Page - 44 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

Table8. PredictionAccuracyComparisonofFiveModels. NO. Model Accuracy NO. Model Accuracy 1 SVM 0.650 4 LSTM 0.625 2 GBDT 0.683 5 Perceptron 0.945 3 XGBoost 0.740 accuracy is not high.At the same time, it can also be concluded thatwhen choosing a widevariety of complexmachine learning, the specific choice ofwhatmethoddepends on the size of thedataset and the complexity of theproblem itself. In the experiment of thispaper, theeffectof simpleperceptron is themost effective. The following is a detailed introduction to the experimental situationof the simple perceptron, and the test resultsof theperceptronare shown inTable9. The ‘before’ column indicates that themodel only uses structureddata for training andtesting,and its testingaccuracy is92.5%; the‘after’columnrepresents thecombina- tionof structureddata, news text data anddefinedKeywordGroups (Table 6), and then training and testingwith an accuracy of 94.5%. It can be seen that adding unstructured textdatawill behelpful in improvingaccuracy. Table9. PredictingResultsofPerceptron. NO. before after NO. before after 1 0.92792793 0.95495495 6 0.927927928 0.954954955 2 0.92792792 0.936936937 7 0.90990991 0.927927929 3 0.90990991 0.963963965 8 0.918918919 0.945945948 4 0.927927928 0.936936937 9 0.927927928 0.945945947 5 0.936936937 0.936936938 10 0.936936937 0.945945946 Average 0.925225225 0.945045045 Thefollowingisadeterminationof theprosandconsof themodel throughtheROC (Receiver Operating Characteristic) curve, which is often used to evaluate the merits of aBinaryClassifier [11]. As can be seen fromFig.2, theROCcurve (before) of one classifier is completely ‘below’ the curve (after) of another [12], then it canbe asserted that the performance of the latter is better than the former, that is, by adding the news text caneffectively improve theaccuracyof thepredict. 7. Conclusion In thispaper, inorder to improve theaccuracyofpredicting the levelofcapitaladequacy ofcommercialbanks,worksare summarizedas follows: • Thecombinationof structuredandunstructureddata. • Themethodofextractingkeyword features. • Multiplemodel comparisonandselection. The average accuracy rate of predicting the level of capital adequacy of the com- mercialbank in thenext sevendays ismore than94.5%,whichproves thevalidityof the model. Y.Duetal. /Predicting the InterbankCapitalAdequacyLevelBasedonFinancialDataAnalysis44