WWW'22【曝光度偏差】《具有潜在混杂因素的无偏序贯推荐》
需积分: 30 70 浏览量
2022-11-02
15:25:49
上传
评论
收藏 1.4MB PDF 举报
Unbiased Sequential Recommendation with
Latent Confounders
Zhenlei Wang
1,2,3
, Shiqi Shen
3
, Zhipeng Wang
3
, Bo Chen
3
, Xu Chen
1,2,∗
,Ji-Rong Wen
1,2
1
Gaoling School of Articial Intelligence, Renmin University of China, Beijing, China
2
Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing, China
3
WeChat, Tencent, Beijing, China
wang.zhenlei@foxmail.com,shiqishen@tencent.com,markrocwang@tencent.com
jennychen@tencent.com,successcx@gmail.com,jrwen@ruc.edu.cn
ABSTRACT
Sequential recommendation holds the promise of understanding
user preference by capturing successive behavior correlations. Ex-
isting research focus on designing dierent models for better tting
the oine datasets. However, the observational data may have been
contaminated by the exposure or selection biases, which renders
the learned sequential models unreliable. In order to solve this
fundamental problem, in this paper, we propose to reformulate
the sequential recommendation task with the potential outcome
framework, where we are able to clearly understand the data bias
mechanism and correct it by re-weighting the training instances
with the inverse propensity score (IPS). For more robustness model-
ing, a clipping strategy is applied to the IPS estimation to reduce the
variance of the learning objective. To make our framework more
practical, we design a parameterized model to remove the impact of
the potential latent confounders. At last, we theoretically analyze
the unbiasedness of the proposed framework under both vanilla
and clipping IPS estimations. To the best of our knowledge, this
is the rst work on debiased sequential recommendation. We con-
duct extensive experiment based on both synthetic and real-world
datasets to demonstrate the eectiveness of our framework.
CCS CONCEPTS
• Information systems → Recommender systems.
KEYWORDS
Sequential Recommendation, Unbiased Recommendation, Potential
Outcome Framework
ACM Reference Format:
Zhenlei Wang
1, 2, 3
, Shiqi Shen
3
, Zhipeng Wang
3
, Bo Chen
3
, Xu Chen
1, 2, ∗
,Ji-
Rong Wen
1, 2
. 2022. Unbiased Sequential Recommendation with Latent
Confounders. In Proceedings of the ACM Web Conference 2022 (WWW ’22),
April 25–29, 2022, Virtual Event, Lyon, France. ACM, New York, NY, USA,
10 pages. https://doi.org/10.1145/3485447.3512092
∗ Corresponding author.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specic permission and/or a
fee. Request permissions from permissions@acm.org.
WWW ’22, April 25–29, 2022, Virtual Event, Lyon, France
© 2022 Association for Computing Machinery.
ACM ISBN 978-1-4503-9096-5/22/04.. . $15.00
https://doi.org/10.1145/3485447.3512092
1 INTRODUCTION
Sequential recommendation has recently attracted increasing at-
tention due to its more practical settings [
33
,
37
,
40
,
46
,
47
] and
higher recommendation performance [
28
,
36
,
41
]. It advances the
general recommendation [
3
,
7
,
9
,
21
] by explicitly modeling the cor-
relations between user successive behaviors. In the past few years,
quite a lot of promising sequential recommender models have been
proposed. For example, FPMC [
22
] estimates the current user pref-
erence based on their latest behaviors. GRU4Rec [
16
] introduces
recurrent neural network (RNN) to model more comprehensive
history information. NARM [
17
] leverages attention mechanism
to distinguish the importance of dierent user behaviors. A ma-
jor assumption held by these models is that the training data can
perfectly represent the user sequential preferences. However, this
assumption can be arguable in practice, since the observational
data may have been biased by the former recommender system (i.e.,
exposure bias) or the user intrinsic tendency (i.e., selection bias). As
exampled in Figure 1, according to the interaction history, we can
infer that the user likes sports brands (e.g., “Nike” and “Adidas”), and
is recently interested in red clothes. As a result, a red Adidas T-shirt
is observed in the dataset as the next item. Basically, the observation
of an item is jointly determined by the underlying user sequential
preference and the item recommendation probability. In the above
example, the red Nike tops (C) or other red Adidas clothes (B) are
also reasonable next items; the correlations between them and the
history behaviors can reveal important user sequential preferences.
However, these items are not recommended to the user, and thus
have less opportunities to be observed. This example manifests
that the observed item correlations is skewed for revealing the real
user sequential preference. If the model is directly trained on the
observational data, it can be biased from the true user preference.
For more robust, fair and unbiased optimization, recent years
have witnessed many studies on incorporating causal inference
into traditional machine learning algorithms [
4
,
6
,
44
]. Basically,
these models design novel learning objectives by assuming dierent
causal graph for data generation [
23
,
24
,
31
]. In this paper, we follow
the same idea to combine sequential recommendation with causal
inference for removing the above mentioned data bias problem.
The major challenges of such combination lie in two folds: to begin
with, the causal graph assumed by existing debiased recommender
models [
24
,
31
,
44
] is usually like Figure 1(b), where the feedback
is jointly inuenced by the user and whether the target item is
selected, and the user simultaneously inuences the item selection
and feedback. In such a formulation, dierent items are assumed to
be independent. However, in sequential recommendation, capturing
2195
资源评论
