Learning from Comparisons and Choices
Authors: Sahand Negahban, Sewoong Oh, Kiran K. Thekumparampil, Jiaming Xu
JMLR 2018 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | This is accompanied by numerical simulations on synthetic and real data sets, confirming our theoretical predictions. ... We present experimental results on both synthetic and real-world data sets confirming our theoretical predictions and showing the improvement of the proposed approach in predicting users choices. |
| Researcher Affiliation | Academia | Sahand Negahban EMAIL Statistics Department Yale University; Sewoong Oh EMAIL Department of Industrial and Enterprise Systems Engineering University of Illinois at Urbana-Champaign; Kiran K. Thekumparampil EMAIL Department of Electrical and Computer Engineering University of Illinois at Urbana-Champaign; Jiaming Xu EMAIL Krannert School of Management Purdue University |
| Pseudocode | No | The paper describes the algorithms in prose (e.g., Section 3.4.1 'Algorithm' and Section 4.4.1 'Algorithm') outlining steps like 'we apply the following two operations on the current estimate, Θ t, of Θ L1/2' but does not present them in a structured pseudocode block or algorithm environment. |
| Open Source Code | Yes | Code for our experiments are available at https://github.com/POLane16/Nucnorm-Ranking. |
| Open Datasets | Yes | To showcase the practicality of our nuclear norm based algorithm (9) we apply our algorithm to the Food100 Data set2 Wilber et al. (2014). ... Jester data set3 Goldberg et al. (2001) has 24, 982 users... The Irish Election data set4 is an opinion poll conducted among 1083 participants... Extended Bakery data set 6 Benson et al. (2018) consists of details of 75,000 purchases... |
| Dataset Splits | Yes | For the synthetic experiments, we generate random rank-r matrices of dimension d d... For the nuclear norm minimization, we estimate the Laplacian L using the empirical distribution of the triplets and λ is chosen to be 0.1 p log(d)/2 d xn. In the Fig. 4(b) (to the left) on the testing data set, we see that our MNL model based nuclear norm regularized algorithm clearly outperforms both unregularized algorithm and the Placket-Luce model estimator, especially when there is less training data. If x fraction of the data is used for training, we use the rest (1 x) of the data for testing. ... For each user, k = 100x jokes were randomly selected uniformly at random for training, rest of the 100 k = 100(1 x) jokes where used for testing, where x is the fraction of jokes selected for training. |
| Hardware Specification | No | The paper describes the methods and results but does not specify any hardware used for running the experiments, such as CPU or GPU models. |
| Software Dependencies | No | The paper mentions using 'proximal gradient descent' and 'modified Barzilai-Borwein (BB) rule' but does not specify any particular software libraries, frameworks, or their version numbers. |
| Experiment Setup | Yes | We make a choice of λ = p (log d)/(kd2). ... We use λ = 0.45 p (log d)/(kd2) and λ = 0.1 p (log d)/(kd2) for k-wise and pairwise rank breaking algorithms respectively. ... We use λ = 0.7 p (0.5 log(d1d2))/(kd1 d1d2) and λ = 0.16 p (0.5 log(d1d2))/(kd1 d1d2) for k-wise and pairwise rank breaking algorithms respectively. ... We use λ = 0.8 p (0.5 log(d1d2))/(kd1 d1d2). |