【计算机科学】【2017.09】推荐系统的深度学习：相似与偏好的联合学习

最新推荐文章于 2022-08-27 11:57:44 发布

梅花香——苦寒来

最新推荐文章于 2022-08-27 11:57:44 发布

阅读量217

点赞数

在这里插入图片描述

本文为德国卡尔斯鲁厄应用技术大学（作者：Marcel Kurovski）的硕士论文，共124页。

推荐系统是一种过滤引擎，它汇集各种意见，以个性化建议支持决策过程。它们被广泛使用并影响着几乎每个人在不同领域的日常生活，如电子商务、社交媒体或娱乐。在大规模系统中高效地生成相关建议是一项非常复杂的任务。为了提供个性化服务，引擎和算法需要捕捉用户的不同品味，并找出他们和许多物品之间的非线性依赖关系。巨大的数据稀疏性和大量的实时需求进一步加剧了这一挑战。同时，深度学习的广泛应用也证明了在传统机器学习失败或表现平庸的情况下，可以解决诸如对象或语音识别等复杂任务。尽管如此，在推荐系统中应用深度学习的研究仍处于起步阶段。

本文概述了推荐系统在深度学习方面的最新进展，阐述了不同深度神经网络类型在各个领域的成功应用，指出了对应的挑战及发展趋势。此外，结合最新研究成果，构建了一个基于深度学习的推荐系统：大型在线市场上的车辆推荐。扩展了谷歌提出的深度学习方法，因此，除了偏好属性外，还引入了相似性的优化准则。开发了有效的候选生成策略，将局部优化的产品量化应用于用户和物品种类嵌入。该方法在相关性方面优于协同过滤和基于内容的混合式协同过滤。

Recommender Systems are ltering engines that aggregate opinions to support decision making processes with personalized suggestions. They are widely used and in-uence the daily life of almost everyone in dierent domains like e-commerce, social media, or entertainment. The ecient generation of relevant recommendations in large-scale systems is a very complex task. In order to provide personalization, engines and algorithms need to capture users’ varying tastes and nd mostly nonlinear dependencies between them and a multitude of items. Enormous data sparsity and ambitious real-time requirements further complicate this challenge. Meanwhile, the pervasive application of deep learning proves to solve complex tasks like object or speech recognition where traditional machine learning failed or showed mediocre performance. Despite this fact, research on the application of deep learning on recommender systems is still in its infancy. My master thesis provides an overview of the state of art in deep learning for recommender systems. I illustrate the successful application of dierent deep neural network types in various domains, point out the challenges and derive trends. Furthermore, I combine latest research to build a recommender system based on deep learning for another domain: the recommendation of vehicles on a large online market. I extend the wide and deep learning approach proposed by Google. Therefore, I augment the optimization criterion with embedding similarity besides preference. I develop strategies for ecient candidate generation applying Locally Optimized Product Quantization on user and item embeddings. The proposed approach outperforms collaborative ltering and hybrid collaborative-content-based ltering in terms of relevance.

1 引言

1.1 研究动机

1.2 研究的问题与目标

1.3 研究内容

2 理论基础

2.1 推荐系统

2.2 深度学习

2.3 相关技术

3 当前最新技术水平回顾

3.1 面向应用的研究

3.2 面向理论的研究

3.3 调研与小结