arrow_back DevOps for Data Science: Experiences from building a cloud-based data science platform
Display prospecting using explore-exploit strategy
Submitted by Akshita Sukhlecha (@akshita-sukhlecha) on Saturday, 31 March 2018
In display advertising domain, prospecting aims to build brand awareness and drive new users to the site. Due to absence of any prior user intent or user history, the task of product selection for a prospecting user from the huge item catalog becomes a great challenge. Traditionally, strategies like showcasing bestsellers, discounted products, or manually curated products have been used by marketers. But such strategies are neither efficient, nor scalable. As such items may not be attractive enough to induce a new user to click. Also, the click-performance of items changes with time due to factors like item seasonality and fatigue due to over-exposure.
Learning performance of all items is also not an efficient strategy as it would incur more time and ad-spend, due to low CTR rates of prospecting ads.
Along with ‘what to show’, ‘whom to show’ is another aspect of the prospecting problem. Targeting the entire population with items from all the categories would result in poor ad performance. The incoming user and inventory signals from the DSPs are often too sparse to be used directly.
To tackle these challenges, we use an explore-exploit strategy for item selection to maximize the clicks. The vast online and store data is leveraged to initialize the target-audience and item set. The approach can be further augmented to make use of the ad content attributes, despite their sparsity.
• Problem Statement: Prospecting advertising
• Audience curation using offline data
• Ad products pool generation using site data
• Batched multi-armed bandit strategy for ad-item selection to maximize clicks
• Future experiments
Member of data science team at @WalmartLabs. Have worked on analysing extensive Walmart user and product data and building scalable machine learning solutions to display advertising problems.