A Transformer-Powered Recommendation Engine for Personalized Online Advertising
, Senior Manager Data Science, Capital One
, Principal Data Scientist, Capital One
, Principal Data Scientist, Capital One
Practitioners of automated, personalized online advertising are often challenged with a conversion attribution problem. It's commonly discussed in the context of ad performance reporting, but misattribution can limit the predictive performance of the ad recommendation engine itself. If, for example, an algorithm can't learn which impressions were likeliest to have motivated a particular visitor to convert, then it'll struggle to serve the most relevant ads to visitors like them. We present a state-of-the-art personalized recommendation architecture for Capital One advertising, powered by the ALBERT Transformer algorithm and NVIDIA’s Merlin Transformers4Rec package (de Souza Pereira Moreira et al., 2021), to demonstrate that for repeat visitors coming to the Capital One homepage, the transformer outperforms an ensemble-based alternative model by a wide margin. We contend that it achieves superior performance by recruiting its self-attention mechanism to surmount the attribution problem.