优化循环神经网络在会话推荐中的应用

需积分: 9 132 浏览量更新于2024-09-07 收藏 397KB PDF 举报

"Improved Recurrent Neural Networks for Session-based Recommendations" 这篇论文深入探讨了使用改进的循环神经网络（RNNs）进行基于会话的推荐任务。RNNs最近在这一领域显示出对传统推荐方法的显著提升。作者提出并研究了两种技术来增强RNN模型在会话推荐中的性能：数据增强和输入数据分布变化的处理方法。同时，他们还实验性地研究了泛化蒸馏的应用，并提出了一种新颖的直接预测项目嵌入的替代模型。 1. 循环神经网络（RNNs）在会话推荐中的应用：RNNs因其在处理序列数据上的优势，如能捕获用户行为的时序模式，被引入到会话推荐系统中。它们可以学习和理解用户在一个会话中的行为序列，以预测下一个可能的交互项目。 2. 数据增强：为了提高模型的泛化能力，作者提出了数据增强策略。这种方法通过创建合成的训练样本，增加了模型看到的不同类型会话的多样性，从而使其能够更好地处理实际环境中可能出现的各种情况。 3. 输入数据分布变化的处理：由于真实世界的数据分布可能会随着时间推移而发生变化，作者设计了一种方法来处理这种输入分布的漂移。这有助于模型适应不断变化的用户行为模式，保持其推荐的准确性和相关性。 4. 泛化蒸馏：这是一种知识蒸馏技术，用于将大型复杂模型（通常称为教师模型）的“智慧”转移到更小、更高效的模型（学生模型）。在这个研究中，泛化蒸馏被用来优化模型的效率，同时保持高推荐性能。 5. 预测项目嵌入的替代模型：传统的RNN推荐模型可能直接预测项目的概率。然而，该论文提出了一种新方法，即直接预测项目的嵌入向量，这可能提供了更丰富的上下文信息，从而提高推荐的精确度。在RecSys Challenge 2015数据集上的实验表明，这些改进带来了显著的性能提升，相对于先前报告的结果，召回率@20和平均倒数排名@20分别提高了12.8%和14.8%。 CCS概念涵盖了监督学习和神经网络等相关计算方法，这表明该研究对于理解和改进基于机器学习的推荐系统具有重要的理论和实践意义。

Improved Recurrent Neural Networks for Session-based

Recommendations

Yong Kiam Tan

tanyongkiam@gmail.com

Xinxing Xu

xuxinx@ihpc.a-star.edu.sg

Yong Liu

liuyong@ihpc.a-star.edu.sg

Institute of High Performance Computing, A*STAR

1 Fusionopolis Way, Singapore 138632

ABSTRACT

Recurrent neural networks (RNNs) were recently proposed

for the session-based recommendation task. The models

showed promising improvements over traditional recommen-

dation approaches. In this work, we further study RNN-

based models for session-based recommendations. We pro-

pose the application of two techniques to improve model

performance, namely, data augmentation, and a method to

account for shifts in the input data distribution. We also

empirically study the use of generalised distillation, and a

novel alternative model that directly predicts item embed-

dings. Experiments on the RecSys Challenge 2015 dataset

demonstrate relative improvements of 12.8% and 14.8% over

previously reported results on the Recall@20 and Mean Re-

ciprocal Rank@20 metrics respectively.

CCS Concepts

•Computing methodologies → Supervised learning;

Neural networks; •Information systems → Recom-

mender systems;

Keywords

Recurrent neural networks; Recommender systems; Session-

based recommendations

1. INTRODUCTION

Users of e-commerce websites are often inundated by the

huge number of items available for sale. Recommender sys-

tems can be used to enhance user experience by making

personalized and useful recommendations for each user. For

example, the system could automatically display items of

interest, or suggest new discounts relevant to each user. In

order to personalize recommendations, traditional recom-

mender systems often need to build up a user proﬁle. Col-

laborative ﬁltering approaches [15, 14, 21] can deﬁne user-

user similarity based on their history of purchases, or they

Publication rights licensed to ACM. ACM acknowledges that this contribution was

authored or co-authored by an employee, contractor or afﬁliate of a national govern-

ment. As such, the Government retains a nonexclusive, royalty-free right to publish or

reproduce this article, or to allow others to do so, for Government purposes only.

DLRS ’16, September 15 2016, Boston, MA, USA

ACM ISBN 978-1-4503-4795-2/16/09. . . $15.00

DOI: http://dx.doi.org/10.1145/2988450.2988452

could rely on matrix factorization to build latent factor vec-

tors for each user. Crucially, these approaches require the

user to be identiﬁed when making recommendations. This

may not always be possible: new users to the site will not

have any proﬁle, or users may not be logged in, or they

may have deleted their tracking information. This leads to

the problem of cold-start for recommendation methods that

require user history.

An alternative to relying on historical data is to make

session-based recommendations [23]. In this setting, the rec-

ommender system makes recommendations based only on

the behaviour of users in the current browsing session. This

avoids the aforementioned cold-start issue but we must en-

sure that the system remains accurate and responsive (i.e. the

predictions do not take too long to make). Recurrent Neu-

ral Networks (RNNs) were recently proposed in [10] for the

session-based recommendation task. The authors showed

signiﬁcant improvements over traditional session-based rec-

ommendation models using an RNN. The proposed model

utilizes session-parallel mini-batch training, and also em-

ploys ranking-based loss functions for learning the model.

In this work, we further study the application of RNNs for

session-based recommendations. In particular, we examine

and adapt various techniques from the literature for this

task. These include:

• Data augmentation via sequence preprocessing and em-

bedding dropout to enhance training and reduce over-

ﬁtting.

• Model pre-training to account for temporal shifts in

the data distribution.

• Distillation using privileged information to learn from

small datasets.

Additionally, we propose a novel alternative model that re-

duces the time and space requirements for predictions by

predicting item embeddings directly. This makes RNNs

more readily deployable in real-time settings.

Our proposed techniques were evaluated on the RecSys

Challenge 2015 data set. The eﬀectiveness of our data aug-

mentation strategy is evidenced by relative model perfor-

mance improvements of 12.8% and 14.8% over previously re-

ported results on the Recall@20 and Mean Reciprocal Rank

@20 (MRR@20) metrics respectively. We also showed that

distillation could be successfully applied for performance

gains on small datasets. Finally, our novel item embedding

output approach signiﬁcantly reduces the time and space

requirements of the RNN model.

arXiv:1606.08117v2 [cs.LG] 16 Sep 2016

下载后可阅读完整内容，剩余5页未读，立即下载

Jayxp

粉丝: 6
资源: 137

优化循环神经网络在会话推荐中的应用

ES9038Q2M-Datasheet-v1.4.pdf：ESS TECHNOLOGY, INC. 32位立体声低功率音频DAC 数据手册

"2019版ITU-T H.265标准文档：高效视频编码推荐

LPDDR4x标准及修订：JESD209-4-1.pdf详解 (27字)

Improved Residual Networks for Image and Video Recognition.pdf

IS-GPS-200M.pdf

NIST SP800-163r1.pdf

6月移动市场报告（中英双语）-爱立信-6-54页.pdf

Vi iMproved _VIM_vimbook-opl-EN.pdf

Improved relation classification by deep recurrent neural networks

Graph Neural Networks_ A Review of Methods and Applications----清华大学周杰.pdf

最新资源