One Class Collaborative Filtering

one-class collaborative filtering

• Pan, R., Zhou, Y., Cao, B., Liu, N. N., Lukose, R., Scholz, M., & Yang, Q. (2008, December). One-class collaborative filtering. In 2008 Eighth IEEE international conference on data mining (pp. 502-511). IEEE.

• one-class collaborative filtering addresses the signal ambiguity problem that arises because only positive instances are observed in implicit feedback datasets, meaning missing data cannot simply be treated as negative.

  \[\mathcal{O}:=\left\{(u,i)\mid y_{u,i}=1\right\}\]

  • AMAU(All Missing As Unknown) trigger trivial solution:

    \[\mathcal{D}:=\left\{(u,i)\mid(u,i)\in\mathcal{O}\right\}\]

  • AMAN(All Missing As Negative) causes signal distortion (obs do not mean pos., unobs do not mean neg.):

    \[r_{u,i} =\begin{cases} 1,\quad(u,i)\in\mathcal{O}\\ 0,\quad\mathrm{otherwise} \end{cases}\]

• negative sampling strategy is approach designed to enhance training efficiency by sampling only a subset of unobserved items and utilizing them as negative observations.

• ranking optimization strategy is approach that do not aim to maximize the conditional likelihood of binary categories, but rather focus on establishing a relative preference order between observed and unobserved items.

• latent variable modeling strategy models the process of interaction generation by introducing a latent variable that mediates the relationship between observation and preference, rather than directly linking the two.

negative sampling strategy

• Ma, H., Xie, R., Meng, L., Feng, F., Du, X., Sun, X., … & Meng, X. (2026). Negative sampling in recommendation: A survey and future directions. ACM Transactions on Information Systems, 44(4), 1-44.

• static negative sampling (SNS) is strategy with fixed rules. (e.g. uniform sampling, popularity sampling, in-batch sampling, etc.)
  • Rendle, S., Freudenthaler, C., Gantner, Z., & Schmidt-Thieme, L. (2012). BPR: Bayesian personalized ranking from implicit feedback. arXiv preprint arXiv:1205.2618.
• dynamic negative sampling (DNS) is strategy for selecting samples that are difficult for the current model to predict.
  • Zhang, W., Chen, T., Wang, J., & Yu, Y. (2013, July). Optimizing top-n collaborative filtering via dynamic negative item sampling. In Proceedings of the 36th international ACM SIGIR conference on Research and development in information retrieval (pp. 785-788).
• adversarial negative generation (ANG) is strategy that employs contrastive learning with a generative model (e.g. gan) to generate speech samples that are difficult for the model to predict.
  • IRGAN Wang, J., Yu, L., Zhang, W., Gong, Y., Xu, Y., Wang, B., … & Zhang, D. (2017, August). Irgan: A minimax game for unifying generative and discriminative information retrieval models. In Proceedings of the 40th International ACM SIGIR conference on Research and Development in Information Retrieval (pp. 515-524).

ranking optimization strategy

• pairwise ranking optimization is strategy that maximizes the gap in predicted preference logits between observed and unobserved items, ensuring that the obs. is ranked higher than the unobs. (maximize AUROC).
  • BPR Rendle, S., Freudenthaler, C., Gantner, Z., & Schmidt-Thieme, L. (2012). BPR: Bayesian personalized ranking from implicit feedback. arXiv preprint arXiv:1205.2618.
• listwise ranking optimization is strategy that directly optimizes the ordering of the entire ranked list, aiming to maximize ranking metrics such as MRR, MAP, or NDCG instead of enforcing pairwise preference constraints.
  • CLiMF Shi, Y., Karatzoglou, A., Baltrunas, L., Larson, M., Oliver, N., & Hanjalic, A. (2012, September). Climf: learning to maximize reciprocal rank with collaborative less-is-more filtering. In Proceedings of the sixth ACM conference on Recommender systems (pp. 139-146).

latent-variable modeling

• confidence weight allocation is strategy that assigns weights to losses in proportion to the number of observations.
  • wALS Hu, Y., Koren, Y., & Volinsky, C. (2008, December). Collaborative filtering for implicit feedback datasets. In 2008 Eighth IEEE international conference on data mining (pp. 263-272). Ieee.
• exposure modeling is a method that defines non-preference as a state of having been exposed but not observed, and models the exposure probability as a latent variable.
  • ExpoMF Liang, D., Charlin, L., McInerney, J., & Blei, D. M. (2016, April). Modeling user exposure in recommendation. In Proceedings of the 25th international conference on World Wide Web (pp. 951-961).