Continuously Optimizing Contextual Ad Targeting
Increased emphasis on Contextual
In digital advertising, Behavioral and Contextual are two prominent targeting strategies used to serve relevant ads. Behavioral targeting focuses on user’s online behavior and utilizes personally identifiable information to reach the audience digitally with high relevance whereas, contextual targeting solely focuses on the content of the page which represents the context.
With the augmenting privacy concerns in the post-GDPR era, behavioral targeting has taken a major hit making it almost implausible to reach audiences using PII’s. Hence contextual targeting has grown both in prominence and relevancy.
Contextual Strategies
Contextual targeting reaches a set of the audience representing a context instead of using personal identifiers to target users. Here context can consist of features like URL, keywords, postcode, browser, OS, etc. MIQ’s proprietary framework besides ameliorating business ROI’s has also been designed to suggest the top-performing contextual strategies for any specific advertiser.
Input and Processing
The input to the framework primarily consists of three things:
- Campaign KPI: CVR, CPA, etc.
- Bidder data: data about all ad auctions happened at Demand-side platforms (DSPs).
- First-Party Data: Pool of users (called converted users), who has performed some desired action like user visited product detail page on your website.
Our highly robust feature engineering pipelines crunches heaps of data to extract features for our log-level feeds.
Segment Creation and Bid Allocation
This module computes a Utility function using the number of converted users visited and the total number of users visiting a context. It ranks contextual strategies i.e features using an overlap ratio.
Our framework makes use of a linear bidding function and allocates bids to feature proportional to their overlap ratio and in accordance with the base bid set for the campaign.
Here is one of the linear bidding functions and extracted contextual strategies for a given campaign.
Below is sample data for a campaign and corresponding bids for features considering max bid =10 and min bid =1.
Feedback
Feedback module analyses historical data and calculates win rate which measures the number of impressions won over the number of impressions bid. Win rates, in turn, help us in forecasting the available opportunities for a certain feature.
The Feedback module monitors the performance of features and improves it using proportional feedback by controlling bids and deliveries. It’s custom rules, and optimization technique dynamically learns and Optimizes the CVR and CPA for a specific campaign to keep in mind the budget and delivery constraints.
In the above table for strategy, id=5
feedback engine calculated bid update parameter = +0.4, which increases bid from 3 to 4.2.
Similarly, fro strategy id=6
feedback engine calculated bid update parameter = -0.2, which reduces bid from 3 to 2.4.
Conclusions and Future Work
A comparative study between our feedback loop-based optimization approach as compared to a heuristic-based approach helped in ameliorating the campaign performance metrics resulting in augmented savings on the overall campaign budget. This again highlights the benefits of our approach in optimizing spends in a world obsessed with privacy.
In future work, we plan to further investigate the utility and bidding function. Utility function will be using machine learning to predict KPI’s like CVR and CPA, given feature vector. will also be experimenting with non-linear bidding functions and improving feedback engine.
Glossary:
Cost per thousand impressions (CPM) — cost or expense incurred for every thousand visitors view the ad.
Cost per acquisition (CPA) — also known as cost per action is an online advertising pricing model where the advertiser pays for a specified acquisition — for example, a sale click.
Conversion Rate (CVR) — is the percentage of visitors that complete a desired goal (a conversion) out of the total number of visitors who viewed the advertisement.
