Generative AI has a broad variety of use cases and this is precisely why it is being adopted widely and is rapidly transforming many sectors. However, its application in many areas remains unclear.
We have worked with numerous customers across industries like retail, SaaS, finance and more to determine avenues to generate value with Generative AI. In some circumstances, Generative AI can even augment data analytics capabilities that are generally approached with traditional AI and machine learning technologies.
In this blog, we will discuss a project we conducted with a confidential client, in which we developed a proof of concept (PoC) to test the use of GenAI in recommendation systems. Spoiler alert: the concept was actually unproven, and surprisingly, that was great news.
We will explore how GenAI enhances traditional recommendation systems, ponder the benefits and challenges, and provide an overview of how GenAI will set new benchmarks for personalized services.
First, let's look at how we conducted the PoC.
PoC: Personalized Promotions and Customer Experiences
Understanding the Context
In the current digital marketplace, the demand for personalized customer experiences has never been higher. So, the client wanted to adopt GenAI as the core of their recommendation platform to find new data patterns, improve recommendation quality, and develop new product offerings, as traditional promotional strategies often lack the targeted engagement necessary for optimizing conversion rates. This gap between customer expectations and the actual shopping experience highlights a critical need for innovation in how promotions are crafted and delivered.
Business Objectives
The objective was to leverage AI to enhance the personalization of promotional content. By doing so, the client aimed to:
- Enhance customer engagement and loyalty by providing customers with promotions that match their individual needs.
- Improve conversion rates by increasing the effectiveness of promotion targeting and consequentially maximizing marketing ROI.
Strategic Goals for the PoC
To achieve the business objectives given the context above, we started a PoC to personalize promotions for each client. The PoC goals were:
- Validating the concept by demonstrating if GenAI-driven promotions are more appropriate than the ones created using other approaches.
- Assess the best architecture for a scalable implementation of the recommendation system.
PoC Implementation
We relied on the following datasets:
- Historical transaction data
- Customer demographics
- Mobility data
- Past promotions
Also, we used pre-trained LLMs from Amazon Bedrock, enabling us to jumpstart the personalization process. The development was divided into five phases:
1. Data Collection & Preparation
The initial phase involved filtering data, selecting relevant columns, and merging tables to create a unified dataset that was ready for analysis and model training.
Amazon S3 hosted the compressed CSV files containing the data tables for customers, visits, transactions, and promotions, while SageMaker Studio Notebook accessed the data from S3, performed necessary preprocessing, and prepared the data for prompt engineering.
2. Prompt Engineering
We added context using demographics and past transaction data, enriching the model's ability to generate personalized content.
We re-coded demographics to ensure the model would interpret them accurately and converted the aggregated data from the previous step into a natural language format that Claude v2.1 large language model (LLM) could process. Then, the LLM generated the top-k recommendations.
3. Setup the LLM
We set up the Claude v2.1 model in the SageMaker environment via the Bedrock API, to run inferences using the prompt engineered in the previous step. Utilizing Claude v2.1, we did multi-step queries to dive deep into customer individuality.
4. Generating Recommendations
Finally, we had a model inference feeding the formatted customer data into the LLM to generate promotion recommendations.
For example, the model generates: "For customer 123, offer a 20-off promo code which gives them a 20% discount on electronics, valid for the next two weeks."
Then, we post-process it, converting the model's natural language output into a structured format for practical use. For example, parsing the model's text output to extract promotion details and map them to existing offer IDs in the database.
5. Evaluation and Refinement
The final step is setting up metrics to measure the effectiveness of the recommendations (e.g., conversion rate, customer engagement, ROI, etc). We recommended testing and Iterating. This means, running tests with a control group and the recommended promotions to evaluate the performance. Based on the results, refine the model.