Awesome

Automatic Continuation of User-Generated Item Lists

Introduction

This is the implementation of the paper as well as datasets and their splitting used in the paper:

Consistency-Aware Recommendation for User-Generated ItemList Continuation. WSDM, 2020. Yun He, Yin Zhang, Weiwen Liu and James Caverlee. Department of Computer Science and Engineering, Texas A&M University. Contact: yunhe@tamu.edu Personal Page: http://people.tamu.edu/~yunhe/

User-generated item lists are popular on many platforms. Examples include song-based playlists on Spotify, image-based lists (or“boards”) on Pinterest, book-based lists on Goodreads, and answer-based lists on question-answer forums like Zhihu. In these platforms, user-generated item lists are manually created, curated, and managed by users. Typically, users must first identify candidate items, determine if they are a good fit for a list, add them to a list, and then potentially provide ongoing updates to the list (e.g., by adding or deleting items over time). To accelerate this process and assist users to explore more related items for their lists, we study the important yet challenging problem of user-generated item list continuation. That is, how can we recommend items that are related to the list and fit the user’s preferences?

We also have another very related work in CIKM 2019, which recommends user-generated item lists to right users:

A Hierarchical Self-Attentive Model for Recommending User-Generated Item Lists. CIKM, 2019. Yun He, Jianling Wang, Wei Niu and James Caverlee. Department of Computer Science and Engineering, Texas A&M University. Link: https://github.com/heyunh2015/AttList

Usage of CAR (WSDM 2020)

We propose a novel model CAR to predict the next item that a user will possibly add into a list.

Input

The input of CAR includes: (1) the containing relationship between lists and items; (2) the creating relationship between users and lists.

(1) The containing relationship between lists and items.

listId itemId

which means that those items are curated in this list. These interactions are stored in \data folder, like AotM.txt.zip (unzip this file when you run our code).

Note that the last item is regarded as test data, the item before the last item is regarded as validation data and the rest of items are treated as training data.

(2) The containing relationship between lists and items.

key: listId, value: userId

which means that the list is created by this user. These creating relationships are stored in \data folder, like AotM_creator_list.dict (python dictionary object).

Output

The output are the evaluation results comparing the ranked items for each list from CAR and the groundtruth (e.g., the last item of AotM.txt). For each test item, we sample 100 negative items. CAR predicts scores for the 100 negative items and 1 test item, and rank these 101 items based on their scores. The 100 negative items for each list are stored in \data folder, like AotMListItems_len5_item5_cut1000.negativeEvalTest.zip (unzip this file when you run our code).

Run

python main.py --dataset Zhihu --train_dir test

For more hyper-parameters, please see main.py.

Datasets

Except to a public dataset AOTM-2011 (https://bmcfee.github.io/data/aotm2011.html), we also self-collected three datasets (in data folder): playlists from Spotify, book lists from Goodreads and answer collections from Zhihu, providing a wide range of varying item types for evaluating related methods like automatic playlists continuation. You are welcome to use our datasets and cite our paper.

Citation

TBD

Acknowledgement

The technique of self-attention in this paper is learned from https://github.com/kang205/SASRec. Th first author is Wang-Cheng Kang. Thanks to them! And my labmate Jiangling Wang (http://people.tamu.edu/~jwang713/) helps me to construct the playlist dataset from Spotify, thanks!