A Long Sequence Modeling Benchmark for CTR Prediction
In the following, we present benchmarking results for long sequence models across various sequence lengths in CTR prediction. Specifically, we establish the DIN model with a sequence length of 50 as the global baseline and report the relative improvements of other models compared to this baseline. Each row illustrates the performance variation of a specific model as the sequence length increases. Each column compares the performance of different models at a given sequence length, with the best performance in each column highlighted by underlining and the global best results highlighted in bold.
| Models | AUC | gAUC | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| 50 | 100 | 200 | 500 | 1000 | 50 | 100 | 200 | 500 | 1000 | |
| DCNv2 | 68.89/+0.2% | 70.37/+2.4% | 70.53/+2.6% | 70.75/+2.9% | 71.19/+3.6% | 70.07/+0.5% | 70.42/+1.0% | 70.58/+1.2% | 70.63/+1.3% | 70.73/+1.4% |
| FinalMLP | 68.76/+0.0% | 70.05/+1.9% | 70.33/+2.3% | 70.71/+2.9% | 71.05/+3.4% | 69.98/+0.3% | 70.09/+0.5% | 70.27/+0.8% | 70.61/+1.2% | 70.44/+1.0% |
| DIN | 68.74/+0.0% | 70.20/+2.1% | 70.57/+2.7% | 71.62/+4.2% | 71.80/+4.5% | 69.74/+0.0% | 70.15/+0.6% | 70.01/+0.4% | 71.18/+2.1% | 71.31/+2.3% |
| DIEN |
You can’t perform that action at this time.
| |||||||||