The Fifth Elephant 2023 Monsoon

On AI, industrial applications of ML, and MLOps



Vikram Vij


Forecasting @ Samsung Ads

Submitted May 12, 2023

Samsung Ads is an intuitive audience platform that delivers meaningful experiences reaching the right audience across screens, formats and devices. With more than 900M Mobiles and 150M Smart TVs, and the largest first party data set powered by ACR, we help marketers reach targets and enhance experiences that span digital landscapes. The business has grown 10x since 2015. Our foundation is based on Samsung’s strength as a manufacturer in two key connected consumer device spaces: Mobiles & Smart TVs ...from which we derive two critical components that power our Samsung Ads businesses today: Data and Ad Impressions. We combine these assets to create powerful Ad offerings that drive reach, performance, and return on Ad spend for the world’s leading marketers.
Samsung Ads needs to sell to advertisers in advance to show advertisements. The Ad opportunities depend on user behaviour – users turning the TV on and going to specific screens. Having an automated way to forecast availability opportunities is critical to:
• Ensure we do not over-commit to advertisers (monetary implications + hurts reputation)
• Ensure we do not under-commit (opportunities are wasted, potential revenue wasted)
• Ensure users are able to self service the forecast process (tap into a larger market segment that wants self service
Our goal is to predict how many impression events will be received for a specific campaign over the duration of the campaign. The campaigns are setup to target opportunities based on various criteria including Location, Time of day, TV Model, Type of ad opportunity and User identifiers.
Forecasting is a complex problem that typically involves a single time series, and to predict for one step ahead. We have many different time series patterns, and multi-step forecasts with a long forecast horizon (over 90 days). This requires the use of sequence to sequence models and modern techniques such as transformer architectures. We are building this solution using the state of the art Temporal Fusion Transformer models. We will go over the different type of Ads such as Roadblocks, Audience Take overs, Rotationals and Video Ads and the factors affecting forecasting.
We will go over the key challenges faced in coming up with a working model architecture, such as erroneous ground truth data, data availability & quality issues and data understanding gaps, and our approaches to deal with these challenges. We will go over the use of data sketches and an OLAP DB like druid to get past data and use that and other features as inputs to a TFT model. For modeling external competition, we will explain how we estimate price dependence of win rates using survival models. We will also introduce the evaluation framework built to evaluate the forecasting accuracy at three different phases - during development, pre-release and post release. Earlier, the analysis was done manually which had many challenges like lack of consistency, delays, lack of historical data etc. which were solved with the evaluation framework.


{{ gettext('Login to leave a comment') }}

{{ gettext('Post a comment…') }}
{{ gettext('New comment') }}
{{ formTitle }}

{{ errorMsg }}

{{ gettext('No comments posted yet') }}

Hybrid access (members only)

Hosted by

All about data science and machine learning

Supported by

E2E Cloud is India's first AI hyper scaler, a cloud computing platform providing accelerated cloud-based solutions at maximum optimization and lowest pricing