Recommenders: Deep Learning for News Recs

[nikhilrj]

Deep Learning Based Recommender for News
Model: DKN
Pipeline: Training on DLVM+GPU via AML, O16N on AKS+GPU
Dataset: News (includes user or item features)

[miguelgfierro]

TODO:

• Find a knowledge graph
• After the knowledge graph, try to find a dataset that has text (maybe not news) and matches knowledge graph

[anargyri]

I'd like to start a discussion about what are feasible scenarios with data in the public domain, that can be implemented. The objective is to have a scenario using DKN, maybe we could be flexible whether the data are about news or other types of text.

@Leavingseason do you think we could repeat the setup from this paper?
https://www.kdd.org/kdd2016/papers/files/adf0066-zhangA.pdf
The first data used is Movielens + Satori; is the Satori data publicly available? if not could we replace Satori with a public knowledge graph?
Is the second data set (IntentBooks) publicly available?

It looks like the first approach (Movielens + DBpedia) has been followed already in this paper
http://giusepperizzo.github.io/publications/Palumbo_Rizzo-RecSys2017.pdf

Instead of news, we could use other text associated to the movies.

[anargyri]

Jester is another text recommendations data set
http://eigentaste.berkeley.edu/dataset/
Amazon reviews
http://jmcauley.ucsd.edu/data/amazon/

[gramhagen]

https://github.com/zygmuntz/goodbooks-10k is an interesting dataset, it would be nice to scrape text / images from each book to be able to leverage content features

[anargyri]

https://github.com/zygmuntz/goodbooks-10k is an interesting dataset, it would be nice to scrape text / images from each book to be able to leverage content features

There may be IP issues with text / images from the book that may be difficult to sort out.

[almudenasanz]

Possible Knowledge Graphs:

• YAGO (wordNet + wikipedia): https://www.mpi-inf.mpg.de/departments/databases-and-information-systems/research/yago-naga/yago/
• Unigraph: https://github.com/unigraph/docs/wiki/Datadump
• DBPedia: https://wiki.dbpedia.org (Can't used due to license)
• Wikidata: https://www.wikidata.org/wiki/Wikidata:Main_Page (CAN USE, approved by CELA)
• Google Knowledge Graph: https://developers.google.com/knowledge-graph/
• Freebase (deprecated): https://developers.google.com/freebase/

News datasets:

• Webhose: https://webhose.io/free-datasets/ (link from @miguelgfierro)
• Yahoo News: https://webscope.sandbox.yahoo.com/catalog.php?datatype=r
• This is an interesting approach, building their own dataset from twitter info: https://github.com/huangy22/NewsRecommender
• To obtain just a list of news and titles, we could use MediaCloud's API: https://mediacloud.org

[almudenasanz]

I think this other dataset could be used for news recommendation: https://www.kaggle.com/gspmoreira/articles-sharing-reading-from-cit-deskdrop

[anargyri]

I think this other dataset could be used for news recommendation: https://www.kaggle.com/gspmoreira/articles-sharing-reading-from-cit-deskdrop

Do they have a link to the data outside Kaggle?

[anargyri]

A github I saw related to knowledge graphs
https://github.com/SANSA-Stack

[miguelgfierro]

today, after a meeting with Jianxun, we discussed that we can create something similar to a news dataset, by merging Movielens with the text summary of the movie.

cc @almudenasanz @leavingseason

[miguelgfierro]

MSRA added 4 methods