DeskAI: Automated Campaign Management for Advertisers

The bulk of my time at the Trade Desk has been spent conceptualizing, testing, and prototyping a machine learning algorithm which automates the process of managing an online advertising campaign.

Summary

The Trade Desk uses a linear model to compute bids for online advertisement space auctions. Currently, professional traders adjust the coefficients of this model based on their intuition, knowledge, and expertise. I build a platform that connects this model to TensorFlow, and allows it to be automatically trained in mini-batches, which can be updated as the campaign progresses.

We tested this approach on a campaign which posted advertisements for triathalon coaching. Our automated approach outperformed a campaign using their current optimization tool by a factor of almost two-to-one. We measure performance by cost per click, and overall our campaign showed a CPC of $1.63, while the control campaign had a CPC of $2.73.

Read on for more details.

Background

Online advertising space is sold through a bidding process, which occurs as the user loads a webpage. The bidders have access to information such as IP address, the users browsing history (if the appropriate cookies have been enabled), geographic location, site on which the ad will be presented, shape of available space, and so on. The goal of the bidding process is to use this information to evaluate whether a user presents a particularly valuable advertising opportunity.

Before the Trade Desk, most companies used what are called “line items”; these are basically case statements, instructing the system to bid a certain amount if a set of criteria are met; for example, “bid $3 if the user is from California and has previous visited the page.” The innovation of the Trade Desk is the development of independent bid adjustments. Each factor (located in California, site is yahoo.com, etc) has its own individual bid adjustment, and the total bid is calculated by multiplying these adjustments together.

Currently, these adjustments are primarly manipulated by traders, professionals whose job it is to manage and optimize campaigns. Some limited statistical optimization also takes place, but the scope of this is quite limited.

Automation

The foundational realization behind DeskAI is that the Trade Desk is fundamentally running a linear model. That is to say, they are calculating bids in the same way that a logistic regressor would calculate odds; by multiplying together individual odds from a variety of factors.

The concept of the Trade Desk’s platform as a linear model raises a natural question: What is the loss function associated with this platform? That is to say, how is the performance of a given set of bids evaluated? Once such a loss function is defined, then this model can be trained. This insight allows for computation of optimial bid factors across a variety of fields (time, location, site, device, etc.) very rapidly.

DeskAI establishes a loss function, extracts training data, and trains the linear model to establish optimal bid adjustments. The training (stochatic gradient descent) is performed in a streaming fasion using TensorFlow. The library which implements DeskAI is built in a modular, object-oriented fashion to allow for simple implementation and flexibility.

Testing

We tested the model with an advertising campaign that promoted triathalon training. Our metric of success was user clicks on the advertisement shown. We ran both an experimental campaign (with DeskAI enabled) and a control campaign (in which the typical TTD campaign planning tool was used).

The campaign optimized by DeskAI achieved twice the click-through rate of the control campaign. Note that this came at effectively no additional cost; the cost-per-click of the DeskAI campaign was half that of the control. We also observed that the DeskAI campaign won a large fraction of the spaces it bid on, indicating that it is bidding in a strategic manner; DeskAI is only bidding on space that is desirable, but when it does so it bids generously to ensure that the space is secured.

The Future

I look forward to watching this project grow and develop at the Trade Desk. Automation is the way of the future, and we have established yet another scenario in which it is highly efficient to employ machine learning techniques to analyze and optimize performance.