Metric learning with adversarial hard negative samples for tag recommendation

Tag recommendation can suggest a collection of tags to users and effectively describe the characteristics of resources or users’ preferences, thereby enhancing the accuracy of information retrieval. In the field of recommendation, metric learning models are employed to measure distances among different categories. However, classifiers are susceptible to noise interference, and it is difficult to mine hard negative samples that lie close to the decision boundary in the training set. Resulting in the hard negative samples is incorrectly recommended to users. In this paper, we propose an Adversarial Tag Hard Negative samples framework to synthesize hard negative samples from observed training data, enhancing the diversity of data. We learn the distribution of the raw training data and the features that come from the hard negative samples. Specifically, we utilize the original metric learning model to preserve the effect of learning on normal samples. We also design an adversarial metric learning model, improving the model’s ability to adapt to unseen samples and potentially complex contexts, especially preventing the model from misclassifying a specific type of data in a classification. Finally, we conduct extensive experiments on the hetrec2011-MovieLens-2k and hetrec2011-LastFm-2k datasets, as well as an effective ablation experiment to quantify the contributions of adversarial components. The results demonstrate that the proposed method outperforms existing approaches in Top-N tag recommendation. Compared to the most competitive baselines, our method shows improvements of 2.1%, 8.3%, 10.9% and 5.3%, 5.7%, 7% in terms of F1@3, F1@5, F1@10 and NDCG@3, NDCG@5, NDCG@10 evaluation metrics on hetrec2011-MovieLens-2k dataset, respectively.