xlnetflixprize.blogspot.com
xlvector's solution of NetflixPrize: 五月 2009
http://xlnetflixprize.blogspot.com/2009_05_01_archive.html
Xlvector's solution of NetflixPrize. 感谢王元涛同学的帮助,今天实现了RBM模型,用50个hidden unit得到Probe上的RMSE=0.924。将这个结果融合进以前的预测器集合中,得到Quiz上的RMSE从0.8694改进到0.8688。 看了RBM确实是一大类方法,能够弥补其他方法的不足。目前我在实验更多hidden unit的RBM,已经用Gauss hidden unit 的RBM。 订阅: 帖子 (Atom). Hector :go 语言机器学习库. Just a guy in a garage. Supercomputing made super easy. Netflix Grand Prize technical presentation. Follow me on Twitter. 中关村, 北京, China.
minchechiu.blogspot.com
Michael's Blog: April 2009
http://minchechiu.blogspot.com/2009_04_01_archive.html
Michael(Min-Che, Chiu)'s personal blog, about programming, funny things . Wednesday, April 29, 2009. NAPOLEON DYNAMITE for Netflix Prize. If You Liked This, You’re Sure to Love That. Published: November 23, 2008, New York Times. Mathematically speaking, “Napoleon Dynamite” is a very significant problem for the Netflix Prize. If we can’t understand our own preferences, can computers really be any better at it? Legend('svd','useraverage','movieaverage');. Xlabel('User Id(not sequential)');. MIN EPOCHS = 60.
xlnetflixprize.blogspot.com
xlvector's solution of NetflixPrize: 关于评分分布的思考
http://xlnetflixprize.blogspot.com/2009/04/blog-post.html
Xlvector's solution of NetflixPrize. 对于一部电影,它被一堆人评分了,这些评分具有一些属性,均值,方差,偏差,等等。 目前我在设计一个新的模型,对于每一部电影,我们计算他的平均得分,记为m(i),那么对于一个用户-电影(u,i),. R(u,i) = mu bu bi dot(p[u], q[i]) * h[k]. 其中,k = m(i). 这个模型可以不断的变换,比如我们也可以令k = var(i),也就是说. Dot(p[u], q[i]) * h[k]. 表示了用户u对具有k属性的电影i的看法。 我仍然用梯度法训练这个模型。结果稍后公布。 Other than SVD, what other algorithms are you currently using? KNN, clustering, global effects. What does FLC stand for? FLC融合 you mentioned in March) Is this a special Chinese terminology? 订阅: 帖子评论 (Atom).
xlnetflixprize.blogspot.com
xlvector's solution of NetflixPrize: 四月 2009
http://xlnetflixprize.blogspot.com/2009_04_01_archive.html
Xlvector's solution of NetflixPrize. Http:/ todwang.blogspot.com/2009/04/technical-details-in-rbm-for-cf.html. Http:/ todwang.blogspot.com/2009/03/understanding-rbm-for-cf.html. 因为我用的计算机有4个核,为了充分利用计算机资源,今天将基于SVD的模型重新用多线程进行实现,速度大大提升。 SVD很容易并行化,我的策略是在扫描数据的循环上并行化处理,在修改模型时用锁进行互斥。 Clustering items and users by latent factors? By SVD model, we can calculate latent factors for users and items. p(u) is latent factor for u, while q(i) is latent factor for i. S(u,v) = f(p(u), p(v)? 其中,k = m(i). 在预测某个电影的评分时,...
xlnetflixprize.blogspot.com
xlvector's solution of NetflixPrize: 推荐两篇关于RBM的中文介绍
http://xlnetflixprize.blogspot.com/2009/04/rbm.html
Xlvector's solution of NetflixPrize. Http:/ todwang.blogspot.com/2009/04/technical-details-in-rbm-for-cf.html. Http:/ todwang.blogspot.com/2009/03/understanding-rbm-for-cf.html. 订阅: 帖子评论 (Atom). Hector :go 语言机器学习库. Just a guy in a garage. Supercomputing made super easy. Netflix Grand Prize technical presentation. Follow me on Twitter. 中关村, 北京, China. Clustering items and users by latent factors? An improved item-based KNN predictor.
xlnetflixprize.blogspot.com
xlvector's solution of NetflixPrize: clustering items and users by latent factors?
http://xlnetflixprize.blogspot.com/2009/04/clustering-items-and-users-by-latent.html
Xlvector's solution of NetflixPrize. Clustering items and users by latent factors? By SVD model, we can calculate latent factors for users and items. p(u) is latent factor for u, while q(i) is latent factor for i. Recently, I thought about calculate user similarity by latent factor,. S(u,v) = f(p(u), p(v)? I am testing this idea now, and I hope this idea can improve prediction accuracy. 订阅: 帖子评论 (Atom). Hector :go 语言机器学习库. Just a guy in a garage. Supercomputing made super easy. Follow me on Twitter.
xlnetflixprize.blogspot.com
xlvector's solution of NetflixPrize: An improved item-based KNN predictor
http://xlnetflixprize.blogspot.com/2009/04/improved-item-based-knn-predictor.html
Xlvector's solution of NetflixPrize. An improved item-based KNN predictor. Today, I revise the item-based KNN predictor and get RMSE = 0.8730 in quiz with other 39 predictors. The classical item-based kNN will firstly calculate similarity between item i and j by:. The, the rating r(u,i) will be predicted by:. However, I revise this predictor by:. This predictor can produce more accurate prediction by choosing adequate alpha. 订阅: 帖子评论 (Atom). Hector :go 语言机器学习库. Just a guy in a garage. Follow me on Twitter.
xlnetflixprize.blogspot.com
xlvector's solution of NetflixPrize: 程序并行化 multi-thread
http://xlnetflixprize.blogspot.com/2009/04/multi-thread.html
Xlvector's solution of NetflixPrize. 因为我用的计算机有4个核,为了充分利用计算机资源,今天将基于SVD的模型重新用多线程进行实现,速度大大提升。 SVD很容易并行化,我的策略是在扫描数据的循环上并行化处理,在修改模型时用锁进行互斥。 订阅: 帖子评论 (Atom). Hector :go 语言机器学习库. Just a guy in a garage. Supercomputing made super easy. Netflix Grand Prize technical presentation. Follow me on Twitter. 中关村, 北京, China. Clustering items and users by latent factors? An improved item-based KNN predictor.
xlnetflixprize.blogspot.com
xlvector's solution of NetflixPrize: 三月 2009
http://xlnetflixprize.blogspot.com/2009_03_01_archive.html
Xlvector's solution of NetflixPrize. 最近一直在写一篇关于如何使用时间信息的论文,准备投4月份的WI,今天将论文中的一些结果加入到以前的预测器集合中,RMSE从 0.8770. 12290;以前的预测器集合中已经有了很多考虑时间信息的预测器,不过这个这个预测器效果是最好的。 根据我的估计,时间信息对最终结果的影响大概在0.007 - 0.01之间,在我的论文中将会阐述这个结果的由来。 Global Effects in Netflix Prize. 忙活了几天,终于把那个global effects给搞定了,不容易啊,嘿嘿. The BigChaos Solution to the Netflix Prize 2008. Scalable Collaborative Filtering with Jointly Derived Neighborhood Interpolation Weights. B0(u,i) = global mean. X(u,i) = average(i) = sum u {r(u,i) - b0(u,i)} / n(i). X(u,i) =...