Qcon 2016

Mobvista

Mobvista

www.mobvista.com
2016-4-22


2007-2008 2008-2014 2014-2015

AdCenter
MSNuser segment
LBS


50%+(5kw+/)

Wechat:dustinseant

SSP
AD Network
DSP
DMP



&user
segment http://semocean.com/


1.
2.
3. &
4. &
 Mobvista

AFFILIATES
Offer
CVREPC

M

Mobvista
230+
100+/
 -

1.
2.
3.
4. Banner
5. Native
6.
7. Video

Native APPWALL

Video

CPM
-

CTR
/

GooglePlay

app

CVR
 -

ecpm
e.g. India VS South Korea
.
e.g. 200

VS

ecpm
traffic

From:akamais [state of the internet]

Mobvista



Mobvista ecpm

ecpm = 1000 * ctr * cvr * price

Win

f(x)

Features Lables Predict

ecpm
CTRCVR
A/B test

Activity Logging
 Activity Tracking Data Pipeline
All device activity (events) stored in Mongo&Dynamo as a JSON string
e.g., Fetches, Impressions, Clicks, Conversions, etc.

Model Training Pipeline
Modeling Pipeline
Ad Sort


&



Nominal
Scale

ID Feature

ID




Interval
Scale

&




Binning Methods

Equal Interval

Equal Total(Each bin with 3 elements)

e.g. RecencyClicks,TotalRevenue,RecencyConversion

Equal Positive/Negative(Each bin with 3 positive/negative elements)

 Binning Methods

Original Features Transposition

Scan Boundaries Sort for Binning

e.g. RecencyClicks,TotalRevenue,RecencyConversion
 Variable Selection

e.g. RecencyClicks,TotalRevenue,RecencyConversion


&

Global
 Global

Specific A regular segment consists of a

concatenation of specific combination
of values of the model's regular
Regular segmentation variables, plus a
Segment predicted conversion rate.

A fallback segment prediction is

looked for when a regular segment
Fallback prediction isn't found.
Segments

A single default predicted conversion

rate value which is used as the
Default prediction when regular and fallback
Prediction segment predictions aren't found.

General


GBDTCASE

##Calculating an Offers Predictive Score with a DT Model

1.BiasOffset: -7.02642680869964
2.Decision Tree Results
- Tree 1: 0.00576330069452524
- Tree 2: -0.0100230621173978
3.Logit of Predicted CR
= BiasOffset + SUM( [result of each tree] ) = -
6.966206
4.Predicted Conversion Rate
= 1 / ( EXP( -1 * Logit of Predicted CR) + 1 )
= 0.000942278186909201
5.Predictive Score for Sorting (Predicted Revenue)
= Predicted Conversion Rate * Bid Price
= 0.000942278186909201 * $2.00
= 0.00188455637381840 (* 1000 = $1.88 eCPM)
 Smothing CTR Estimation

Bayes Methods

 Exploration&Exploitation

Parameters for each offer:

Predicted Conversion Rate CRp and total number of Impressions
Given true conversion rate . The likelihood is a binomial distribution:

N imp Nimp CRp N imp N imp CR p

CR (1 CR )
N CR
imp p
Conjugate Prior of true conversion rate

Posterior Distribution of true conversion rate CR is:

CR ~ Beta( , )
Draw Conversion rate CRs from Beta (add reshaping parameter to control randomness)
CR ~ Beta( Nimp CRp , Nimp Nimp CRp )

Here is the reshaping parameter.

the
Select the offer which has imp CRp N imp N imp CRp
Nlarges
CR ~ ,

CRS bid price

 &

e.g. SDK&

Ecpm SDK1

Ecpm FB/MV Ecpm SDK2

&

DEBUG
bug
Thank You
Wechat:dustinsea

www.mobvista.com