Sequential Recommendation Datasets

Provide a tool for helping 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
• Retailrocket
• TaFeng
• Taobao
• Tmall
• Yelp

Install this tool

pip install -U srdatasets —-user

Download datasets

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

srdatasets download --dataset=[dataset_name]

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

srdatasets info

Process datasets

The generic processing command is

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

Splitting options

Two dataset splitting methods are provided: user-based and time-based. User-based means that splitting is executed on every user behavior sequence given the ratio of validation set and test set, while time-based means that 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)
--test-split   Proportion of test set to full dataset (default: 0.2)
--dev-split    Proportion of validation set to full training set (default: 0.1)

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 and 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 (default: 5)
--target-len        Number of target items (default: 1)
--pre-sessions      Number of previous sessions (default: 10)
--pick-targets      Randomly or lastly pick items from current session (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. The config id shown in output is a required argument of DataLoader.

srdatasets info --dataset=[dataset_name]

DataLoader

DataLoader is a built-in class that makes loading processed datasets easy. Practically, once initialized a dataloder by passing the dataset name, processed version (config id), batch_size and a flag to load training data or test data, you can then loop it to get batch data. Considering that some models use rank-based learning, negative sampling is intergrated into DataLoader. The negatives are sampled from all items except items in current data according to popularity. By default it (negatives_per_target) is turned off. Also, the time of user behaviors is sometimes an important feature, you can include it into batch data by setting include_timestmap to True.

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)

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:
        # Shape
        #   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)
        #
        # DataType
        #   numpy.ndarray or torch.LongTensor
        pass

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

For long and 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:
        # Shape
        #   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)
        #
        # DataType
        #   numpy.ndarray or torch.LongTensor
        pass

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

Disclaimers

This repo does not host or distribute any of the datasets, it is your responsibility to determine whether you have permission to use the dataset under the dataset's license.