Facebook 是一个社交网络服务网站,于 2004 年 2 月 4 日上线。从 2006 年 9 月到 2007 年 9 月间,该网站在全美网站中的排名由第 60 名上升至第 7 名。同时 Facebook 是美国排名第一的照片分享站点。 2012年 2 月 1 日,Facebook向美国证券交易委员会提交集资规模为 50 亿美元的上市申请。

VIP内容

论文:Embedding-based Retrieval in Facebook Search

地址:https://www.zhuanzhi.ai/paper/471ae2edf9adf0e766b4fd8cf95ca986

相对于传统的网页搜索来说,社交网络中的搜索问题不仅需要关注输入query的信息,还需要考虑用户的上下文信息,在Facebook搜索中用户的社交图网络便是这种上下文信息中非常重要的一部分。虽然embedding的检索技术在传统的搜索引擎中得到了广泛应用,但是Facebook搜索之前主要还是使用布尔匹配模型。本文讨论了如何将embedding检索技术应用在Facebook搜索的技术方案,我们提出了一套统一的embedding框架用于建模个性化搜索中的语义embedding,以及基于经典的倒排索引进行在线embedding检索的系统。同时讨论了整个系统中很多端对端的优化技巧,例如ANN调参经验、全链路的优化等。最后,我们在FaceBook垂直搜索场景下验证了本文方法的有效性,在线A/B实验取得了显著的收益。

成为VIP会员查看完整内容
0
15

最新内容

Power laws are a characteristic distribution that are ubiquitous, in that they are found almost everywhere, in both natural as well as in man-made systems. They tend to emerge in large, connected and self-organizing systems, for example, scholarly publications. Citations to scientific papers have been found to follow a power law, i.e., the number of papers having a certain level of citation x are proportional to x raised to some negative power. The distributional character of altmetrics has not been studied yet as altmetrics are among the newest indicators related to scholarly publications. Here we select a data sample from the altmetrics aggregator Altmetrics.com containing records from the platforms Facebook, Twitter, News, Blogs, etc., and the composite variable Alt-score for the period 2016. The individual and the composite data series of 'mentions' on the various platforms are fit to a power law distribution, and the parameters and goodness of fit determined using least squares regression. The log-log plot of the data, 'mentions' vs. number of papers, falls on an approximately linear line, suggesting the plausibility of a power law distribution. The fit is not very good in all cases due to large fluctuations in the tail. We show that fit to the power law can be improved by truncating the data series to eliminate large fluctuations in the tail. We conclude that altmetric distributions also follow power laws with a fairly good fit over a wide range of values. More rigorous methods of determination may not be necessary at present.

0
0
下载
预览

最新论文

Power laws are a characteristic distribution that are ubiquitous, in that they are found almost everywhere, in both natural as well as in man-made systems. They tend to emerge in large, connected and self-organizing systems, for example, scholarly publications. Citations to scientific papers have been found to follow a power law, i.e., the number of papers having a certain level of citation x are proportional to x raised to some negative power. The distributional character of altmetrics has not been studied yet as altmetrics are among the newest indicators related to scholarly publications. Here we select a data sample from the altmetrics aggregator Altmetrics.com containing records from the platforms Facebook, Twitter, News, Blogs, etc., and the composite variable Alt-score for the period 2016. The individual and the composite data series of 'mentions' on the various platforms are fit to a power law distribution, and the parameters and goodness of fit determined using least squares regression. The log-log plot of the data, 'mentions' vs. number of papers, falls on an approximately linear line, suggesting the plausibility of a power law distribution. The fit is not very good in all cases due to large fluctuations in the tail. We show that fit to the power law can be improved by truncating the data series to eliminate large fluctuations in the tail. We conclude that altmetric distributions also follow power laws with a fairly good fit over a wide range of values. More rigorous methods of determination may not be necessary at present.

0
0
下载
预览
Top