Introduction
Search method
Related work
Gradient Boosted Decision Trees
CatBoost Gradient Boosted Trees Implementation
LightGBM Support for Categorical Variables
CatBoost applications by field
Tables of works studied
Title | CatBoost: unbiased boosting with categorical features |
Description | Paper introducing CatBoost algorithm |
Performance metric | logloss, zero-one loss |
Winner | CatBoost |
Reference | [2] |
Title | Benchmarking and optimization of gradient boosting decision tree algorithms |
Description | Compare CatBoost, LightGBM, and XGBoost run on GPU’s, using four benchmark tasks |
Performance metric | AUC ROC and Normalized discounted cumulative gain (\(\text {NDCG}\)) |
Winner | CatBoost wins AUC for Epsilon DataSet, LightGBM wins AUC for the Higgs dataset, XGBoost wins (NDCG) for Microsoft and Yahoo Datasets |
Reference | [8] |
Title | A Semi-Supervised Tri-CatBoost method for driving style recognition |
Description | Combine labeled and unlabeled data, use CatBoost as a base classifier to identify driving style |
Performance metric | N/A CatBoost used for semi-supervised learning not compared to other classifiers |
Winner | N/A |
Reference | [36] |
Title | Reconstructing commuters network using machine learning and urban indicators. |
Description | Construct graph on human movement between cities, extract features, apply CatBoost among other algorithms to reconstruct graph |
Performance metric | Accuracy |
Winner | CatBoost wins but training time is long compared to XGBoost, so authors use XGBoost for remainder of study |
Reference | [7] |
Title | Comparison between XGBoost, LightGBM and CatBoost using a home credit dataset |
Description | Evaluate of XGBoost, LightGBM, and CatBoost performance in predicting loan default |
Performance metric | AUC, running time |
Winner | LightGBM |
Reference | [19] |
Title | Short term electricity spot price forecasting using CatBoost and bidirectional long short term memory neural network |
Description | CatBoost for feature selection for time-series data |
Performance metric | Mean absolute percentage error |
Winner | CatBoost not a competitor, used for feature selection |
Reference | [21] |
Title | Research on personal credit scoring model based on multi-source data |
Description | Use “Stacking&Blending” with CatBoost, Logistic Regression, and Random Forest to calculate credit score in a regression technique |
Performance metric | Model is ensemble of no direct comparison between algorithms; performance measured in AUC |
Winner | N/A |
Reference | [22] |
Title | Predicting loan default in peer-to-peer lending using narrative data. |
Description | Evaluate CatBoost against other classifiers on the task of predicting loan default using Lending Club data |
Performance metric | Accuracy, AUC, H measure, type I error rate, type II error rate |
Winner | CatBoost |
Reference | [20] |
Title | KiDS-SQuaD II. Machine learning selection of bright extragalactic objects to search for new gravitationally lensed quasars |
Description | Use CatBoost to classify astronomical data |
Performance metric | AUC |
Winner | CatBoost |
Reference | [18] |
Title | Attack detection in enterprise networks by machine learning methods |
Description | Compare CatBoost, LightGBM, SVM, and logistic regression in multi-class and binary classification task of identifying computer network attacks. |
Performance metric | AUC, CV balanced accuracy, balanced accuracy, F1, precision, recall |
Winner | CatBoost |
Reference | [37] |
Title | Short-term weather forecast based on wavelet denoising and catboost |
Description | Use CatBoost to predict weather-related observations, and compare to other machine learning algorithms doing the same task |
Performance metric | unique method, based on root mean squared error |
Winner | CatBoost |
Reference | [51] |
Title | Evaluation of CatBoost method for prediction of reference evapotranspiration in humid regions |
Description | compare CatBoost, \(\text {SVM}\), and \(\text {RF}\) ability to predict amount of water lost through evaporation and transpiration |
Performance metric | MAPE, RSME, R2 |
Winner | Results do not indicate clear overall-winner |
Reference | [33] |
Title | The use of data mining methods for the prediction of dementia: evidence from the english longitudinal study of aging |
Description | Classify dementia on imbalanced data, maximum cardinality of feature is 50, compare CatBoost to other classifiers |
Performance metric | Normalized Gini coefficient |
Winner | Convolutional neural network |
Reference | [26] |
Title | A novel fracture prediction model using machine learning in a community-based cohort |
Description | Use CatBoost to predict fragility fracture |
Performance metric | AUC |
Winner | CatBoost |
Reference | [24] |
An efficient novel approach for iris recognition based on stylometric features and machine learning techniques | |
Description | Use CatBoost after doing feature extraction from image data converted to base-64 encoded data |
Performance metric | AUC |
Winner | multiboostAB |
Reference | [23] |
Title | CT-based machine learning model to predict the Fuhrman nuclear grade of clear cell renal cell carcinoma |
Description | Classify kidney cancer images into instances of high-grade or low-grade cancer, presents opportunities for research at Big Data scale |
Performance metric | Used only CatBoost |
Winner | N/A |
Reference | [28] |
Title | diseases spread prediction in tropical areas by machine learning methods ensembling and spatial analysis techniques |
Description | Use CatBoost to predict spread of dengue fever |
Performance metric | Mean absolute error |
Winner | LSTM and XGBoost ensemble |
Reference | [27] |
Title | Performance analysis of boosting classifiers in recognizing activities of daily living |
Description | Compare CatBoost with XGBoost in ability to identify human physical activity types from sensor data |
Performance metric | f-measure |
Winner | Friedman stochastic gradient boosting, ada-decision trees |
Reference | [25] |
Title | Predicting online shopping behavior from clickstream data using deep learning |
Description | CatBoost is part of ensemble that is best clickstream predictor |
Performance metric | AUC |
Winner | GRU—CatBoost Ensemble |
Reference | [39] |
Title | Construction and analysis of molecular association network by combining behavior representation and node attributes. |
Description | Leverage graph representation of association network of biological entities to predict associations as input for classifier, compare CatBoost with other popular classifiers as association predictor |
Performance metric | Accuracy, sensitivity, specificity, precision, Matthew’s Correlation, Coefficient, AUC, |
Winner | CatBoost (except Sensitivity) |
Reference | [38] |
Title | Prediction model of aryl hydrocarbon receptor activation by a novel QSAR approach, deepSnap–deep learning |
Description | Compare CatBoost to other learners in image processing task related to study relationship between genes and liver function |
Performance metric | AUC, accuracy |
Winner | DeepSnap-DL (deep learning algorithm) |
Reference | [5] |
Title | Bridging the gap between energy consumption and distribution through non-technical loss detection |
Description | Use CatBoost for predicting non-technical loss in power distribution networks, authors report little in terms of quantitative results |
Performance metric | Performance metric not explicit |
Winner | Not clear, authors do not give exact numbers |
Reference | [29] |
Title | Performance Analysis of Different Types of Machine Learning Classifiers for Non-Technical Loss Detection |
Description | Compare CatBoost with 14 other classifiers |
Performance metric | Precision, recall, F-Measure |
Winner | CatBoost has highest precision and F-measure, \(\text {ANN}\) has 0.003 higher recall |
Reference | [52] |
Title | Energy theft detection using gradient boosting theft detector with feature engineering-based preprocessing |
Description | Technique for using CatBoost with highly imbalanced data |
Performance metric | True positive rate, false positive rate |
Winner | CatBoost, has lowest false positive rate, LightGBM wins true positive rate, CatBoost has longest total train and test time, LightGBM has shortest total train and test time |
Reference | [31] |
Title | Impact of feature selection on non-technical loss detection |
Description | Use incremental feature selection, compare performance of CatBoost, Decision Tree and K-Nearest Neighbors classifiers |
Performance metric | Precision, recall, F-Measure |
Winner | CatBoost, except for recall of models trained with 9 features, where K-NN wins |
Reference | [30] |
Astronomy
CLASS_STAR
, that takes a value from 0 to 1, that gauges how point-like a light source is. Khramtsov et al. write that they perform classification in a 37 dimensional features space, which implies they discard the 9 original color intensity features after they derive the 36 features from pairs of color intensity features. The total number of instances in the \(\text {KiDS}\) data that they classify is approximately 9.6 million. Khramtsov et al. report that their \(\text {KiDS}\) data is imbalanced, since their final classification finds 5,665,586 galaxies, 3,660,368 stars, and 145,653 quasars, and 122,306 instances of indeterminate class. For in-depth coverage of techniques for addressing class imbalance, please see [55]. Khramtsov et al. apply threshold values to the list of output probabilities of the CatBoost classifier to partition \(\text {KiDS}\) data into classes. The indeterminate instances are those where the probability that the instance is a quasar or a galaxy are approximately the same.CLASS_STAR
feature makes the data heterogeneous. In the next section, we focus on the subject of applications of CatBoost to Finance.Finance
iterations
, that specifies the maximum number of Decision Trees CatBoost will construct. A low value for this parameter can impact CatBoost’s performance relative to other GBDT implementations. Since Daoud does not document this value, it is difficult to conclude that LightGBM is the best performer. We recommend as a best practice that researchers document hyper-parameters used in GBDT implementations they compare.Softer dataset | |||
---|---|---|---|
Model | Accuracy | AUC | H-measure |
LR-softer | 0.7516 [0.7508, 0.7523] | 0.6151 [0.6139, 0.6163] | 0.0843 [0.0827, 0.0860] |
RT-softer | 0.6952 [0.6911, 0.6996] | 0.5444 [0.5391, 0.5493] | 0.0124 [0.0095, 0.0153] |
BNN-softer | 0.7496 [0.7480, 0.7516] | 0.6120 [0.6095, 0.6151] | 0.0801 [0.0766, 0.0843] |
RF-softer | 0.7436 [0.7415, 0.7456] | 0.6043 [0.6013, 0.6073] | 0.0695 [0.0659, 0.0733] |
GBDT-softer | 0.7504 [0.7488, 0.7520] | 0.6132 [0.6107, 0.6158] | 0.0818 [0.0784.0.0853] |
XGBoost-softer | 0.7511 [0.7496, 0.7526] | 0.6143 [0.6120, 0.6167] | 0.0833 [0.0801, 0.0866] |
CatBoost-softer | 0.7523 [0.7511, 0.7535] | 0.6162 [0.6144, 0.6180] | 0.0859 [0.0834, 0.0885] |
Model | Type I rate | Type II rate | |
---|---|---|---|
LR-softer | 0.1557 [0.1550, 0.1565] | 0.6142 [0.6123, 0.6160] | |
RT-softer | 0.2024 [0.1978, 0.2072] | 0.7087 [0.6994, 0.7198] | |
BNN-softer | 0.1569 [0.1557, 0.1580] | 0.6190 [0.6141, 0.6231] | |
RF-softer | 0.1617 [0.1599, 0.1639] | 0.6298 [0.6241, 0.6346] | |
GBDT-softer | 0.1564 [0.1554, 0.1574] | 0.6171 [0.6130, 0.6211] | |
XGBoost-softer | 0.1560 [0.1550, 0.1569] | 0.6153 [0.6115, 0.6190] | |
CatBoost-softer | 0.1552 [0.1545, 0.1560] | 0.6124 [0.6095, 0.6152] |
Medicine
XGB | LGB | CatBoost | K-CNN | RF | RGF | LR |
---|---|---|---|---|---|---|
0.9234 | 0.9153 | 0.9218 | 0.9307 | 0.9295 | 0.9276 | 0.9069 |
E1 | E2 | E3 | E4 | E5 | E6 |
---|---|---|---|---|---|
0.9332 | 0.9331 | 0.9325 | 0.9322 | 0.9332 | 0.9333 |
Method | Accuracy | Precision | Recall | F1 | AUC |
---|---|---|---|---|---|
All features | |||||
OneR | 0.9982 ± 0.003 | 1.00 ± 0.01 | 0.99 ± 0.01 | 0.99 ± 0.01 | 1.00 ± 0.01 |
J48 | 0.9926 ± 0.006 | 0.99 ± 0.02 | 0.96 ± 0.04 | 0.98 ± 0.02 | 0.98 ± 0.02 |
SMO | 0.9927 ± 0.005 | 0.99 ± 0.02 | 0.96 ± 0.03 | 0.98 ± 0.02 | 0.98 ± 0.01 |
SVC | 0.9955 ± 0.004 | 0.97 ± 0.03 | 1.00 ± 0.01 | 0.98 ± 0.02 | 0.99 ± 0.00 |
RandomForest | 0.9980 ± 0.003 | 1.00 ± 0.01 | 0.99 ± 0.02 | 0.99 ± 0.01 | 1.00 ± 0.00 |
MultiboostAB | 0.9998 ± 0.001 | 1.00 ± 0.00 | 1.00 ± 0.00 | 1.00 ± 0.00 | 1.00 ± 0.00 |
CatBoost | 0.9993 ± 0.001 | 1.00 ± 0.01 | 1.00 ± 0.00 | 1.00 ± 0.00 | 0.99 ± 0.00 |
RFE-16 | |||||
OneR | 0.9978 ± 0.003 | 1.00 ± 0.01 | 0.99 ± 0.02 | 0.99 ± 0.01 | 0.99 ± 0.01 |
J48 | 0.9947 ± 0.005 | 0.99 ± 0.01 | 0.97 ± 0.03 | 0.98 ± 0.02 | 0.99 ± 0.01 |
SMO | 0.9966 ± 0.004 | 0.99 ± 0.01 | 0.98 ± 0.02 | 0.99 ± 0.01 | 0.99 ± 0.01 |
SVC | 0.9951 ± 0.002 | 0.97 ± 0.02 | 0.99 ± 0.01 | 0.98 ± 0.01 | 0.99 ± 0.00 |
RandomForest | 0.9983 ± 0.002 | 1.00 ± 0.01 | 0.99 ± 0.01 | 0.99 ± 0.01 | 1.00 ± 0.00 |
MultiboostAB | 0.9988 ± 0.002 | 1.00 ± 0.01 | 0.99 ± 0.01 | 1.00 ± 0.01 | 1.00 ± 0.00 |
CatBoost | 0.9979 ± 0.002 | 0.99 ± 0.01 | 1.00 ± 0.01 | 0.99 ± 0.01 | 0.99 ± 0.00 |
RRF-8 | |||||
OneR | 0.9971 ± 0.003 | 1.00 ± 0.01 | 0.98 ± 0.02 | 0.99 ± 0.01 | 0.99 ± 0.01 |
J48 | 0.9960 ± 0.004 | 1.00 ± 0.01 | 0.98 ± 0.02 | 0.99 ± 0.01 | 0.99 ± 0.01 |
SMO | 0.9995 ± 0.002 | 1.00 ± 0.01 | 1.00 ± 0.00 | 1.00 ± 0.00 | 1.00 ± 0.00 |
SVC | 0.9997 ± 0.001 | 1.00 ± 0.00 | 1.00 ± 0.01 | 1.00 ± 0.00 | 0.99 ± 0.00 |
RandomForest | 0.9982 ± 0.003 | 1.00 ± 0.01 | 0.99 ± 0.01 | 0.99 ± 0.01 | 1.00 ± 0.00 |
MultiboostAB | 0.9977 ± 0.003 | 1.00 ± 0.01 | 0.99 ± 0.02 | 0.99 ± 0.01 | 1.00 ± 0.00 |
CatBoost | 0.9986 ± 0.002 | 0.99 ± 0.01 | 1.00 ± 0.01 | 1.00 ± 0.01 | 0.99 ± 0.00 |
Electrical utilities fraud
XGBoost | w/o | w/ | w/ | w/ | w/ | w/ |
---|---|---|---|---|---|---|
Synth | Mean | Std | Min | Max | All 4 | |
DR(%) | 94 | 95 | 95 | 95 | 95 | 96 |
FPR(%) | 6 | 5 | 4 | 4 | 4 | 4 |
CatBoost | w/o | w/ | w/ | w/ | w/ | w/ |
Synth | Mean | Std | Min | Max | All 4 | |
DR(%) | 97 | 97 | 97 | 97 | 97 | 97 |
FPR(%) | 5 | 6 | 5 | 5 | 5 | 3 |
Light | w/o | w/ | w/ | w/ | w/ | w/ |
GBM | Synth | Mean | Std | Min | Max | All 4 |
DR(%) | 97 | 97 | 97 | 97 | 97 | 97 |
FPR(%) | 7 | 7 | 6 | 5 | 6 | 5 |
Features | CatBoost (%) | DT (%) | KNN (%) | |
---|---|---|---|---|
Precision | 71 | 98.11 | 97.23 | 94.18 |
9 | 97.40 | 96.8 | 96.58 | |
Recall | 71 | 99.27 | 97.80 | 45.10 |
9 | 98.68 | 98.24 | 99.12 | |
F-Measure | 71 | 98.69 | 97.51 | 61.00 |
9 | 98.04 | 97.53 | 97.83 |
Meteorology
Psychology
Traffic Engineering
Time per tree | |
---|---|
CatBoost Plain | 1.1 s |
CatBoost Ordered | 1.9 s |
XGBoost | 3.9 s |
LightGBM | 1.1 s |
Cyber-security
Bio-chemistry
Method | Accuracy (%) | Sensitivity (%) | Specificity (%) |
---|---|---|---|
MAN-HOPE-LR | 83.75 ± 0.11 | 83.21 ± 0.47 | 84.30 ± 0.32 |
MAN-HOPE-Ada | 84.73 ± 0.18 | 85.53 ± 0.29 | 83.93 ± 0.22 |
MAN-HOPE-RF | 92.66 ± 0.12 | 92.03 ± 0.15 | 93.29 ± 0.22 |
MAN-HOPE-XGB | 89.56 ± 0.41 | 90.60 ± 0.28 | 88.51 ± 0.95 |
Proposed method | 93.30 ± 0.12 | 91.50 ± 0.14 | 95.10 ± 0.11 |
Method | Precision (%) | MCC (%) | AUC (%) |
---|---|---|---|
MAN-HOPE-LR | 84.13 ± 0.20 | 67.52 ± 0.22 | 91.58 ± 0.13 |
MAN-HOPE-Ada | 84.19 ± 0.18 | 69.48 ± 0.36 | 92.07 ± 0.13 |
MAN-HOPE-RF | 93.21 ± 0.20 | 85.33 ± 0.24 | 97.12 ± 0.05 |
MAN-HOPE-XGB | 88.75 ± 0.81 | 79.13 ± 0.79 | 96.02 ± 0.24 |
Proposed method | 94.91 ± 0.11 | 86.66 ± 0.24 | 97.93 ± 0.08 |