【论文笔记(0)】A case study of batch and incremental recommender systems in supermarket data under concept

Ⅰ 论文信息

论文为 A case study of batch and incremental recommender systems in supermarket data under concept drifts and cold start ，于2021年8月发表在 Expert Systems With Applications 期刊上。

该期刊为JCR二区，近年来影响因子有所提升，处于二区中等位置。

本文的研究内容主要为streaming-based models与recommender systems的结合，使得在提升performance的同时可以减轻concept drift与cold start问题。

Ⅱ 论文框架

This paper is divided as follows. Section 2 describes recommender systems, types of feedback, the problem of concept drift, and the challenge of cold start. Section 3 describes existing works on positive-only recommender systems in both batch and stream learning scenarios. Section 4 details a new dataset we make available regarding supermarket transactions, which exhibits concept drift and cold start characteristics. Section 5 describes the experiments undertaken to perform the proposed analysis of recommender systems. Section 6 analyzes existing works in recommender systems to answer whether streaming recommender systems overcome batch approaches w.r.t. concept drifts and cold start. Finally, Section 7 concludes this paper and describes future works.

1. Introduction

• CF由于所需信息少（只要past user-item interactions）成为研究主流
  In CF, the only data required is a list of user-item interactions, while in CBF, items’ details are required. Consequently, CF is less restrictive and has been the target of many works over the years.

• 概念漂移在推荐系统中如何体现
  In recommender systems, concept drift reflects changes in the interactions between customers and items, either because (i) customers’ preferences change, (ii) new items become available for purchase, etc.

• batch fashion & streaming fashion
  BATCH fashion:
  Recommender systems are traditionally trained in a batch fashion, which means that given a training set composed of interactions between users and items, a static model is learned and deployed ad eternum.
  STREAMING fashion:
  Consequently, it is relevant to tailor recommender systems that can be incremented over time, assuming that the interactions between users and items are made available as a stream of events.

• The incremental ability of the streaming-based recommender systems allows a better recovery when cold start is present.

2. Recommender systems

2.1 Concept drift

The most efficient way to deal with potentially drifting scenarios is to increment the model as user-item interactions are made available.

2.2 The cold start problem

In recommender systems, the cold start problem includes three cases:
(1) cold start of users (how to recommend items to a user recently entered the system);
(2) cold start of items (how to recommend a new item recently introduced into the system to interested users);
(3) cold start of the system (how to realize accurate recommendation in a new system).

3. Positive-only approaches for recommender systems

【positive-only】The feedback is positive only, for example 0-5.

3.1 Batch approaches

这里主要介绍了传统的MF模型 - SVD，BPRMF。此外还介绍了近期结合神经网络的方法 - GMF，MLP，NeuMF。

3.2 Streaming approaches

在streaming approaches中，我们更新推荐系统in a single pass manner，保持数据的自然到达顺序，这样不仅减少了计算成本，还能进行drift adaptation。

这一小节主要介绍了两种work incrementally的方法：Incremental Stochastic Gradient Descendent (ISGD) (Vinagre et al., 2014) 和 Incremental Bayesian Personalized Ranking for MF (IBPRMF) (Rendle et al., 2012)。

5. Experimental protocol

下图展示了对于batch方法和streaming方法，本文是如何进行数据集的分割和训练的。

四个月的数据集，前两个月用于training，后两个月用于testing，而不论什么阶段，stream protocol都会从输入的数据流中不断学习。

本文使用了RECALL@N metric来评估获得的推荐的质量。

同时，用了两种方法进行评估：
（1）basic evaluator：用整个test set进行测试，允许recommender systems与hypothesis testing之间的比较。
（2）window-based evaluator：基于window（test set的1%）汇报recall。背后的原理在于允许the assessment of recommender systems over time。

6. Experimental results and analysis

• The discrepancy between streaming algorithms and the corresponding batch counterpart depicts the importance of constantly updating the recommender system as new data becomes available.
• recommender models based on neural networks exhibit interesting behavior in cold-start scenarios even though they are not continuously updated
  可能是神经网络中学到的higher-order embeddings能够generalize用户的潜在行为，使其性能比传统的MF更好。

Ⅲ 启发

1. Streaming recommender systems 可以有效地处理 cold start issue.
2. 可以尝试结合 explicit drift detection on matrix factorization 和 neural models in terms of recommendation techniques.