wantedly / machine-learning-round-table

Gather around the table, and have a discussion to catch up the latest trend of machine learning 🤖
https://www.wantedly.com/projects/391912
305 stars 2 forks source link

[2022/12/07]Machine Learning 輪講 #174

Open hakubishin3 opened 1 year ago

hakubishin3 commented 1 year ago

Why

Machine Learning 輪講は最新の技術や論文を追うことで、エンジニアが「技術で解決できること」のレベルをあげていくことを目的にした会です。

prev. https://github.com/wantedly/machine-learning-round-table/issues/172

What

話したいことがある人はここにコメントしましょう! 面白いものを見つけた時点でとりあえず話すという宣言だけでもしましょう!

hakubishin3 commented 1 year ago

Generative Adversarial Framework for Cold-Start Item Recommendation

Generation Adversarial Recommendation: GARというフレームワークを提案。アイテムのコンテンツ情報からembeddingを生成する機構と推薦機構を敵対的に訓練することで、cold itemの質の高いembeddingを生成器から獲得して推薦で利用する。

スクリーンショット 2022-12-07 3 03 21

nogawanogawa commented 1 year ago

Modeling Two-Way Selection Preference for Person-Job

論文URL

https://arxiv.org/abs/2208.08612

著者

Chen Yang, Yupeng Hou, Yang Song, Tao Zhang, Ji-Rong Wen, Wayne Xin Zhao

会議

RecSys '22

背景

person-job fitは採用マッチングプラットフォームにおける中核技術となっている。

相互推薦や職務推薦、job-oriented skill measurementなど、人と仕事に関する研究は様々行われているが、既存の研究は1方向の選択プロセス、または、全体のマッチングのモデリングに焦点を当てている。しかし、本質的に採用は双方向の期待に答えることが重要と考えられる。

目的

人と仕事の適合性の向上

アプローチ

image

実装

https://github.com/RUCAIBox/DPGNN

memo

https://github.com/nogawanogawa/paper_memo/issues/68

Hayashi-Yudai commented 1 year ago

Towards Reliable Item Sampling for Recommendation Evaluation

推薦システムを評価するときに一部をサンプリングして計算した評価値はデータ全体で計算した評価値とずれるという問題があった。各ユーザーについて全てのアイテムを使って評価値を出すのが可能であることは少ないので、サンプリングして得られた評価値をグローバル評価値に近づける方策がほしい。

image image

\~10 Kのアイテムを含むデータセットから各ユーザーについて100アイテムをサンプリングしてランキングを作成して評価。全てのアイテムを使ってランキングを作成して評価した場合と比較。

image

提案手法において最も、サンプリングデータでの評価値がデータセット全体での評価値に近いという結果が得られた。

zerebom commented 1 year ago

The Difference Between a Click and a Cart-Add: Learning Interaction-Specific Embeddings