improvements are similarly strong: HR@10 0.548→0.582 (+0.034, ≈+6.2%), NDCG@10 0.354→0.380 (+0.026, ≈+7.3%), and MRR 0.264→0.286 (+0.022, ≈+8.3%). These consistent lifts, especially in MRR and NDCG, demonstrate that the proposed cognitive alignment facilitates better early rank placement of true items in both transfer directions. Overall, the symmetry of gains suggests the method learns a domain-agnostic user representation that generalizes reliably, rather than relying on idiosyncrasies of a particular source domain.

•   Cold-Start and Sparsity: Cold-start and sparsity represent two complementary stress tests for recommendation systems. Cold-start evaluates recommendations for users without target-domain interactions, while sparsity tests robustness when training data is reduced. In both settings, Cognitive-CDR consistently outperforms the baselines. As shown in Table 4, the cold-start scenario is particularly challenging because the model must infer preferences solely from the source domain. Cognitive-CDR achieves the best results across all metrics, improving over the strongest baseline CoNet from 0.469 to 0.508 in HR@10 (+8.3%), from 0.296 to 0.322 in NDCG@10 (+8.8%), and from 0.221 to 0.241 in MRR (+9.0%).

The sparsity experiment further confirms the robustness of the proposed method. As interactions are removed, all models degrade, but Cognitive-CDR declines more slowly and maintains a larger margin. Compared with CoNet, HR@10 improves by +0.027, +0.029, and +0.034 at 20%, 40%, and 60% drop rates, respectively. This widening gap suggests that cognitive coupling helps preserve meaningful user representations when interaction data is scarce.

•   Ablation and Sensitivity: Table 5 shows that removing any cross-domain component reduces performance compared with the full model (NDCG@10 =0.386).

Removing the cognitive similarity term decreases NDCG@10 to 0.363 (−5.96%), indicating that cognitive user similarity improves the embedding structure. Removing domain alignment further drops performance to 0.355 (−8.03%), highlighting the importance of aligning representations across domains. The largest decline occurs without the cross-domain adapter (NDCG@10 =0.347, −10.10%), suggesting that modeling domain-specific variations while sharing global structure is critical. Overall, the components complement each other and jointly improve recommendation quality. Sensitivity analysis shows that varying the cognitive weight α produces a shallow inverted-U trend in HR@10, peaking at α=0.3 with 0.579. Compared with α=0.0 (HR@10 =0.551), performance improves for α=0.1, 0.3, and 0.5, but slightly declines at α=0.7, suggesting over-regularization when cognitive coupling becomes too strong. The stable range α∈[0.3,0.5] indicates that the model is robust to moderate changes in this parameter.

4.4 Visualization

To create the visualization in Fig. 1a, we first constructed a similarity matrix in which each cell represents the cognitive similarity between two users. These scores are calculated using cosine similarity in Eq. (18). The output is a 2D scatter plot where each point represents a user, and the distance between points corresponds to the cognitive similarity between those users. Users with similar preferences will appear closer together on the plot. This visualization helps visualize complex relationships in high-dimensional data by projecting them into a lower-dimensional space, making it easier to identify clusters or patterns in user behavior across different domains. Another visualization is shown in Fig. 1b, which illustrates pairwise user similarities. Each cell in the matrix represents the similarity between two users, with higher values (closer to 1) indicating greater similarity. The color intensity reflects the magnitude of the similarity: lighter colors indicate higher similarity, while darker colors indicate lower similarity. Since the matrix is symmetric, the diagonal will always have a similarity of 1 (each user is most similar to themselves). Bright squares or rectangles indicate the clusters within the matrix where users share a higher similarity. This suggests that these users share similar cognitive preferences, possibly stemming from overlapping interests across both domains. The darker regions represent users with low similarity. These may indicate distinct user preferences across domains or little overlap in their cognitive patterns.

Figure 1: Visualization of user cognitive similarity. (a) shows the relative positions of users based on cognitive similarity, with points closer together indicating more similar users. The axes represent the two dimensions obtained from the MDS projection. (b) The heatmap illustrates pairwise user similarities, where higher values indicate stronger similarity.

4.5 Discussion on Complexity Analysis and Limitations

Let N denote the number of users and F the dimensionality of the cognitive feature space. Computing cognitive similarity requires O(NF) operations per target user. Selecting the top-K neighbors can be done in O(Nlog⁡K) using a priority queue. Recommendation generation incurs O(KM) complexity, where M is the number of candidate items. Since user features and similarity matrices can be computed offline, the online recommendation stage remains computationally efficient. Furthermore, similarity computations are naturally parallelizable and can be accelerated using distributed processing frameworks for large-scale deployment.

Despite the encouraging results, several limitations remain. First, the current framework relies on handcrafted cognitive features derived from ratings, genres, and review keywords, which may not capture all aspects of user preferences. Second, the evaluation focuses on two Amazon domains (Books and Movies), and additional experiments on more heterogeneous domains would further validate generalizability. Third, cosine-based similarity assumes a linear relationship between user profiles and may not fully capture complex nonlinear preference patterns. Future work will investigate transformer-based user representations, graph-enhanced cognitive modeling, and broader multi-domain evaluations.

5  Conclusions

This study explored cross-domain collaborative filtering by leveraging user cognitive similarity to improve recommendation accuracy across domains. Using Amazon Product Reviews datasets from Books and Movies, the results show that cross-domain approaches outperform domain-specific methods, with the proposed Cognitive-CDR model achieving the best performance. By incorporating external knowledge and adversarial learning, the model improved recommendation relevance, achieving lower RMSE and higher Precision@10. In addition, MDS was used to visualize cognitive similarity between users. By projecting the high-dimensional similarity matrix into a 2D space, MDS revealed clusters of users with similar preferences across domains. These visualizations provide intuitive insights into user relationships and demonstrate how cognitively similar users from different domains can be grouped together, facilitating more effective cross-domain knowledge transfer and improved personalization.

Acknowledgement: Thanks to AIT Laboratory at Faculty of Artificial Intelligence, FPT University.

Funding Statement: This research is funded by FPT University under grant number DHFPT2026-08.

Author Contributions: The authors confirm contribution to the paper as follows: Luong Vuong Nguyen: Conceptualization, Methodology, Software; Luong Vuong Nguyen, Hoang Tran and Thuy-Trang Pham: Data curation, Writing-Original draft preparation; Hoang Tran and Thuy-Trang Pham: Visualization, Investigation; Luong Vuong Nguyen: Supervision, Software, Validation; Luong Vuong Nguyen, Hoang Tran and Thuy-Trang Pham: Writing-Reviewing and Editing. All authors reviewed and approved the final version of the manuscript.

Availability of Data and Materials: The data that support the findings of this study are available from the Corresponding Author, Luong Vuong Nguyen, upon reasonable request.

Ethics Approval: Not applicable.

Conflicts of Interest: The authors declare no conflicts of interest.

References

1. Nguyen LV, Hong MS, Jung JJ, Sohn BS. Cognitive similarity-based collaborative filtering recommendation system. Appl Sci. 2020;10(12):4183. doi:10.3390/app10124183. [Google Scholar] [CrossRef]

2. Nguyen LV, Vo ND, Jung JJ. DaGzang: a synthetic data generator for cross-domain recommendation services. PeerJ Comput Sci. 2023;9(3):e1360. doi:10.7717/peerj-cs.1360. [Google Scholar] [PubMed] [CrossRef]

3. Nguyen LV, Jung JJ. Sabre: cross-domain crowdsourcing platform for recommendation services. In: International Symposium on Intelligent and Distributed Computing. Berlin/Heidelberg, Germany: Springer; 2022. p. 213–23. [Google Scholar]

4. Cantador I, Fernández-Tobías I, Berkovsky S, Cremonesi P. Cross-domain recommender systems. In: Recommender systems handbook. New York, NY, USA: Springer; 2015. p. 919–59. [Google Scholar]

5. Zhao C, Zhao H, He M, Zhang J, Fan J. Cross-domain recommendation via user interest alignment. In: Proceedings of the ACM Web Conference 2023; 2023 Apr 30–May 4; Austin, TX, USA. p. 887–96. [Google Scholar]

6. Zhu F, Wang Y, Chen C, Zhou J, Li L, Liu G. Cross-domain recommendation: challenges, progress, and prospects. arXiv:2103.01696. 2021. [Google Scholar]

7. Zang T, Zhu Y, Liu H, Zhang R, Yu J. A survey on cross-domain recommendation: taxonomies, methods, and future directions. ACM Trans Inf Syst. 2022;41(2):1–39. doi:10.1145/3548455. [Google Scholar] [CrossRef]

8. Fernández-Tobías I, Cantador I, Kaminskas M, Ricci F. Cross-domain recommender systems: a survey of the state of the art. In: Spanish Conference on Information Retrieval. Valencia, Spain: ACM; 2012. [Google Scholar]

9. Khan MM, Ibrahim R, Ghani I. Cross domain recommender systems: a systematic literature review. ACM Comput Surv. 2017;50(3):1–34. doi:10.1145/3073565. [Google Scholar] [CrossRef]

10. Zhang H, Cheng M, Liu Q, Jiang J, Wang X, Zhang R, et al. A comprehensive survey on cross-domain recommendation: taxonomy, progress, and prospects. arXiv:2503.14110. 2025. [Google Scholar]

11. Zhao Y, Ju J, Gong J, Zhao J, Chen M, Chen L, et al. Cross-domain recommendation via adaptive bi-directional transfer graph neural networks. Knowl Inf Syst. 2025;67(1):579–602. doi:10.1007/s10115-024-02246-9. [Google Scholar] [CrossRef]

12. Wang Q, Zhao Y, Zhang Y, Zhang Y, Deng S, Yang Y. Federated contrastive learning for cross-domain recommendation. IEEE Trans Serv Comput. 2025;18(2):812–27. doi:10.1109/tsc.2025.3528325. [Google Scholar] [CrossRef]

13. Liu Q, Zhao X, Wang Y, Zhang Z, Zhong H, Chen C, et al. Bridge the domains: large language models enhanced cross-domain sequential recommendation. In: Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval; 2025 Jul 13–18; Padua, Italy. p. 1582–92. [Google Scholar]

14. Loni B, Shi Y, Larson M, Hanjalic A. Cross-domain collaborative filtering with factorization machines. In: Advances in Information Retrieval: 36th European Conference on IR Research, ECIR 2014; 2014 Apr 13–16; Amsterdam, The Netherlands. Berlin/Heidelberg, Germany: Springer; 2014. p. 656–61. [Google Scholar]

15. Sahu AK, Dwivedi P. Knowledge transfer by domain-independent user latent factor for cross-domain recommender systems. Future Gener Comput Syst. 2020;108(4):320–33. doi:10.1016/j.future.2020.02.024. [Google Scholar] [CrossRef]

16. Gao S, Luo H, Chen D, Li S, Gallinari P, Guo J. Cross-domain recommendation via cluster-level latent factor model. In: Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2013; 2013 Sep 23–27; Prague, Czech Republic. Berlin/Heidelberg, Germany: Springer; 2013. p. 161–76. [Google Scholar]

17. Zhang L, Ge Y, Ma J, Ni J, Lu H. Knowledge-aware neural collective matrix factorization for cross-domain recommendation. arXiv:2206.13255. 2022. [Google Scholar]

18. Kanagawa H, Kobayashi H, Shimizu N, Tagami Y, Suzuki T. Cross-domain recommendation via deep domain adaptation. In: European Conference on Information Retrieval. Berlin/Heidelberg, Germany: Springer; 2019. p. 20–9. [Google Scholar]

19. Tiroshi A, Kuflik T. Domain ranking for cross domain collaborative filtering. In: User Modeling, Adaptation, and Personalization: 20th International Conference, UMAP 2012; 2012 Jul 16–20, Montreal, QC, Canada. Berlin/Heidelberg, Germany: Springer; 2012. p. 328–33. [Google Scholar]

20. Gao C, Chen X, Feng F, Zhao K, He X, Li Y, et al. Cross-domain recommendation without sharing user-relevant data. In: Proceedings of the World Wide Web Conference; 2019 May 13–17; San Francisco, CA, USA. p. 491–502. [Google Scholar]

21. Hao X, Liu Y, Xie R, Ge K, Tang L, Zhang X, et al. Adversarial feature translation for multi-domain recommendation. In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining; 2021 Aug 14–18; Singapore, Singapore. p. 2964–73. [Google Scholar]

22. Liu H, Guo L, Li P, Zhao P, Wu X. Collaborative filtering with a deep adversarial and attention network for cross-domain recommendation. Inf Sci. 2021;565(6):370–89. doi:10.1016/j.ins.2021.02.009. [Google Scholar] [CrossRef]

23. Liu M, Li J, Li G, Pan P. Cross domain recommendation via bi-directional transfer graph collaborative filtering networks. In: Proceedings of the 29th ACM International Conference on Information & Knowledge Management; 2020 Oct 19–23; Virtual. p. 885–94. [Google Scholar]

24. Liu J, Huang W, Li T, Ji S, Zhang J. Cross-domain knowledge graph chiasmal embedding for multi-domain item-item recommendation. IEEE Trans Knowl Data Eng. 2022;35(5):4621–33. doi:10.1109/tkde.2022.3151986. [Google Scholar] [CrossRef]

25. Zhang Q, Wu D, Lu J, Zhang G. Cross-domain recommendation with probabilistic knowledge transfer. In: Neural Information Processing: 25th International Conference, ICONIP 2018; 2018 Dec 13–16; Siem Reap, Cambodia. Berlin/Heidelberg, Germany: Springer; 2018. p. 208–19. [Google Scholar]

26. Man T, Shen H, Jin X, Cheng X. Cross-domain recommendation: an embedding and mapping approach. In: Proceedings of the 26th International Joint Conference on Artificial Intelligence; 2017 Aug 19–25; Melbourne, VIC, Australia. p. 2464–70. [Google Scholar]

27. Zhu Y, Tang Z, Liu Y, Zhuang F, Xie R, Zhang X, et al. Personalized transfer of user preferences for cross-domain recommendation. In: Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining; 2022 Feb 21–25; Tempe, AZ, USA. p. 1507–15. [Google Scholar]

28. Sahebi S, Walker T. Content-based cross-domain recommendations using segmented models. In: Proceedings of the CBRecSys 2014; 2014 Oct 6; Silicon Valley, CA, USA. p. 57–64. [Google Scholar]

29. Shi Y, Larson M, Hanjalic A. Tags as bridges between domains: improving recommendation with tag-induced cross-domain collaborative filtering. In: User Modeling, Adaption and Personalization: 19th International Conference, UMAP 2011; 2011 Jul 11–15; Girona, Spain. Berlin/Heidelberg, Germany: Springer; 2011. p. 305–16. [Google Scholar]

30. Xin X, Liu Z, Lin CY, Huang H, Wei X, Guo P. Cross-domain collaborative filtering with review text. In: Proceedings of the Twenty-Fourth International Joint Conference on Artificial Intelligence; 2015 Jul 25–31; Buenos Aires, Argentina. p. 1827–33. [Google Scholar]

31. Rendle S, Freudenthaler C, Gantner Z, Schmidt-Thieme L. BPR: Bayesian personalized ranking from implicit feedback. In: Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence; 2009 Jun 19–21; Montreal, QC, Canada. p. 452–61. [Google Scholar]

32. He X, Liao L, Zhang H, Nie L, Hu X, Chua TS. Neural collaborative filtering. In: Proceedings of the 26th International Conference on World Wide Web (WWW ’17); 2017 Apr 3–7; Perth, WA, Australia. p. 173–82. [Google Scholar]

33. Wang X, He X, Wang M, Feng F, Chua TS. Neural graph collaborative filtering. In: Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’19); 2019 Jul 21–25; Paris, France. p. 165–74. [Google Scholar]

34. He X, Deng K, Wang X, Li Y, Zhang Y, Wang M. LightGCN: simplifying and powering graph convolution network for recommendation. In: Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR’20); 2020 Jul 25–30; Virtual. p. 639–48. [Google Scholar]

35. Singh AP, Gordon GJ. Relational learning via collective matrix factorization. In: Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD’08); 2008 Aug 24–27; Las Vegas, NV, USA. p. 650–8. [Google Scholar]

36. Hu G, Zhang Y, Yang Q. CoNet: collaborative cross networks for cross-domain recommendation. In: Proceedings of the 27th ACM International Conference on Information and Knowledge Management (CIKM ’18); 2018 Oct 22–26; Turin, Italy. p. 667–76. [Google Scholar]