Presentation from the International Conference on Data Mining 2013. Dallas, USA.

1. MINING USER LIFECYCLES
FROM ONLINE COMMUNITY
PLATFORMS AND THEIR
APPLICATION TO CHURN
PREDICTION
DR. MATTHEW ROWE
SCHOOL OF COMPUTING AND COMMUNICATIONS
@MROWEBOT | M.ROWE@LANCASTER.AC.UK
International Conference on Data Mining 2013
Dallas, USA

2. Identity Development: Offline
1
Development
happens through
stages
Development =
conflicts
Mining User Lifecycles from Online Community Platforms and their Application to Churn Prediction

3. User Development: ‘Online’
2
¨
Recently studied in isolated dimensions:
¤ Socially
(Telecoms Networks: Miritello et al. 2013)
n Communication
networks tend to a capacity
¤ Lexically
(Online Communities: Danescu-Niculescu-Mizil
et al. 2013)
n Language
¨
adapts to the community, before diverging
Without analysing development:
a)
b)
Relative to earlier signals
Relative to the community of interaction
Mining User Lifecycles from Online Community Platforms and their Application to Churn Prediction

4. Understanding User Development
enables…
3
work (more later)
Jul
Sep
Nov
A
(b) T
0.8
Entropy
Period Entropy
Community Entropy
In−degree
Out−degree
Lexical
All
0.2
0.4
0.6
Figure 3: Average rat
moving average of the
categories.
0.0
of this talk
True Positive Rate
n Focus
churners from development signals
1.0
Churn Prediction
¤ Forecast
Mar
Time
(a) Lens
2.
7.0
8.0
May
6.0
Average Rating
3.8
3.6
3.4
Directorial Debut Films
1990s Comedy Films
5.0
n Current/future
Average Rating
Stage-based user
neighbourhoods (e.g. user-kNN)
¤ Modelling taste evolution (e.g. biases in MF)
3.2
¤ Developmental
4.0
Recommender Systems
3.0
1.
for MovieLens the scores re
Movie Tweetings ‘Independe
rating and ‘Directorial Debu
rating over time. Such info
the biases of the recommen
stability of a given bias in
made: i.e. considering the
and how this relates to pre
0.0
Mining User Lifecycles from Online Community Platforms and their Application to Churn Prediction
5.
0.2
0.4
0.6
0.8
1.0
ANALYSING TA
False Positive Rate
Analysing the evolution a
allows one to understand h

5. Outline
4
Datasets: Online Community Platforms
¨ Defining User Lifecycles and Properties
¨ Mining Lifecycle Trajectories
¨ Predicting Churners
¨ Findings and Conclusions
¨ Future Work
¨
Mining User Lifecycles from Online Community Platforms and their Application to Churn Prediction

6. examination of user lifecycles we used data collected from Facebook, the SAP
Community Network (SAP) and Server Fault. Table 1 provides summary statistics of the datasets where we only considered users who had posted more than 40
times within their lifetime on the platform.1 The Facebook dataset was collected
Datasets: Online Community Platforms
from groups discussing Open University courses, where users talked about their
5
issues with the courses and guidance on studying. The SAP Community Network
is a community question ‘Open University’related to SAP technologies where
1. Facebook answering system Groups
users post questions and provide answers related to technical issues. Similarly,
¤ Containing discussions about courses and degrees
Server Fault is a platform that is part of the Stack Overflow question answering
2.
site collection2SAP Community Network related to server-related issues. We
where users post questions
divided each platform’s users up into 80%/20% splits for training (and analysis)
¤ Question-answering system for SAP technologies
and testing, using the former in this section to examine user development and
3.
the latter splitServer Fault detection experiments.
for our later
¤ Stack
Overflow subsidiary site for server-related issues
Table 1. Statistics of the online community platform datasets.
Platform
Time Span
Post Count User Count
Facebook
[18-08-2007,24-01-2013] 118,432
4,745
SAP
[15-12-2003,20-07-2011] 427,221
32,926
Server Fault [01-08-2008,31-03-2011] 234,790
33,285
3.1
Defining Lifecycle Periods
Mining User Lifecycles from Online Community Platforms and their Application to Churn Prediction
In order to examine how users develop over time we needed some means to

7. User Lifecycles: Derivation
6
Offline Lifecycle Periods
Primary School
High
School
University
Postgrad
Postdoc
Lecturing
Time
First Post
Last Post
Lifecycle Periods of a potential Question-Answering System user (conjecture!)
Novice Users
Asking Questions
Asking & Answering
Questions
Answering
Questions
In reality: do not know the labels, however we can split by equal time intervals:
1
2
3
…
n
Yet, users non-uniformly distribute their activity across lifecycles
1
2
3
…
Mining User Lifecycles from Online Community Platforms and their Application to Churn Prediction
n

8. User Lifecycles: Properties
7
We set n=20
1
2
1
#posts
¨
3
2
=
…
n
Divide lifetime into equal activity periods
#posts
Capture period-specific user properties (in period s):
¤
In-degree distribution
n
¤
Out-degree distribution
n
¤
Relative frequency distribution of senders to user u in period s
Relative frequency distribution of recipients from user u in s
Term distribution
n
Relative frequency distribution of terms used by u in s
Mining User Lifecycles from Online Community Platforms and their Application to Churn Prediction
s

9. they develop in the community (for SAP and Facebook),
however Server Fault users remain relatively stable. This
could be due to the relatively minor interaction effects that
take place on ServerFault: users largely lurk on the platform
Analysing Development: Period not contribute
to seek answers to questions, and thus do Entropy
unless it is necessary (i.e. they feel that their expertise is
(3)
8
sufficient to answer a question or that a new question is
¨ required), asin users’itproperties across periods
Variation a result is likely that users have an implicit
understanding of how one should formulate a post and thus
¨ Computed period entropy for each property
ghout their
the language that should be used.
using three
tribution in
change in
Facebook
with earlier
SAP
Server Fault
ng relative
ion in one
es over the
0 0.2
0.5
0.8 1
0 0.2
0.5
0.8 1
0 0.2
0.5
0.8 1
G
G
G
G
GG
G
GGGG
GG
G
GG
GG
G
Lifecycle Stages
G
GG
GGG
GG
G
G
GGGGG GG
Lifecycle Stages
G
G
G
Distribution Entropy
2.5 3.0 3.5 4.0 4.5
G
Distribution Entropy
0.6
0.8
G
0.4
Distribution Entropy
0.1
0.3
0.5
0.7
: C[t,t ] →
sage by the
conditional
t, t ] as:
G
GGGGGGGGG GGGGG
GGGG
G
Lifecycle Stages
(a) In-degree
(b) Out-degree
(c) Lexical
tropy): To
hin a given
Generally stable trends: of lifetime-stage distributions formed from users’ terms
Figure 1. Entropies consistent variance in communication and
probability
in-degrees, out-degrees and lexical their Application to Churn Prediction
Mining
y describes User Lifecycles from Online Community Platforms andterms.
riable, and

10. that consistently across the platforms, users are contacted
by people who have contacted them before and that fewer
novel users appear. The same is also true for the out-degree
distributions: users contact fewer new people than they did
before. This is symptomatic of community platforms where
despite new users arriving within the platform, users form
sub-communities in which they interact and communicate
Changes in properties relative to earlier
with the same individuals. Figure 2(c) also demonstrates that
Computed the minimised over time and thus produce a
users tend to reuse language cross-entropy for each
gradually
propertydecaying cross-entropy curve.
users form
tently perfor
We find a
where diver
the latter st
demonstrate
SAP we fi
initially bef
while for Se
cross-entrop
suggesting t
Convergence on prior properties diverge f
to
This effect
[2] where u
begin with,
Cross Entropy
0.10
0.20
G
Facebook
SAP
Server Fault
1.2
G
G
G
G
G
G
G
0
G
GGGGGGGGGGGGGGG
0.2
0.5
0.8
Lifecycle Stages
1
0.00
0.00
GG
0
G
G
GG
GG
GGG
GGG
GG
G
GG
0.2
0.5
0.8
Lifecycle Stages
1
GGG
GGGGGG
GGGGGG
0.0
0.30
¨
Cross Entropy
0.4
0.8
¨
Cross Entropy
0.05
0.10
9
0.15
Analysing Development: Period CrossEntropy
0
0.2
0.5
0.8
Lifecycle Stages
1
V.
Inspecting
concentrated
Convergence: lack of communication with new people, or use of new terms
platform, ex
Figure 2. Cross-entropies derived from comparing users’ in-degree, outnamics of co
Mining User Lifecycles from Online Community Platforms and theirwith previous lifecycle periods. We
degree and lexical term distributions Application to Churn Prediction
now turn to
see a consistent reduction in the cross-entropies over time.
(a) In-degree
(b) Out-degree
(c) Lexical

11. Analysing Development: Community CrossEntropy
10
Difference in properties relative to the community
¨ Computed cross-entropy for each property between
user @ [t,t’] and community @ [t,t’]
¨
G
G
GGGG
GGG
GGGGGGG
0
2.0
G G GGGGGGG
GGGGGG G G
(a) In-degree
G
Cross Entropy
7.0
8.0
G
G
G
G
GG
6.0
G
Cross Entropy
3.0
4.0
5.0
Cross Entropy
1
2
3
4
lexical en
Facebook
entropy re
SAP
Server Fault
increase. W
here due t
users R2 >
0 0.2
0.5
0.8 1
0 0.2
0.5
0.8 1
0 0.2
0.5
0.8 1
Lifecycle
Lifecycle Stages
Convergence onLifecycle Stages properties Stages
community
Divergence from the community
B. Modell
(b) Out-degree
G
GG
G
GG
G
G
G
G
G
GG
G
GG
(c) Lexical
G
Inspecti
Convergence-divergence: first, adapt to community; second, separate
earlier, by
Figure 3. Cross-entropies derived from comparing users’ in-degree, outMining User Lifecycles from Online Community Platforms and their Application to Churn Prediction
paring use
degree and lexical term distributions the community platform over the same
time periods. We see a increased divergence towards the end of lifecycles.
decreasing

12. How can we model the evolution of individual users?
Solution: Mine Lifecycle Trajectories
i.e. fit a curve for each user’s development measure (property and indicator)
Properties: in-degree, out-degree, terms
Indicators: period entropy, period cross-entropy, community cross-entropy
Measures: property and indicator (e.g. in-degree period entropy)
11
Mining User Lifecycles from Online Community Platforms and their Application to Churn Prediction

13. opment of user properties, setting the explanatory variable
to be the lifecycle period of the user and the response
variable to be the user property’s entropy. In modelling
entropy development we can characterise each user using the
slope (β) of the model, thus indicating the rate of change of
entropy throughout the lifecycle periods. We induced user¨ specific entropy models for each platform’s users and then
Fitted per-user linear regression models
examined the cumulative frequency distribution [0,1] β¤ Ind’ var: entropy. Dep’ var: lifecycle period of the
values for the different user properties and platforms, these
¤ >80% of users R2 > 0.4
are shown in Figure 4.
−4
0
2 4
β
6
(a) In-degree
8
0.0
F(x)
0.4
0.0
F(x)
0.4
0.0
F(x)
0.4
0.8
Facebook
SAP
Server Fault
0.8
12
0.8
Lifecycle Trajectories: Period Entropy
propertie
the avera
decay ov
users had
than 0, th
model. T
to be pro
x (e.g. i
λ = 1/¯.
x
model u
[t0 , t0.05 ]
model as
the perio
out-degre
−2 −1
0
1
2
β
(b) Out-degree
3
−3
−1 0
β
1
2
3
(c) Lexical
Mining User Lifecycles from Online Community Platforms and their Application to Churn Prediction
Figure 4. Cumulative frequency distributions of linear regression models’
As we
model fo
users alo
In Figure

14. user throughout their lifecycles, letting
By
se the in-degree, out- derivingbythe distribution cross-entropy when proportional f (ui , [t, t ])
earlier,
deriving the minimum of average commparing users’ represent changes in user
model to
munity platform over the same values user properties with past properties, platforms and user
change paring ( ) then
across the users converge indicated clear
different on their past
devel-the end ofdevelopment, beforetrends. Thatfunction that returns the period cross-entropy of an
towards
community lifecycles. decreasing
is, property (e.g. in-degree) for a given user
propertiesbehaviour over time. the proportion of users for
we examined user
riable users. We begin this section This suggests that an exponential decay whom
individual
interval:
model process. suitable for than 0, Cross-Entropy
would be
Lifecyclechange was greater describing such reductions
Trajectories: Period and thus indicating
yatforms differtheterms mining
trajectories in average
and the
ponse
throughout user’s lifecycles. Applying such a model requires f (ui , [t, t ]) − f (ui , [
1
e development of users overall. We found that = cross-entropy values over
decay
for all tested measures, all
elling 13
period
y Trajectoriestrajecto- that users reduce in their δui
|T | − 1
f (u , [t, t ])
ng the lifecycle
time. average proportional changethe case]∈T, greater i
To examine whether this was indeed ],[t ,t we
[t,t value
users
an
ngof the entropy haduser properties
the
s (in-degree, out-degree,of defined the converged on past behaviour of
t<t <t
¨ Earlier: users measure δu that returns the average propor(entropy,
nge of period-cross- thus suggesting the period cross-entropy for a given growth
exhibited athan 0, tional change value in suitability of a decaying
generally stable entropy
Mining is performed by
By deriving [t, t one
model. Thewe chose their terms, letting requires denote a
exponential
ycle periods. in ¤ I.e. previously seendecay model f (ui , the ])distribution of average pro
user-changes Thereforeuser throughoutthe lifecycles,relationships, etc. parameter
resent
user
period cross-entropy of
the
del as a suitableFirst, examined the the change decay exponential decay platforms
modelfunction that returns potential for rateacross given value
for the that defines the values (δ) an arbitrary different
develof a time
d then beforebe provided
evelopment, to then
¨
user property (e.g. in-degree) for a given user and
properties we examined the proportion
erties, begin this the explanatory variable
rs. We setting section in-degree period cross-entropy) over time, where of users f
x (e.g. interval:
he the model:
nd
the average change was greater than 0, and thus i
periodmining process. and the response
of the user
1
f (u [t, t ]) − (ui [t , t ])
= 1/¯. We defined the lifecyclei ,period ffor, the exponential tested mea
x δu =
these
decay overall. ,We ])
found that for all
ser property’s entropy. In modelling
|T | − 1
f (ui [t, t
using an integer ,tusers s an average . , 20}, hence
[t,t ],[t
]∈T,
we can user model each user using the t<t value had = {1, 2, . .proportional change value o
opy of characterise
properties
Average proportional
<t
Feature value for interval [t,t’]
[t value in
enerally change0 , t thefeature change of
stable entropy rate
than 0, Feature: property and development indicator
thus suggesting the(6)
suitability of
l, thus indicating 0.05 ] ⌘of s1 , and then defined the exponential decay a decayin
By deriving the
Therefore we chose the We induced lettingdistribution) beexponential decay returns
he lifecycle model as follows, a proportionalof average proportional model requires one p
periods. users had user- fmodel. The a function that
(s, ui change value <0,
¨ All
change values (δ) across the different platforms and user
le model for the develbe arbitrary feature (in-degree,
els explanatory platform’s properties wethen ofthe proportion ofλ thatfor whom the decay rate of a giv
for eachthe periodusers and examinedto an provided users defines
cross-entropy
the
variable
ative and the response the average change wasx (e.g.than 0, and thus indicating
frequency hence fitted exponential decay model:
distribution of the βgreater in-degree period cross-entropy) over tim
user
out-degree, terms) for a given 1/¯. We defined the lifecycle period for the ex
user and lifecycle period:
Average of user’s features
λ= x
i
i
As we induce a per-user parameter, and thus derive a
0.8
0.8
ntentropy.properties and platforms, these found that for all tested measures, all
user In modelling decay overall. We
Exponential Decay Model
erise
model using an s
integer value
4. each user using the users had an average proportional change value of greater s = {1, 2, . . . , 20
g(ui , s) the ,suitability≡ a , and growth
(u
than 0, thus suggesting= f t i ,]s1 )es decaying then defined(7) exponenti
ng the rate of change of
[t0 0.05 of 1
the
model. Community Platforms decay Application to Churn Prediction
riods. WeMining User Lifecycles from Online The exponential and theirmodel requires one parameter
induced user2 3
model decay rate of letting f (s,
to be provided λ that defines the as follows, a given value ui ) be a function tha
latform’s users and then

15. Lifecycle Trajectories: Community CrossEntropy
14
n Divergence
linear regression
●
●
●
●
●
●●●●
0
●
●
●●●
●●●●●●●
●
● ● ●●●●●●●
●●●●●● ● ●
0.2
0.5
0.8
Lifecycle Stages
●●
●●
1
●
●
●
●
●●●●
●
6.0
0.2
0.5
0.8
Lifecycle Stages
●●●
●●●●●●●
●
●
Cross Entropy
7.0
8.0
Cross Entropy
3.0
4.0
5.0
0
6.0
0
1
2.0
●
● ●
● ● ●●●●●●●
● ●●●● ● ●
●●
2.0
0
Cross Entropy
1
2
3
4
●
0.2
0.5
0.8
Lifecycle Stages
●●
0
●
●
●
●●
●
●●
●
●
●
(b
lex
1
n Facebook, SAP: quadratic regression
Facebook
en
SAP
Figure 3. Cross-entropies deri
n Server Fault: linearIn-degree
(a)
(b) Out-degree
(c) Lexical inc
Server Fault regression
degree and lexical term distribut
he
time periods. We see a increase
>73% of users have R2 > 0.4
Figure 3. Cross-entropies derived from comparing users’ in-degree, use
out
0
(a) In-degree
2.0
Facebook
SAP
Server Fault
● ●
● ● ●●●●●●●
● ●●●● ● ●
¤ Lexical:
¨
●
Cross Entropy
3.0
4.0
5.0
¤ Out-degree:
Cross Entropy
1
2
3
4
n Convergence-divergence
Facebook
SAP
Server Fault
Cross Entropy
7.0
8.0
quadratic regression
●
0
¤ In-degree:
Cross Entropy
1
2
3
4
Identified differences between platforms and
properties’ trajectory models
Cross Entropy
3.0
4.0
5.0
¨
1
0
●●
●
●●
0.2
0.5
0.8
Lifecycle Stages
●
●
●
●
●
●●
●
●●
●
●
0.2 degree and lexical term0.2
0.5
0.8 1
0 distributions the community platform over the sam
0.5
0.8 1
0 0.2
0.5
0.8 1
Lifecycle periods. We see a Lifecycle Stages Prediction
Lifecycle Stages
Mining User Lifecycles from Online CommunityStages
time Platforms and their Application to Churn
increased divergence towards the end of lifecycles
0
(a) In-degree
(b) Out-degree
B.
informs how online com
(c) Lexical

16. Mining lifecycle trajectories enables users to be
categorised by their behaviour…
Facilitating Churn Prediction
15
Mining User Lifecycles from Online Community Platforms and their Application to Churn Prediction

17. entr
F EATURES USED FOR THElabel of PREDICTION EXPERIMENTS .B. Experimental4 Setup
CHURN the user from one of two T HE
values: y 2 {0, 1},
the closed inter
In this section we INDICATORS OF LIFECYCLExiTRAJECTORIESbinary R-valued feature vector for mod
define churn prediction as an 11-element TO
while
denotes a ARE USED
formance eac
CHARACTERISE USER EVOLUTION ALONG or DIFFERENT USER 10-element feature of and
either examined indicaclassification task and use the previously a FacebookTHE SAP user, and a For our experiments
PROPERTIES .
sures: (i) fina
vector for a Server Fault user - given that we use a linear precis
by
tors of lifecycle trajectories to predict whether a for each user’s lexical combining the test
user is a
mod
regression model
community cross- charac
operator
Property
Model Feature(s)
Platform
setting feat
ea
churner or not. As we confine Indicator
user lifecycle periodsWe model thetogether and ranked the u
entropy development. from
feature vector of each
we
In-degree
Period Entropy
Linear Regression
All
sele
user using thetrajectory indicators Alland a standard deviatio
the trajectories
from the previous section,
Period Cross-Ent
the start of their lifecycle to the end we useExponential Decay 2
16
the induced perf
mo
in short Quad’ Regress’ a1 ,our set ofAllagain where we place
Table II defines a
features into the respectiv
Comm’ Cross-Ent
mined from this period to characterise howLinearset depending on the dynamics it captures.ranks, set
users develop. All
top-k
A
Out-degree
Period Entropy within a Regression
each
¨
be
Period Cross-Ent
Exponential Decay
the mean asthe
We define churners as any user who posts for the last time Allthe same instancesof dict
Comm’ Cross-Ent
Linear Regression Table II All
dom
observingto .which the u
different user
F
Period window of our datasets,PREDICTION EXPERIMENTS T HE
Linear FOR THE CHURN
All
before the ¨
final 10%Lexicalthe time EntropyEATURES USEDRegression
of
the
Period Cross-Ent INDICATORS OF LIFECYCLE TRAJECTORIES AREchurn prediction
Exponential Decay
Allics on USED TO
correct.
use
cutoff points are: 2012-07-09 Comm’Facebook,Quad’ Regress’ EVOLUTION Fb, SAPTHE DIFFERENT USER We form
for Cross-Ent
2010-05-11 2
CHARACTERISE USER a1 , afor ALONG
¨
a randomly sele
sure
Comm’ Cross-Ent
Linear RegressionPROPERTIES .
SAP, and 2010-12-23 for ServerFault. Our dataset is of the SFerty in isolation, for in
oper
to the probabil
Property
Indicator
Modeland the entropy, period
Feature(s)
Platform
following form: D = {(xi , yi )}, where yi denotes the class Linear Regression
In-degree
Period Entropy
All
(setting p =we
|ch
Period Cross-Ent 4 Exponential Decay
All
entropy trajectory indi
the
label of the user from one of two values: y Comm’ {0, 1}, Quad’ Regress’ a1 , a2 the receiver op
2 Cross-Ent
All
A. Prediction Model Definition
Out-degree
Period Entropy
Linear Regression
All
model in confidence of a
isolation, topfor
while xi denotes an 11-element R-valued feature vector for Exponential Decay
Period Cross-Ent
All
the
Facebook, SAP: 11 features
Comm’
All
and examining in-degre
observed
userfeature ) contains
ui
to w
settings of confi
either a Facebook The SAP user,featurea vector of Period Cross-Ent (xiLinear Regression
or features and 10-element Entropy Linear Regression
Lexical
All
Server Fault: 10
Period along
corr
the indicator trajectories of we use a linear Exponential Decay
finally thereby setting
combined a
vector for a Server Fault user - given that the user Cross-Ent the different, a2 we All
Comm’ Cross-Ent
Quad’ Regress’ a1
Fb, SAP
properties. We use the logistic regression modelLinear predict In SF
to Regression
Comm’ Cross-Ent
model. follows: soa ran
doing to w
regression model for each user’s lexical community crosst
the conditional probability of user ui churning as follows:
features maximum
p
(sett
¨
Induce
entropy development. We model the feature vector of each coefficients via on prediction =
f (x)
the
1 Definition
likelihood estimation
selection for specific
A. Prediction Model
user using the trajectory indicators from |the)previous section,
P r(Y = 1 xi =
(9)
|x
confi
i
Probability of user churning
1+e
model dif
The where we place
user ui (xi ) For each setti
in short Table II defines our set of featuresobserved feature vector of performingcontains for F
the indicator trajectoriesweight user As the used
along
different
Mining User Lifecycles TheOnline Community Platforms and their)Application to Churn Predictionattached we positivethe therP
from model’s coefficients (
define the of the
log
(T
each within a set depending on the dynamics We captures. regression model to predictrate follo
it use the logistic
properties.
to each identity trajectory feature within the linear model
Predicting Churners
Binary classification task: is user u a churner?
Dataset churners: who last posted before final 10%
Dataset attributes from trajectory model features
Induced Logistic regression model:
and from are
diction model we these

18. Evaluation: Setup
17
User-wise dataset split: 80% training, 20% testing
¨ Experiments:
¨
¤ Isolated
user properties, isolated development indicator
features, all features together
¨
Evaluation measures:
1.
2.
¨
Precision@k (P): Avg over k={1,5,10,20,50,100}
Area Under the Receiver Operator Curve (AUC)
Baseline: Success probability in single Bernoulli trial
¤ I.e.
randomly selecting a churner
Mining User Lifecycles from Online Community Platforms and their Application to Churn Prediction

19. Table III
¯ ) AND A REA U NDER THE RECEIVER OPERATOR
P RECISION @ K (P
CHARACTERISTIC C URVE (AU C) VALUES FOR FACEBOOK , SAP AND
S ERVER FAULT WHEN TESTING DIFFERENT: ( I ) USER PROPERTIES , ( II )
DEVELOPMENT INDICATORS , ( III ) ALL FEATURES TOGETHER .
Evaluation: Results
AU C) is preferable (thus achieving a value
baseline for this measure is 0.5.
18
nts the performance of the different models
¨ Variance in features
atforms, showing variation in the optimum
ation measures. Interestingly, we find that
depending on:
ures combined together does not yield the
¤ Accuracy preference
y of the tested platforms. For Facebook the
hat the prediction model using community
n I.e. precision ¯ recall
>
icators performed best in terms of both P
sted the difference between this model and
¤ Platform
ming model (Full) using a Mann-Whitney
n Different detection
he difference to be significant (at the 5%
signals for different
found differences in the communities
best performing
Platform
Facebook
SAP
Server Fault
Feature
Entropy
Period Cross Entropy
Community Cross Entropy
In-degree
Out-degree
Lexical
Full
Baseline
Entropy
Period Cross Entropy
Community Cross Entropy
In-degree
Out-degree
Lexical
Full
Baseline
Entropy
Period Cross Entropy
Community Cross Entropy
In-degree
Out-degree
Lexical
Full
Baseline
to the evaluation measure used: in-degree
¯
exical features ¨ FullThese differences
for P . model is never
entrating on top ranks and thus informing
the best
ners with high-levels of confidence can be
assessing the term distributions of users
dynamics, while for preferring recall the
Mining User Lifecycles from Online Community Platforms and their Application to Churn Prediction
distributions is preferable.
o Server Fault, the results also indicate
¯
P
0.761
0.624
0.791
0.648
0.781
0.681
0.730
0.629
0.434
0.321
0.334
0.351
0.250
0.438
0.363
0.342
0.392
0.300
0.352
0.232
0.293
0.459
0.421
0.319
AU C
0.500
0.485
0.617
0.511
0.570
0.557
0.573
0.500
0.549
0.568
0.549
0.592
0.503
0.539
0.539
0.500
0.526
0.555
0.538
0.475
0.512
0.546
0.554
0.500

20. s are salient
gh precision
their in-degree distributions, and the extent to which they
are contacted during one time period relative to their past
communications, reduces at a much faster rate than on
ServerFault.
Evaluation: Churner Patterns
Table IV
19
B EST PERFORMING PREDICTION MODEL COEFFICIENTS FOR FACEBOOK
icting churn( COMMUNITY CROSS - ENTROPY ), SAP (I N - DEGREE ) AND S ERVER
nspecting the
Reduced quadratic coefficients: churnersLL FEATURES ARE SIGNIFICANT
FAULT ( PERIOD CROSS - ENTROPY ). A exhibit steep
. One of the
WITHIN THEIR RESPECTIVE MODELS (↵ < 0.05)
cross-community curves towards the end of their lifecycles
as our churn
Feature
Facebook
SAP
Server Fault
s that can be
In-degree Entropy
0.0532
dual features
In-degree Period Cross-Ent
0.0139
-0.1826
1
In-degree Comm’ Cross-Ent a
-0.1057
-0.1878
y inspecting
2
In-degree Comm’ Cross-Ent a
-0.0510
-1.5104
odel we can
Out-degree Comm’ Cross-Ent
0.3173
Out-degree Period Cross-Ent
0.0210
ase/decrease)
Lexical Period Cross-Ent
Lexical Comm’ Cross-Ent a1
Lexical Comm’ Cross-Ent a2
0.3253
-0.0541
-
0.0557
-
nts from the
g the AU C,
Variance in decay coefficient: degree of communication decays
and SAP we
VII. D ISCUSSION a lot faster forW ORK
AND F UTURE SAP than Server Fault
n model for
distributions
Prior work on social network evolution by Panzarasa et al.
Mining
[6] from Miritello et al. [1] their Application to Churn social
has a vertexUser Lifecyclesand Online Community Platforms andfound that users’Prediction networks
sed and that
tend to a limit in terms of their communication capacity.

21. Conclusions
20
1.
Users communicate with a fixed-set of users
¤ Similar
2.
to findings from (Miritello et al. 2013)
Convergence-divergence effect: users converge on
community ‘norms’ before diverging
¤ (Erikson.
1959) theorised that younger people are
susceptible to social norms
¤ (Danescu-Niculescu-Mizil et al. 2013) found users to
converge on lexical norms, before diverging
3.
Variance in churner signals
¤ No
common best model was found across platforms
Mining User Lifecycles from Online Community Platforms and their Application to Churn Prediction

22. Current & Future Work
21
1.
Regularised Linear Models
¤ Achieved ~30% AUC boost with growth and magnitude
that users tend to converge in their reviewing behaviour and
features
u,s,c
that previous profiles allow one to gauge how the user will
Dtrain )
(4)
rate items in the future given their category information.
u,s,c0
ng(Dtrain )
2.
Conversely, for MovieLens and Movie Tweetings we see an
opposite e↵ect: users’ taste profiles become less predictable
¤ Used lifecycle model (n=5) to form category-ratings profiles
as they develop; users rate items in a way that renders unassess the relative
certainty variance from previous information.
n user and lifecycle ¤ Identified in profiling in taste evolution across platforms
mapping function
categories they are
Dissimilarity
categories ( g ) we
in taste profile
o di↵erent categorfrom previous
the former profile
profile
gories, would lead
ficity that the cat1
2
3
4
5
1
2
3
4
5
1
2
3
4
5
Lifecycle Stages
Lifecycle Stages
Lifecycle Stages
type, formed from
ries would lead to
(a) Lens
(b) Tweetings
(c) Amazon
uenced byMiningprior
the User Lifecycles from Online Community Platforms and their Application to Churn Prediction
thors consider only
0.220
0.290
0.275
●
●
0.215
●
●
0.210
●
●
Conditional Entropy
0.285
●
0.205
●
0.280
●
Conditional Entropy
0.235
0.245
●
●
0.225
Conditional Entropy
Evolving-Taste Recommender System
●

23. 22
Questions?
@mrowebot
m.rowe@lancaster.ac.uk
http://www.lancaster.ac.uk/staff/rowem/
Mining User Lifecycles from Online Community Platforms and their Application to Churn Prediction