Contextual Combinatorial Bandits With Changing Action Sets Via Gaussian Processes
Authors: Andi Nika, Sepehr Elahi, Cem Tekin
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We experimentally show that both algorithms exploit inter-base arm outcome correlation and vastly outperform the previous state-of-the-art UCB-based algorithms in realistic setups. ... We perform semi-synthetic crowdsourcing simulations on a real-world dataset and compare our algorithm with the previous state-of-the-art. Moreover, we illustrate how the time complexity of O CLOK-UCB can be improved by using the sparse approximation method, where we only use a subset of the past history of selected contexts to update the posterior. We experimentally show that our method outperforms the previous state-of-the-art UCB-based algorithms in this setting. |
| Researcher Affiliation | Academia | Andi Nika EMAIL Max Planck Institute for Software Systems Germany Sepehr Elahi EMAIL Department of Computer and Communication Sciences EPFL Switzerland Cem Tekin EMAIL Department of Electrical and Electronics Engineering Bilkent University Turkey |
| Pseudocode | Yes | Algorithm 1 O CLOK-UCB Input: X, K, M; GP prior: µ0 = µ, k0 = k. Initialize: µ0 = µ, k0 = k. for t = 1, . . . , T do Observe base arms in Mt and their contexts Xt. for xt,m : m Mt do Calculate µJt 1K(xt,m) and σJt 1K(xt,m) as in equation 1 and equation 3. Compute index it(xt,m) as in equation 5. end for St Oracle(it(xt,1), . . . , it(xt,Mt)). Observe outcomes of base arms in St and collect the reward. end for |
| Open Source Code | Yes | All simulations were run using Python, with source code provided on our Git Hub,2 on a PC running Ubuntu 16.04 LTS with an Intel Core i7-6800K CPU, an Nvidia GTX 1080Ti GPU, and 32 GB of 2133 MHz DDR4 RAM. ... 2Link: https://github.com/Bilkent-CYBORG/OCLOK-UCB |
| Open Datasets | Yes | We perform semi-synthetic crowdsourcing simulations using the Foursquare dataset. The Foursquare dataset (Yang et al., 2015) contains check-in data from New York City (227,428 check-ins) and Tokyo (573,703 check-ins) for a period of 10 months from April 2012 to February 2013. ... We use the Movie Lens 25M dataset (Harper & Konstan, 2015). |
| Dataset Splits | No | The paper describes simulation setups where data is generated or sampled per round (e.g., "250 tasks arrive", "expected number of which is sampled from a Poisson distribution with mean 100", "we sample Lt movies from the movie set"), rather than defining fixed train/test/validation splits of a static dataset. |
| Hardware Specification | Yes | All simulations were run using Python, with source code provided on our Git Hub,2 on a PC running Ubuntu 16.04 LTS with an Intel Core i7-6800K CPU, an Nvidia GTX 1080Ti GPU, and 32 GB of 2133 MHz DDR4 RAM. |
| Software Dependencies | No | The paper mentions using Python and the GPflow library (Matthews et al., 2017), but does not provide specific version numbers for either of them. For example, it states: "All simulations were run using Python..." and "We use the GPflow library (Matthews et al., 2017)..." |
| Experiment Setup | Yes | In our simulation, 250 tasks arrive (i.e., T = 250)... expected number of which is sampled from a Poisson distribution with mean 100. ... We also set a fixed budget constraint of K = 5... we define the expected outcome of each base arm as f(x) = E[r(x)] = Ag x1 x2 x3, where g is a Gaussian probability density function with mean 0 and standard deviation 0.4, and A = 2π 0.42 is the scaling constant. ... where η is zero mean noise with a standard deviation of 0.1. ... We set δ = 0.05 and use a squared exponential kernel with both variance and lengthscale set to 1. ... We run simulations with three different inducing points: 10, 20, and 50. ... We set v1 = 3, v2 = 1, ρ = 0.5, and N = 8... we set h T = 300 1 3 1+3 = 3. |