2020 Shopee Code League

---

Week 1: Order Brushing

Task

Identify all shops and buyers that are deemed to have conducted order brushing by concentrate rate is greater than or equal to 3 at any instance.

Concentrate rate = 
Number of Orders within 1 hour / Number of Unique Buyers within 1 hour

Data format

orderid	shopid	userid	event_time
31076582227611	93950878	30530270	2019-12-27 00:23:03

Submission format

shopid	userid
162014252	183926374
321014322	19233237&23421231
22754767	0

My strategy

1. Use sliding window to count every shop-buyer transaction frequency by hour
2. Select the transaction frequency >= 3
3. Arrange userid by shopid

Score

Out of time Late Submission: 0.96712

Post-match review

1. The use of pandas is not skilled enough
2. Didn't deal with time series data before
3. Running by my computer, computing power too low

---

Week 2: Product Detection

Task

Build a commodity classification model by ~100k images within 42 different categories.

Data format

Images dataset

Submission format

filename	category
fd663cf2b6e1d7b02938c6aaae0a32d2.jpg	20
c7fd77508a8c355eaab0d4e10efd6b15.jpg	27
127f3e6d6e3491b2459812353f33a913.jpg	04
5ca4f2da11eda083064e6c36f37eeb81.jpg	22
46d681a542f2c71be017eef6aae23313.jpg	12

My strategy

Use pre-trained model from ImageNet (Transfer Learning)

Score

0.67643

Post-match review

1. Lack of relevant knowledge to build CNN model and adjust parameters by my own

2. Don't know how to enable GPU and TPU, so can only training with MobileNet which has minimal parameters

---

Week 3: Short Algorithm Contest #1

Task

Solve algorithm problems

Score

Out of time

Too difficult to me Q_Q

---

Week 4: Title Translation

Task

Translate product title in Traditional Chinese to English

Data format

product_title	category
Gucci Gucci Guilty Pour Femme Stud Edition 罪愛女性淡香水限量版 50ml T	Health & Beauty
（二手）PS4 GTA 5 俠盜獵車手5 Grand Theif Auto V繁體 中文版	Game Kingdom
百獸卡	Life & Entertainment
nac nac活氧全效柔衣素	Mother & Baby
#Nike耐吉官方F.C. 男子足球長褲新款標準型 拒水 拉鏈褲腳\nCD0557	Men's Apparel

My strategy

NLP is not my area of interest, so I just skip it

---

Week 5: Logistics

Task

Identify all the orders that are considered late depending on the Service Level Agreements (SLA).

1. 1st attempt = 1st_deliver_attempt - pick, judge to be late if > SLA
2. 2nd attempt = 2nd_deliver_attempt - 1st_deliver_attempt, judge to be late if > 3 days regardless of origin to destination route
3. If no 2nd_deliver_attempt means 1st_deliver_attempt successful
4. All time formats are stored in epoch time based on Local Time (GMT+8)
5. Only consider the date when determining if the order is late; ignore the time
6. Only consider working days, excluding Sunday and public holidays (2020-03-08, 2020-03-25, 2020-03-30, 2020-03-31)
7. Both attempts need to be on time

SLA matrix:

from_to	Metro Manila	Luzon	Visayas	Mindanao
Metro Manila	3 working days	5 working days	7 working days	7 working days
Luzon	5 working days	5 working days	7 working days	7 working days
Visayas	7 working days	7 working days	7 working days	7 working days
Mindanao	7 working days	7 working days	7 working days	7 working days

Data format

orderid	pick	1st_deliver_attempt	2nd_deliver_attempt	buyeraddress	selleraddress
2215676524	1583138397	1583384958		Baging ldl BUENAVISTA,PATAG.CAGAYAN Buagsong,cordova,cebu Mt.VERNON Buolding, Habagat Lordman NATL Metro Manila	Pantranco vill. 417 Warehouse# katipunan 532 (UNIT Metro Manila
2219624609	1583309968	1583463236	1583798576	coloma's quzom CASANAS Site1 Masiyan 533A Stolberge 10,Baloy eastt away 041banahaw street,Tuguegarao agro, Metro Manila	BLDG 210A Moras C42B 2B16,168 church) Complex JUNKSHOP. 22-c Metro Manila
2220979489	1583306434	1583459779		21-O LumangDaan,Capitangan,Abucay,Bataan .Bignay Office,Buhanginan saBrgy186, 34i (bayanihan MALARIA, Alindahaw, Rm401, st.ngry p.pasubas metro manila	#66 150-C, DRIVE, Milagros Joe socorro Metro Manila
2221066352	1583419016	1583556341		616Espiritu MARTINVILLE,MANUYO #5paraiso kengi 12nn-9pm Brgy,Milagrosa 6Putohan,Tramo #18saint вєrnαвє st,CAA Metro Manila	999maII 201,26 Villaruel Barretto gen.t number: 70-B 7A. MALL kanto- 1040 Metro Manila
2222478803	1583318305	1583480500		L042 Summerbreezee1 L2(Balanay analyn Lot760 Cluster3-2T seppina UPPERG/L luzon	G66MANILA Hiyas Fitness MAYSILO magdiwang Lt.4C lot6 2F-48 st.,Binondo 1188Mall2M01 carnation Mae Metro Manila

Submission format

orderid	is_late
1955512445	0
1955598428	1

is_late: assign value 1 if the order is late, otherwise 0

My strategy

1. Split location from address
2. Assign SLA by location
3. GMT+8 and transfer to date time
4. Working days count (np.busday_count)
5. Late judgement

Score

0.63885

Late Submission: 1.0

Post-match review

1. Don't know how to count working days

2. Forget GMT+8

3. Need to use vectorization to speed up when deal with large dataset

---

Week6:Sentiment Analysis (Data Science)

Task

Build a customer review rating model

Data format

review_id	review	rating
11576	It's working properly. Very quick heating capability. Good product with this price thanks	5
10293	Excellent service by the staff, helpful and polite. Great experience overall.	5
01820	The delivery was fast but the packaging was not that good, the price is reasonable, overall the product is ok.,	4
32090	Package not that well	2

Submission format

review_id	rating
1156	1
2654	0

My strategy

NLP is not my area of interest, so I just submit a table with all rating are 5

Score

0.40517

---

Week 7:Short Algorithm Contest #2

Task

Solve algorithm problems

Score

10 (done by mp0530)

---

Week 8: Marketing Analytics

Task

Predict whether users will open the marketing emails

Data format

[train/test].csv

• country_code: An integer code for the country where the user lives.
• grass_date: The date when the email was sent.
• user_id: the unique identifier of each user
• subject_line_length: the number of characters in the subject of the email
• last_open_day: How many days ago was the last time the user opened an email
• last_login_day: How many days ago the user last logged in its Shopee account
• last_checkout_day: How many days ago the user last purchased on Shopee
• open_count_last_[10/30/60]_days: the total number of email opens in the last N days.
• login_count_last_[10/30/60]_days: the total number of user logins in the last N days.
• checkout_count_last_[10/30/60]_days: the total number of checkouts (=purchases) by the user in the last N days.
• open_flag: the target variable. Whether or not the email was opened.
• row_id:

users.csv [empty values are simply unknown]

• user_id: the unique identifier of each user
• attr_[1/2/3]: general user attributes. Attr_1 and attr_2 are boolean, attrib_3 is categorical (can be integer [0,1,2,3,4])
• age: The user's reported age.
• domain: The user's top-level email domain. Less common domains are bundled together under the label 'other'.

Submission format

row_id	open_flag
0	1
1	1
2	0
3	0

My strategy

1. EDA
2. data preprocess (data type, time code, date)
3. feature engineering (fill NA, remove outlier)
4. model selection (random forest, GBM, XGBoost, LightGBM)
5. grid search

my notebook

1. EDA by R (data.table, ggplot2)

2. simple gbm model by h2o

Score

0.53353 (NO. 26)

Post-match review

1. train and test dataset are not similar, so it's hard to find the right direction to improve accuracy
2. Lack of sensitivity for processing time series data, missing a lot of information
3. don't get right way to deal with NA
4. don't really understead the heperparameter in model, so can't grid search well
5. That's the best score of all shopee league competition, good job!