Experiments (Weights)

We conducted additional offline experiments that aim to validate the effectiveness of AiRS when equipped with a more-sophisticated DNN-based model, rather than the logistic regression (LR) model, in the ranking stage. Specifically, we compare AiRS employing the logistic regression model with its three variants, i.e., AiRS_{DNN_L2}, AiRS_{DNN_L3}, AiRS_{DNN_L4}, which learn feature weights based on a DNN-based model with different numbers of layers. For instance, AiRS_{DNN_L2} indicates AiRS that employs a DNN-based model with two layers. For comparison, we denote AiRS based on the logistic regression model as AiRS_{Linear}. The table below shows the results for NAVER News and Adressa datasets.

We see that AiRS based on the DNN-based model consistently and universally outperforms AiRS based on the LR-based model in both datasets and all accuracy metrics. The results show that learning the weights for multi-type features in a sophisticated way helps to improve the recommendation accuracy of AiRS significantly. Therefore, we plan to study a way to rank the candidate articles based on a more-sophisticated model such as DNN and XGBoost, rather than a simple linear regression model.