Using Large Language Models to Run Efficient Ad- campaigns: Applications of LLM in Programmatic Advertising

Abstract

One of the most prominent technological advances in Natural Language processing problems has been the advent of the Large Language Model which has changed the ball game in the field of Machine Learning. LLMs are finding applications in a plethora of industries and sectors particularly where there are large dependencies on processing textual data. In this paper, we present some critical applications of LLMs in the domain of Programmatic Advertising[1] which largely relies on dealing with structured numerical data.

Our use cases help the ad campaign managers in expediting the mundane tasks of running the online campaign using the power of LLMs. Pre-trained LLMs suffer from problems such as hallucination and training/fine-tuning can be costly for organisations with the risk of privacy in terms of data leakage, hence we focus on Retrieval Augmented Generation[2].

We have utilised LangChain[3] to enhance accuracy in managing complex tasks such as statistical computations & document retrieval. This helps us integrate context and memory into the completion process by creating intermediate steps and chaining commands together which help LLMs to do better In- Context Learning via richer prompt engineering to produce quality results. We teach the LLM via custom prompts to perform tasks explicitly such as ad-campaign performance summarisation, post campaign analysis presentation retrieval and generation, client email generation and product recommendation in order to create subject matter experts for each task and we tie these experts via a router chain so that LLM can react to a query of particular topic and product the output accordingly.

We also talk about deployment of our model using flask APIs and hosting the web application on our cloud infrastructure. This process can help the ad managers in saving almost 60-70% of their time in performing each of these tasks which can result in higher ROI and managing a higher number of campaigns.

Access The Research Paper: