New Learning Methods for Supervised and Unsupervised Preference Aggregation

Authors: Maksims N. Volkovs, Richard S. Zemel

JMLR 2014 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically validate the models on rank aggregation and collaborative filtering data sets and demonstrate superior empirical accuracy.
Researcher Affiliation Academia Maksims N. Volkovs EMAIL Richard S. Zemel EMAIL University of Toronto 40 St. George Street Toronto, ON M5S 2E4
Pseudocode Yes Algorithm 1 Feature-Based Learning Algorithm Algorithm 2 CRF Learning Algorithm
Open Source Code Yes The code for all models introduced in this paper is available at www.cs.toronto.edu/~mvolkovs.
Open Datasets Yes For rank aggregation problem we use the LETOR (Liu et al., 2007a) benchmark data sets. For collaborative filtering experiments we used the Movie Lens data set (Herlocker et al., 1999).
Dataset Splits Yes Each data set comes with five precomputed folds with 60/20/20 splits for training/validation/testing.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments. It discusses runtime comparisons between methods but does not specify the underlying hardware.
Software Dependencies No The paper mentions specific methods like Lambda Rank, but it does not specify software dependencies with version numbers (e.g., Python version, library versions like scikit-learn, TensorFlow, PyTorch, etc.).
Experiment Setup Yes For all models we found that 100 steps of gradient descent were enough to obtain the optimal results. To avoid constrained optimization we reparametrized the variance parameters as γni = exp(βni) and optimized βni instead. This reparametrization was done for all the reported experiments. Throughout all experiments we used samples from a Gaussian with mean 0 and standard deviation of 0.01 to initialize the parameters and found that the difference in results across multiple restarts was negligible. For the SVD-based model we found through cross-validation that setting p = 1 (SVD rank) gave the best performance which is expected considering the sparsity level of the pairwise matrices. The Lambda Rank training of the scoring function was run for 200 iterations with a learning rate of 0.01, and validation NDCG@10 was used to choose the best model. For the CRF model we used expected NDCG (see Equation 5) as the target objective and set ϵ = 6 ensuring that at least one document of every relevance label was chosen each time.