然后呢细节上的实施 就不太行
细节是魔鬼
我就是个忽略魔鬼得人。
要注意细节!
没有什么成功是不需要细节的,
以后这方面要多多训练自己
1. time series - 看完了kaggle arimax, 要看
- lstm with external
- prasad recommend
- 细节搞懂
发现自己很多assumption是对的
- one hot encoding
- residual analysis
基本 time series analysis 过程
- break date into day, month, year
- line plot to see trend
- do seasonable decompose to see seasonal
- test stationary
- test the right thing for I
- ACF, PACF
- Arima, Sarimax
- Residual Test
- Draw both forecast and actual
- Use Mape and SMape
Question
1. How to do cross validation?
2. Arimax how to take the x into consideration
3. VArimax how the code looks like
4. The DCMM
5. Template code
a. transform the fields into general format so following code can be reused
b. general EDA (ACF, PACF, stationary, seasonal, residual test, plot (forecast/actual, line, normality)
c. VArimax code
d. LSTM code (handling to get the right dimension)
e. DCMM code
f. XGBoost code
h. other known good model
i. Putting external information
j. accuracy metrics - can they diffentiable, which one is good, do/how they used in algorithms.
- prasad recommend
- 细节搞懂
发现自己很多assumption是对的
- one hot encoding
- residual analysis
基本 time series analysis 过程
- break date into day, month, year
- line plot to see trend
- do seasonable decompose to see seasonal
- test stationary
- test the right thing for I
- ACF, PACF
- Arima, Sarimax
- Residual Test
- Draw both forecast and actual
- Use Mape and SMape
Question
1. How to do cross validation?
2. Arimax how to take the x into consideration
3. VArimax how the code looks like
4. The DCMM
5. Template code
a. transform the fields into general format so following code can be reused
b. general EDA (ACF, PACF, stationary, seasonal, residual test, plot (forecast/actual, line, normality)
c. VArimax code
d. LSTM code (handling to get the right dimension)
e. DCMM code
f. XGBoost code
h. other known good model
i. Putting external information
j. accuracy metrics - can they diffentiable, which one is good, do/how they used in algorithms.
2. nlu - 0
3. algo - 0
4. expvis - 看了survey,
几种方法
1. prototype
2. Approximation as trees
3. rules extraction
4. Reverse engineering
5. saliency masks
6. Activation maximizatoin
7. sensitivity
---
目前感觉 approximation最简单
其次activation maximization
其实我是靠猜的
下一步要看看里面的代表作更进一步了解。
几种方法
1. prototype
2. Approximation as trees
3. rules extraction
4. Reverse engineering
5. saliency masks
6. Activation maximizatoin
7. sensitivity
---
目前感觉 approximation最简单
其次activation maximization
其实我是靠猜的
下一步要看看里面的代表作更进一步了解。