# 交友軟體的推薦系統 ###### tags: `MLG` 簡報網址:hackmd.io/@kuouu/MLG_final_proposal --- ## Motivation ---- ### AInimal 人工智慧社群養成 ![](https://i.imgur.com/6SWft3N.jpg) ---- ### 相關連結 - [網頁版](https://ainimal.io)(需調整瀏覽器尺寸至手機比例 - [下載連結](http://onelink.to/kfp7jm) - [FB 粉專](https://www.facebook.com/AInimal) - [IG 主頁](https://www.instagram.com/ainimal_official/) - [其他連結](https://joy.link/ainimal) ---- ### 現況 ---- #### 目前配對機制 ![](https://i.imgur.com/5yy1tik.png =200x) 隨機選擇一位"符合篩選條件"和"在當天有上線"的人 ---- #### 改良--基於人個特質的聊天活絡程度預測 [Google Colab](https://colab.research.google.com/drive/1V33OHXKVuFGInMpeTwKCzmjxGJMZwjyt?usp=sharing) ![](https://i.imgur.com/J21NA2j.png) ---- ### 正確率 ``` train accuracy: 0.78 test accuracy: 0.62 ``` 可能原因: - 模型太過簡單 - 人格特質不適合作為聊天活絡程度的評斷標準 --- ## Problem Statement ---- 希望軟體內能有個推薦系統讓配對達到以下要求: 1. 會回訊息 2. 聊天熱絡 3. 交換其他通訊軟體 ---- 其他期望: - 讓沒有在用的人回來使用軟體 - 在使用者數量增加之後依然適用 ---- ### Input/Output Input: `user`, `chat`, `friend` Output: list of recommended users with ranking --- ## Technical Challenges ---- 1. 分析哪些屬性值得使用 2. 配對成功的評斷標準 3. 資料篩選 --- ## Related Work ---- - [Recommender System for Online Dating Service](https://arxiv.org/pdf/cs/0703042.pdf) 2007 / 162 cites - [RECON: A reciprocal recommender for online dating](https://www.researchgate.net/publication/221140972_RECON_A_reciprocal_recommender_for_online_dating) 2010 / 184 cites - [Recommender Systems for Online Dating](https://core.ac.uk/download/pdf/33736431.pdf) 2015 / 6 cites - [Design of Reciprocal Recommendation Systems for Online Dating](http://web.cs.ucla.edu/~yzsun/papers/snam2016.pdf) 2016 / 24 cites - [Shedding More Light on How Instagram Works](https://about.instagram.com/blog/announcements/shedding-more-light-on-how-instagram-works?fbclid=IwAR0L5kk9JEu3dEmz1VqIETTOevSWoc0IcZceaGTS9Vj_m0LccUGuJGsu_T0) --- ## Dataset ---- | 資料名稱 | 數量 | | -------- | -------- | | user | 5043 | | chat | 290559 | | friend | 22683 | ---- ### user - birth - gender - create_at - ... 可能有些資料必須處理過後取得 ---- ### chat - message - send_from, send_to - time - type (文字、貼圖) ---- ### friend - user_1, user_2 - last_message - created_at - updated_at - is_new - ... --- ## Preliminary Methods ---- 1. rule base 2. CF 3. GNN --- ## Evaluation Plans ---- 與 HW3 相同 - RMSE - Recall - NDCG --- ## Connections to MLG ---- 所有的配對 `friend` 可以看成是一個 Graph ,並且 `chat` 能當作是邊上的權重, `user` 則是 node feature --- ## Expected Time Schedule ---- ```mermaid gantt title Model development axisFormat %m%d section development research: 2021-06-01, 5d data processing: 4d programing : 3d document : 2d ``` ---- ```mermaid gantt title Model development axisFormat %m%d section development research: 2021-06-01, 5d data processing: 4d programing : 3d document : 2d section deploy discussion: 2021-07-01, 7d research: 14d deploy : 7d modify : 7d ``` --- ## End
{"metaMigratedAt":"2023-06-16T00:53:19.466Z","metaMigratedFrom":"Content","title":"交友軟體的推薦系統","breaks":true,"contributors":"[{\"id\":\"35cea605-1599-4b9b-a7f8-610e7a3b3107\",\"add\":5380,\"del\":2428}]"}
    92 views