srdatasets 0.0.2

A collection of research ready datasets for sequential recommendation

Sequential Recommendation Datasets

Provide a tool for help dealing with some common sequential recommendation datasets

Datasets

• Amazon-Books
• Amazon-Electronics
• Amazon-Movies
• Amazon-CDs
• Amazon-Clothing
• Amazon-Home
• Amazon-Kindle
• Amazon-Sports
• Amazon-Phones
• Amazon-Health
• Amazon-Toys
• Amazon-VideoGames
• Amazon-Tools
• Amazon-Beauty
• Amazon-Apps
• Amazon-Office
• Amazon-Pet
• Amazon-Automotive
• Amazon-Grocery
• Amazon-Patio
• Amazon-Baby
• Amazon-Music
• Amazon-MusicalInstruments
• Amazon-InstantVideo
• CiteULike
• FourSquare-NYC
• FourSquare-Tokyo
• Gowalla
• Lastfm1K
• MovieLens20M
• TaFeng
• Taobao
• Tmall
• Yelp

Install this tool

pip install srdatasets

Download datasets

Run the command below to download datasets. Note, since some datasets are not directly accessible, you'll be warned then to download them manually and place them somewhere it tells you.

python -m srdatasets download --dataset=[dataset_name]

To get a view of downloaded and processed status of all datasets, run

python -m srdatasets info

Process datasets

The generic processing command is

python -m srdatasets process --dataset=[dataset_name] [--options]

Splitting options

Two dataset splitting methods are provided: user-based and time-based. User-based means splitting is executed on every user hehavior sequence given the ratio of validation set and test set while time-based means splitting is based on the date of user behaviors. After splitting some dataset, two processed datasets are generated, one for development, which uses the validation set as the test set, the other for test, which contains the full training set.

• --split-by: user or time (default: user)
• --dev-split: proportion of validation set to full training set (default: 0.1)
• --test-split: proportion of test set to full dataset (default: 0.2)

NOTE: time-based splitting need you to manually input days at console by tipping you total days of that dataset, since you may not know.

Task related options

For short term recommnedation task, you use previous input-len items to predict next target-len items. To make user interests more focused, user behavior sequences can also be cut into multiple sessions if session-interval is given. If the number of previous items is smaller than input-len, 0 is padded to the left.

For long-short term recommendation task, you use pre-sessions previous sessions and current session to predict target-len items. The target items are picked randomly or lastly from current session. So the length of current session is max-session-len - target-len while the length of any previous session is max-session-len. If any previous session or current session is shorter than the preset length, 0 is padded to the left.

• --task: short or long-short (default: short)
• --input-len: number of previous items (for short term task) (default: 5)
• --target-len: number of target items (default: 1)
• --pre-sessions: number of previous sessions (for long-short term task) (default: 10)
• --pick-targets: randomly or lastly pick items from current session (for long-short term task) (default: random)
• --session-interval: session splitting interval (minutes) (default: 0)
• --min-session-len: sessions less than this in length will be dropped (default: 2)
• --max-session-len: sessions greater than this in length will be cut (default: 20)

Common options

• --min-freq-item: items less than this in frequency will be dropped (default: 5)
• --min-freq-user: users less than this in frequency will be dropped (default: 5)
• --no-augment: do not use data augmentation (default: False)
• --remove-duplicates: remove duplicated items in user sequence or user session (if splitted) (default: False)

Dataset related options

• --rating-threshold: interactions with rating less than this will be dropped (Amazon, Movielens, Yelp) (default: 4)
• --item-type: recommend artists or songs (Lastfm) (default: song)

Version

By using different options, a dataset will have many processed versions. You can run the command below to get configurations and statistics of all processed versions of some dataset

python -m srdatasets info --dataset=[dataset_name]

DataLoader

DataLoader is a built-in class that helps to load the processed datasets

Arguments

• dataset_name : dataset name (case insensitive)
• config_id : configuration id
• batch_size: batch size (default: 1)
• train: load training dataset (default: True)
• development: load the dataset aiming for development (default: False)
• negatives_per_target: number of negative samples per target (default: 0)
• include_timestamp: add timestamps to batch data (default: False)
• drop_last: drop last incomplete batch (default: False)

Initialization example

from srdatasets.dataloader import DataLoader

trainloader = DataLoader("amazon-books", "c1574673118829", batch_size=32, Train=True, negatives_per_target=5, include_timestamp=True)
testloader = DataLoader("amazon-books", "c1574673118829", batch_size=32, Train=False, include_timestamp=True)

For pytorch users, there is a wrapper implementation of torch.utils.data.DataLoader, you can then set keyword arguments like num_workers and pin_memory to speed up loading data

from srdatasets.dataloader_pytorch import DataLoader

trainloader = DataLoader("amazon-books", "c1574673118829", batch_size=32, Train=True, negatives_per_target=5, include_timestamp=True, num_workers=8, pin_memory=True)
testloader = DataLoader("amazon-books", "c1574673118829", batch_size=32, Train=False, include_timestamp=True, num_workers=8, pin_memory=True)

NOTE: Set negatives_per_target or include_timestamp only when your model needs

Iteration template

For short term recommendation task

for epoch in range(10):
    # Train
    for users, input_items, target_items, input_item_timestamps, target_item_timestamps, negative_samples in trainloader:
        # Shapes
        #   users: (batch_size,)
        #   input_items: (batch_size, input_len)
        #   target_items: (batch_size, target_len)
        #   input_item_timestamps: (batch_size, input_len)
        #   target_item_timestamps: (batch_size, target_len)
        #   negative_samples: (batch_size, target_len, negatives_per_target)
        pass

    # Evaluate
    for users, input_items, target_items, input_item_timestamps, target_item_timestamps in testloader:
        pass

For long-short term recommendation task

for epoch in range(10):
    # Train
    for users, pre_sessions_items, cur_session_items, target_items, pre_sessions_item_timestamps, cur_session_item_timestamps, target_item_timestamps, negative_samples in trainloader:
        # Shapes
        #   users: (batch_size,)
        #   pre_sessions_items: (batch_size, pre_sessions * max_session_len)
        #   cur_session_items: (batch_size, max_session_len - target_len)
        #   target_items: (batch_size, target_len)
        #   pre_sessions_item_timestamps: (batch_size, pre_sessions * max_session_len)
        #   cur_session_item_timestamps: (batch_size, max_session_len - target_len)
        #   target_item_timestamps: (batch_size, target_len)
        #   negative_samples: (batch_size, target_len, negatives_per_target)
        pass

    # Evaluate
    for users, pre_sessions_items, cur_session_items, target_items, pre_sessions_item_timestamps, cur_session_item_timestamps, target_item_timestamps in testloader:
        pass

NOTE: The default DataLoader use numpy.Array to represent data while the pytorch version use torch.LongTensor

Disclaimers

The datasets have their own licenses, this repo only provides a way to use them.