*A state-of-the-art survey on solving non-IID data in Federated Learning*(<font color="#f00">還沒有整理完</font>font>) https://www.sciencedirect.com/science/article/pii/S0167739X22001686 1. non-iid type 1. Feature distribution skew * P_k (x)不同,但是P_k (y|x) 相同 * 例如兩個人寫的5一個粗一個係(feature不同,P_k (x)不同),但寫粗的都是5 2. Label distribution skew * P_k (y)不同,但是P_k (x|y)相同 * 在client A的資料庫中有90% 1 10% 2;在client B的資料庫中則是9% 1 5% 2。 P_A (1)與P_B (1)不同,但是P_A (x|y)與P_B (x|y)相同 3. Same label, different features * P_k (x|y)不同,但是P_k (y) 相同 4. Same features, different label * P_k (y|x)不同,但是P_k (x) 相同 5. Quantity skew * Quantity skew refers to the large difference in the quantity of different client data Pi(x, y) * A有100比data而B有1000000比data 2. 解決non-iid方式 ![image](https://hackmd.io/_uploads/SknpvWs4T.png)