Citation is considered in scientometrics as the reference of one published document in another document. It also is a degree to evaluate the prominence and measure of research outputs because it presents how usually a specific work is referenced in alternative scholarly literature. A high citation count often indicates that a work has significantly impacted its field, influencing subsequent research and scholarship (Blümel & Schniedermann, 2020; Broadus, 1987; Moed, 2006; Khokhlov, 2020). So it can act as an indicator of research quality; works with high citation counts are often thought to be more valuable or credible than less cited peer-reviewed literature (Bornmann & Daniel, 2008; Butler & Visser, 2006). Citation counts can also provide to spot trends in research over time, indicating areas of growing interest, and shifts in the scientific landscape. Besides, citation counts play a key role in resource allocation decisions, as funding agencies and institutions also rely on these metrics to decide which research areas or researchers are worthy of support. Overall, citation counts play a vital role in academic communication, evaluation, and the advancement of knowledge in various fields (Abramo et al., 2023; Belikov & Belikov, 2015; Cao et al., 2016; Durieux & Gevenois, 2010; Gao et al., 2024; Groos & Pritchard, 1969; Sohrabi & Iraj, 2017; Yu et al., 2014).

On the other side, several famous bibliometrics measures, including Impact Factor (IF) and h-index, are also based on citations from publications and journals. In the early 1960s, the IF was introduced by the Institute of Scientific Information (ISI) (Garfield, 2006). The IF is determined by the number of citations it receives throughout the year. The IF is calculated by dividing the current number of citations by the number of articles published in the last two years. As a result, the performance of a journal can be directly estimated based on its IF (Amin & Mabe, 2003; Bornmann & Daniel, 2009; Braun et al., 2006; Durieux & Gevenois, 2010; Hirsch, 2010; Lundberg, 2006). It has been criticized for not taking into account the diversity of individual researchers among different fields of study and for being inaccurate in terms of its purpose. In addition, researchers were given an index called the h-index in order to measure their objective function. To determine an author's h-index, one must determine how many of his or her articles have been cited directly by other researchers at least a similar number of times (Braun et al., 2006; Fassin, 2020; Hirsch, 2010; Khurana & Sharma, 2022). In addition, Q1 to Q4 quartile rankings for journals are provided by two key sources: Clarivate Analytics and Elsevier. Clarivate’s Journal Citation Reports (JCR) ranks journals based on their Impact Factor (IF) through its Web of Science platform, assigning quartiles from Q1 (top 25%) to Q4 (bottom 25%). Elsevier’s Scopus platform uses the SCImago Journal Rank (SJR), which evaluates journals based on citation impact and assigns similar quartiles. Both systems are widely recognized and used for academic journal rankings globally, guiding researchers in assessing and selecting journals for publishing their work (Almas et al., 2021; González-Betancor & Dorta-González, 2017; Kosyakov & Pislyakov, 2024; Moussa, 2023; Okagbue et al., 2020; Okagbue et al., 2021; Teixeira da Silva, 2020; Torres-Salinas et al., 2022).

Gao et al., 2024 leveraged multi-layer academic networks to improve citation count predictions by fusing different relationship types between publications. The authors show how taking this multilayer perspective of academic relations can enhance precision. Their model offers the ability to capture complexity in citation dynamics. Nevertheless, the intricacy of a multi-layered network could impose difficulties in model interpretability, hindering researchers from interpreting the mechanisms that power predictions. Also, the model performance may depend on the quality of the network that it is trained on, which may differ greatly between different sectors.

The process of peer reviewing is widely accepted as a means of evaluating papers (Li et al., 2019). It is important for a reviewer to evaluate the originality, creativity, contribution, integrity, and readability of a paper. Since peer-reviewing data includes the assessment comments of related experts, it may also be possible to predict the future influence of a paper. The algorithm of a comprehensive review paper has made it possible to obtain peer-reviewed data for the Citation count prediction (CCP) task. It is apparent in their comments that they focus on issues that are not directly related to the paper's main contribution. Reviewers may include reminders regarding formatting issues in their reviews. Several people may be reviewing articles at the same time, resulting in differing opinions. Therefore, when determining the impact of a paper, the coverage and divergence of review comments would both be considered (Li et al., 2019).

While citation count prediction is a relatively well-explored area, many existing studies did not comprehensively consider the multitude of factors influencing citations, particularly in specialized fields like Computer Science and Electrical Engineering (Aksnes et al., 2019; Baas et al., 2020; Enduri et al., 2022; Furman & Teodoridis, 2020; He et al., n.d.; Zhang et al., 2025; Hutchins et al., 2016; Okagbue et al., 2020).

Citation patterns can vary across disciplines, so our work aimed at citation count predictions of academic papers for computer science and electrical engineering disciplines. Due to the fact that both Google Scholar and JCR are well-respected sources in the academic community and public availability of data, we suggested mainly gathering raw data from Google Scholar Profiles (GSPs) and public JCR annual reports. The GSP offers a wide range of citation information, such as Citation, h-index, and i10-index, while the JCR reports provide metrics like IF and journal rankings (Q1, Q2, Q3, Q4). Combining these two main data sources can enrich our feature set and help capture more nuanced aspects of citation behavior, making our model more accurate. Raw data like citation counts, h-index, or publication year are useful but limited in their predictive power. Citations are influenced by a variety of factors, and raw data may not fully capture non-linear relationships between these factors. Creating new features can improve the performance of the citation count prediction model by uncovering hidden patterns and relationships that raw data alone cannot reveal. As a result, our approach includes twenty three unique and novel features, offering a fundamental understanding of the factors that may impact citation count. The citation count was then estimated using a number of robust regression techniques, including SVR, Nu-SVR, Linear SVR, K-Nearest Neighbors (KNN), Decision Tree (DT), Bayesian Ridge, and SGD Regressor. We also assessed our method using several performance measures including the mean absolute percentage error (MAPE), Mean Square Error (MSE), Mean Absolute Error (MAE), and R-squared. The study also utilized PCA and t-SNE for dimensionality reduction and examined their impacts on CCP.

The rest of this paper is organized as follows: Section two discusses related works. In section three, our methodology is introduced. Simulation results are presented in section four, and section five is the conclusion.

Citation count prediction (CCP) has been approached from multiple perspectives, including network-based modeling, textual analysis, trend forecasting, and deep learning. While existing methods offer valuable insights, they often focus on isolated aspects of the problem, leaving room for a more holistic and generalizable approach (Bai et al., 2025; He et al., 2025; Zhang et al. 2025, Zhu et al., 2025).

Early work by (Pobiedina & Ichise, 2016) framed CCP as a link prediction problem, leveraging citation networks to model future citations. While this approach captures structural dependencies, it overlooks critical external factors such as author reputation and research competitiveness. Similarly, (Wang et al., 2023) introduced AGSTA-NET, a spatio-temporal fusion model that improves dynamic citation network analysis. However, its computational complexity and reliance on heterogeneous network data may limit scalability. These studies highlight the potential of network-based methods but also reveal their dependence on well-structured citation data, which may not always be available Recent work has explored the role of textual features in CCP. For example, (Li et al., 2019) incorporated peer review text into a neural network model, demonstrating that qualitative feedback can enhance prediction accuracy. However, their reliance on peer review data—which varies widely across disciplines—poses a generalizability challenge. Similarly, (Sohrabi & Iraj, 2017) focused on keyword frequency, showing that strategic keyword use can improve visibility. Yet, their model neglects broader contextual factors, such as journal prestige or research impact. Baba et al. (2019) extended textual analysis to paper abstracts but did not account for external citation influences. These studies suggest that while textual features are valuable, they must be integrated with other predictive factors for robust performance.

Historical citation trends have also been used for prediction. (Li et al., 2015) demonstrated that temporal patterns improve out-of-time forecasts, but their model struggles with disruptive research that defies conventional citation trends. Meanwhile, (Abrishami & Aliakbary, 2019) applied deep learning, achieving superior accuracy over traditional methods. However, their approach requires large datasets and risks overfitting, limiting applicability in low-data scenarios.

Existing CCP methods face key limitations: network models ignore external factors; textual approaches lackgeneralizability; trend-based methods fail with disruptive research; and deep learning requires excessive data. Most critically, no unified framework integrates multi-modal data while ensuring efficiency and interpretability (Nguyen et al., 2025; Zafar et al., 2024).

In this paper, we aim to predict the citation count of an article. An article's citation count is a suitable determinant for the impact assessment. For this purpose, we have created and developed four datasets that are called Author Information Database (AI DB), Journal Information Database (JI DB), Paper Information Database (PI DB), and finally Author & Paper & Journal Information Database (APJ DB). Figure 1 depicts the proposed algorithm. The initial preprocessing step normalizes the data. Then, we can predict the citation count of a published paper using several robust regression algorithms such as K-Nearest Neighbors (KNN), Decision Trees (DT), and Support Vector Regression (SVR), and Bayesian Regression methods, along with some of the most important performance metrics in regression problems such as Mean Absolute Error (MAE), Mean Squared Error (MSE), Mean Absolute Percentage Error (MAPE), and the Coefficient of Determination (R²) for evaluation of the achieved results. To improve interpretability, we also assessed the influence of dimensionality reduction techniques (e.g., t-SNE and PCA) on the results and discussed their comparative effects.

In this paper, seven distinct regression methods, including SVR, Nu-SVR, Linear SVR, kNN, Decision Tree Regression, Bayesian Ridge, and SGD Regression, were trained to find the best Citation count prediction (CCP) solution. We also used several error methods for estimating the errors, such as Mean Absolute Error (MAE), Mean Squared Error (MSE), Mean Absolute Percentage Error (MAPE), and the Coefficient of Determination (R²), to measure the algorithm performance for the Author information database (AI DB), journal information database (JI DB), paper information database (PI DB), and finally author & paper & journal information database (APJ DB), separately. We also tested APJ DB and AI DB in two scenarios, with dimension reduction and without it. All of our simulations have been implemented using Python software and were run on a Lenovo device with a 2.5 GHz Intel Core i7 processor and 16 GB DDR4 RAM. The results obtained from the different methods are presented from Table 5 to Table 17.

In our experiment, Grid search that is a hyperparameter tuning technique was used to systematically work through multiple combinations of parameter values to determine which combination yields the best model performance.

In Table 5, max_depth=24.42 suggests a deep tree for DecisionTree Regressor, which may be harder to interpret than a shallower one. So, we could constrain max_depth to a smaller value about 5 during grid search. Furthermore, KNeighborsRegressor with n_neighbors=1 may highly prone to overfitting and lacks interpretability. As a result, a higher value at 5 was selected. LinearSVR and SGDRegressor use regularization (C, alpha). A low C=0.01672 (LinearSVR) suggests strong L2 regularization, which can simplify the model by shrinking coefficients. In addition, SGDRegressor uses penalty='elasticnet', which can promote sparsity, making feature importance clearer. For Kernel Choices in SVM Models Both SVR and Nu-SVR use the 'rbf' kernel, which is inherently less interpretable than a linear kernel. As a result, for considering interpretability, we could restrict the search to kernel='linear'. For Bayesian Ridge’s Complexity, the high values for lambda_1=1e11 and alpha_1=3.59e6 suggest strong prior assumptions, but the model remains interpretable since it’s a linear regression variant.

Table 6 presents the performance of various regression models on the APJ dataset without any dimensionality reduction. The best-performing models are SVR and Decision Tree Regression, both achieving an MAE of 0.16, MSE of 0.50, and R² of 0.48, indicating strong predictive accuracy and stability. Nu-SVR also performs well with the lowest MAPE (0.36), suggesting better relative error control. In contrast, K-Neighbors Regression and Bayesian Ridge show slightly higher errors, while Linear SVR and SGDRegression exhibit the weakest R² scores (0.35 and 0.43, respectively). Overall, non-linear models (SVR, Decision Tree) outperform linear ones, likely due to their ability to capture complex relationships in the data.

According Table 7, applying PCA-based dimension reduction generally worsened model performance, except for K-Neighbors Regression, which saw slight improvements (e.g., MSE dropping from 0.55 to 0.53). The degradation is most severe for Decision Tree Regression, where MSE nearly doubled (0.50 → 0.92) and R² collapsed to 0.05, because PCA’s linear transformations disrupt the tree-based feature splits. Similarly, SVR and Nu-SVR suffered, possibly due to lost non-linear feature interactions. The improvement in K-Neighbors suggests PCA may have removed noise, aiding its distance-based computations. The overall decline implies that PCA either discarded informative features or failed to preserve structures critical for regression, highlighting that blind dimensionality reduction can harm performance unless the model benefits from noise removal (like kNN).

Comparing both tables show that PCA does not universally improve model performance and even degrades results for certain algorithms. Given this outcome, this study further evaluates t-SNE (t-Distributed Stochastic Neighbor Embedding) as an alternative and compares results with PCA. The Table 8 presents MAE and MSE values for six regression models, with performance measured after applying PCA and t-SNE for dimensionality reduction. While t-SNE excels at visualization, PCA remains almost superior for regression tasks due to its stability, interpretability, and preservation of globally meaningful features. The results shows that t-SNE cannot outperform PCA in regression settings when it was applied to APJ DB.

Table 9 presents the regression results for the AI DB without dimension reduction. The Decision Tree model performs best in terms of MSE (0.86) and R² (0.12), while SVR follows closely with an MAE of 0.24 and R² of 0.1. Nu-SVR and Linear SVR achieve the lowest MAE (0.22), though Linear SVR has a slightly higher MSE (0.95). K-Neighbors Regression performs the worst, with an MSE of 1.08 and a negative R² (-0.11), indicating poor fit. Bayesian Ridge and SGDRegression show moderate performance, with MAPE values ranging from 0.65 to 0.96.

In Table 10, where PCA-based dimension reduction is applied, results vary. Nu-SVR improves slightly, achieving the lowest MAE (0.21) and MAPE (0.55), while its MSE remains stable (0.89). However, Linear SVR deteriorates significantly, with MSE increasing to 0.97 and R² dropping to -0.004. Decision Tree regression, which performed well without PCA, now shows a higher MSE (1.02) and a negative R² (-0.04). K-Neighbors Regression remains poor, with nearly identical metrics as in Table 7. Bayesian Ridge is largely unaffected by PCA, maintaining similar MAE, MSE, and R² values.

Tables 11 to 15 make a brief comparison of different proposed methods based on results of four performance metrics, including MAE, MSE, MAPE, and R2. In these tables, DR stands for dimension reduction. In our implementations, APJ DB has been evaluated with DR and without DR; AI DB follows the same rule as APJ DB with and without DR but both JI DB and PI BI have been used without DR.

Moreover, in Tables 16 and 17, we summarized all information in tables to find out which algorithm leads us to the best result regarding the methods. It can be clearly seen that the dominant algorithm is Nu-SVR, which outperformed all the other ones, and in terms of methods, features included in PI DB led us to the least error rates and the most R2.

Furthermore, Figure 5 and Figure 6 draw a wide comparison among all proposed algorithms and our four introduced datasets in terms of MAE and R2, respectively.

Ultimately, we called our best model found on PI DB-without DR & Nu-SVR as PI-CCP. According to Tables 18, 19, and 20, the simulation results achieved based on the proposed PI-CCP modle are compared with the available studies presented by (Abrishami et al., 2019; Gao et al., 2024; Li et al., 2015; Li et al., 2019). Table 18 shows that PI-CCP performs exceptionally well at minimizing errors, as evidenced by the remarkably low MAE. It is noteworthy that it performs better than esteemed models like (Li et al., 2019), demonstrating its effectiveness in generating precise forecasts. The significant difference with (Gao et al., 2024) emphasizes PI-CCP's superiority even more.

The R2 values demonstrate how well PI-CCP captures the variation in citation counts. surpassing both CCP and T-CCP (Li et al., 2015), and getting near NNCP (Abrishami & Aliakbary, 2019), PI-CCP demonstrates its validity as a predictor, underscoring its usefulness in figuring out the influence of research papers (see Table 19). Besides, Table 20 shows that PI-CCP maintains remarkable precision in citation count predictions, while NNCP (Abrishami & Aliakbary, 2019) displays a reduced MSE. This ensures a balance between accuracy and reliability. As a result, PI-CCP is constantly positioned as a state-of-the-art and highly accurate model for citation count prediction by the thorough examination across MAE, R2, and MSE.

Citation count serves as crucial metrics in evaluating the impact of scientific articles and researchers, playing a pivotal role in scholarly and academic endeavors. The purpose of our research was to implement a high-accuracy citation count prediction (CCP) model based on easy access and available public data. As a first step, we created four datasets (AI DB, JI DB, and PI DB) containing twenty three proposed attributes. The data were collected from about 2000 GSPs in the fields of computer science and electrical engineering. The obtained results of the proposed model, so called PI-CCP, allows us to conclude that the features suggested in the Paper Information Dataset (PI DB) including and OAA are the most crucial variables in predicting the number of citations for scientific works. Besides, the most effective regression technique for forecasting citation count was determined to be the Nu-SVR algorithm. Achieving records such as 0.0930 for MAE and 0.8366 for R2 confirm that among all discussed algorithms, Nu-SVR is capable of managing the intricate, non-linear correlations between the input features and citation counts. Additionally, we examined APJ DB and AI DB in two different configurations: one with PCA/t-SNE-based dimension reduction and the other without them. The results showed that neither PCA nor t-SNE can consistently produce a lower error. Comparative analyses against existing models in the literature affirm the significant advancements achieved by our proposed algorithm, notably outperforming others. Our research not only presents the robust PI-CCP model but also contributes methodologically by offering insights into the selection of parameters, dataset creation, and algorithmic choices. The identification of crucial variables and the superior performance of Nu-SVR underscore the novelty and significance of our work in the domain of citation count prediction.

Even though we introduced twenty three novel features, there may still be other potentially relevant variables that could not be missed. Factors such as historical citation trends (time series), collaboration networks, social media presence, the impact of conference versus journal publications, institutional and funding factors, etc. may also play significant roles in citation counts but were not included in our analysis. Besides, our study focuses specifically on computer science and electrical engineering. The findings may not be directly applicable to other fields, as citation behaviors can vary significantly across disciplines. Future research could explore the applicability of our model in different academic domains.

Disclosure statement: The authors report there are no competing interests to declare.

Funding statement: There is no funding resource for this study.

Conflict of interest: There is no conflict of interest to declare.

Authors contributing to Information Research agree to publish their articles under a Creative Commons CC BY-NC 4.0 license, which gives third parties the right to copy and redistribute the material in any medium or format. It also gives third parties the right to remix, transform and build upon the material for any purpose, except commercial, on the condition that clear acknowledgment is given to the author(s) of the work, that a link to the license is provided and that it is made clear if changes have been made to the work. This must be done in a reasonable manner, and must not imply that the licensor endorses the use of the work by third parties. The author(s) retain copyright to the work. You can also read more at: https://publicera.kb.se/ir/openaccess