In the bustling ambiance of the Machine Learning Developers Summit 2024 in Bengaluru, Humeil Makhija, a distinguished Data Scientist at Pixis, took the stage to share insights into his groundbreaking modeling approach for enhanced marketing expansion in the B2B digital space. The crux of his presentation revolved around addressing the challenges posed by increasing privacy concerns, focusing on creating a robust recommendation system. Let’s delve into the key aspects of his talk, dissecting the agenda he meticulously laid out.
Makhija commenced his talk by illuminating the business problem at hand – the growing difficulty in tracking and targeting users due to privacy concerns surrounding first-party cookies. In response, his team devised a recommendation system, leveraging anonymized first-party data, or ‘Pi data.’ This ingenious approach aimed to create a bridge over the cookie-less abyss, enabling businesses to target users effectively across various platforms.
A pivotal aspect of Makhija’s modeling approach lies in the architectural prowess behind it. He introduced the audience to a two-fold pipeline – the ‘Seat Set Creation Pipeline’ and the creation of a ‘Multigraph.’ The former involved the imputation of missing data, generating a foundational set of users based on given parameters or Ideal Customer Profiles (ICP). The latter, the Multigraph, served as the backbone of the recommendation system. It efficiently recommended a set of users based on the characteristics of the input users, facilitating a personalized and targeted marketing approach.
Creating a multigraph presented its own set of challenges, especially when dealing with diverse features like categorical and high-cardinality data. Makhija elucidated their approach to calculating the distance metric, a fundamental element in graph creation. Leveraging the HNSW (Hierarchy Navigable Small World) graph-building algorithm for precision and timeliness, Makhija’s team managed to achieve an impressive recommendation rate of around 500 queries per second.
Makhija then delved into the intricacies of imputing missing data, especially in scenarios where initial user sets from clients were sparse and noisy. The team employed a similarity measure algorithm, utilizing B models and M LS, to map user data to a fixed list of industries or job titles from platforms like LinkedIn. This meticulous imputation process formed the backbone for creating a comprehensive seat set, setting the stage for effective recommendation generation.
The crux of Makhija’s recommendation approach lies in scoring the candidate users. By calculating feature importance and understanding the information value derived from the seat set, the team established a robust scoring mechanism. This method ensured that users highly similar to the seat set were prioritized, creating a dynamic and personalized recommendation ranking.
The talk seamlessly transitioned to real-world experiments, where Makhija showcased the model’s prowess. The team divided the seat set into multiple parts, assessing precision by predicting users in one part and validating against the others. The results were nothing short of impressive, with precision values ranging from 0.85 to 0.9, outperforming industry benchmarks. Makhija also highlighted the computational efficiency of the model, boasting low latency and scalability with impressive results even in experiments involving large datasets.
In conclusion, Makhija emphasized the data-centric approach his team had pioneered, not only addressing the challenges in B2B marketing but also envisioning its applicability across diverse domains like Healthcare and Finance. The model’s ability to tackle limitations posed by varying platform features, coupled with precision improvements and computational efficiency, positions it as a game-changer in the dynamic landscape of digital marketing. As businesses seek innovative solutions in the post-cookie era, Makhija’s modeling approach emerges as a beacon, illuminating the path to personalized, efficient, and scalable B2B marketing strategies.
