# Gillian (PHD)
25/02/2022
- read and understand the derivation of LDA
- read paper "Nonconvex Optimization Meets Low-Rank MatrixFactorization: An Overview" and explain on next meeting -Section 1
- write proposal
- modify journal

18/03/2022
- find example for applying the algorithm of LDA
- explain the algorithm of LDA
- ask 10 questions about LDA
- use markdown

01/04/2022
- redo slide for LDA and present
- read master's paper
- learn sklearn to apply LDA
- download and learn how to use ghostwriter

08/04/2022
- modify journal (master)
- redo slide for LDA and present
- find some other methods for solving non-negative matrix factorization
- email Dr Khaw and email CV to Dr Wong

22/04/2022
- 2pm academic briefing
- read and understand Lee-Seung algorithm
- redo slide for LDA

25/04/2022
- solve an example of non-negative matrix factorisation by methods proposed in Lee & Seung's paper
- calculated by hand and python

29/04/2022
- using NMF from sklearn to do Non-negative matrix factorisation and compare with methods proposed in Lee & Seung's paper
- amend journal to reduce plagiarism and check with turnitin
- read Lee & Seung's paper- "Algorithms for Non-negative Matrix Factorization"
- read journal- "Nonconvex Optimization Meets Low-Rank Matrix Factorization: An Overview"

12/05/2022
- derivation of Lee Seung algorithm
~~- Find other alternative methods for solving low-rank decomposition~~
~~- Try to use "JabRef"~~
- Compare different of NMF between Lee Seung algorithm and built in function from sklearn
~~- write proposal's outline~~
- modify code by deriving the input data from text file

19/05/2022
- derive NMF (L&S) algorithm
- read PCA, VQ and NMF (from textbooks)

23/05/2022
- derive NMF (L&S) algorithm

31/05/2022
- derive and write properly (hand-written) NMF (L&S) algorithm
- coding: NMF algorithm and optimization (sklearn.decomposition.NMF)
- use example
- alpha_W=0.1 and alpha_H=0.1
- literature review - "nnsvd"...

14/06/2022
- rewrite NMF (L&S) algorithm and type into latex
- write code to decompose matrix with some constraints
- combine NMF and optimization

21/06/2022
- write code to decompose matrix with some constraints
- combine NMF and optimization
- do some literature review
- write proposal
- rewrite NMF (L&S) algorithm and type into latex

05/07/2022
- write code to decompose the matrix and make sure W is orthogonal
- differentiate the cost function $\|V-WH\|+\alpha\|W^TW-I\|_2^2$
- write journal
- add in more PCA and NMF and briefly talk about LDA.
- change the cost function (objective) and research methodology.

12/07/2022
- write code to decompose the matrix and make sure W is orthogonal
- differentiate the cost function $\|V-WH\|+\alpha\|W^TW-I\|_2^2$
- write journal
- add in more PCA and NMF and briefly talk about LDA.
- change the cost function (objective) and research methodology.
- read 'Linear Algebra and Optimization for Machine Learning' Chapeter 8

26/07/2022
- literature review
- write proposal
- do assignment

10/08/2022
- modify journal (master)
- literature review
- write proposal

17/08/2022
- write proposal
- Code for LDA and NMf did not do lemmatization and tokenization
- have tried but failed to implement

20/09/2022
- Paper
- Coding
- add sparsity constraint
- Convergence analysis
- separate into two parts
- Find two applications
- PCa
- Image processing
- Matrix completion

- Literature review
- Augmented lagrange
- Lipchitz constant
- Read Linear Algebra textbook Chapter 1
- need to present

6/10/2022
- Paper
- Convergence analysis
- separate into two parts
- Find two applications
- PCa
- Image processing
- Matrix completion
- Literature review
- Augmented lagrange
- Lipchitz constant
- Read Linear Algebra textbook Chapter 1 and 2
- Pseudo inverse

21/10/2022
- Paper
- Convergence analysis
- separate into two parts
- Find two applications
- PCa
- Image processing
- Matrix completion
- Literature review
- Augmented lagrange
- Lipchitz constant
- Read Linear Algebra textbook Chapter 1 and 2
- Pseudo inverse
- Coding
- add SPM
- find new dataset for application
- plot figure to show '0' in H

11/11/2022
- Questions:
1. why not W is sparse and H is orthogonal? So that can find W which is document-topic matrix.
- Paper
- Convergence analysis
- separate into two parts
- Find two applications
- PCa
- Image processing
- Matrix completion
- Literature review
- Augmented lagrange
- Lipchitz constant
- Read Linear Algebra textbook Chapter 1 and 2
- Pseudo inverse
- Coding
- add SPM
- find new dataset for application
- plot figure to show '0' in H
- update python to do implementation of metric quality

2/12/2022
- Prepare slides for research sharing seminar
- meeting with Prof Leong to discuss on the algorithm of SPM method for matrix
- understand the relation between semi-orthogonal matrix and pseu-do inverse for non-square matrix
- semi-orthogonal matrix is a pseudo-inverse matrix
- but pseudo-inverse matrix not a must to be the semi-orthogonal matrix

15/12/2022
- in progress of developing SPM algorithm
- check on the stopping criteria of the algorithms
- writing paper

21/02/2023
- doing literature review on the metric quality for topic modeling (perplexity, mean coherence, and execution time)
- writing paper
- doing automation for the code
- plot graph for function value

15/06/2023
- write python code to find optimal rank (optimal number of topics)
- save results and read csv file
- append tgt and plot graph (norm of gradient vs rank)
- write data preprocessing steps in the paper
Q:
1. abs or without abs for proximal method
2. stop by iteration or function value (of what) or norm of gradient (of what)
3. How many iteration for nmf and optimisation
4. modify from SPM method since SPM the data is in vector but now using data in matrix form

09/08/2023
- use 3 datasets (no need average)
- Compare terms in topics with the other research projects.
- tidy up the code

05/10/2023
- done on the automation code to find optimal rank
- submitted abstract to The third pacific optimisation conference (POC2023)
- submitted abstract to LKC FES Postgraduate colloquium
- going to add proximal method + conjugate gradient method
- in progress of tidying up the code
- in progress of producing results (without non-negativity constraint) and writing paper

09/04/2024
- coding:
- add in nonnegativity for $W$ and $H$.
- produce a result to compare with and without nonnegativity.
- add SVD method to find the optimal rank before decompostion.
- find r, and try value that near r for comparison.
- (result: the r is too high-20 % info, r reduces from 1000 to 97)
- paper:
- reduce LR part, and move images from results to appendices.

18/04/2024
- coding:
- add nonnegativity in SPM.
- add in conjugate gradient method.
- try on different value for r , from 15-30. Calculate the quality metric and plot graph.

- LR:
- lipchitz constant (for fyp)
- salt and pepper noise (for fyp)
- metric quality

- paper:
- reduce LR part, and move images from results to appendices.
- fyp abstract: modify accordingly.

15/5/2024
- coding:
- save last image
- pop out image for every iteration
- download official python
- use new data set. and do clustering metrics.
- https://www.kaggle.com/datasets/rmisra/news-category-dataset

- paper:
- appendix put behind reference
- result , plot top words, put in table

- literature review:
- read clustering
- https://en.wikipedia.org/wiki/Cluster_analysis
- https://machinelearningmastery.com/clustering-algorithms-with-python/

December 2024:
- literature review: adjusted Rand index (ARI) and clustering accuracy (ACC),

26/10/2025
- still call nonconvex alternating minimization? since is l1-norm
- alternating minimization

- the methods codings are not standardize. norm of gradient per element.diagonal bfgs.

- nmf initially has orthogonal property? since the heatmap shows orthogonality property.
- try use random matrix, min-max scaling (scikit learn),nmf

- compress sensing , continue. how
- topic modelling, more metric
- check diagonal bfgs algorithm
- separate into two py file , call the library when apply to different application. (one file for methods, one file for preprocessing of topic modeling)

4/11/2025
- Noise cancellation

___

- Model evaluation: Evaluate the performance of your trained model using appropriate evaluation metrics, such as perplexity, BLEU score, or human evaluation. This step helps you determine the quality and reliability of your model's responses.
- Model iteration and improvement: This may involve refining the training data, tuning hyperparameters, or updating the model architecture to enhance its performance over time.

next meeting : 19/07/2022 (Tuesday) 9:30am