2. Quick Introduction on Video Advertising
What exactly happens behind the video ad?
Real time bidding occurs on each
ad slot before an ad is shown.
Multiple Buyers compete in this
and whoever bids highest wins by
paying the second highest bid –
second bid price model.
Each buyer reaches ad slot via
multiple paths. The more the
number of players in the paths, the
higher the price usually.
Performance is usually measured
by the number of completions
once the video ad plays.
3. Problem Statement & Introduction to SPO
Auction duplication is getting worse, not better. On
an average an impression was available through
19.8 distinct supply paths.
DSP bias with an exchanger seems to worsen the
already existed problem.
A wise choice of path would mean low CPMs and
greater performance – Goal
Discover better paths based on ML algorithms to
whitelist and blacklist the best and worst paths
respectively.
DSP
SSP1 SSP2 SSP3
$2.55
4. Path Cost Spent
(in USD)
Impressions Video
Completion
Rate
Cost per 1000
impressions
(in USD)
Path 1 15.7 1410 90.7% 11.13
Path 2 67.9 5444 75.6% 12.47
Path 3 86.8 7580 80.8% 11.45
Path 4 50.1 1800 83.3% 27.83
Path 5 123 8120 94.7%
15.14
Path Cost Spent (in
USD)
Impressions Video
Completion
Rate
Cost per 1000
impressions
(in USD)
Path 1 15.7 1410 90.7% 11.13
Path 2 67.9 5444 75.6% 12.47
Path 3 86.8 7580 80.8% 11.45
Path 4 50.1 1800 83.3% 27.83
Path 5 123 8120 94.7%
15.14
Input
Output
In order to reach a particular site domain,
we have achieved impression delivery via
these specified paths.
Clearly, we can see few paths which are
not optimized in terms of cost and
performance (highlighted in red).
The goal we need to achieve is to blacklist
the unoptimized paths as mentioned in the
output.
To reach a particular ad slot on xyz.com
To reach a particular ad slot on xyz.com
5. DEA
Used for ranking the paths and coming up with optimum bid modifier values
Non-parametric method used to rank the Unsupervised dataset of various supply paths.
These efficiencies are calculated by comparing all the possible production frontiers of decision-making units
(here supply paths)
A,B,C,R,S are supply paths
6. Binning
WOE and IV
calculation
•Run Logistic
regression to
come up
coefficients
•Logit, Odds and
Probability
calculation
•Factor and Offset
calculation
•Come up with
scores
How Scoring is done?
Scoring – Classification Modelling
Used for ranking the paths and coming with bid modifiers
Feature Score
Supply Vendor 1 93
Supply Vendor 2 75
Site Domain 1 41
Site Domain 2 53
Site Domain 3 68