Thesis
Large scale methods for kernels, causal inference and survival modelling
- Abstract:
-
Machine Learning is widely applied in industrial applications such as product design and recommendations, sales and churn prediction among many others. This thesis proposes scalable methods geared towards industrial application where we address problems related to dimensionality, uncertainty quantification and inference.
First, we inferring features from images of products that are pertinent to high sales. We propose an additional regularization objective inspired by kernel two-sample testing to separate high-sellers from low-sellers in latent space for generative models. Experiments demonstrate that we can achieve linear separation in latent space and consequently infer important features using lasso methods.
Next, we develop a non-parametric test for causal association. We propose the backdoor-Hilbert Schmidt Independence Criterion with importance weights that encode the interventional distribution. Experiments demonstrate that the proposed test is calibrated and has good power.
We then propose Kernel Fried Tensor (KFT), which is a tensor factorization model augmented with kernelized side information, designed to handle high-dimensional forecasting. To improve the utility KFT, we derive a variational inference scheme to provide uncertainty quantification around our estimates. Experiments demonstrate competitive forecasting results and calibrated uncertainty quantification.Further, we consider right-censored data where existing survival models do not have a tractable likelihood and evaluate on non-proper-scores. We propose survival monotonic neural networks (SuMo-net) to train and evaluate on the right-censored likelihood, a proper score. Experiments demonstrate that SuMo-net achieves improved likelihood scores with lower inference times.
As many of our proposed methods rely on kernel matrix vector multiplication (KMVM), they are bottlenecked by the O(n2) memory and complexity. We propose Faster Fast and Free Memory Method (F3M) and achieve O(n) memory and complexity for tall (n 109) and skinny data (D ≤ 7).
In summary, the methods proposed in this thesis provide a collection of scalable methods for various use cases often encountered in industrial settings and contributes both through novel methodology and statistical characterization of the tasks, which hopefully inspires further research in the machine learning community.
Actions
Access Document
- Files:
-
-
(Preview, Dissemination version, pdf, 20.0MB, Terms of use)
-
Authors
Contributors
- Role:
- Supervisor
- ORCID:
- 0000-0001-5547-9213
- Role:
- Supervisor
- ORCID:
- 0000-0002-9341-1313
- Role:
- Examiner
- DOI:
- Type of award:
- DPhil
- Level of award:
- Doctoral
- Awarding institution:
- University of Oxford
- Language:
-
English
- Keywords:
- Subjects:
- Pubs id:
-
2043261
- Local pid:
-
pubs:2043261
- Deposit date:
-
2023-02-05
- ARK identifier:
Terms of use
- Copyright holder:
- Hu, R
- Copyright date:
- 2022
If you are the owner of this record, you can report an update to it here: Report update to this record