基于隐式反馈与上下文信息的推荐算法研究
文献类型：学位论文
| 作者 | 靳冠坤; 谢闯
|
| 学位类别 | 硕士 |
| 答辩日期 | 2018-05-17 |
| 授予单位 | 中国科学院沈阳自动化研究所 |
| 授予地点 | 沈阳 |
| 导师 | 库涛 |
| 关键词 | 个性化推荐 隐式反馈 上下文建模 排序学习 因子分解机 |
| 其他题名 | Research on recommendation algorithm based on implicit feedback and context information |
| 学位专业 | 机械电子工程 |
| 中文摘要 | 随着互联网与电子商务的快速发展，信息过载问题日益严重，传统的搜索引擎技术已经无法满足人们日益个性化的需求。为了帮助用户快速发现其所需信息，推荐系统应运而生。传统的推荐技术大都基于用户的显式评分数据，而忽略了在工业界更加常见，数据量更大，更符合实际应用场景的隐式反馈数据。另外，随着移动互联网的快速发展，推荐的场景日益复杂和多样化，同时系统也收集到了更多的上下文信息（例如时间、位置、状态等），在推荐的过程中融入用户物品的上下文信息无疑会大大提升推荐的准确性。针对上述两个问题，即在面向隐式反馈的推荐场景下，如何在推荐的过程中融入用户物品的上下文信息，进行了相关研究工作，主要工作内容包含以下几点：（1）调研并分析了面向隐式反馈的推荐和上下文感知推荐下的特点以及现有算法的优缺点，针对隐式反馈下的负反馈缺失问题，考虑从排序的角度出发，针对物品列表进行优化，将个性化推荐问题转化为对于推荐列表排序的优化问题。对于如何针对上下文信息进行建模的问题，调研并发现基于模型的方法更为合理，其中因子分解机模型，更加适用于推荐场景下数据稀疏性问题和效率问题，同时又具有线性的复杂度，从而受到了广泛认可。（2）针对隐式反馈下的负反馈缺失问题和上下文信息建模问题，本文提出了将因子分解机融入到Pairwise排序学习方法中，提出了对级学习因子分解机模型（PLFM）。首先提出对级偏好假设，即用户对有过反馈的物品的偏好大于没有过反馈的物品的偏好，并基于此构造训练样本。排序学习中Pairwise方法的主要思想是将排序问题转化为二分类，通过判断任意两个物品对是否满足偏序关系来进行排序。而对于一个推荐列表而言，排序结果越好，那么有错误关系的物品对就越少，故我们以最小化错误物品对的数量作为目标，使用了交叉熵作为损失函数来度量每个样本对的预测代价。另外本文将因子分解机作为其中的排序函数来建模用户物品的上下文信息，将上下文信息融入到推荐的过程当中来提升推荐的准确度，最后设计实验，对比了本文改进的模型与现有模型，实验也证明了本文算法的有效性。（3）在对级学习方法中所提出的偏好假设，即对于一个用户而言，相比没有过反馈的物品往往会对有过反馈的物品有更强的偏好, 这种假设在推荐算法的学习过程中会衍生出大量的训练样本对并且其中大量的样本对于模型学习的作用很小。另外为应对大规模的数据集，本文所研究算法往往都是基于均匀采样的随机梯度下降法来优化求解，但这种均匀采样的策略会导致模型收敛非常缓慢。基于此本文针对负反馈采样方案进行了研究与设计，提出了两种采样方案：一种为静态全局采样，一种为动态自适应采样，最后通过实验也证明，所设计的采样方案可以较好的加快模型的收敛速度。 |
| 英文摘要 | With the rapid development of the Internet and e-commerce, the problem of information overload is becoming increasingly seriously. How to help users quickly find the information they need, the recommendation system has been paid attention increasingly. A good recommendation system has an indispensable role in improving user experience and product sales. Most of the traditional recommendation technologies are based on the user's explicit ratings, and ignore the implicit feedback that is more common in the industry, has a larger amount of data, and is more in line with the actual application scenario. In addition, with the rapid development of the mobile Internet, the recommended scenarios are increasingly complex and diverse. At the same time, the system also collects more contextual information (such as time, location, status, etc.), and The incorporation of contextual information of user-items in the recommendation process undoubtedly will greatly improve the accuracy of the recommendation. Aiming at the above two problems, that is, recommendation based on implicit feedback and how to incorporate contextual information of user-items into the recommendation process, we conduct related research work, the main work content includes the following points: (1) Investigate and analyze the characteristics of recommendation based on implicit feedback and context-aware recommendation and the advantages and disadvantages of existing algorithms. For the problem of negative feedback missing under implicit feedback, from the perspective of optimizing the ranking of item list, which translates personalized recommendation questions into optimization questions of ranking the recommended list. For the problem of how to model the contextual information, we investigated and found that the model-based method is more reasonable, in which the factorization machines (FM) is more suitable for data sparsity and efficiency problems in the recommended scenario, while it also has a linear complexity and has been widely applied. (2) In order to solve the problem of negative feedback missing and context information modeling, we proposes to integrate the factorization machines into the pairwise learning method and proposes the pairwise learning factorization machines model (PLFM). First, we propose the pair preference assumption that the user's preference for the item that has selected is greater than the preference of the item that has not selected, and construct the training sample based on this assumption. The main idea of the pairwise method in Learning to Rank (LtR) is to convert the ranking problem into binary classification problem, and ranking by determining whether any two items meet the partial order relationship. For a recommendation list, the better the ranking result, the fewer the pairs of items that have the wrong relationship, so we use the cross-entropy as the loss function to measure the price of each pair of samples with the goal of minimizing the number of pairs of incorrect items. In addition, the factorization machine is used as the ranking function to model the context of user- items, and the contextual information is incorporated into the recommendation process to improve the accuracy of the recommendation. Finally, we design experiments to compare the proposed model and the existing models, results also prove the effectiveness of the proposed algorithm. (3) The preference hypothesis proposed in the pairwise learning method means that for a user, compared to items without selected, there is often a stronger preference for items that have selected. This assumption leads to a large number of training samples in the learning process of the recommendation algorithm, and a large number of these samples have little effect on model learning. In addition, in order to deal with large-scale data sets, the algorithms studied in this paper are often based on the uniform sampling of the stochastic gradient descent method to optimize the solution, but this uniform sampling strategy will lead to very slow convergence of the model. Based on this, this paper studies and designs the negative feedback sampling scheme, and proposes two sampling schemes: one is static global sampling and the other is dynamic adaptive sampling. Finally, experiments have also proved that the designed sampling scheme can be better speed up the convergence of the model. |
| 语种 | 中文 |
| 产权排序 | 1 |
| 页码 | 63页 |
| 源URL | [http://ir.sia.cn/handle/173321/21770]  |
| 专题 | 沈阳自动化研究所_工业控制网络与系统研究室 |
| 推荐引用方式 GB/T 7714 | 靳冠坤,谢闯. 基于隐式反馈与上下文信息的推荐算法研究[D]. 沈阳. 中国科学院沈阳自动化研究所. 2018. |
入库方式： OAI收割
来源：沈阳自动化研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。