An implementation of logistic regression using google's FTRL-proximal online predicting and adaptive learning method with python3 to solve a CTR prediction problem with features in a very high dimension space.
Origin paper:
https://www.eecs.tufts.edu/~dsculley/papers/ad-click-prediction.pdf
The code is based on the following codes from kaggle:
https://www.kaggle.com/jiweiliu/ftrl-starter-code
Dataset used: https://www.kaggle.com/c/springleaf-marketing-response/data
- It is an online learning method, samples are fed in stream (the model learns each sample only once, which means there is only one endless training "epoch".) and the model is updated in real-time.
- The learning rate is adaptive per-coordinate, which means the learning rates of all weights are varying adaptively.
- Using the strong L1 and L2 regularization, the model is going to be sparse, saving lots of memory.
- A good balance between model performance (e.g. AUC) and model sparsity.