https://github.com/rmsnow/adtracking
TalkingData AdTracking Fraud Detection Challenge.
https://github.com/rmsnow/adtracking
Last synced: over 1 year ago
JSON representation
TalkingData AdTracking Fraud Detection Challenge.
- Host: GitHub
- URL: https://github.com/rmsnow/adtracking
- Owner: RMSnow
- Created: 2018-04-21T15:45:06.000Z (about 8 years ago)
- Default Branch: master
- Last Pushed: 2018-05-01T17:06:12.000Z (about 8 years ago)
- Last Synced: 2025-01-30T10:30:44.233Z (over 1 year ago)
- Language: Python
- Size: 1.04 MB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Feature Engineering
## Data Fields
| Features | Descriptions |
| --------------- | ------------ |
| ip | Not Null |
| app | Not Null |
| device | Not Null |
| os | Not Null |
| channel | Not Null |
| click_time | Not Null |
| attributed_time | |
| is_attributed | 0 / 1 |
## Feature Selection Categories
### 1 Basic Features (#5)
ip, app, device, os, channel
### 2 Basic Features' attributed contributors (#5 * 2)
For each **unique value** of every basic feature, count the number of click that **is attributed or not**.
### 3 Frequencies of Basic Features (#5)
For each **unique value** of every basic feature, calculate the value's frequency in the whole dataset.
### 4 Conversion Rate (#5)
For each **unique value** of every basic feature, calculate the values conversion rate (i.e. the fraction, `#is_attributed clicks / #clicks`).
### 5 Correlated Features' Combination (#n)
Select the features' combination whose features own a high value of correlation.
### 6 Temporal Extraction
As for different time span(whole time, minute, and hour), calculate `raw` , `average` and `standard deviation` of all the features above.
So the amount of all above features is `3 * (20 + n) * 3`.
### 7 Temporal Conversion Rate (#1)
Calculate every hour's conversion rate.
### 8 Others
eg: temporal interval
## Experiments
### Performance
| Num | Category | Features | AUC on dev set | AUC on test set | AUC in Real World |
| ---- | :------: | :------------------------------------------------: | :----------------: | :-------------: | :---------------: |
| 1 | 1 | Basic #5 | 0.9774025893305988 | | 0.9583 |
| | | Basic #5 | 0.959279 | 0.963003 | 0.9559 |
| 2 | 1 | Basic #6 (hour) | 0.9765065318797589 | | 0.9563 |
| | | Basic #7(hour, day) | 0.960369 | 0.964231 | 0.9563 |
| 3 | | "Basic" | 0.974860783943336 | | **0.9684** |
| 4 | 1,2 | Add count | 0.9759619912016608 | | |
| 5 | 1,2 | Add attributed count | 0.9840863114093297 | | 0.6239 |
| 6 | 1,2 | Add attributed count (no hour's effect) | 0.9840863114093297 | | 0.6114 |
| 7 | 1,2 | Add count and attributed count | 0.9842108186300558 | | |
| 8 | 1,3 | Add Frequency | 0.9759619912016608 | | |
| 9 | 1,2,3 | Add count, attributed count, frequency | 0.9843591094736229 | | |
| 10 | 1,4 | Add count, conversion rate | | | |
| 11 | 1,2,3,4 | Add count, attributed count, frequency, conversion | 0.9842607399131638 | | |
| 12 | 1,6 | Add hour count | 0.975100704348888 | | |
| 13 | 1,6 | Add hour attributed count auc | 0.9952688459421872 | | 0.7051 |
| 14 | 1,5 | Add ip_channel | 0.979456127558631 | | 0.9622 |
| 15 | 1,5 | Add app_channel | 0.9783578247039497 | | 0.9592 |
| 16 | 1,5 | Add ip_device | 0.9801894479874732 | | 0.9628 |
| 17 | 1,5 | Add All two degree of features | 0.9813457066441867 | | 0.9646 |
| 18 | 1,5 | Add All two degree of features(100,000,000) | | | 0.9656 |
| 19 | 1,5 | "Basic" + some features of **high_importance** | 0.9863726922286226 | | **0.9684** |
**PS**: In `Basic#5`, `click_time` is transfered into `hour` and `day`.
### Plot Importance of Features
#### Num 3: `"Basic"`
#### Num 18: `Basic + All two degree of features`
Select `device_ip`, `channel_app`, `app_ip`, `device`, `os_ip` to add into `Num3: "Basic" `.
#### Num 19
### Correlations of features
#### Mutual Information
`Normalized corr_rate = I(X;Y) / H(X,Y) = I(X;Y) / (H(X) + H(Y) - I(X;Y))`
Mask = 0.1
```
app & channel: 0.269795263371
app & ip_channel: 0.104454990715
app & device_channel: 0.263802502515
app & os_channel: 0.174031886398
channel & ip_app: 0.105000757011
channel & app_device: 0.256371969166
channel & app_os: 0.148765039685
ip_app & device_channel: 0.109073593121
ip_app & os_channel: 0.150919655929
ip_channel & app_device: 0.105314833255
ip_channel & app_os: 0.136361783362
app_device & os_channel: 0.169265661127
app_os & device_channel: 0.149640128183
```
Mask = 0.2
```
app & channel: 0.269795263371
app & device_channel: 0.263802502515
channel & app_device: 0.256371969166
```
Mask = 0
```
ip & app: 0.00927286991461
ip & device: 0.00694314823037
ip & os: 0.0295303991247
ip & channel: 0.01910882176
ip & app_device: 0.0116509685587
ip & app_os: 0.0352817805537
ip & app_channel: 0.0269119027086
ip & device_os: 0.0319996095279
ip & device_channel: 0.0232395465844
ip & os_channel: 0.0687097193232
ip & hour: 0.00842828568901
app & device: 0.0278221472785
app & os: 0.0138230238585
app & channel: 0.269795263371
app & ip_device: 0.0147259939819
app & ip_os: 0.0259135610616
app & ip_channel: 0.104454990715
app & device_os: 0.0146956128532
app & device_channel: 0.263802502515
app & os_channel: 0.174031886398
app & hour: 0.00196613604054
device & os: 0.0292843937519
device & channel: 0.0190570407725
device & ip_app: 0.010875061967
device & ip_os: 0.0123530449439
device & ip_channel: 0.0109356373285
device & app_os: 0.0164432432512
device & app_channel: 0.0215004570006
device & os_channel: 0.0157067148459
device & hour: 0.00218613697383
os & channel: 0.0111228058317
os & ip_app: 0.0392854321075
os & ip_device: 0.036077091733
os & ip_channel: 0.0587870031998
os & app_device: 0.0155230901264
os & app_channel: 0.0127829864518
os & device_channel: 0.0152616221357
os & hour: 0.000521350840747
channel & ip_app: 0.105000757011
channel & ip_device: 0.0241785886481
channel & ip_os: 0.0498727607888
channel & app_device: 0.256371969166
channel & app_os: 0.148765039685
channel & device_os: 0.0125333204848
channel & hour: 0.00297988945705
ip_app & device_os: 0.0420245202447
ip_app & device_channel: 0.109073593121
ip_app & os_channel: 0.150919655929
ip_app & hour: 0.0116179491756
ip_device & app_os: 0.0412224118725
ip_device & app_channel: 0.0327801472947
ip_device & os_channel: 0.0755055123239
ip_device & hour: 0.00926587918155
ip_os & app_device: 0.0292665954443
ip_os & app_channel: 0.0654641563309
ip_os & device_channel: 0.0560870467462
ip_os & hour: 0.0302603369959
ip_channel & app_device: 0.105314833255
ip_channel & app_os: 0.136361783362
ip_channel & device_os: 0.0618911697116
ip_channel & hour: 0.0232839554938
app_device & os_channel: 0.169265661127
app_device & hour: 0.00226247457287
app_os & device_channel: 0.149640128183
app_os & hour: 0.00113844687176
app_channel & device_os: 0.0147866713056
app_channel & hour: 0.00478454682739
device_os & hour: 0.000885698518659
device_channel & hour: 0.00329861922203
os_channel & hour: 0.00227965522251
```