Regularisation via Von Neumann Entropy for Matrix Completion
Lead Research Organisation:
University College London
Department Name: Computer Science
Abstract
Matrix completion is an area of machine learning with very immediate and practical applications in a business setting. The most notable being recommendation systems such as the so-called 'Netflix problem' in which a user's preferences for a given set of objects are inferred under the assumption that the data may be represented by a partially-known low-rank matrix. The aim of this research is to develop new, efficient algorithms with low mistake bounds for matrix completion in the context of online learning with a focus on entropic regularisation techniques. There's a possibility that parts of this research may be done in collaboration with Siemens UK. This research aligns very closely with EPSRC's research area of Artificial Intelligence Technologies.
Organisations
People |
ORCID iD |
| James Robinson (Student) |
Studentship Projects
| Project Reference | Relationship | Related To | Start | End | Student Name |
|---|---|---|---|---|---|
| EP/N509577/1 | 01/10/2016 | 24/03/2022 | |||
| 1782548 | Studentship | EP/N509577/1 | 01/10/2016 | 15/06/2021 | James Robinson |
| Description | In this research I have developed several adaptive algorithms for adaptive online machine learning, where it is assumed that the nature of the learning problem is changing overtime and the algorithm must be able to adapt. An example is an algorithm for predicting the labels of nodes of a network, when the labelings are changing over time. The flag-ship algorithm of this work is exponentially faster than competing algorithms. The algorithms developed all have proofs of their guarantee of success. This work has been published in a peer-reviewed paper in a leading AI conference (Neural Information Processing Systems 2019). I have also developed adaptive algorithms for a setting known as prediction with expert advice. One instance is an algorithm for adaptive machine learning in this setting, where the algorithm learns to track a small pool of predictors. The algorithm employs a new type of projection update which has many interesting properties including guarantees for its efficiency in the sense of reducing transaction costs when applied to the real-world problem of online portfolio selection. |
| Exploitation Route | There are several practical applications of my findings, for example in a smart city with many urban sensors, algorithms must be able to make predictions 'on the fly' as data is streamed in. Similarly the environment from which this data is streaming is changing over time, and thus the algorithms must me able to adapt to track these changes. My findings for an efficient update for online portfolio selection in the presence of transaction costs may have applications in finance as well as in very large models where updates may cost considerable time and/or money. |
| Sectors | Communities and Social Services/Policy,Electronics,Energy,Environment,Financial Services, and Management Consultancy,Transport |
| URL | https://papers.nips.cc/paper/8923-online-prediction-of-switching-graph-labelings-with-cluster-specialists |
| Title | Cluster Specialists for Models |
| Description | My research has introduced a novel method of prediction on graphs/networks using a framework we call "Cluster Specialists". The framework offers a flexible way of solving several machine learning problems on graphs. In our published paper on the work we describe an algorithm using cluster specialists which is exponentially faster than competing algorithms. The algorithms also have proofs of guarantees on their performance. |
| Type Of Material | Computer model/algorithm |
| Year Produced | 2019 |
| Provided To Others? | Yes |
| Impact | I am extending the framework and developing new methods of building sets of cluster specialists. |
| URL | https://papers.nips.cc/paper/8923-online-prediction-of-switching-graph-labelings-with-cluster-specia... |
| Title | Efficient relative entropy projection algorithm |
| Description | This research has produced an efficient (O(N) time) algorithm to compute exact relative entropy projection onto the simplex with non-uniform box constraints. |
| Type Of Material | Computer model/algorithm |
| Year Produced | 2021 |
| Provided To Others? | Yes |
| Impact | This algorithm can have a wide range of uses and be of general interest, and applicable to many fields in computer science. |
| URL | https://proceedings.neurips.cc/paper/2021/hash/3e9f7c16bd1cdea78f8e2eea72dfdfbe-Abstract.html |
| Description | Talk/Presentation |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | National |
| Primary Audience | Other audiences |
| Results and Impact | Around 40-50 people from various fields (mathematics, computer science, neuroscience, technology industry) from various levels (undergraduates, postgraduates, researchers, industry practitioners) attended a series of presentations and discussions on the "Mathematics of Networks". I gave a talk on my research which sparked an interesting discussion with many good questions and ideas. |
| Year(s) Of Engagement Activity | 2018 |
| URL | http://www.monmeetings.org/meeting18/ |