CASIA OpenIR
基于用户行为序列的推荐算法研究
康柳
Subtype硕士
Thesis Advisor李学恩
2022-05-18
Degree Grantor中国科学院自动化研究所
Place of Conferral中国科学院自动化研究所
Degree Discipline计算机技术
Keyword序列推荐,逻辑回归,自注意力机制,用户长期兴趣,对比学习
Abstract

随着中国网络科技产业的持续高速发展, 网络数据量规模也随之出现了指数级的上升。在海量的数据中,为用户快速定位目标信息变得愈加困难,信息过载问题愈发严重。近些年来,推荐算法成为帮助用户精确定位所需信息并解决信息过载问题的有效手段,其核心思想为基于用户的历史行为挖掘用户兴趣,进而根据每个用户的兴趣为其推荐最匹配的内容。由于在实际的应用场景中,用户兴趣呈现动态性变化的特点,即每个用户的兴趣会随着时间发生改变。因此,利用序列推荐算法动态学习用户兴趣,从而提供个性化的推荐是成为当前研究的热点问题之一。

然而,当前的序列推荐算法依然存在诸多问题。例如,在数据处理过程中,用户正向行为的筛选方法粒度过粗;对辅助特征的利用不够充分、缺乏用户长期 兴趣等,使得难以精准表征用户兴趣,推荐性能的提升依然存在较大空间。本文 主要针对序列推荐算法的现存问题,提出从数据处理、模型构建、和模型训练三个方面联合优化,将本文的主要工作和创新点归纳如下:

1)在数据处理方面,本文提出一种基于逻辑回归的用户行为判定方法,提升了数据处理的准确性。

在推荐领域,精准筛选用户正向行为对个性化推荐至关重要,但当前工作通 常使用权重分配法进行正向行为筛选,这种方法粒度较粗且缺乏针对性。因此, 本文提出使用逻辑回归模型自适应判断用户正向行为。即根据内容视频的互动 次数、观看次数、以及用户的年龄、学历、性别等特征,自适应学习不同特征的权重,从而实现精准、高效、细粒度的用户正向行为筛选。

2)在模型构建层面,提出一种基于保留 Value融合物品特征信息与用户长期兴趣的建模方法,进一步提升推荐的准确率。

当前工作通常使用自注意力机制基于内容信息建模或直接引入物品的边信息进行建模。为了更加高效地利用特征信息,本文通过优化边信息的融合方式,提出保留 Value的融合法来提升模型的推荐性能。并且,本文还参考 LSTM 的门控机制构建了联合用户长期兴趣的推荐模型,实现了更有效的用户兴趣征,达到了提升推荐性能的效果。  

3)在模型训练过程中,针对样本稀疏问题,本文提出使用对比学习的训练方法提升模型的泛化性。

具体而言,通过将视频内容经过不同的数据增强方式作为模型输入,利用对比学习约束模型优化过程,使得模型对同一物品的特征表征靠近,从而更好地学 习物品间的潜在关系,进一步提升了推荐模型的泛化能力。

Other Abstract

With the sustained and rapid development of the Chinese network technology in dustry, the scale of network data has also increased exponentially. It becomes more andmore difficult to quickly locate the target information for users among the massive data, and the problem of information overload becomes more and more serious. In recent years, recommendation algorithm has become an effective means to help users accurately locate the desired information and solve the problem of information overload.

The core idea of the recommendation algorithm is to mine users’ interests based on their historical behavior and then recommend the best matching content for each user according to their interests. In practice, the user’s interest presents the characteristics of dynamic change; that is, the interest of each user will change over time. Therefore, using the sequence recommendation algorithm to dynamically learn user interests and provide personalized recommendations has become a hot issue in current research.

However, the current sequence recommendation algorithm still has many problems. For example, the filtering method of user’s positive behavior is too coarse in data processing, insufficient use of additional features and lack of users’ long-term interest make it difficult to characterize users’ interests accurately. There is still much room to improve recommendation performance. To tackle the above problems, this paper proposes joint optimization from three aspects: data processing, model construction, and model training. The main work and innovations of this paper are summarized as follows:

1) In terms of data processing, this paper proposes a user behavior judgment method based on logistic regression, which improves the accuracy of data processing.

In the field of recommendation, accurately selecting users’ positive behavior is critical for personalized recommendation. Yet, previous works usually adopt the weight distribution method for positive behavior selection, which is coarse-grained and lacks pertinence. Therefore, this paper proposes to use the logistic regression model to adaptively judge users’ positive behavior according to the interaction times and viewing times of content videos, as well as the user’s age, education, gender, and other characteristics, to achieve accurate, efficient and fine-grained user positive behavior selecting.

2) For model construction, a modeling method based on ”reserved value” to integrate item feature information and users’ long-term interests is proposed to improve recommendation performance.

Previous works usually use a self-attention mechanism to extract features based on content information or directly introduce the edge information of items to model the relationship. To better utilize feature information, this paper proposes a ”value preserving” fusion method to improve the model performance by optimizing the fusion mode of edge information. In addition, referring to the gating mechanism of LSTM, this paper constructs a recommendation model of joint users’ long-term interests and realizes a more effective representation of users’ interests to further improve the recommendation performance.

3) In terms of model training, aiming at the problem of sparse samples, this paper proposes a contrastive learning method to improve the model generalization ability.

Specifically, we take the video content through different data augmentation methods as model inputs and use contrastive learning to shepherd the model optimization process. the model to extract similar feature representations of the same item, making it better learn the potential relationship between items and further improve its generalization ability.

Pages64
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/48473
Collection中国科学院自动化研究所
Recommended Citation
GB/T 7714
康柳. 基于用户行为序列的推荐算法研究[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2022.
Files in This Item:
File Name/Size DocType Version Access License
基于用户行为序列的智能推荐算法研究.pd(2438KB)学位论文 开放获取CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[康柳]'s Articles
Baidu academic
Similar articles in Baidu academic
[康柳]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[康柳]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.