Why game users drop out of blade & soul? - 2018 big contest

Why Users Drop out of
Blade & Soul ?
o Leo HyeongNam Lee
o Sep 14, 2018

1. Introduction
2. EDA & Data Preparing
- Preparing Analysis
- Party Data
- Activity Data
- Trade Data
- Payment Data
- Guild Data
3. Modeling
- Model Selection
- Feature Engineering
- Grid Search
4. Analysis : why users drop out?
5. Conclusion

EDA &
Data PreParing
1. Preparing Analysis
2. Party Data
3. Activity Data
4. Trade Data
5. Payment Data
6. Guild Data

Create Custom Functions
1. Counting Functions : Count number of things.
EX) participation in party and guild.
2. Converting Functions : Convert original forms to target forms.
Usually use string to list, time and etc. label to score also.
3. Connecting Functions : Find target information from their index or
id list.
4. Association Functions : Analyze relationship between users, party
and guild. Especially use party, guild and trade data.
5. Visualization Functions : ECDF, bar and etc.

Party Data
1. Data Structure
2. Number of Participation
3. Duration of party
4. Scale of party
5. Week, Day, Hour

Data Structure
1. About 7M rows and 7 columns
2. No NA
3. No duplicates

Number of participation
1. ECD Graph
2. Cut to 1000 (except outlier)
3. Week = black, month = red, 2month = blue, retained = green

Duration of Party
1. ECD Graph
2. Count number of participation in party by DURATION(UNIT: 5min)
EX) 0~5, 5~10, … , 55~60, 60~
3. 2 kinds of shape

Scale of Party
1. ECD Graph
2. Count number of participation in party by
NUMBER OF PARTY MEMEBERS. EX) 1, 2, … , 10, 11~
3. 1, 6, 11~ are interesting.

Week
1. ECD Graph
PARTY START WEEK. EX) 1, 2, … , 7, 8 AND mean, median, max
3. 1, 7, 8, summaries are interesting.

1. ECD Graph
PARTY START WEEK. EX) 1, 2, … , 7, 8 AND mean, median, max
3. 1, 7, 8, summaries are interesting.
Week

1. ECD Graph
PARTY START DAY. EX) 1, 2, … , 6, 7 AND mean, median, max
3. 1, 7, max are interesting.
Day

1. ECD Graph
2. Count number of participation in party by PARTY START HOUR.
EX) 0~2, 3~5, … , 18~20, 21~23 AND mean, median, max
3. 6~8, , 18~20, max are interesting.
Hour

Activity Data
1. Data Structure
2. Data Preparing by Week
3. Interesting Weeks
4. Interesting Columns

Data Structure
1. About 440,000 rows and 38 columns
2. If users didn’t log in the game a whole week,
activity data don’t have been logged that week
3. No duplicates
4. Almost columns are regularized.
Irregular columns are below
wk : activity week
acc_id : user id
cnt_dt: 1~7, int

Data Preparing by Week
1. Define Function
2. Seperate Dataset
3. Extract Values : 1st~8th week & sum, mean, median, max

Interesting Week
1. ECD Graph
2. 2 kinds of shape
3. Anyway, HARD TO CLASSIFY MONTH AND 2MONTH.
They have extremely similar features.
1st type : almost 1st ~ 7th week 2nd type : 8th week

Interesting Columns
1. ECD Graph
2. 2 kinds of shape
3. Labels are extremely same as each others in many cases.
1st type : about 5 features are meaningful 2nd type : others are 0 or outlier

Trade Data
1. Data Structure
2. Association between Users
3. Scoring Association
4. Counting Trade by Week
5. Counting Trade by Day
6. Counting Trade by Hour

Data Structure
1. About 10M rows and 7 columns
2. No NA
3. 45% of rows are duplicates

Association between Users
1. Bar Graph
2. Counting number of users who trade more than twice with each users.
3. 2month label is special
In Sell Cases In Buying Cases In Both Cases

Scoring Association
1. Define Function
2. In Sell Cases 2. In Buy Cases

Counting Trades by Week
1. ECD Graph
2. Counting number of trades by Week
3. Never meaningful (In sell cases, very slightly meaningful)

Counting Trades by Day
1. ECD Graph
2. Counting number of trades by Day

Counting Trades by Hour
1. ECD Graph
2. Counting number of trades by Hour

Payment Data
1. Data Structure
2. Weekly Payment
3. Paying or Not

Data Structure
1. 800,000 rows and 3 columns
2. No NA: even if NO PAY
3. No duplicates
4. Payment Amounts are regularized.

Weekly Payment
1. ECD Graph
2. A few users had paid
3. Only retained users had paid

Paying or Not
1. Bar Graph
2. Retained users had paid more

Guild Data
1. Data Structure
2. Association between Guild and Users
3. Scale of Guild
4. Joining Guild or Not

Data Structure
1. About 10,000 rows and 2 columns
2. No NA
3. No duplicates

Association between Guild and Users
1. ECD Graph
2. Scoring Guild: if guild members belong to train data,
they take a score depending on their label.
And summarize it.
3. Retained users stand out against others

Scale of Guild
1. Scale : Number of guild members
2. Sometimes users join multiple guilds. So summarize it.
4. ECD Graph, Week = black, month = red, 2month = blue, retained = green

Joining in Guild or Not
1. If users join in guild, 1. Or if not, 0.
3. Bar Graph, Week = black, month = red, 2month = blue, retained = green

Modeling
1. Model selection
2. Feature Engineering
3. Grid Search
4. Prediction

Model Selection
1. Select Candidates: Regression, SVM, DNN and Tree
2. Features of Dataset
- Hundreds of Features : many features are very similar each class
- 100,000 rows
- Multi-Class Classification
3. Regression : Not good at hundreds of features
4. SVM : Not good at 100,000 rows
5. DNN : I have only laptop
6. Tree : Tree ensemble model can handle this problem.
So, I chose light-gbm.

Feature Engineering
1. Based on EDA, validate the feature importance.
2. Feature importance follows EDA results
Unmeaningful things are badMeaningful features perform excellent

Grid Search
1. Start light-gbm from default option
2. Learning rate : default is best
3. Loss function : don’t have meaningful differences
4. Number of rounds & number of tree leaves : Search
Default Option Grid Search

Why Users Drop out of
Blade & Soul ?

1. Retained Users
- They seems to Blade & Soul is a part of their life.
- They enjoy various game contents actively.
- They actively interact others
2. Others
- They play game extreme way. Not a routine.
- They rarely play hardcore contents
- They do not interact others

I Failed This Way
1. Create pseudo rows :
Assumed each 1st ~ 7th week is the last week of recoding,
not 8th week, so can infer label of new data.
2. Over sampling
3. Predict each class : Not multi-class classification
4. Ensemble model vertically

Copyright
1. Image : NCSOFT Corp.
2. Fonts
- DX신사임당Bold : DXKOREA

Why game users drop out of blade & soul? - 2018 big contest

More Related Content

Similar to Why game users drop out of blade & soul? - 2018 big contest (20)

More from HYEONGNAM LEE (6)

Recently uploaded (20)

Why game users drop out of blade & soul? - 2018 big contest