Added notes from last meeting and ideas from papers

README.md

 # ML Project - Operator Decision
+## TODOs
+- [x] Import and Read data
+- [x] Read report from student
+- [ ] Check the statistical model (methods used for the first classification)
+- [x] Check other paper for methods for false alarm detection (See project proposal)
+- [ ] Reproduce the MLP model on our own
+- [ ] Explore the output data
+- [ ] Implement better visualisation/metrics to understand results
 
-## 1st Meeting
-TODO:
+Next steps:
+- [ ] New version of MLP code
+- [ ] Access to our Gitlab
+- [ ] Ask for access for the PakoPak Gitlab
+- [ ] Talk with tech-guys about how to setup the environment (Docker, own JupyterLab on server or similar)
 
-- Read data
-- Read report from student
-- Reproduce the MLP model on our own
-- Check the statistical model (methods used for the first classification) -> see paper 
-- Check other paper for methods for false alarm detection (See project proposal)
-
-Additional notes:
+## Meetings
+### 1st Meeting
 - we use features calculated from data rather than fill data with 0 when we have different resolutions (better performance) 
-    - option for later: add own features
 - database: only alarms -> classified as good/bad
 - data unit: the full profile -> reject/accept the whole profile as bad/good
-- use Pytorch
-- Meeting One day/week -> 16h friday 12.04
 - problem with database : not large
 	- careful by splitting -> require class balance (reduces number of data used)
 	- depending on splitting -> different performance
 	- solutions: replicate database, class balance (but for now just implement it like this)
-- historical data -> from boyes; what about the other datasets ? 
+- historical data -> from boyes	
 - in-situ -> field study 
+- use Pytorch
+- Meeting One day/week 
+- Next Meeting: friday 12.04 16h
+
+### 2nd Meeting
+- For sampling for training: use undersampling (min(\# true, \# false))
+- Choice of architecture: from some paper. Link ? 
+- Model results in the report: only for the temperature (best results, but training was done for salinity/temperature separately)
+- Task: investigate which data is well modeled (or not) & Analysis of the results
+- Think of more metrics (other than confusion matrix) for the results
+- Investigate categories for false positives (spike, sensor drift - time correlated errors .. etc)
+	- investigate distribution of categories and search patterns
+	- go back to input data and visualize the original measurements
+- Output should be: a value of confidence (probability of error) between 0 and 1
+	- high values of confidence -> some type of error
+	- medium -> another type 
+	- low -> a third type (or false error)
+
+
+## Ideas
+- Feature engineering ideas:
+	- add more features (25%, 75%, std, etc ...)
+	- sample more points rather than just min/max/mean
+	- let a network learn the features or some embedding or the measures
+	- use contrastive learning to learn similarity between false/true alarms [Link to paper 1]
+- Explain the results
+	- Visualize the alarms detected (or not) as False Positive (just go over them manually to see patterns)
+	- use SHAP value for expalinable AI (which features are most relevant) [Link to paper 2]
+	- Investigate possible reliability measures (online search TODO)
+
+## Questions
+- what about the other datasets mentioned in paper ? 
+- How about other historical datasets with less/more alarams ? 
+
+