公开(公告)号：US20140373047A1

公开(公告)日：2014-12-18

申请号：US13916132

申请日：2013-06-12

Applicant: NETFLIX, Inc.

Inventor： Harald STECK

IPC: H04N21/25 ,  H04N21/81

CPC classification number: H04N21/251 ,  G06Q30/0255 ,  H04N21/812

Abstract: Techniques are described for promoting original media titles. Given metadata tags associated with the original title and other media titles, a tag data matrix is generated and factored into two matrices, one of which includes vectors representing the media titles in a first latent space. Similarity scores are computed between a vector representing the original title and each of the other media title vectors to determine a set of media titles most similar to the original title. Then, a play data matrix is factorized, and an average of vectors representing the most similar titles in a second latent space is taken to be a vector representation of the original title in the second latent space. This representation is compared with representations of users in the second latent space to generate similarity scores, and the original title is then promoted to users associated with the highest similarity scores.

Abstract translation: 描述了促进原始媒体标题的技术。 给定与原始标题和其他媒体标题相关联的元数据标签，生成标签数据矩阵并将其分解为两个矩阵，其中之一包括表示第一潜在空间中的媒体标题的向量。 在表示原始标题的矢量和每个其他媒体标题向量之间计算相似性分数，以确定与原始标题最相似的一组媒体标题。 然后，对游戏数据矩阵进行因式分解，并且在第二潜在空间中表示最相似的标题的向量的平均值被认为是第二潜在空间中原始标题的向量表示。 将该表示与第二潜在空间中的用户的表示进行比较，以生成相似性分数，然后将原始标题提升为与最高相似性分数相关联的用户。

公开(公告)号：US20170024391A1

公开(公告)日：2017-01-26

申请号：US15044020

申请日：2016-02-15

Applicant: NETFLIX, INC.

Inventor： Harald STECK

IPC: G06F17/30 ,  G06N99/00

CPC classification number: G06F17/3053 ,  G06F17/30702 ,  G06F17/30867 ,  G06N99/005

Abstract: In one embodiment of the present invention, a training engine teaches a matrix factorization model to rank items for users based on implicit feedback data and a rank loss function. In operation, the training engine approximates a distribution of scores to corresponding ranks as an approximately Gaussian distribution. Based on this distribution, the training engine selects an activation function that smoothly maps between scores and ranks. To train the matrix factorization model, the training engine directly optimizes the rank loss function based on the activation function and implicit feedback data. By contrast, conventional training engines that optimize approximations of the rank loss function are typically less efficient and produce less accurate ranking models.

Abstract translation: 在本发明的一个实施例中，训练引擎教导矩阵分解模型，以基于隐含反馈数据和秩丢失函数来对用户排列项目。 在运行中，训练引擎将分数近似为近似高斯分布。 基于这种分布，训练引擎选择平滑地映射分数和等级之间的激活函数。 为了训练矩阵分解模型，训练引擎基于激活函数和隐式反馈数据直接优化秩损失函数。 相比之下，优化秩损失函数近似的常规训练引擎通常效率较低，并且产生不太准确的排名模型。