#1 - 2017-8-11 21:31
epix ((((;゚Д゚)))))))
这个工具是针对单个用户推荐动画
即根据大家在bgm上的打分和你在bgm上的打分, 来猜测你可能会给哪些你没看过的动画打高分

使用方法:
点击右上角头像, 打开你的个人页面, 比如我的话就是 https://bgm.tv/user/epix
把域名 bgm.tv 替换为 bgm.exz.me 并回车, 即打开网址为 https://bgm.exz.me/user/epix 的页面
此页面即是推荐的动画列表, 按推荐程度从高到低排列, 100部

由于是通过对你已有的打分数据进行推测的, 所以如果你在bgm上很少打分, 则效果很差(或者根本没有).
你已经添加到想看/在看/看过/搁置/抛弃的动画将不会出现在列表中

其他接口: https://bgm.exz.me/api/epix 返回JSON格式数据

==========

使用数据: 2017年8月之前, 所有在bgm有排名的动画的标记为看过并打分的记录, 约2.38M. 动画4.2k, 用户36k.

使用算法: SVD [1]
制作人员: @epix  @剪下回忆 @Andy

未来计划:
使用改进的算法(SVD feature[2], FFM[5], etc.)
增加其他因素(时间, tag, staff)
提供其他信息(用户平均打分/用户打分变化趋势/用户相似度/条目相似度, etc.)
拓展数据来源(其他类似网站)
定时刷新数据

源码稍后整理开源
数据私聊

引用文献
[1] Koren Y, Bell R, Volinsky C. Matrix factorization techniques for recommender systems[J]. Computer, 2009, 42(8).
[2] Chen T, Zhang W, Lu Q, et al. Svdfeature: a toolkit for feature-based collaborative filtering[J]. Journal of Machine Learning Research, 2012, 13(Dec): 3619-3622.
[5] Yan P, Zhou X, Duan Y. E-Commerce item recommendation based on field-aware factorization machine[C]//Proceedings of the 2015 International ACM Recommender Systems Challenge. ACM, 2015: 2.
#2 - 2017-8-11 21:37
谢谢~~
#3 - 2017-8-11 21:37
看到 poi 了(bgm38)
#4 - 2017-8-11 21:56
(Enjoy your (real) life!)
Looks interesting! May I know how is your train/valid/test set splitted? How did you perform cv? What's your RMSE on test set?
#4-1 - 2017-8-11 22:39
剪下回忆
The training is conducted via 3-fold cross validation. The best RMSE achieved was 0.96459. It should be noted that this system is designed for users that have watched many animes and made enough comments. In that case, test set is not necessary as far as we know.
#4-2 - 2017-8-11 23:59
Genius、小乖
剪下回忆 说: The training is conducted via 3-fold cross validation. The best RMSE achieved was 0.96459. It should...
Well, I know collaborative filtering well and understand your concern, but what I want to address here is CORRECT CV. Traditionally, when using SVD, people would randomly select some entires in a matrix as test set -- at least in one fold -- but this should be avoided. In your task, you are to predict what anime a user is likely to watch later, so if I were asked to do this prediction, I would sort user's watch record by time, and mask those entries after a certain date. Besides, you predict a user's preference based on user's historical data, which means user's data should be observed fully. Alternatively speaking, this requires one to partition data by user during cv.

I think they are very trivial stuff and you may have already considered that, but many people also neglect these things. Feel free to neglect me if you have already considered the points I addressed above.🙂
#4-3 - 2017-8-12 00:55
剪下回忆
Genius、小乖 说: Well, I know collaborative filtering well and understand your concern, but what I want to address he...
Thanks for your generous suggestion! Actually we are new to recommendation and don’t have much knowledge on it. I think you mean we should split every user’s data to be two parts: before and after a certain date, during the cross-validation. Then that may indicate you suggest us to consider the time component. I should say for now we haven’t taken that into consideration. I personally think Svdfeature[2] may be able to solve this problem, will it?
#4-4 - 2017-8-12 09:16
Genius、小乖
剪下回忆 说: Thanks for your generous suggestion! Actually we are new to recommendation and don’t have much knowl...
I studied SVDfeature a little bit before, but this is somewhat outdated from my current view point. I suggest you to use Factorization Machine or Field-aware Factorization Machine. They originates from the core idea of SVD but jumps out of the scope of matrix factorization, using a pure supervised learning method to learn parameters. I'm very sure they will offer you a higher level view on "how to build a recommender".
#4-5 - 2017-8-12 13:38
剪下回忆
Genius、小乖 说: I studied SVDfeature a little bit before, but this is somewhat outdated from my current view point. ...
I have read the paper of FFM briefly. It’s really great and exciting! We decided to take your advice. However, I found that FM has a similar performance on MovieLens 1M with SVDfeature[3][4]. Can you give us some insights of the advantages of FFM?
[3] Yamada M, Lian W, Goyal A, et al. Convex Factorization Machine for Regression[J]. arXiv preprint arXiv:1507.01073, 2015.
[4] Strub F, Gaudel R, Mary J. Hybrid recommender system based on autoencoders[C]//Proceedings of the 1st Workshop on Deep Learning for Recommender Systems. ACM, 2016: 11-16.
#4-6 - 2017-8-12 18:02
Genius、小乖
剪下回忆 说: I have read the paper of FFM briefly. It’s really great and exciting! We decided to take your advice...
I suggest you to read the paper about FM then FFM, because FFM is based on FM and introduces the concept of field in features to capture different characteristics of features.
#4-7 - 2017-8-12 19:59
剪下回忆
Genius、小乖 说: I suggest you to read the paper about FM then FFM, because FFM is based on FM and introduces the con...
OK, I've read papers about FM[6] and FFM[7] and know the advantages. We will update the system using FFM later. Thanks again!
[6] Rendle S. Factorization machines[C]//Data Mining (ICDM), 2010 IEEE 10th International Conference on. IEEE, 2010: 995-1000.
[7] Juan Y, Zhuang Y, Chin W S, et al. Field-aware factorization machines for CTR prediction[C]//Proceedings of the 10th ACM Conference on Recommender Systems. ACM, 2016: 43-50.
#5 - 2017-8-11 22:38
(DD集まれ! (๑•̀ㅂ•́)و✧)
这个不错
假如能根据收藏的人(主要是声优)来推荐就好了
因为别人推荐的话我只参考作品里出演的声优(bgm93)
#5-1 - 2017-8-11 22:42
剪下回忆
考虑到收藏不一定人人会做,之后会学习用户关注的作品里的tag来辅助推荐。
#5-2 - 2017-8-11 22:44
NevaR
剪下回忆 说: 考虑到收藏不一定人人会做,之后会学习用户关注的作品里的tag来辅助推荐。
哦哦 很厉害的样子
#5-3 - 2017-8-11 22:50
epix
或许可以做个高级点的搜索来部分实现?
像图站那样"+声优:茅野爱衣 -脚本:大河内一楼 sort:rating"
#5-4 - 2017-8-12 12:33
Neko_Aria
epix 说: 或许可以做个高级点的搜索来部分实现?
像图站那样"+声优:茅野爱衣 -脚本:大河内一楼 sort:rating"
-脚本:大河内一楼。。这个可以有。(bgm38)
#6 - 2017-8-11 23:12
(没什么想说的)
又一个(又?(bgm38)
#7 - 2017-8-11 23:20
看标题还以为是这个的续作(
感觉对我而言更有用的用法是看自己or其他人哪些片没点过(bgm38)
果然对纯酱无效(bgm38)
#8 - 2017-8-12 00:21
(24歳、学生です。)
有一定的参考价值 不错 mark了
#9 - 2017-8-12 00:23
看了下推荐效果似乎跟我以前跑着玩的那个差不多..其实像这种总共条目数量也不多,主流喜好又比较统一的情感,我感觉直接根据tag来效果会更好
#10 - 2017-8-12 01:15
emmm,可以参考着补番
#11 - 2017-8-12 01:17
很强
#12 - 2017-8-12 10:35
(ちょっと..異世界旅行中)
书签+1
#13 - 2017-8-12 11:39
厉害了
#14 - 2017-8-12 11:54
(大破的那珂酱好萌!)
命中率够高,期待不断更新。
#15 - 2017-8-12 14:05
(Zu jeder Zeit, an jeder Ort, bleibt das Tun, der M ...)
这个准确率可以啊,居然还有之前忘记标记的(bgm38)
#16 - 2017-8-12 15:01
(さくら咲く〜君に会いたい)
面白
#17 - 2017-8-12 15:25
看了一下,挺合理的,非常感谢。
#18 - 2017-8-12 16:33
(透心凉 心飞扬)
这排序纯看分数的吧,好蠢,基片给排第一
#19 - 2017-8-12 20:32
(已滚)
http://bgm.exz.me/user/asm13177806
纯酱不显示 这就很爆笑了(bgm38)
#20 - 2017-8-12 20:39
哇 一直想看自己哪些片还没点,真是实用啊
#21 - 2017-8-12 21:21
(原来bgm有这么多人)
厉害
#22 - 2017-8-12 21:28
(愿意给我5分钟的编辑时间吗?)
很懷疑自己的結果,但畢竟都是沒看過的作品(bgm38)

天元突破相當突破天際的拿到了獨自一檔的9.71真是太神奇了......(排名第二9.08)
#22-1 - 2017-8-12 21:31
剪下回忆
天元突破好喜欢的=v=浪漫得不行
#22-2 - 2017-8-12 21:34
小T
剪下回忆 说: 天元突破好喜欢的=v=浪漫得不行
仍在研究你做的工具的功能,第一反應是感覺list不同於預期......推薦的打分感覺很有趣哈哈


不好意思回錯人啦!!!我也很期待哪天能去看天元~
#23 - 2017-8-12 22:18
(时光机已补全)
偏爱校园日常,但最近看了几部之前不会看的番,比如一拳超人、星际牛仔。只能说是不好意思打低分,这样的风格真的喜欢不起来。
所以看到这个结果真是一脸懵逼(bgm38)
综合tag的分析可能会好一点,虽然bgm的tag太随意
#23-1 - 2017-8-13 13:05
小T
樣本太少,不容易準
#24 - 2017-8-12 22:31
(崩坏の节操)
不行啊,我要怀疑人生了(bgm38)感觉自己应该很少很少看校园,日常类的,还中出了一部里番
https://bgm.exz.me/user/soviet
#24-1 - 2017-8-12 22:40
神之迷徒
好吧,我想看在看点了600多格
#24-2 - 2017-8-13 13:04
小T
神之迷徒 说: 好吧,我想看在看点了600多格
想看被忽略嗎?
#25 - 2017-8-12 23:08
emmmm,有一点最好考虑进去,就是抛弃和搁置list,比如我因为不喜欢虫师的节奏抛弃了,但是还是出现在了推荐上
我看错了,虫师一期没有出现在列表上,但是二期等等还是有的
#26 - 2017-8-12 23:26
没几个想看的感觉.
#27 - 2017-8-12 23:41
感觉大部分都是排行榜前几页里我没标的…我觉得可以把想看的也算进去,像我这样把排行榜前几页都点想看的人应该也有…
#28 - 2017-8-13 00:06
(请不要在未经许可时在本站外引用我在本站的留言 ... . ...)
em...按评分来推荐的结果就是推了一堆很NB但我还没力气去补的作品
上来就催我补CB,然后是攻壳,银英 etc...(bgm38)
#28-1 - 2017-8-13 00:11
#29 - 2017-8-13 15:30
(Death is the gateway to Eternity, my dear.)
好像给我推荐了很多美漫,比如马男2……
#30 - 2017-8-13 16:13
(愿意给我5分钟的编辑时间吗?)
如果數據能有更新的話,倒是願意調整一下自己的條目收藏情況(特別是想看)來玩玩看(bgm37)
#30-1 - 2017-8-13 18:30
剪下回忆
更新是应该会更新的,毕竟我自己的都没标注多少(,以后还想搞在线学习但是没啥头绪。。。。
#31 - 2017-8-13 19:26
(思い通りにならない現実なんて、つまらないゲームと同 ...)

什么鬼。。。我一点都不基的啊(bgm38)
#31-1 - 2017-8-14 14:07
剪下回忆
下次更新会把tag偏好纳入考虑范围=v=估计就不会这样了
#32 - 2017-8-14 14:15
(媛娇系是检验大法的唯一标准)
发现不少忘记标的了(bgm38)
#33 - 2017-8-14 14:19
(about:blank)
建议加上抛弃的动画。。。感觉给我的推荐完全不对口味(bgm38)
#34 - 2017-8-14 15:58
(この勝利を、近所のおばさんに捧げる! ... ... ... ... ...)
给我推荐全是硬核,明明我已经萌豚了多年。。
刚从满眼deep的ICML回来,看到lz用的算法觉得,真好(bgm38)(bgm38)(bgm38)
#34-1 - 2017-8-14 16:01
剪下回忆
这个也有deep的算法,不过看过感觉纯拼数字,不一定好,算力要求高,能考虑的因素不够多。