scholarlyimpact

Abstract

Identifying highly-cited scholarly literature at an early stage is a vital endeavor to the academic research community and to other stakeholders, such as technology companies and government bodies. Due to the sheer amount of research published and the growth of ever-changing interdisciplinary areas, researchers need an effective approach to identifying important scholarly studies if they are to read or even skim all the new studies published in their respective fields. The number of citations that a given research publication has accrued has been used to help researchers in this regard. However, citations take time to occur and longer to accumulate. In this article, we used Altmetrics to predict citations that a scholarly publication could receive. We built various classification and regression models and evaluated their performance. We found that tree-based models performed best in classification. We found that Mendeley readership, publication age, post length, maximum followers, and academic status were the most important factors in predicting citations.

Data

The dataset used for the experiments comprises of social media and scholarly indicators for scientific articles. The size of the dataset is 130,745 and for all the experiments 70 percent was used for training and 30 percent for test. Furthermore, for all the neural network models 20 percent of the training data was used as validation. There are three target variables in the dataset respectively for the three experiments. These are :

target_exp_1: Binary label saying if citations exist or not.
target_exp_2: Binary label saying if existing citations are more than median number of citations or not.
target_exp_3: Discrete values for log(1 + citation).

Methodology

The project comprises of three experiments. The first two experiments are classification problems while the third being a regression problem. A combination of approaches were used for solving the classification and regression problems. Neural networks, supervised learning algorithms and support vector machines were used for training the models on the data. The features used for all of the experiments were the same. While training the supervised learning models three algorithms had Randomized search and Grid search to obtain best hyper parameters for training the models. All of the supervised learning algorithms used 10 fold cross validation approach for training the models. The neural network models were implemented using TensorFlow. The supervised learning models and the support vector machine algorithms were implemented using scikit-learn.

Results

1. Classification experiment for predicting if citations exist or not

1.a. Optimum parameters for the Neural network

Model	Optimum hyper parameters
Epochs	10
Batch size	64
Loss Function	binary cross-entropy
Hidden Layers	1 layer with 512 neurons
Optimization function	RMS with 0.001 learning rate
Activation function(s)	SeLU for the hidden layer, Softmax for the o/p layer

1.b Accuracy, Precision, Recall and F-1 for the neural network

Metric	Value
Training Loss	2.07495
Training Accuracy	0.8646
Validation Loss	2.0891
Validation Accuracy	0.86368
Test Accuracy	0.8655
Precision	0.866
Recall	1.0
F-1	0.9279

1.c Accuracy, Precision, Recall and F-1 for the supervised learning algorithms

Model	Train Accuracy	Test Accuracy	Precision	Recall	F-1
Random Forest	0.865	0.862	0.863	1.0	0.927
Decision Tree	0.865	0.863	0.863	1.0	0.927
Gradient Boosting	0.865	0.863	0.863	1.0	0.927
AdaBoost	0.87	0.866	0.87	0.993	0.928
BernouliNB	0.84	0.836	0.876	0.943	0.908
KNN	0.85	0.851	0.883	0.953	0.917

1.d Optimum tuning parameters for the tree based and ensemble algorithms

Model	Optimum hyper parameters
Random Forest	n_estimators: 2, min_samples_split: 0.9, min_samples_leaf: 0.3, features: 18, max_depth: 9, criterion: gini-index
Decision Tree	min_samples_split: 0.5, min_samples_leaf: 0.3, max_features: 10, max_depth: 32, criterion: gini-index
Gradient Boosting	n_estimators: 200, min_samples_split: 0.6, min_samples_leaf: 0.1, max_features: 9, max_depth: 4, learning rate: 0.001

1.e Optimum tuning parameters for the C-support vector machine algorithm

Parameter	Value
Kernel	Sigmoid
Degree of the kernel	3
Tolerance	0.001
Gamma	0.045

1.f Accuracy, Precision, Recall and F-1 for the C-support vector machine algorithm

Metric	Value
Training Accuracy	0.86 (+/- 0.00)
Validation Accuracy	0.86 (+/- 0.01)
Test Accuracy	0.861
Precision	0.864
Recall	0.997
F-1	0.925

2. Classification experiment for predicting if citations are more than median or not

2.a Optimum parameters for the Neural network

Model	Optimum hyper parameters
Epochs	10
Batch size	64
Loss Function	binary cross-entropy
Hidden Layers	3 layers 64, 128, 64 neurons for respective layers
Optimization function	RMS with 0.001 learning rate
Activation function(s)	SeLU for the second hidden layer, Sigmoid for remaining layers

2.b Accuracy, Precision, Recall and F-1 for the neural network

Metric	Value
Training Loss	0.4710
Training Accuracy	0.78006
Validation Loss	0.4726
Validation Accuracy	0.7797
Test Accuracy	0.7794
Precision	0.81918
Recall	0.69692
F-1	0.75312

2.c Accuracy, Precision, Recall and F-1 for the supervised learning algorithms

Model	Train Accuracy	Test Accuracy	Precision	Recall	F-1
Random Forest	0.778	0.784	0.799	0.737	0.767
Decision Tree	0.773	0.768	0.725	0.834	0.776
Gradient Boosting	0.802	0.80	0.810	0.767	0.788
AdaBoost	0.80	0.797	0.806	0.760	0.782
BernouliNB	0.67	0.674	0.742	0.495	0.594
KNN	0.75	0.752	0.777	0.680	0.725

2.d Optimum tuning parameters for the tree based and ensemble algorithms

Model	Optimum hyper parameters
Random Forest	n_estimators: 100, min_samples_split: 0.1, min_samples_leaf: 0.1, features: 3, max_depth: 30, criterion: entropy
Decision Tree	min_samples_split: 0.4, min_samples_leaf: 0.1, max_features: 15, max_depth: 32, criterion: entropy
Gradient Boosting	n_estimators: 200, min_samples_split: 0.1, min_samples_leaf: 0.1, max_features: 12, max_depth: 20, learning rate: 0.005

2.e Optimum tuning parameters for the C-support vector machine algorithm

Parameter	Value
Kernel	Sigmoid
Degree of the kernel	3
Tolerance	0.001
Gamma	0.045

2.f Accuracy, Precision, Recall and F-1 for the C-support vector machine algorithm

Metric	Value
Training Accuracy	0.52 (+/- 0.00)
Validation Accuracy	0.52 (+/- 0.01)
Test Accuracy	0.519
Precision	0.478
Recall	0.007
F-1	0.014

3. Regression experiment for predicting log(1+citations)

3.a Optimum parameters for the Neural network

Model	Optimum hyper parameters
Epochs	500
Batch size	128
Loss Function	mean squared error
Hidden Layers	7 layers 32, 64, 64, 128, 64, 64, 32 neurons for respective layers
Optimization function	RMS with 0.001 learning rate
Activation function(s)	ReLU for all layers, linear activation for o/p layer

3.b MSE, MAE and R-squared for the neural network

Metric	Value
Training MSE	1.24965
Training MAE	0.84157
Test MSE	1.29756
Test MAE	0.85583
Test MSE	0.52284

3.c MSE and R-squared for the supervised learning algorithms

Model	Train MSE	Test MSE	R-squared
Random Forest	0.26	1.32	0.512
Decision Tree	1.647	1.663	0.389
Linear	1.75	1.758	0.354

3.d Optimum tuning parameters for the tree based and ensemble algorithms

Model	Optimum hyper parameters
Random Forest	n_estimators: 16, min_samples_split: 0.4, min_samples_leaf: 0.2, features: 16, max_depth: 24, criterion: mse
Decision Tree	min_samples_split: 0.4, min_samples_leaf: 0.1, max_features: 13, max_depth: 32, criterion: friedman mse

License

The MIT License

Author(s)

Akhil Pandey, Hamed Alhoori

akhilpandey95 / scholarlyimpact