Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

DKN: Deep Knowledge-Aware Network for News Recommendation, Wang+, WWW'18 #363

Open
AkihikoWatanabe opened this issue Jun 1, 2021 · 6 comments

Comments

@AkihikoWatanabe
Copy link
Owner

https://arxiv.org/pdf/1801.08284.pdf

@AkihikoWatanabe
Copy link
Owner Author

AkihikoWatanabe commented Jun 1, 2021

Overview

Contents-basedな手法でCTRを予測しNews推薦。newsのタイトルに含まれるentityをknowledge graphと紐づけて、情報をよりリッチにして活用する。
CNNでword-embeddingのみならず、entity embedding, contextual entity embedding(entityと関連するentity)をエンコードし、knowledge-awareなnewsのrepresentationを取得し予測する。
※ contextual entityは、entityのknowledge graph上でのneighborhoodに存在するentityのこと(neighborhoodの情報を活用することでdistinguishableでよりリッチな情報を活用できる)

CNNのinputを[[word_ embedding], [entity embedding], [contextual entity embedding]](画像のRGB)のように、multi-channelで構成し3次元のフィルタでconvolutionすることで、word, entity, contextual entityを表現する空間は別に保ちながら(同じ空間で表現するのは適切ではない)、wordとentityのalignmentがとれた状態でのrepresentationを獲得する。

image

Experiments

BingNewsのサーバログデータを利用して評価。
データは (timestamp, userid, news url, news title, click count (0=no click, 1=click))のレコードによって構成されている。
2016年11月16日〜2017年6月11日の間のデータからランダムサンプリングしtrainingデータセットとした。
また、2017年6月12日〜2017年8月11日までのデータをtestデータセットとした。

word/entity embeddingの次元は100, フィルタのサイズは1,2,3,4とした。loss functionはlog lossを利用し、Adamで学習した。

image

image

DeepFM超えを達成。
entity embedding, contextual entity embeddingをablationすると、AUCは2ポイントほど現象するが、それでもDeepFMよりは高い性能を示している。
また、attentionを抜くとAUCは1ポイントほど減少する。

1ユーザのtraining/testセットのサンプル
image

@AkihikoWatanabe
Copy link
Owner Author

AkihikoWatanabe commented Jun 1, 2021

#365 によって経験的にRNN, Recursive Neural Network等と比較して、sentenceのrepresentationを獲得する際にCNNが優れていることが示されているため、CNNでrepresentationを獲得することにした模様(footprint 7より)

@AkihikoWatanabe
Copy link
Owner Author

Factorization Machinesベースドな手法(LibFM, DeepFM)を利用する際は、TF-IDF featureと、averaged entity embeddingによって構成し、それをuser newsとcandidate news同士でconcatしてFeatureとして入力した模様

@AkihikoWatanabe
Copy link
Owner Author

content情報を一切利用せず、ユーザのimplicit feedbackデータ(news click)のみを利用するDMF(Deep Matrix Factorization)の性能がかなり悪いのもおもしろい。やはりuser-item-implicit feedbackデータのみだけでなく、コンテンツの情報を利用した方が強い。

@AkihikoWatanabe
Copy link
Owner Author

(おそらく)著者によるtensor-flowでの実装: https://github.com/hwwang55/DKN

@AkihikoWatanabe
Copy link
Owner Author

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

No branches or pull requests

1 participant