Semi-Parametric Contextual Pricing Algorithm using Cox Proportional Hazards Model
Authors: Young-Geun Choi, Gi-Soo Kim, Yunseo Choi, Wooseong Cho, Myunghee Cho Paik, Min-Hwan Oh
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the performance of the Cox CP algorithm through Monte Carlo simulation experiments. For comparison, we included other semi-parametric algorithms for unknown baseline valuation (Shah et al., 2019; Luo et al., 2022; Fan et al., 2022). |
| Researcher Affiliation | Collaboration | 1Department of Mathematics Education, Sungkyunkwan University, Seoul, South Korea; 6Shepherd23 Inc., Seoul, South Korea. |
| Pseudocode | Yes | Algorithm 1 Cox Contextual Pricing (Cox CP) algorithm; Algorithm 2 The Cox Contextual Pricing (Cox CP) algorithm with ϵ-greedy heuristic |
| Open Source Code | Yes | Codes for the simulation are available at https://github.com/younggeunchoi/Cox Contextual Pricing. |
| Open Datasets | No | We generated vt following the PH model (3) with the dimension of context as d = 5. For true β Rd, we considered β = 4 d1d and β = 0d, where 1d and 0d are d-dimensional vectors of ones and zeros, respectively. For sampling distributions of xt Rd, we considered a uniform distribution on d-dimensional ball with radius 1 2, and multivariate t-distribution with the degree of freedom as 3 and the scale parameter as 1 4 3(d+2)Id d, where Id d is the d d identity matrix. As for true baseline valuation F0, we considered two mixture distributions: F0 = 1 2U[1, 4] + 1 2U[4, 10] and F0 = 3 4TN(3.25, 0.52, 1, 10) + 1 4TN(7.75, 0.52, 1, 10), where TN(µ, σ2, a, b) is the truncated normal distribution with support [a, b], location parameter µ and scale parameter σ2. The paper generates synthetic data for its simulations rather than using a pre-existing public dataset. |
| Dataset Splits | No | We conducted a grid search for t0 = 3,000 rounds, with τ1 {64, 128, 256, 512, 1024} and γ {2 4, 2 3, 2 2, 2 1, 2 0}. The best hyperparameter was identified in the sense of the cumulative realized revenue Pt0 t=1 ytpt. We continued the algorithm with the best hyperparameter for the remaining T t0 rounds. While the paper uses an initial period for hyperparameter tuning, it does not explicitly define training, validation, and test dataset splits in the traditional sense, as the data is generated synthetically throughout the simulation. |
| Hardware Specification | No | The paper discusses computation times (e.g., '20 seconds for running the entire T = 30,000 horizon'), but it does not specify any hardware details like CPU, GPU models, or memory. |
| Software Dependencies | No | Anderson-Bergman (2016) s algorithm is available in R package icen Reg (Anderson-Bergman, 2017), which we employed to implement our algorithm in Section 7. The paper mentions the R package 'icenReg' and its associated publication year, but does not specify a version number for the package itself. |
| Experiment Setup | Yes | The total horizon was set to T = 30,000. We generated vt following the PH model (3) with the dimension of context as d = 5... For example, the exploration of the Cox CP algorithm can be controlled through the first-epoch length τ1 and the forced sampling frequency αk = min{γ2 (k 1)/3, 1}. We conducted a grid search for t0 = 3,000 rounds, with τ1 {64, 128, 256, 512, 1024} and γ {2 4, 2 3, 2 2, 2 1, 2 0}. |