Tutorial on metrics of user engagement -- Applications to Search & E- commerce

Tutorial on Metrics of User Engagement
Applications to Search & E-Commerce
1
Mounia Lalmas & Liangjie Hong

Outline
2
1. Introduction and scope
2. Towards a taxonomy of metrics
3. Experimentation and evaluation of metrics
4. Optimisation for metrics
5. Applications
a. Search
b. E-commerce
6. Recap and open challenges
7. References … to come

Acknowledgements
● This tutorial uses some material from a tutorial “Measuring User
Engagement” given at WWW 2013, Rio de Janeiro (with Heather O’Brien
and Elad Yom-Tov).
● M. Lalmas, H. O’Brien and E. Yom-Tov. “Measuring User Engagement”,
Synthesis Lectures on Information Concepts, Retrieval, and Services,
Morgan & Claypool Publishers, 2014.
3

Introduction and scope … Outline
Who we are
What is user engagement
Approaches to measure user engagement
The focus of this tutorial
5

Who we are
● Mounia Lalmas, Research Director at Spotify, London
○ Research interests: user engagement in areas such as advertising, digital
media, search, and now music
○ Website: https://mounia-lalmas.blog/
● Liangjie Hong, Head of Data Science at Etsy, New York
○ Research interests: search, recommendation, advertising
and now hand-craft goods
○ Website: https://www.hongliangjie.com/
6

What is user engagement? … Some definitions
User engagement is regarded as a persistent and pervasive cognitive affective
state, not a time-specific state (Schaufeli et al., 2002)
User engagement refers to the quality of the user experience associated with the
desire to use a technology (O’Brien and Toms, 2008)
User engagement is a quality of the user experience that emphasizes the positive
aspects of interaction – in particular the fact of wanting to use the technology
longer and often (Attfield et al., 2011).
7
All the above can translate into the “emotional, cognitive and behavioural
connection that exists, at any point in time and over time, between a user
and a technological resource” (O’Brien, Lalmas & Yom-Tov, 2013)

Why is it important to engage users?
Users have increasingly enhanced expectations about their interactions with
technology
… resulting in increased competition amongst the providers of (online)
services.
utilitarian factors (e.g. usability) → hedonic and experiential factors of
interaction (e.g. fun, fulfillment) → user engagement
8
(O’Brien, Lalmas & Yom-Tov, 2013)

Is this site engaging?
9
leisure
-Aesthetics
aesthetics

Janette Lehmann
10
shopping
usability

Janette Lehmann
11
news
trust

What influences user engagement?
Context
• Trends
• Awareness
• Novelty
User
• Demographics
• Social
environment
• Branding
Site
• Presentation
• Content
• Functionality
Attributes of the user experience
Aesthetic
Usability
User engagement
Satisfaction
Usefulness
Three main types of characteristics
12
Manyconnections

Considerations in measuring user engagement
● short term ←→ long term
● laboratory ←→ “in the wild”
● subjective ←→ objective
● qualitative ←→ quantitative
● large scale ←→ small scale
13(O’Brien, Lalmas & Yom-Tov, 2013)

Methods to measuring user engagement
14
User study (lab/online)
mostly qualitative
User study (lab/online)
mostly quantitative,
scalability an issue
Data study
quantitative, large scale
Cognitive
engagement
objective
Interaction
engagement
objective
Analytics
Data science
Task-based methods
Neurological
Physiological
Questionnaire,
interview, report,
product reaction
cards
Self-reported
engagement
subjective

Scope of this tutorial
Assume that applications are “properly designed”.
Based on “published” work and our experience.
Focus on applications that users “chose” to engage with, widely used by
“anybody” on a “large-scale” and on a mostly regularly basis.
This tutorial is not an “exhaustive” account of all existing works.
16

Towards a taxonomy of
metrics
17

Towards a taxonomy of metrics … Outline
Terminology, context & consideration
Facets of user engagement
Sessions and metrics
Intra-session metrics
Inter-session metrics
Other metrics
Proposed taxonomy
18

Measures, metrics & key performance indicators
19
Measurement:
process of obtaining
one or more quantity
values that can
reasonably be attributed
to a quantity
e.g. number of clicks on
a site
Metric:
a measure is a number
that is derived from
taking a measurement
… in contrast, a metric
is a calculation
e.g. click-through rate
Key performance
indicator (KPI):
quantifiable measure
demonstrating how
effectively key business
objectives are being
achieved
e.g. conversion rate
https://www.klipfolio.com/blog/kpi-metric-measure
a measure can be used as metric but not all metrics are measures
a KPI is a metric but not all metrics are KPIs

Three levels of metrics
Business metrics -- KPIs
Behavioral metrics -- online metrics, analytics
Optimisation metrics -- metrics used to train machine
learning algorithms
20
our focus in this section
Thesethreelevelsareconnected

Why do we need several metrics of online behaviour?
21
Games
Users spend much
time per visit
Search
Users come frequently
and do not stay long
Social media
Users come
frequently and stay
long
Niche
Users come on average
once a week e.g. weekly
post
News
Users come periodically,
e.g. morning and evening
Service
Users visit site, when
needed, e.g. to
renew subscription
Sites differ in their patterns of engagement

A basic taxonomy of metrics … A starting point
22
Popularity #Users Number of distinct users
#Visits Number of visits
#Clicks Number of clicks
Activity Click Depth Average number of page views per visit.
Dwell Time Average time per visit
Loyalty #Active Days Number of days a user visited the site
Return Rate Number of times a user visited the site
(Lehmann et al., 2012)
Capture various facets of engagement

interest-specific
search
media (periodic)
e-commerce
23(Lehmann et al., 2012)
Sites differ in their patterns of engagement … Indeed
Some observations made as part of this study (nothing unexpected but metrics aligned well):
Activity depends on the structure and freshness of the site
Loyalty influenced by external and internal factors (e.g. freshness, current interests, bugs, events)
media (daily)
80 sites, 2M users, 1 month sample

What may impact user engagement?
When?
Temporality
Usage level
24
Why?
Task
Segmentation
Who?
Demographics
Recency
Where?
View
Platform
What?
Function

Engagement varies from weekdays to weekends
hobbies,
interest-specific
weather
daily news
work-related
Temporality … When?

26
Engagement is periodic or may contain peaks
Engagement is influenced by internal (e.g. crash) and external factors (e.g.
major events)
Social network shopping
Temporality … When?
site crash
payday

Periodicity (day) … When?
weekday:
peak during morning
users active later during weekend
than during week
weekend: stable during day
27230K mobile apps, 600M daily unique users, 1 month sample (Van Canneyt etal, 2017)

The UK in shock
The pound crashing
app engagement increase by
114% for finance
43% for news
Dayofthereferendum
result
24June,2016(UK)
External factors (news) … When?
28
(Van Canneyt etal, 2017)

Lower usage on
weekends
EU referendum result (24 June)
Back to normal after 4 weeks
External factors (news) … When?
29
Finance apps (Van Canneyt etal, 2017)

Users take photos
on New Year day
Increased use
of social media
N
ew
yearday
(UK)
External factors (social) … When?
30

● Engagement varies by task
○ A user who accesses a site to check for emails (goal-specific task) has
different engagement patterns from one browsing for leisure.
○ Task has an effect on periodicity
● In one study (Yom-Tov et al, 2013), sessions in which 50% or more of the visited
sites belonged to the five most common sites (for each user) were classified
as goal-specific.
○ Goal-specific sessions accounted for 38% of sessions
○ 92% of users have both goal-specific and non-goal-specific sessions
○ Average “downstream engagement” in goal-specific sessions was lower
compared to non-goal-specific ones
31
Task … Why?

Task (day) … Why?
32

Week: productivity
Weekend: sports
Task (week) … Why?
33

Usage level (how often we see a user) … When?
34
Various definitions of usage level (e.g. #days per month/week, #sessions per week)
Discard tourist users in analysis → unless the focus is on new users
(Lehmann etal, 2012)

Facets of engagement … Several proposals
Factor
focus attention; positive affect; aesthetics; endurability; novelty; richness & control; reputation,
trust & expectation; motivation, interests, incentives & benefits
Degree
involvement, interaction, intimacy, influence
Process
point of engagement, period of engagement, disengagement, re-engagement
Index
click depth, duration, recency, loyalty, brand, feedback, interaction
35

Factor of user engagement (I)
36
Focused attention (Webster
& Ho, 1997; O’Brien, 2008)
•Emotions experienced by user are intrinsically motivating
•Initial affective “hook” can induce a desire for exploration, active
discovery or participation
Positive Affect
(O’Brien & Toms, 2008)
Aesthetics
(Jacques et al, 1995; O’Brien, 2008)
•People remember enjoyable, useful, engaging experiences and
want to repeat them
•Repetition of use, recommendation, interactivity, utility
Endurability
(Read, MacFarlane, & Casey, 2002; O’Brien,
2008)
•Users must be focused to be engaged
•Distortions in subjective perception of time used to measure it
•Sensory, visual appeal of interface stimulates user and promotes
focused attention; perceived usability
•Linked to design principles (e.g. symmetry, balance, saliency)

• Novelty, surprise, unfamiliarity and the unexpected; updates &
innovation
• Appeal to user curiosity; encourages inquisitive behavior and
promotes repeated engagement
Novelty
(Webster & Ho, 1997; O’Brien, 2008)
• Richness captures the growth potential of an activity
• Control captures the extent to which a person is able to achieve
this growth potential
Richness and control
(Jacques et al, 1995; Webster & Ho, 1997)
• Trust is a necessary condition for user engagement
• Implicit contract among people and entities which is more than
technological
Reputation, trust and
expectation (Attfield et al, 2011)
• Why should users engage?
• Friends using it
Motivation, interests,
incentives, and benefits
(Jacques et al., 1995; O’Brien & Toms, 2008)
37
Factor of user engagement (II)

Involvement •Presence of a user on the site
•Measured by e.g. number of visitors, time spent, revisit rate
Interaction •Action of a user on the site
•Measured by e.g. CTR, online transaction, uploaded photos
Intimacy
•Affection or aversion of a user
•Measured by e.g. satisfaction rating, sentiment analysis on social media
&, comments, surveys, questionnaires
Influence •Likelihood that a user advocates
•Measured by e.g. forwarding & sharing, invitation to join
(Forrester Research, June 2008)
38
Degree of engagement

Process of user engagement
• How engagement starts
• Aesthetics & novelty in sync with user interests & contexts
Point of engagement
(O’Brien & Toms, 2008)
• Ability to maintain user attention and interests
• Main part of engagement and usually the focus of study
Period of engagement
• Loss of interests lead to passive usage & even stopping usage
• Identifying users that are likely to churn often undertaken
Disengagement
• Engage again after becoming disengaged
• Triggered by relevance, novelty, convenience, remember past
positive experience sometimes as result of campaign strategy
Re-engagement
(Webster & Ahuja, 2006; O’Brien & Toms, 2008)
39

Point of engagement … Process
Point of engagement relates to acquisition → how users arrive at a site
Which channels users are originating from:
organic search, direct targeting, paid search, referral,
social media, advertising campaign
● is about attracting & acquiring new users
● understand acquisition cost (e.g. important for marketing)
40

Period of engagement … Process
Relates to user behavior with site → per page, per visit, per session
Involvement
pageview, dwell time, playtime (e.g. video)
Interaction
click-through rate, #shares, #likes, conversion rate, #save, bounce rate
Contribution
#blog posts, #comments, #create (e.g. playlist), #replies, #uploads (e.g. photo)
some metrics (e.g. #clicks) are aggregated across visits & sessions → popularity
some metrics (e.g. dwell time) are used as optimisation metrics → optimise the page/visit/session
41
Note that
Interaction
(e.g. share) &
Contribution
(e.g. post) may
have an effect
on Influence

Disengagement … Process
Churn rate measures the percentage of users not returning to site (the
opposite is retention rate)
From day 1 → focus on new users
● Calculated over day (d7, d14), week (w1, w2), and month (d30, d60, d180)
● Apps on average have retention rate of 40% after a month, which can → use as benchmark
● Retaining users over acquiring new ones
Over units of time → all users
● WoW, MoM, YoY
Churn prediction
Treatment (e.g. reduce ads)
Notification & alerts (e.g. email)
42

Re-engagement … Process
Notification
Email
Offer
Marketing
Advertising
...
43https://www.wordstream.com/blog/ws/2016/05/13/targeted-email-marketing-strategies

Index of user engagement
Click Depth Index: page views
Duration Index: time spent on site
Interaction Index: user interaction with site or product (click, upload, transaction)
Recency Index: rate at which users return (frequency)
Loyalty Index: level of long-term interaction the user has with the site or product
Brand Index: apparent awareness of the user of the brand, site, or product (e.g.
search terms, social media posts)
Feedback Index: qualitative information including propensity to solicit additional
information, supply direct feedback (e.g. rating, review)
(Peterson et al., 2008)
44

Time … From visit to session
45
visit visit visit visit visit visit
session session session
Dwell time is time spent on site (page) during a visit
Session length is amount of time user spent on site within the session
Session frequency captures time between two consecutive sessions
session length shows how engaged users are while session frequency shows how often users
are coming back (loyalty)
often 30mn is used as threshold for session boundary

Metrics and their relation to sessions
intra-session metrics
● page level or less
● visit level
● session level
46
visit visit visit visit visit
session session session
visit
● return soon
● remain engaged later on
inter-session metrics
long-term value (LTV) metrics

Intra-session (within → activity) inter-session (across → loyalty)
Involvement
• Dwell time
• Session duration
• Page view (click depth)
• Revisit rate
• Bounce rate
Interaction
• Click through rate (CTR)
• Number of shares & likes (social
& digital media)
• Conversion rate (e-commerce)
• Streamed for more that x
seconds
Contribution
• Number of replies
• Number of blog posts
• Number of uploads
Granularity
Module
↓
Viewport
↓
Page
↓
Visit
↓
Session
From one session to the next session (return soon)
• Time between sessions (absence time)
From one session to a next time period such next week,
or in 2 weeks time (remain engaged later on)
• Number of active days
• Number of sessions
• Total usage time
• Number of clicks
• Number of shares
• Number of thumb ups
• ...
Intra- vs inter-sessions metrics
• intra-session engagement measures user activity on the site during the session
• inter-session engagement measures user loyalty with the site

Intra- vs inter-sessions metrics … Granularity
Intra-session metrics
Module → Viewport → Page → Visit → Session
Optimisation mostly with these metrics, with
increasing complexity from “Module” to “Session”
Next session → Next Day → Next Week → Next Month, etc.
48

Examples of intra-session metrics
● Measures success in keeping user engaged during the session
○ clicking, spending time, adding content, making a purchase
○ user may leave the site but return within the same session
● Involvement, Interaction, Contribution
49

Ratio of users who click on a specific link to the number of total
users who view a page, email, or advertisement
Translates to play song/video at least x seconds
for music/video sites/formats
● Relate to abandonment rate
● Issues include clickbait, site design
50
Click-through rates (CTR) … Interaction

The contiguous time spent on a site
or web page
Similar measure is play time for video
and music sites
● Not clear what user is actually
looking at while on page/site
● Instrumentation issue with last
page viewed and open tabs
51
Dwell time … Involvement
Distribution of dwell times on 50 websites

● Dwell time varies by
site type: leisure sites
tend to have longer dwell
times than news,
e-commerce, etc.
● Dwell time has a
relatively large
variance even for the
same site
52
Dwell time … Involvement
Average and variance of dwell time of 50 sites

Pageview … Involvement
Page view is request to load a single page
Number of pages viewed (click depth): average number of contiguous pages
viewed during a visit
Reload after reaching the page → counted as additional pageview
If same page viewed more than once → a single unique pageview
53
Can be problematic with
ill-designed site as high click depth
may reflect users getting lost and
user frustration.
https://www.slideshare.net/timothylelek/google-analytics-for-dummies

Social media metrics
54
Conversations
#comments, #posts,
#replies, #edits
Applause
#like, #thumbs up or
down, #hearts, +1
Amplification
#share, #mail
Metrics specific to user generated
content sites such as social
platforms, including social
networks, blogs, wiki, etc.
…
interaction
…
interaction …
contribution
https://www.kaushik.net/avinash/best-social-media-metrics-conversation-amplification-applause-economic-value/

Conversion rate … Interaction
Fraction of sessions which end in a desired user action
particularly relevant to e-commerce (making a purchase) … but also include subscribing, free to
premium user conversion
Online advertising using conversion as cost model to charge
advertisers
Not all sessions are expected to result in a conversion, so this
measure not always informative
dwell time often used as proxy of satisfactory experience as may reflect affinity with the brand
55

Number of returns to the website within a
session
Common in sites which may be browser
homepages, or contain content of regular
interest to users.
Useful for sites such as news aggregators,
where returns indicate that user believes
there may be more information to glean
from the site
56
Revisit rates … Involvement

Goal-oriented sites (e.g.,
e-commerce) have lower
revisits in a given time
range observed → revisit
horizon should be
adjusted by site
57
Revisit rates … Involvement

Revisit rate … Session length
2.5M users, 785M page views, 1 month
sample
Categorization of the most frequent
accessed sites
11 categories (e.g. news), 33 subcategories
(e.g. news finance, news society)
60 sites from 70 countries/regions
short sessions: average 3.01 distinct sites visited with revisit rate 10%
long sessions: average 9.62 distinct sites visited with revisit rate 22%
58

Time between each revisit
50% of sites are
revisited after less
than 1 minute
59

Some final words on intra-session metrics
Metrics for smaller granularity levels such as
viewport or specific section → attention
Metrics for scroll → important for stream and
mobile
60
Whether an intra-session metric belongs to
Involvement, Interaction, or Contribution may depend
on the expected type of engagement of the site
(e.g. sharing may mean very different things on social media vs news
sites)
scrollingdown
viewport

Inter-session metrics → Loyalty
How many users and how fast they return to the site
Total use measurements → Popularity
Total usage time
Total number of sessions
Total view time (video)
Total number of likes (social networks)
Direct value measurement → Lifetime value
Lifetime value, as measured by ads clicked, monetization, etc.
61
Non intra-session metrics

Examples of inter-session metrics
62
Loyalty is about having users return to the site again and again, and to
perceive the site as beneficial to them
● Return soon
● Remain engaged later on

Intra-session measures can easily
mislead, especially for a short time
● Consider a very poor ranking
function introduced into a search
engine by mistake
● Therefore, bucket testing may
provide erroneous results if only
intra-session measures are used
63
Why inter-session metrics?
(Kohavi et al, 2012)

64
vsit
nextsession
visit
visit
visit
visit
visit
absence time next day, next
week, next
month, etc
Total number of visits or sessions
Total number of days active
Total number of clicks
Total amount of time spent ...
a day, a week, 2 weeks, a month, etc

Select intra-session
metrics
Decide the
inter-session
metric
System
Models
Features
Which intra-session
metric is good
predictor of
inter-session metric
Optimize for the
identified
intra-session
Lots of data
required to
remove noise
What is a signal?
What is a metric?
What is a feature?
65
Intra- vs inter-sessions metrics … Otimisation

Related off-site content
Example I: Focused news reading
Off-site link → absence time
Providing links to related off-site content has a
positive long-term effect (for 70% of news sites,
probability that users return within 12 hours increases by
76%)
66

Example II: Ad landing page quality
Dwell time → ad click
Positive post-click experience
(“long” clicks) has an effect on
users clicking on ads again
(mobile)
(Lalmas etal, 2015)
User click on an ad → ad landing page
Dwell time is time until user returns to publisher
and used as proxy of quality of landing page
67

Other metrics
68
● Popularity
● Long-term value (LTV)

Popularity metrics
With respect to users
● MAU (monthly active users), WAU (weekly active users), DAU (daily active users)
● Stickiness (DAU/MAU) measures how much users are engaging with the product
● Segmentation used to dive into demographics, platform, recency, …
With respect to usage
● Absolute value metrics (measures) → aggregates over visits/sessions
total number of clicks; total number of sessions; total number of time spent per day,
month, year
● Usually correlate with number of active users
69

Long-term value (LTV) metrics
70
How valuable different users are based on lifetime performance → value that
a user is expected to generate over a given period time, e.g. such as 12
months
● Services relying on advertising for revenue:
○ based on a combination of forecasted average pageviews per user, actual retention &
revenue per pageview
● E-commerce relying on actual purchases (CLV):
○ based on total amount of purchases
Help analyzing acquisition strategy
(customer acquisition cost) and
estimate further marketing costs
https://www.apptamin.com/blog/lifetime-value-mobile-customer/

Taxonomy of metrics …. in two slides
71
User journey
day 1, day 2, … , week 1, ... now
Acquisition→
retainingnewusers
Intra-session
Involvement
Interaction
Contribution
Optimisation
Aggregates → popularity
Intra-session
Involvement
Interaction
Contribution
Optimisation
Aggregates → popularity
inter-session → loyalty
Disengagement?
Re-engagement?
correlation
prediction

Taxonomy of metrics …. in two slides
72
Metrics to use to
optimize machine
learning algorithms
Key performance
indicators (KPIs)
Long-term value
(LTV) metrics
Popularity metrics

Experimentation and
Evaluation of Metrics
73

learning algorithms
74

Why experiments
76
Common reasons of not having experiments
● Let’s launch and see what happens and compare metrics before & after.
Usually in the context of all kinds of product innovations, aiming fast iterations.

Why experiments
77
● Too risky.
Usually in the context of ads, exploration & exploitation and etc.

Why experiments
78
● Too risky.
Usually in the context of ads, exploration & exploitation and etc.
● Historical data can’t represent future.
Usually in the context of offline experiments
...

Why experiments
79
Main benefits of having experiments
● Metrics can be measured, tracked and compared.

Why experiments
80
● We can learn, improve and optimize.

Why experiments
81
● We can learn, improve and optimize.
● Save time and faster iterations.
...

Experiments and Non-Experiments
82

83
Sometimes, experiments may not be feasible or practical.

84
● Example 1:
We want to test which “Add to Cart” button may lead to more Monthly-Active-Users (MAUs).

85
● Example 2:
We want to test which search ranking algorithm may lead to higher Year-Over-Year Changes of
user search sessions.

86
Intra-Session Metrics Inter-Session Metrics
Experimentable Non-Experimentable

Experiments
87
Summary
● Run experiments as much as possible.
● Understand experimentable and non-experimentable.

Experiments
88
Summary
● Run experiments as much as possible.
● Understand experimentable and non-experimentable.
● Bias: almost always indicates temporal, spatial and population sampling.
● Conclusions: almost always needs inference.

Types of experiments
90
● Online Experiment
● Offline Experiment
● Offline A/B Experiment

Online experiment
91
A/B Tests or Bucket Tests or Online Controlled Experiments

Online experiment
92
● Have deep roots in classic statistics, with new challenges.
e.g., “always need more traffic”

Online experiment
93
● Can derive causal relationships easier.
e.g., complex user behaviors

Online experiment
94
● Have direct impact on users.
e.g., users may decide not to come back

Online experiment
95
● Have direct impact on users.
e.g., users may decide not to come back
● Cannot easily be reused.
e.g., need to re-launch the experiment

Online experiment
96

Online experiment
97

Online experiment
98
Metrics for Online Experiments
● Directional
Have correlations with inter-session metrics and KPIs.

Online experiment
99
Metrics for Online Experiments
● Directional
Have correlations with inter-session metrics and KPIs.
● Sensitivity
Easily detect changes.

Online experiment
100
Summary
● Direct and dynamic
● Causality
● Metrics for online experiments
● Impacts (e.g, user engagement, traffic, set-up and etc.)
● Cannot re-use
References:
[1] Ron Kohavi, Roger Longbotham, Dan Sommerfield, and Randal M. Henne. 2009. Controlled
Experiments on the Web: Survey and Practical Guide. DMKD 18, 1 (February 2009).
[2] Alex Deng and Xiaolin Shi. 2016. Data-Driven Metric Development for Online Controlled
Experiments: Seven Lessons Learned. KDD 2016.
[3] Pavel Dmitriev, Somit Gupta, Dong Woo Kim, and Garnet Vaz. 2017. A Dirty Dozen: Twelve
Common Metric Interpretation Pitfalls in Online Controlled Experiments. KDD 2017.

Offline experiment
101
Traditional Offline Dataset/Collection Experiment
● High risk experiments.
It may drive users away.

Offline experiment
102
● Learn more insights & highly reusable.
Easy to gather data and easy to compute metrics and compare.

Offline experiment
103
● Learn more insights & highly reusable.
Easy to gather data and easy to compute metrics and compare.
● Machine learning theory of generalization.
Textbook scenario.

Offline experiment
104
Train Validation Test
Offline Online

Offline experiment
105
● Selection/sampling bias
e.g. presentation bias, system bias
● Correlation
e.g. hard to control everything
● Static
e.g., temporal dynamics, lacking “new” user behaviors

Offline experiment
106
Summary
● Indirect and can be reused
● Good machine learning theories
● Correlation
● Static
References:
[1] Mark Sanderson (2010). Test Collection Based Evaluation of Information Retrieval Systems.
Foundations and Trends® in Information Retrieval: Vol. 4: No. 4.
[2] Donna Harman (2011). Information Retrieval Evaluation. Synthesis Lectures on Information
Concepts, Retrieval, and Services 3:2.

Offline A/B Experiment
107
Counterfactual Offline Reasoning/Experiment

108
Logging Policy
● Uniform-randomly show items.
● Gather user feedbacks (rewards).
New Policy
● Show items according to a model/algorithm.
● Accumulate rewards if item matches history pattern.
References:
[1] Lihong Li, Wei Chu, John Langford, and Xuanhui Wang. Unbiased Online Evaluation of
Contextual-bandit-based News Article Recommendation Algorithms. In WSDM 2011.
[2] Alexander L. Strehl, John Langford, Lihong Li, and Sham M. Kakade. Learning from Logged Implicit
Exploration data. In NIPS 2010.

109
References:
Contextual-bandit-based News Article Recommendation Algorithms. WSDM 2011.

110
References:

111
● Address data bias
● Causality
● Reusable
● Some good theories

112
● Generalization to Non-uniform Logging/Exploration

113
● Generalization to Non-uniform Logging/Exploration

114
● Need logging and an exploration strategy
● In development, emerging topic

115
How to effectively gather data that minimize hurting user engagement metrics?
Reference:
[1] Liangjie Hong, Adnan Boz. An Unbiased Data Collection and Content Exploitation/Exploration
Strategy for Personalization. CoRR abs/1604.03506 (2016).

116
● Uniform-random greatly hurts user engagement and nobody is doing this.
● Classic Thompson Sampling and Upper-Confidence-Bound would eventually converge.

117
Requirements:
● Provide randomness and do not converge.
● User-friendly.

118
Requirements:
● Provide randomness and do not converge.
● User-friendly.

119

120

121

122

123
Summary
● Causality
● Reusable
● Need logging and an exploration strategy
● In development, emerging topic
References:
[1] Lihong Li, Jinyoung Kim, Imed Zitouni: Toward Predicting the Outcome of an A/B Experiment
for Search Relevance. WSDM 2015.
[2] Adith Swaminathan et al. Off-policy evaluation for slate recommendation. NIPS 2017.
[3] Tobias Schnabel, Adith Swaminathan, Peter I. Frazier, and Thorsten Joachims. 2016. Unbiased
Comparative Evaluation of Ranking Functions. ICTIR 2016.

Evaluation of Metrics
124
● Hypothesis Testing
● Causal Inference

Hypothesis Testing
125
Statistical Comparison
● Well grounded theory for classic cases
● Not well studied in a lot of online settings
● Gold standard for statistical difference
● Weak for practical difference
References:
[1] Ben Carterette. Statistical Significance Testing in Information Retrieval: Theory and Practice.
SIGIR 2017 Tutorial.
[2] Tetsuya Sakai. Statistical Significance, Power, and Sample Sizes: A Systematic Review of
SIGIR and TOIS, 2006-2015. SIGIR 2016.
[3] Tetsuya Sakai. The Probability that Your Hypothesis Is Correct, Credible Intervals, and Effect
Sizes for IR Evaluation. SIGIR 2017.
[4] Benjamin A. Carterette. Multiple Testing in Statistical Analysis of Systems-based Information
Retrieval Experiments. ACM Trans. Inf. Syst. 30, 1, Article 4 (March 2012), 34 pages.

Causal Inference
126
Statistical Relationship
● Emerging topics between statistics and machine learning
● Well grounded theory for classic cases
● Easy for simple cases
● Not well studied in a lot of online settings
● Difficult for complex scenarios
References:
[1] David Sontag and Uri Shalit. Causal Inference for Observational Studies. ICML 2016 Tutorial.
[3] Lihong Li, Jin Young Kim, and Imed Zitouni. Toward Predicting the Outcome of an A/B
Experiment for Search Relevance. WSDM 2015.

Metrics, Evaluation and Experiments
127
The relationships between metrics, evaluation and experiments
● Requiring certain user behaviors
○ e.g., NDCG, AUC, Precision, Recall,...
● Decomposition assumption
○ e.g., Conversion Rate, Click-Through-Rate,...
● Naturally missing/partial data
○ e.g., Dwell-time, View, Scroll,...

learning algorithms
129

Optimisations for Metrics
130
● Offline Experiments → Offline Optimization
● Online Experiments → Online Optimization
● Offline A/B Experiments → Counterfactual Optimization
● From Intra-Session to Inter-Session Metrics Optimization

Offline Optimization
131
● Supervised Learning
● Cross-validation
● View online experiments as extension to offline optimization (testset)
Offline Online

132
It doesn’t work or it doesn’t work smoothly.

● Bias
Examples: presentation bias, system bias...
133
Offline Online

● Concept Drifts
Examples: seasonal, interest shift...
134
Offline Online

● Different of offline metrics and online metrics
Examples: AUC/nDCG versus DAU...
135
Offline Online

136
● Bias
● Concept Drift
● Different of offline metrics and online metrics
Offline Online

Online Optimization
138
● Online Learning
● Contextual Multi-armed Bandit
● Reinforcement Learning

Online Optimization
139
How to Online Optimize User Intra-Session Metrics and Inter-Session Metrics
Reference:
[1] Qingyun Wu, Hongning Wang, Liangjie Hong, and Yue Shi. 2017. Returning is Believing:
Optimizing Long-term User Engagement in Recommender Systems. In CIKM 2017.

Online Optimization
140
● Most algorithms focus on intra-session effects (e.g., clicks, dwell, etc.).
[1] Abhinandan S. Das, Mayur Datar, Ashutosh Garg, and Shyam Rajaram. Google News
Personalization: Scalable Online Collaborative Filtering. In WWW 2007.
[2] Y. Koren, R. Bell, and C. Volinsky. Matrix Factorization Techniques for Recommender
Systems. Computer 42, 8 (2009), 30–37.

Online Optimization
141
● Most algorithms focus on intra-session effects (e.g., clicks, dwell, etc.).
[1] Abhinandan S. Das, Mayur Datar, Ashutosh Garg, and Shyam Rajaram. Google News
Personalization: Scalable Online Collaborative Filtering. In WWW 2007.
[2] Y. Koren, R. Bell, and C. Volinsky. Matrix Factorization Techniques for Recommender
Systems. Computer 42, 8 (2009), 30–37.
● Users may leave because of boredom from popular items.
[1] Komal Kapoor, Karthik Subbian, Jaideep Srivastava, and Paul Schrater. Just in Time
Recommendations: Modeling the Dynamics of Boredom in Activity Streams. In WSDM
2015.

Online Optimization
142
● Users may have high immediate rewards but accumate linear regret after they leave.
● Predict a user’s immediate reward, but also project it onto future clicks, making
recommendation decisions dependent over time.
● Rapid change of environment requires this kind of decisions online.

Online Optimization
143
Some more related work about modeling users’ post-click behaviors:
[1] Nicola Barbieri, Fabrizio Silvestri, and Mounia Lalmas. Improving Post-Click User Engagement
on Native Ads via Survival Analysis. In WWW 2016. 761–770.
[2] Mounia Lalmas, Jane.e Lehmann, Guy Shaked, Fabrizio Silvestri, and Gabriele Tolomei. Promoting
Positive Post-Click Experience for In-Stream Yahoo Gemini Users. In KDD 2015.
[3] Nan Du, Yichen Wang, Niao He, Jimeng Sun, and Le Song. Time-Sensitive Recommendation
From Recurrent User Activities. In NIPS 2015.
[4] Komal Kapoor, Mingxuan Sun, Jaideep Srivastava, and Tao Ye. A Hazard Based Approach to
User Return Time Prediction. In KDD 2014.

Online Optimization
144
Balance between
1. Maximize immediate reward of the recommendation

Online Optimization
145
Balance between
2. Explore other possibilities to improve model estimation.

Online Optimization
146
Balance between
2. Explore other possibilities to improve model estimation.
3. Maximize expected future reward by keeping users in the system.
To maximize the cumulative reward over time, the system has to make users click more and return
more often.

Online Optimization
147
Some more related work about multi-armed bandit:
[1] Lihong Li, Wei Chu, John Langford, and Robert E Schapire. A contextual Bandit Approach to
Personalized News Article Recommendation. In WWW 2010.
[2] Lihong Li, Wei Chu, John Langford, and Xuanhui Wang. 2011. Unbiased Online Evaluation of
Contextual-bandit-based News Article Recommendation Algorithms. In WSDM 2011.

Online Optimization
148
Main Idea

Online Optimization
149
Main Idea
● Model how likely an item would yield an immediate click:
[1] At iteration i, if we recommend item ai
, how likely it is going to be clicked by user u.

Online Optimization
150
Main Idea
● Model future visits after seeing this item and their expected clicks:
[2] At iteration i+1, what do we recommend.
[3] How that decision would impact the click behavior at i+1
[4] Future return probability at i+2, and
So on…

Online Optimization
151
Main Idea
● Model future visits after seeing this item and their expected clicks:
[2] At iteration i+1, what do we recommend.
[3] How that decision would impact the click behavior at i+1
[4] Future return probability at i+2, and
So on…
Can be formulated in a reinforcement learning setting.

Online Optimization
152
A Major Challenge:
future candidate pool undefined, thus standard reinforcement learning can’t apply.

Online Optimization
153
A Major Challenge:
future candidate pool undefined, thus standard reinforcement learning can’t apply.
Need approximations.

Online Optimization
154
Approximations

Online Optimization
155
Approximations
1. Future clicks depend on users. (Strong? or not)

Online Optimization
156
Approximations
2. Only model finite steps in future, or even just one step ahead.

Online Optimization
157
Approximations
3. Only model whether the user return in a finite horizon.

Online Optimization
158
Approximations
3. Only model whether the user return in a finite horizon.
New Objective:

Online Optimization
159
Model Summary
1. Use Generalized Linear Model (Bernoulli) to model how likely a user u would click on an item
ai
at iteration i.

Online Optimization
160
Model Summary
ai
at iteration i.
2. Use Moving Average to model a user u’s marginal click probability.

Online Optimization
161
Model Summary
ai
at iteration i.
3. Use Generalized Linear Model (Exponential) to model a user u’s return time intervals.

Online Optimization
162
Model Summary
ai
at iteration i.
4. Use Upper Confidence Bound (UCB) on top of [1-3].

Online Optimization
163
Model Summary
ai
at iteration i.
4. Use Upper Confidence Bound (UCB) on top of [1-3].
Note that both [1] and [3]’s coefficients are personalized.

Online Optimization
164

Online Optimization
165

Online Optimization
166
Simulations
1. Type 1: items with high click probability but short expected return time;
2. Type 2: items with high click probability but long expected return time;
3. Type 3: items with low click probability but short expected return time;
4. Type 4: items with low click probability and long expected return time.

Online Optimization
167
Simulations

Online Optimization
168
Simulations

Online Optimization
169
Simulations

Online Optimization
170
Real-World Dataset
● Collect 4 weeks of data from Yahoo news portal.
● Reduce features into 23 by PCA.
● Sessionized the data by 30 mins.
● Return time is computed by time interval between two sessions.
● Total:
-- 18,882 users,
-- 188,384 articles
-- 9,984,879 logged events, and
-- 1,123,583 sessions.

Online Optimization
171
Real-World Dataset

Online Optimization
172
Real-World Dataset: Evaluation
● Cumulative clicks over Time
● Click-through Rate (CTR)
● Average Return Time
● Return Rate
● Improved User Ratio
● No return Count

Online Optimization
173

Online Optimization
174
Real-World Dataset: Word Cloud

Counterfactual Optimization
175
● Emerging topics
● Optimization under counterfactual setting, simulating A/B testing
References:
[1] Xuanhui Wang, Michael Bendersky, Donald Metzler, Marc Najork. Learning to Rank with
Selection Bias in Personal Search. SIGIR 2016.
[2] Thorsten Joachims, Adith Swaminathan, and Tobias Schnabel. Unbiased Learning-to-Rank with
Biased Feedback. WSDM 2017.
[3] Thorsten Joachims, Adith Swaminathan. Counterfactual Evaluation and Learning for Search,
Recommendation and Ad Placement. SIGIR 2016 Tutorial.
[4] Adith Swaminathan, Thorsten Joachims. Counterfactual Risk Minimization: Learning from
Logged Bandit Feedback. ICML 2015.

Counterfactual Optimization
176
Generic Idea:
1. Rewrite the objective function with inverse propensity scoring.
2. Try to optimize or approximate the new objective.
References:
[1] Xuanhui Wang, Michael Bendersky, Donald Metzler, Marc Najork. Learning to Rank with
Selection Bias in Personal Search. SIGIR 2016.
[2] Thorsten Joachims, Adith Swaminathan, and Tobias Schnabel. Unbiased Learning-to-Rank with
Biased Feedback. WSDM 2017.
[3] Thorsten Joachims, Adith Swaminathan. Counterfactual Evaluation and Learning for Search,
Recommendation and Ad Placement. SIGIR 2016 Tutorial.
[4] Adith Swaminathan, Thorsten Joachims. Counterfactual Risk Minimization: Learning from
Logged Bandit Feedback. ICML 2015.

Optimization Inter-Session Metrics
177

178
Approach I
If inter-session metrics can be explicitly modeled or write them down in their clear form, you can
use online optimization tools to directly optimize them.

179
Approach I
If inter-session metrics can be explicitly modeled or write them down in their clear form, you can
use online optimization tools to directly optimize them.
● This is usually difficult or impossible because of
○ Complexity of inter-session metrics (you can’t really write them down or hard).
○ You don’t have data.
○ Your have extremely sparse data.
○ Hard to deploy such systems.
...

180
Approach II
Optimization
Correlation/Causation

Approach II
1. Intra-Session and Inter-Session Correlation
2. Optimization Intra-Session as Surrogate
3. Finding (Better) Proxy Metrics
Optimization
Correlation/Causation
181

182
Optimization Correlation/Causation

183
Beyond Clicks: Dwell Time in Personalization
Reference:
[1] Xing Yi, Liangjie Hong, Erheng Zhong, Nanthan Nan Liu, and Suju Rajan. 2014. Beyond Clicks:
Dwell Time for Personalization. In RecSys 2014.

184

185

186

187

188

189
● Optimizing Dwell-Time becomes the de-facto method to drive user engagement in Yahoo News
Stream.
● The inter-session user engagement metric is a variant of dwell-time on sessions, considering
the depth of the session.
● They correlate very well in quarterly basis.

190
Summary
● Approach I, Direct Optimization
● Approach II, Correlation and Optimization

Is this a good search engine?
192There is a rich history in evaluating ranking algorithms in information retrieval and web search

How to evaluate a search engine
Coverage
Speed
Query language
User interface
User happiness
○ Users find what they want and return to the search engine for their next
information need → user engagement
But let us remember:
○ In carrying out a search task, search is a means, not an end
Sec. 8.6
(Manning, Raghavan & Schütze, 2008; Baeza-Yates & Ribeiro-Neto, 2011) 193

Evaluating the relevance of a search engine result
Most common proxy: relevance of retrieved results
Sec. 8.1
Relevant
Retrieved
all items
User information need translated into a
query
Relevance assessed relative to
information need not the query
Example:
Information need: I am looking for tennis holiday in a
beach resort with lots of places to eat seafood
Query: tennis academy beach seafood
Evaluation measures:
● precision, recall, R-precision; precision@n;
MAP; F-measure; …
● bpref; nCDG; rank-biased precision,
expected reciprocal rank,, ...
precision
recall
194

Most common proxy: relevance of retrieval results
Sec. 8.1
Explicit signals
Test collection methodology (TREC, CLEF, NCTIR, …)
Human labeled corpora
Crowdsourcing
Implicit signals
User behavior in online settings (clicks, skips, dwell time)
Explicit and implicit signals can be used together
195
Evaluating the relevance of a search engine result
An important question:
when is signal a metric and when is it a feature of the ranking (machine learning) algorithm?

Examples of implicit signals … measures … metrics
Number of clicks
SAT click
Quick-back click
Click at given position
Time to first click
Skipping
Abandonment rate
Number of query reformulations
Dwell time
Hover rate
196
An important question:
when is signal a metric and when is it a feature of the ranking (machine learning) algorithm?

What is a happy user in search?
1. The user information need is satisfied
2. The user has learned about a topic and even about other topics
3. The system was inviting and even fun to use
Intra-session
The actual search session
Inter-session
Users come back soon and frequently
197

Evaluating the actual search session … Metrics
198
Mean average precision (MAP)
Number of clicks or CTR
Dwell time
Well established metrics of engagement with search results
Used as metrics to optimize in ranking algorithms
Also can be used as features in ranking algorithms
But how do they relate to user engagement?
→ inter-session consideration

MAP … User satisfaction
(Turpin & Scholer, 2006)
Similar results obtained with P@2, P@3, P@4 and P@10
Precision-basedsearch
199

(Turpin & Scholer, 2006)
Recall-basedsearch
200
MAP … User satisfaction

I just wanted the phone number … I am totally happy
No click … User satisfaction
201

Cickthrough rate:
% of clicks when URL
shown (per query)
Hover rate:
% hover over URL
(per query)
Unclicked hover:
Median time user hovers over
URL but no click (per query)
Max hover time:
Maximum time user hovers
over a result (per SERP)
(Huang et al, 2011)
202

Abandonment is when there is no click on the search result page
User is dissatisfied (bad abandonment)
User found result(s) on the search result page (good abandonment)
858 queries (21% good vs. 79% abandonment manually examined)
Cursor trail length
Total distance (pixel) traveled by cursor on SERP
Shorter for good abandonment
Movement time
Total time (second) cursor moved on SERP
Longer when answers in snippet (good abandonment)
(Huang et al, 2011) 203
Cursor speed
Average cursor speed (pixel/second)
Slower when answers in snippet (good abandonment)

“reading” cursor heatmap of relevant document vs “scanning” cursor
heatmap of non-relevant document (both dwell time of 30s) (Guo & Agichtein, 2012)
Dwell time … User satisfaction
204

“reading” a relevant long document vs “scanning” a long non-relevant document
(Guo & Agichtein, 2012) 205
Dwell time … User satisfaction

From intra- to inter-session metrics … We recall
What you want
to optimize for
each task,
session, query
intra-session
metrics
inter-session
metrics
What you want
to optimize
long-termSearch
system
Models
Features
206

From intra- to inter-session metrics
207
● Number of clicks
● Time to 1st
lick
● Skipping
● Dwell time
● Click through rate
● Abandonment rate
● Number of query
reformulations
● Hover rate
● …
Intra-session metrics for search
(Proxy: relevance of search results)
userssatisfiedwiththesearchsessionare
likelytoreturnsoonerandfrequentlytothe
searchengine
Inter-session metrics for search
● Absence time
● Number of search
sessions in next 2 weeks
● Number of queries next
day
● ...
Absence time on Yahoo Japan (Dupret & Lalmas, 2013)
Absence time on Bing (Chakraborty etal, 2014)
Dwell time & search engine re-use (Hu etal, 2011)

Search result page for “asparagus” … Study I
208

Another search result page for “asparagus”
209

Absence time and survival analysis
SURVIVE
DIE
DIE = RETURN TO SITE ➔SHORT ABSENCE TIME
hours
210

Absence time applied to search … Study I
Ranking functions on Yahoo Answer Japan
Two-weeks click data on Yahoo Answer Japan search
One millions users Six ranking functions
Session boundary:
30 minutes of
inactivity
211(Dupret & Lalmas, 2013)

DCG versus absence to evaluate five ranking functions
DCG@1
Ranking Alg 1
Ranking Alg 2
Ranking Alg 3
Ranking Alg 4
Ranking Alg 5
DCG@5
Ranking Alg 1
Ranking Alg 3
Ranking Alg 2
Ranking Alg 4
Ranking Alg 5
Absence time
Ranking Alg 1
Ranking Alg 2
Ranking Alg 5
Ranking Alg 3
Ranking Alg 4
(Dupret & Lalmas, 2013)
212

survival analysis: high hazard rate (die quickly) = short absence
5 clicks
control=noclick
Absence time and number of clicks
3 clicks
No click means a bad user search session … in Yahoo Japan search
Clicking between 3-5 results leads to same user search experience
Clicking on more than 5 results reflects poor user search session; users cannot find what they are looking for
(Dupret & Lalmas, 2013)
213

Absence time and search session … What else?
● Clicking lower in the ranking (2nd
, 3rd
) suggests more careful choice
from the user (compared to 1st
)
● Clicking at bottom is a sign of low quality overall ranking
● Users finding their answers quickly (time to 1st
click) return sooner
to the search application
● Returning to the same search result page is a worse user
experience than reformulating the query
intra-session search metrics → absence time
(Dupret & Lalmas, 2013) 214

Absence time and search experience … Study II
From 21 experiments carried out through A/B testing, using absence
time agrees with 14 of them (which one is better)
(Chakraborty et al., 2014)
Positive
One more query in session
One more click in session
SAT clicks
Query reformulation
Negative
Abandoned session
Quick-back clicks
215

Absence time and search experience … Studies I & II
Demonstrated that absence time is an appropriate inter-session metric for
search because of the correlation & predictive power of known indicators of
a positive search experience
These known indicators could act as intra-session metrics, which could be
optimised by the ranking algorithms
They can also be used as features in the ranking algorithms themselves
216

Application: E-commerce
219
● Search
● Recommendation
● Advertising

220
● Search
● Recommendation
● Advertising
● Shopping
● Discovery
...

222
● Search
○ Generic search v.s. E-commerce search
○ Relevance
○ Revenue
○ Diversity
○ Discovery
● Recommendation
○ Rating/favorite prediction
○ Clicks and purchase funnel
○ Revenue
○ Seasonal
○ Occasion
○ Inventory
● Advertising
○ Two-sided marketplace

223
● Search
● Recommendation
● Advertising
● How to measure
● How to optimize

224
● Discovering Styles for Recommendation in E-Commerce
● Reference:
[1] Diane J. Hu, Rob Hall, and Josh Attenberg. Style in the Long Tail: Discovering Unique
Interests with Latent Variable Models in Large Scale Social E-commerce. In KDD 2014.

225

226

227

228

229

230

231

232

233

234

235

236
● How to measure the success of recommender systems in E-commerce?
● How to construct unified framework to optimize recommendation in different modules/pages?
● How to measure style, quality, artistic…?
...

237
Optimizing Gross-Merchandise-Value (GMV) in E-commerce Search

238
● Liang Wu, PhD Student from Arizona State University
● Diane Hu, Staff Data Scientist at Etsy
● Liangjie Hong, Head of Data Science at Etsy

239
● Expected GMV

240
● Purchase Decision Process

241
● Click Decision(s) from Search-Result-Page (SERP)
● Purchase Decision(s) from Listing Page

242
● Click Decision(s) from Search-Result-Page (SERP)
fc
is learned by a neural-network model through back-prop.

243
● Purchase Decision from Listing Page
Price-Weighted Logistic Regression

244

245

246

247

248

249
● This work is about optimizing GMV in Session
○ How about long-term GMV?
○ How about other discovery?
…
● First step in optimizing user engagements in E-commerce search.

Recap
251
● Introduction and Scope
● Towards a Taxonomy of Metrics
● Experimentation and Evaluation of Metrics
● Optimisation for Metrics
● Applications
○ Search
○ E-commerce

Challenges
252
● How to systematically discover new metrics?
● How to measure metrics (metrics of metrics)?
● How to quantify users’ holistic feelings?
● Can we learn metrics?
● Advance methodologies to optimize intra- and inter-session metrics.

Tutorial on metrics of user engagement -- Applications to Search & E- commerce

More Related Content

What's hot (20)

Similar to Tutorial on metrics of user engagement -- Applications to Search & E- commerce (20)

More from Mounia Lalmas-Roelleke (20)

Recently uploaded (20)

Tutorial on metrics of user engagement -- Applications to Search & E- commerce