Speeding Up CTR Prediction Models with Extreme Learning Machines

Extreme Learning Machines (ELMs) offer a rapid alternative to standard gradient-based neural network training, delivering fast learning and strong generalization. Theoretically, ELMs are proven to be universal approximators. In this work, we investigate their use for Click-Through Rate (CTR) estimation — an area where ELMs remain underexplored due to the problem's high dimensionality. We propose an ELM approach augmented with embedding layers to boost performance on CTR tasks, representing a novel contribution. Experiments on standard benchmarks like Avazu and Criteo show that our embeddingenhanced ELM attains competitive F1 scores while cutting training times from hours to minutes compared to advanced methods like Masknet. Overall, our results suggest that ELMs are a practical option for CTR modeling when rapid training is required.