Continuous br Table br UCI datasets used for evaluating perf

Continuous 5.19 31.73 1102.50
Table 4
UCI datasets used for evaluating performances of different online learning models.
Dataset
Data Optimization Training & Number of
Size Set
Testing Set Batches
WDBC
Table 5
Comparison of 5 online learning models on 3 UCI datasets.
Performance
Ionosphere
WBCD Spambase
Metric Model Average SD Average SD Average SD
Fig. 3. Performances of online models on SEER BC dataset.
The GAOGB model shows a testing accuracy of 75.0% in aver-age and 75.2% in median. The AUC with GAOGB USP7/USP47 inhibitor 75.1% in average and 75.2% in median. As demonstrated in Fig. 3, GAOGB outper-forms other online models significantly. In addition, GAOGB shows the lowest variations with Accuracy and AUC. Even though OSELM was the second best model in terms of Accuracy and AUC with UCI datasets, its advantage faded with SEER data, noted by a 71.9% testing accuracy. Instead, OLRGB model, with 73.8% testing accu-racy, performs second best among all models with SEER data. Still, GAOGB is more accurate than the OLRGB from a two sample test (p-value < 0.01).
To understand the advantages of GAOGB compared to conven-tional o ine learning algorithms, Fig. 4 illustrates a comparison
Table 6
pretation of results, the dataset is balanced. The survival cases are under sampled to gain 5,991 cases equally for survival and non-survival cases, amounting to 11,982 observations for compu-tation. The balanced SEER dataset is split to obtain 10% (1200 sam-ples) as the optimization set, 72% (8627 samples) as the training set and 18% (2155 samples) as the testing set. The training set is further split into 10 batches to formulate an online learning environment.
4.5. Evaluation on SEER dataset
The GAOGB model is comprehensively evaluated by comparing to both online and o ine models. Fig. 3 (online models) and Fig. 4 (o ine models) represent the evaluation results in 5 replications on SEER BC dataset. The averaged results for all replications of on-line models are given in Table 7. 
Range of each attributes for preprocessed SEER Breast Cancer dataset.
Categorical variable Number of unique values
Race
Marital status
Primary site code
Histology
Behavior
Grade
Extension of disease
Lymph node involvement
Stage of cancer
Site specific surgery code
Continuous variable Mean SD Range
Table 7
Experimental results on the SEER BC prognosis dataset with different models.
Specificity (%) Sensitivity (%) Retraining Time (Sec.)
Average SD
Average SD Average SD Average SD Average SD
Fig. 4. Comparison between GAOGB and batch learning algorithms as the number of training batches increase.
between GAOGB and o ine algorithms, including the SVM using the Radial Basis Kernel Function, the MLP and the AdaBoost meth-ods. By dividing the training data into 5 batches, Fig. 4 represents the changes in accuracy and retraining time with the number of batch data used for training. The GAOGB model shows insignifi-cant difference to the AdaBoost model in terms of accuracy. With a 75.9% accuracy after training 5 batches, GAOGB shows a trade-off of 0.7% in accuracy compared to the final 76.6% accuracy of the SVM and MLP models. However, the dominating advantages of GAOGB in retraining e ciency is significant in Fig. 4(b). As the number of batches increase, GAOGB maintains a retraining time around 30 s, while the retraining times of the o ine models keep increasing. At Iteration 5, the required retraining time of GAOGB is 24 times lower than the SVM model, 4 times lower than the MLP model, and 3 times lower than the AdaBoost model. The weighted area under the ROC curve ensemble (WAUCE) model proposed by Wang et al. (2017) provides another source of comparison. As a batch learning algorithm, with the same data preprocessing meth-ods, the WAUCE model achieved 76.4% accuracy (Wang et al., 2017). With a 1% average trade-off in accuracy, GAOGB significantly reduced the retraining redundancy of batch learning models.