Open hakubishin3 opened 1 year ago
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。
https://arxiv.org/abs/2208.08612
Chen Yang, Yupeng Hou, Yang Song, Tao Zhang, Ji-Rong Wen, Wayne Xin Zhao
RecSys '22
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。
人と仕事の適合性の向上
https://github.com/RUCAIBox/DPGNN
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。
\~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。
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。
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!