---
title: Activate learning
tags: paper reading
description: description.
---

[toC]

<!-- Put the link to this slide here so people can follow -->
slide: https://hackmd.io/0E2n-BuvQT-_W0ellUGn6w

# MESA: Boost Ensemble Imbalanced Learning with MEta-SAmpler

## background
imbalanced learning,i.e., learning unbiased models from class-imbalanced
data, is a challenging problem.
typical IL methods have following drawbacks:
- unstable/unrobust performance, perform well in terms of global accuracy but poorly on the minority class.
- poor applicatility because of the prerequisite hand-craft matrix
- high cost of computing the distance between instances

FYI, Canonical classification algorithms usually induce the bias, i.e., perform well in terms of global accuracy but poorly on the minority class, in solving class imbalance problems.

## related work
### Resampling
resample training set to balance the class distribution, (i.e.: over/under sampling), or filter noice(i.e.:cleaning resampling).
resampling methods are often guided by the distance information, but culculate the distances between instances of large-scale dataset is computationally expensive.

### Reweighting
Reweighting methods assign different weights to different instances to alleviate a classifier’s bias towards majority group.
this is infeasible because of prerequisite by expert.

### Ensemble Methods
it is effectively imporve IL solutions, but it is simply combinations of resampling/reweighting solution, it suffer from unsatisfactory performance.

### Meta-learning Methods
it is focus on DNN but fails in other deep learning fields like computer vision or natural language processing or tree-based models.

## summarize in a sentence
this paper takes advantage of ensemble learning and meta-learning, and proposed a general, low-computaion-consuming resample method based on the dataset(training and validation) decoupled

from specific models,thus it can be applied to every task without limitaion.

## framework



**Meta-state** is a task-independent state representaion.

**Meta-sampling** is used for sampling D_sample_train from D_all_train.

**Ensemble Training** trains multiple base classifiers in a series way.

**Meta training** using reinforcement learning to learn the most appropriate sampling strategy(sampler parameters) directly from the data to optimize the final generalization performace of the ensemble model.


## Experiments

To thoroughly assess the effectiveness of MESA, two series of experiments are conducted: one on controlled synthetic toy datasets for visualization and the other on real-world imbalanced datasets to validate MESA’s performance in practical applications. We also carry out extended experiments on real-world datasets to verify the robustness and cross-task transferability of MESA

synthetic toy datasets:


real-world dataset:
( In order to verify the effectiveness of MESA in practical applications, we extend the experiments to real-world imbalanced classification tasks from the UCI repository [10] and KDD CUP 2004. )




# Meta-AAD: Active Anomaly Detection with Deep Reinforcement Learning

## background
high false-positive rate is a challenge for anomaly detection. analyst usually use re-rank to solve this problem: select the top-ranked instances and labeled by human, then the label information can help re-rank process.And they often greedily select top-1 instance for query which is sub-optimal since some low-ranked instance cound be more helpful in long-term.

But long-term & short-term re-rank have following 3 challenges:
- First, it is unclear how we can quatify the long-term performance.
- Second, the decision space is quiet large since we need to examine all the instances and select one of them for the query.
- Third, difference datasets have various distributions and different sizes of decision spaces.

## related work
- **Anomaly Detection:** e.g., density-based approach, distance-based approach and ensembles. Most of these algorithms are un-supervised with strong assumptions about the anomaly patterns, and they may not work when the assumptions do not hold.
- **Semi-Supervised Anomaly Detection**:assumes that a small set of labeled instances can help inprove the performance, eg:re-weight the anomaly scores , improves representations.However, these methods are designed for batch setting, which could be sub-optimal in the active learning.
- **Active Anomaly Detection** interacts with humans and recomputes the anomaly scores based on the feedback.they all adopt a greedy strategy to select the top-1 anomalous instance in each iteration, which may fails to model long-term performance.
- **Learning Meta-policy**is to train a reinforcement learning agent to make decisions with the objective of optimizing the overall performance of the task.

## summarize in a sentence
Add deep reinforcement learning to active anomaly detection by helping to select instance in re-rank periods to take advantage of short- and long-term contributions of instance.

## framework


This article want to build a general method, but because of the various distribution and dimension of datasets, they came up with an approch to fit every datasets.
First, extracting transferable meta-features:

thus, the dimention of data features is uniformed. But the lengths of datasets are still different, so they propose to insted operate on data streams.



## expriment






# A Comparative Survey of Deep Active Learning

## background

## related work

## summarize in a sentence

## framework

## expriment

# Templete

## background

## related work

## summarize in a sentence

## framework

## expriment