#EC503 Minutes of Meeting

Firstly, we use several algorithms and try to classify the data. We used DRO, Logistic Regression, KNN, SVM, Decision Tree. 🌲 And we found that the result was bad. Althouth the accuracy were high for all algorithms, that was because the dataset has large bias: 90% of the data are labeled as 0, only 10% of data are labeled as 1. Thus, the accuracy will be 90% even if the classifier predict all test data as 0. And the AUC was significantly low: only 0.5 Even if we normalize the data, the results had nothing change: still very bad.

Plan for the second week:

1. 把label为1的数据提取出来，让label为1的数据的比例大一些，然后看一下结果是否有好转
2. 想办法检测outlier并去除：identify outlier
3. 需要降维吗？
4. 能否试一下：不以整体精确度为导向，而是以“把1都预测对，然后再考虑0”为导向
5. 看一下Kaggle里面的kernel https://www.kaggle.com/c/santander-customer-transaction-prediction/kernels
6. 查一下对数据有bia的情况下的检测方法

Suggestions:

• Survey on the problem of bias in dataset.

  • Undoing the Damage of Dataset Bias</ol
  • Unbiased Look at Dataset Bias

• Survey on outlier detection problems

  • Toward Supervised Anomaly Detection
  • K-Means+ID3: A/ Novel Method for Supervised Anomaly Detection by Cascading K-Means Clustering and ID3 Decision Tree Learning Methods
  • Anomaly detection: A survey

• Imbalanced Learn

• 参考文章ADASYN: Adaptive Synthetic Sampling Approach for Imbalanced Learning. 用 Synthetic data generation来产生新数据点做训练，适用于所有machine learning方法。

• 对于SVM, 可以参考以下两篇结合active learning和svm的文章。

• S. Ertekin, J. Huang, and C. L. Giles, “Active Learning for Class Imbalance Problem,” in Proc. Annual Int. ACM SIGIR Conf. Research and development in information retrieval, pp. 823 - 824, Amsterdam, Netherlands, 2007.

• S. Ertekin, J. Huang, L. Bottou, C. L. Giles, “Learning on the Bor- der: Active Learning in Imbalanced Data Classification,” in CIKM’07, November 6-8, 2007, Lisboa, Portugal.

• 对于NN，可以参考以下文章

• Z. H. Zhou and X. Y. Liu, “Training Cost-Sensitive Neural Networks with Methods Addressing the Class Imbalance Problem,” IEEE Trans- actions on Knowledge and Data Engineering, vol. 18, no. 1, pp. 63-77, 2006.