Analysis of BlogFeedback Data Set
Predict the number of comments in a blog after the basetime
The data used here is BlogFeedback Data Set from UCI https://archive.ics.uci.edu/ml/datasets/BlogFeedback
The data set contatins 1 training data set and 60 test data set.
| Features | |
|---|---|
| Rows | 60021 |
| Columns | 281 |
The metric used to evaluate the performace of this model is the log transformed mean squared error. (1 / n)∑[ln(1 + y)−ƒ̂(x)]
- Feature Creation - Creating new features by combining different original features, PCA, extracting feature interactions from random forest tree model, etc.
- Feature selection - Filtering features by group lasso, feature inportance of random forest model
- Feature transformation - Transform features by scaling with normalization
| Model | training time | Loss |
|---|---|---|
| RandomForest | Moderate | 0.39763 |
| KNN | Fast | 0.63239 |
| Lasso | Fast | 0.49092 |
| Neural Network | Very Slow | 0.42592 |
| GBM | Moderate | 0.37127 |
After experiments with feature engineering, we evaluated the performace of Gradient Boosting Tree model
We further ensemble the gradient boosting tree model by model averaging to improve the performance.
| Model | Error |
|---|---|
| Single | 0.37128 |
| Ensemble | 0.36985 |
Peiran Cao
Zhoutao Pei
Boyang Liu