Temperature Prediction Data Analysis

1. Data Preprocessing

1. Select only June, July, August and September

   

   • Because May and Oct. have very less data, we are trying to focus on mid-summer.

2. Remove outliers from categorical and continuous variables

   

   • For categorical variables, when ‘A’ is OFF, ‘B’ is F and A, and ‘C’ is W
   • For continuous variables, remove outliers using Z-score and Box plot

3. Convert float USER_SET_TEMPERATURE to int values

   

   • Because we have small amount of float USER_SET_TEMPERATUREs

4. Solve multicollinearity

   

   • Remove p2 and p10 , remain only p1 because p1-p2-p10 has very high correlation

5. Separate into June/July/August and September to secure the train/validation/test data

   • Use June, July and August data for training and validation
   • Use September for testing

6. Drop less informative columns because these data are not conductive to learn model

   • Drop ‘Date’, ‘A’ and ‘G’ columns

7. One-hot encode the categorical variables to apply categorical data into our model

8. Do Standard Scaling for continuous variables except the temperature data

   • Model performance may be improved by normalizing values having different units
   • ‘L’ is between 0 and 615,605
   • ‘M’ is between 0 and 3,726

2. Modeling

1. Try Randomized Search

   Model	Parameters	Best Value
   Lasso Regression	alpha	0.059
   Decision Tree	max_depth	20
   	max_features	None
   	min_samples_leaf	15
   	min_samples_split	12
   Random Forest	max_depth	20
   	max_features	auto
   	min_samples_leaf	9
   	min_samples_split	8
   	n_estimators	100
   Gradient Boosting	max_depth	9
   	learning_rate	0.1
   	min_samples_leaf	1
   	min_samples_split	10
   	n_estimators	50
   ADA Boosting	n_estimators	100
   	learning_rate	0.01

2. Try Validation

   Model	Dataset	MAE	RMSE
   LASSO Regression	Valid	1.07	1.49
   	K-Fold Cross Validation	1.07	1.48
   Decision Tree	Valid	0.97	1.41
   	K-Fold Cross Validation	0.98	1.42
   Random Forest	Valid	0.90	1.32
   	K-Fold Cross Validation	0.90	1.32
   Gradient Boosting	Valid	0.91	1.32
   	K-Fold Cross Validation	0.90	1.31
   ADA Boosting	Valid	1.18	1.61
   	K-Fold Cross Validation	1.17	1.60

3. Final Metrics

   Model	MAE	RMSE	Model Size	Model Latency	E2E Latency
   LASSO Regression	1.15	1.59	0.64KB	0.63ms	11ms
   Decision Tree	1.15	1.63	457KB	1.05ms	12ms
   Random Forest	1.05	1.50	48,270KB	9.41ms	19ms
   Gradient Boosting	1.08	1.54	919KB	0.90ms	12ms
   ADA Boosting	1.27	1.78	31KB	13.79ms	23ms