DIGMN Dynamic Intent Guided Meta Network for Differentiated User Engagement Forecasting in Online Professional Social Platforms

2025-05-03 1 0 1.51MB 9 页 10玖币
侵权投诉
DIGMN: Dynamic Intent Guided Meta Network for
Dierentiated User Engagement Forecasting in Online
Professional Social Platforms
Feifan Li∗
feifan@mail.dlut.edu.cn
Dalian University of Technology
Lun Du†
lun.du@microsoft.com
Microsoft Research
Qiang Fu
qifu@microsoft.com
Microsoft Research
Shi Han
shihan@microsoft.com
Microsoft Research
Yushu Du
yusdu@linkedin.com
LinkedIn Corporation
Guangming Lu
glu@linkedin.com
LinkedIn Corporation
Zi Li
zili@linkedin.com
LinkedIn Corporation
ABSTRACT
User engagement prediction plays a critical role in designing inter-
action strategies to grow user engagement and increase revenue
in online social platforms. Through the in-depth analysis of the
real-world data from the world’s largest professional social plat-
forms, i.e., LinkedIn, we nd that users expose diverse engagement
patterns, and a major reason for the dierences in user engagement
patterns is that users have dierent intents. That is, people have dif-
ferent intents when using LinkedIn, e.g., applying for jobs, building
connections, or checking notications, which shows quite dierent
engagement patterns. Meanwhile, user intents and the correspond-
ing engagement patterns may change over time. Although such
pattern dierences and dynamics are essential for user engagement
prediction, dierentiating user engagement patterns based on user
dynamic intents for better user engagement forecasting has not
received enough attention in previous works. In this paper, we pro-
posed a
D
ynamic
I
ntent
G
uided
M
eta
N
etwork (DIGMN), which
can explicitly model user intent varying with time and perform
dierentiated user engagement forecasting. Specically, we derive
some interpretable basic user intents as prior knowledge from data
mining and introduce prior intents to explicitly model dynamic
user intent. Furthermore, based on the dynamic user intent repre-
sentations, we propose a meta-predictor to perform dierentiated
user engagement forecasting. Through a comprehensive evaluation
of LinkedIn anonymous user data, our method outperforms state-
of-the-art baselines signicantly, i.e., 2.96% and 3.48% absolute error
reduction, on coarse-grained and ne-grained user engagement
∗Work done during an internship at LinkedIn.
†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 prot 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 specic permission and/or a
fee. Request permissions from permissions@acm.org.
WSDM ’23, February 27-March 3, 2023, Singapore, Singapore
©2023 Association for Computing Machinery.
ACM ISBN 978-1-4503-9407-9/23/02. . . $15.00
https://doi.org/10.1145/3539597.3570420
prediction tasks, respectively, demonstrating the eectiveness of
our method.
CCS CONCEPTS
•Information systems →Enterprise applications
;
•Comput-
ing methodologies →Neural networks;
KEYWORDS
User Intent, User Engagement Forecasting, Meta Learning
ACM Reference Format:
Feifan Li, Lun Du, Qiang Fu, Shi Han, Yushu Du, Guangming Lu, and Zi
Li. 2023. DIGMN: Dynamic Intent Guided Meta Network for Dierentiated
User Engagement Forecasting in Online Professional Social Platforms. In
Proceedings of the Sixteenth ACM International Conference on Web Search and
Data Mining (WSDM ’23), February 27-March 3, 2023, Singapore, Singapore.
ACM, New York, NY, USA, 9 pages. https://doi.org/10.1145/3539597.3570420
1 INTRODUCTION
Online professional social platforms like LinkedIn have become
a signicant part of today’s lives. People use these platforms to
socialize, apply for jobs, read industry news, etc. Maintaining a high-
level user engagement is vital for these platforms, which can lead to
more revenue (e.g., more ad exposure). For the purpose of increasing
user engagement in the future, these platforms need to formulate
appropriate user interaction strategies, such as delivering content
that satises user intents or interests. Accurate user engagement
prediction is one of the core technologies for developing these
strategies, which can help the platform conduct user modeling,
understand user needs, and provide personalized services.
Real-world users often exhibit dierent behaviors in online so-
cial platforms, leading to diverse engagement patterns. Through
data mining and analysis in the scenario of LinkedIn, we found that
the diversity of user engagement patterns is related to the multiple
intents of users using LinkedIn, as shown in Figure 1. For instance,
some users intend to look for jobs recently, and they will frequently
visit LinkedIn to seek and apply for jobs, increasing their engage-
ment level rapidly in a period. As another example, some users
use LinkedIn to view industry news. Their engagement pattern
arXiv:2210.12402v2 [cs.LG] 22 Feb 2023
WSDM ’23, February 27-March 3, 2023, Singapore, Singapore Feifan Li et al.
Check Notifications Find New Jobs View Industry News
Time
# sessions
⋯ ⋯
Day 1 Day 2 Day 3 ⋯
Day 4 Day 1 Day 2 Day 3 ⋯
Day 4 Day 1 Day 2 Day 3 ⋯
Day 4
Figure 1: Example of dierent user engagement patterns
with dierent user intents.
is usually maintained at a relatively high level because they regu-
larly check the industry news on LinkedIn. These insights suggest
that user intents can be signals to dierentiate user engagement
patterns.
However, user intents are usually not directly observable be-
cause they exist implicitly in human consciousness. How to extract
implicit user intents is a challenging problem. At the same time,
the user’s intent may change over time. For example, some users
initially use LinkedIn to look for jobs, and when they nish looking
for jobs, they may use LinkedIn to socialize (e.g., make new connec-
tions at a new company). Explicitly modeling dynamic user intent
is vital. It can help the platform understand users’ recent intents (or
interests) and provide users with content that matches their intents
to increase user engagement.
Recently, there have been many works on user engagement fore-
casting in social network platforms. [
1
] rstly groups new users
into some clusters and then uses an LSTM-based model to predict
user churn rate. [
2
] constructs a user action graph to characterize
and forecast new user engagement. [
3
] considers user interaction
actions and builds a user graph that evolves to predict user engage-
ment. Although engagement patterns vary between users, these
works use a model with static parameters for all users to predict
their future engagement, which cannot suciently model diverse
user engagement patterns and perform dierentiated user engage-
ment forecasting. Meanwhile, some works show that user intent
can impact user engagement. [
4
] shows that user intent can inu-
ence user engagement (e.g., usage time and return time) at Pinterest.
[
5
] demonstrates that user primary intents are associated with how
likely the user is to re-engage in activity-tracking applications. On
the one hand, these works have limitations in extracting user intent.
In [
4
], the user’s intent is obtained through a survey when the app
is just opened, which may aect the user’s subsequent behaviors
[
6
]. [
5
] adopts the activity that the user most commonly uses as a
proxy for the user’s intent. However, the user’s intent may be di-
verse when using applications [
7
]. On the other hand, these works
do not explicitly model changes in user intent over time.
To address the above challenges, we rst use Latent Dirichlet
Allocation (LDA) [
8
] to perform user intent mining on large-scale
session data and identify basic user intents. Then, we propose a
D
ynamic
I
ntent
G
uided
M
eta
N
etwork (DIGMN), which can capture
the user’s dynamic intent and perform dierentiated user engage-
ment prediction. Specically, DIGMN infers multiple user intents
during each session based on similarity computation with the basic
user intents and captures the variation of user intents over time by
the sequence model. Besides, DIGMN contains a prediction network
based on meta-learning, which adopts a dynamic intent guided at-
tention mechanism to adjust network parameters by performing
a linear combination of basic parameters shared by all users for
dierentiated user engagement forecasting. Extensive experiments
conducted on coarse-grained and ne-grained user engagement
forecasting tasks verify the eectiveness of our DIGMN method.
The major contributions of this paper can be summarized as
follows:
•
We nd that user intent can be benecial for dierentiated user
engagement forecasting in online professional social platforms.
•
We develop a
D
ynamic
I
ntent
G
uided
M
eta
N
etwork (DIGMN)
which explicitly model user intent’s evolution over time and
leverage intent guided attention mechanism to adjust model
parameters for dierentiated user engagement modeling and
forecasting.
•
Through evaluation experiments on anonymous data from Linked-
In, our proposed model DIGMN has improvements of 2.96%
(Macro F1-score) and 3.48% (AUROC) on coarse-grained and
ne-grained user engagement prediction tasks when compared
to the state-of-the-art model, showing the eectiveness of our
proposed method.
2 RELATED WORK
In this section, we present related work on user engagement fore-
casting, user intent modeling, and meta-learning for dynamic net-
work parameters.
User engagement forecasting.
Recently, there have been many
works on user engagement predicting from dierent perspectives
on the social platform. Such as user behaviors and social attributes
[
1
], user action graphs [
2
], interaction actions between users [
3
], pe-
riodicity of user behaviors [
9
], and causal eects of social inuence
[
10
]. These works learn a model with static parameters for all users
to make predictions, which cannot suciently model dierentiated
user engagement patterns. [
11
] leverages a decision tree model to
divide users into disjoint groups and then learns a separate Logistic
regression model for each group of users to predict user churn.
However, such separate modeling compromises the model’s ability
to capture similarities between users. [
12
] adopts the matrix factor-
ization to predict the personalized user’s participation in mobile
video. However, our scenario has a large amount of user behavior
sequence data. This method can not eectively deal with the user’s
behavior sequence and its change over time.
User intent modeling.
Some previous works exploit LDA [
2
,
13
,
14
], n-gram [
2
,
15
] and deep learning model [
16
,
17
] to mine
user intent from user behavior. At the same time, modeling user
intents can help us understand user needs better and is signicant
in many scenarios: for instance, web searching [
18
,
19
], e-commerce
application [
4
], image sharing social platform [
20
], activity tracking
application [
5
] and recommender systems [
21
–
25
]. However, to our
best knowledge, no related work has explicitly modeled user intent
and its variation for dierentiated user engagement forecasting in
online social platforms.
Meta-learning for dynamic network parameters.
Meta-lear-
ning (also known as learning to learn) can be used to learn dynamic
model parameters, which is widely used in scenarios and tasks with
摘要：

DIGMN:DynamicIntentGuidedMetaNetworkforDifferentiatedUserEngagementForecastinginOnlineProfessionalSocialPlatformsFeifanLi∗feifan@mail.dlut.edu.cnDalianUniversityofTechnologyLunDu†lun.du@microsoft.comMicrosoftResearchQiangFuqifu@microsoft.comMicrosoftResearchShiHanshihan@microsoft.comMicrosoftResearc...

展开>> 收起<<
DIGMN Dynamic Intent Guided Meta Network for Differentiated User Engagement Forecasting in Online Professional Social Platforms.pdf

共9页,预览2页

还剩页未读， 继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

相关推荐

更多
立即下载
分类：图书资源 价格：10玖币 属性：9 页 大小：1.51MB 格式：PDF 时间：2025-05-03

开通VIP享超值会员特权

  • 多端同步记录
  • 高速下载文档
  • 免费文档工具
  • 分享文档赚钱
  • 每日登录抽奖
  • 优质衍生服务

作者详情

相关内容

更多

热门标签

人际关系 动力学连接体 力的合成 配电装置 高考理综 全宋诗作者索引 公务员考试
/ 9
评分 收藏
分享
立即下载
客服
关注