Thanks my teammate @Ryan, we got 16th place. Thanks to H&M and Kaggle. This is a wonderful competition.

Framework

Data-splitting

We split the data into 3 groups: cg1(customer group 1) is the users with transactions in last 30 days; cg2(customer group 2) is the users without transactions in last 30 days, but with transactions in history. cg3(customer group 3) is the users without transactions in history.

• cg1 and cg2: multi-recalls + rank.
• cg3: popular items recall.

Recalls

• popular items recall
• repurchase recall
• binaryNet recall
• ItemCF recall
• UserCF recall
• W2V content recall
• NLP content recall
• Image content recall
• Category content recall

Each recall method will recall 100 items for every user, then drop duplicates.

Rank

Feature Engineer

Item Feature

groupby article_id agg cols calculate statistics

• cols: customer_id, price, sales_channel_id, and so on.
• op: 'min', 'max', 'mean', 'std', 'median', 'sum', 'nunique'

User Feature

groupby customer_id agg cols calculate statistics

• cols: price, article_id, sales_channel_id, and so on.
• op: 'min', 'max', 'mean', 'std', 'median', 'sum', 'nunique'

Interaction Feature

• count of user-item purchased in different window(1day, 3days, 1week, 2weeks, 1month).
• The time-diff since the user last purchased the item

other features

• ItemCF score
• BinaryNet score

Model

• lightgbm ranker
• lightgbm binary

Ensemble

Ref to this link